Poster Session and Exhibits - Salon Frontenac, Petite Frontenac, Bellevue, and Verchers
June 23, 2022 from 6:00pm EST to 7:00pm EST
Scientific Session – Poster Session
Thursday, June 23, 2022, 18:00-19:00
Physics considerations from ten-patient experience with single-fraction treatment of ventricular tachycardia
Elsayed Ali, Graham Cook, David Tiberi, Manon Lacelle, Greg Fox, Steve Andrusyk, Kim Charbonneau, Katie Lekx-Toniolo
The Ottawa Hospital
Purpose: To summarize physics considerations from a two-year experience with the first ten patients treated with an ablative dose (25Gy/1fx) for ventricular tachycardia.
Methods: Scar-CT imaging and electrocardiographic mapping were used to determine the myocardial tissue to be targeted for radio-ablation. Flat couchtops and abdominal compression devices were standardized between the scar-CT and 4D treatment-planning CT to improve image registration accuracy. Nothing-by-mouth was required for 6hrs before all imaging and treatment for more consistent stomach position. An ITV approach was used for the target, with a 3mm PTV expansion. Two or more CBCT images were used for image guidance with translational and rotational corrections applied. FFF delivery was introduced for the most-recent patient. All patients had ICD devices and were monitored during and after treatment. Patient and treatment characteristics were analyzed for the 10 patients.
Results: The consistent setup between the scar-CT and planning CT substantially improved image registration. For the 10 patients (8M/2F, 66+/-8yrs), the PTV volume was 128+/-45cc. The PTV D95% was 2453+/-86cGy (objective: 2375cGy). The most-challenging OAR was typically the stomach, and for those 6 patients D0.035cc for the 3mm stomach PRV was 1872+/-262cGy. Introducing FFF delivery reduced beam-ON time by 45% (from 20- to 11-minute average). Mid-treatment CBCT adjustments were 1.9+/-1.5mm and 0.2+/-0.2degrees.
Conclusion: Our centre has treated 10 patients with ventricular tachycardia using radio-ablation. Plan optimization and FFF delivery have the promise of decreasing treatment times, leading to reduced patient inconvenience and intra-fraction motion. A clinical trial comparing radiotherapy to catheter ablation is imminent.
Evaluation of miniature x-ray imaging detector and its application to focal spot assessment
Eva T. E. Anderson, Paul C. Johns
Purpose: The x-ray tube focal distribution is a key determinant of imaging performance. Standard IEC 60336 edition 5 (2020-12) allows use of a digital imager to measure focal spots. We assessed the performance of a miniature imager and in this works in progress are investigating its utility for focal spot imaging.
Methods: Hamamatsu CMOS sensor S11684-12 (1000 x 1500 array, 20 micron square elements, 14-bits) was evaluated. Pinhole images (Fluke Biomedical 75 micron diameter pinhole, model 07-617) were captured using manual mode with 1500 ms integration time, and 60 kVp, 200 ms exposure time. Ten exposures were dark corrected and added. Pinhole image size is corrected using the pinhole magnification m. In practice focal spot locations can be difficult to measure from the housing exterior. Therefore, we obtain m by placing a scale object at a known distance from the sensor.
Results: Our detector had 0 bad pixels. The x-ray signal above the dark signal ranges from 0 to ~7500 LSB. The small imager necessitates placing the scale object closer to the sensor than the pinhole plane is. Placing the 1.02 cm scale object 45.26 cm from the sensor yielded m = 1.67 ± 0.04, in agreement within error tolerance to that from geometry. The pinhole was at 53.11 cm. From profiles through the focal spot images the widths at 15% of maximum were obtained, and corrected for m, pinhole size, and image receptor blur.
Conclusion: A high-resolution digital imager is a practical means to assess focal spots.
Diffusion Anisotropy Imaging in Temporal Lobe Epilepsy
Nico Arezza, Jorge Burneo, Ali Khan, Corey Baron
University of Western Ontario
Purpose: Temporal lobe epilepsy (TLE) is a disorder characterized by seizures originating in the temporal lobe. Some patients with unilateral TLE undergo surgery to remove the seizure focus, but prognosis depends on the localization of brain abnormalities. Here, we investigated whether two diffusion magnetic resonance imaging techniques, fractional anisotropy (FA) and the newly introduced microscopic fractional anisotropy (Î¼FA), could detect hippocampal abnormalities in TLE. Both techniques estimate water diffusion anisotropy as a surrogate for neurodegeneration, but FA can confound true pathology with axon orientation whereas Î¼FA is insensitive to axon orientation.
Methods: Nine patients with unilateral TLE (4 female, 33Â±12 years) underwent four MRI scans: a T1-weighted scan, a T2-weighted scan, and diffusion scans for FA and Î¼FA. The hippocampus was segmented into four subregions: the subiculum, the cornu ammonis 1 (CA1), the CA2/3, and the CA4 plus dentate gyrus (CA4/DG). Mean FA and Î¼FA were measured in the ipsilateral and contralateral sides of each subregion.
Results: A significant difference was observed between ipsilateral and contralateral Î¼FA in the CA4/DG region (p<0.05). Ipsilateral CA4/DG Î¼FA was reduced in all nine patients while FA was reduced in six, with mean percent differences of -12.9% and -6.6%, respectively, between hemispheres.
Conclusion: Î¼FA may be sensitive to abnormalities in the CA4/DG region that is often affected by severe neuron loss in TLE. FA may be less specific to these abnormalities due to its sensitivity to axon orientation, as crossing axons are present in the hippocampus.
Evaluation of Varian iCBCT CT Number Accuracy For Radiation Dose Calculation
Devin Baillie, Dr. Satyapal Rathee
Alberta Health Services - Cross Cancer Institute
Purpose: To investigate the feasibility of using iCBCT for dose calculation in adaptive radiotherapy.
Methods: A custom-built phantom was scanned using Pelvis and Pelvis Large CBCT protocols on a TrueBeam. Scans were reconstructed using filtered back projection (FBP) and iterative reconstructions. Positions of the standardized, electron density inserts within the phantom were varied, and the scans repeated. The same arrangements of inserts were also scanned on a Philips’ Big Bore CT scanner for comparison. The position dependence of CT numbers of each insert was measured and compared across systems and reconstruction methods. The average CT number of each insert formed a calibration table for radiation dose calculation in our radiotherapy treatment planning system.
A CATPHAN 604 (Phantom Lab) was scanned with and without an elliptical collar to add scatter to measure the change in CT number due to scatter.
Pelvis- and lung-mimicking phantoms were similarly scanned and the electron density calibration tables for each case were used to calculate radiation dose for identical treatment plans.
Results: For the large elliptical phantom, iCBCT increases position-dependence of CT numbers of some soft-tissue-mimicking plugs for both Pelvis and Pelvis large protocols, but decreases this dependence for bone-mimicking plugs. The CATPHAN images showed smaller change in CT number due to collar using iCBCT compared to conventional CBCT, consistent with previous work on this subject. Preliminary radiation dose analysis indicates significant improvement in dose calculation accuracy in lung-type phantom using iCBCT.
Conclusion: FBP and iCBCT both allow accurate dose calculation for adaptive radiotherapy in pelvis.
Northern Perspectives: Commissioning and Clinical Operation of the Northernmost Linear Accelerators in Canada
Radim Barta, Andrei Ghila
Grande Prairie Cancer Center, Alberta Health Services
Purpose: Offering radiation therapy services closer to northern communities is an exciting step forward in improving patient quality of care in Canada. Thanks to close collaboration between Alberta cancer centers, expert and high quality treatment is now possible closer to home for more Canadians.
Methods: The development of an excellent program in a short time frame was possible thanks to strong collaborations within Alberta as well as with the wider Medical Physics community. Virtualization of training and day-to-day communication allows a center, 500 km away from the next nearest radiation therapy program, to still draw on expertise of dozens of medical physicists, radiation therapists, and radiation oncologists.
Results: A collaborative and competent local team has come together around a patient centric program that treats patients close to home. The team is treating 14 patients daily. With staffing to fully utilizing one unit at a time, the team has treated over 60 patients since opening. The response from the community is overwhelmingly positive and there is high demand for treatments.
Conclusion: A new radiation therapy treatment center was opened in a northern community during a global pandemic. Lessons learned have application to the opening of radiation therapy centers across Canada especially as Covid continues to be an ebbing and flowing global threat. The response from the community and the benefits to patient quality of life is also motivation for other jurisdictions to bring radiation therapy closer to all Canadians.
Predicting Tumour Response with Radiomics and Machine Learning in MR-Guided Cervix Brachytherapy
Robert Bellis, Dr. Ananth Ravi
Health Sciences North - Northern Ontario School of Medicine University, MOLLI Surgical Inc.
Purpose: This study seeks to determine if radiomic features extracted from the gross tumour volume of locally advanced cervical cancer (LACC) patients can be used to reliably predict tumour responses, using modified RECIST tumour response criteria, prior to brachytherapy treatment start.
Methods: 12 machine learning algorithms were tested with 5-fold cross validation using 1183 radiomic features extracted from 20 patients from 3 MR sequences (T1, T2 and diffusion-weighted). The most predictive radiomic features from each MR sequence were further combined into a single dataset and models were retrained with this data. The highest accuracy models were compared with models formed solely from clinical data to determine the necessity of radiomics features as inputs to the models.
Results: The models were created and assessed based on their accuracy in predicting binary tumour response which was defined as >30% reduction in tumour diameter. Several models performed with accuracies of up to 85%. Combining radiomic features from multiple MR sequences into a single dataset yielded an increase in accuracy, up to 93%. Models trained exclusively on clinical data (e.g. age, stage, dose) achieved maximum accuracies of 65%.
Conclusion: This study showed that machine learning models coupled with radiomic features extracted from GTVs are capable of accurately predicting LACC tumour response prior to administering the first fraction of brachytherapy.
Evaluation of variable bladder filling for daily online-adaptative therapy of the prostate: A case study
Jean-Guy Belliveau, Dennis Stanley, Carlos Cardenas, Richard Popple
University of Alabama at Birmingham
Purpose: We present a case study of a patient treated in 28 fractions to the prostate and nodal chain using Varian Medical System’s Ethos platform to assess the impact of bladder filling and other metrics on the original scheduled and online adapted plan.
Methods: We retrospectively analyzed the data for one patient treated on our Varian Ethos platform for prostate and nodal treatment and assessed single point metrics such as PTV_High V100%, PTV_Low V100%, Bladder V60Gy, and Rectum V70Gy. We extracted and compared the scheduled and adapted plans for each daily session for metric comparison.
Results: Plan adaptation resulted increased PTV coverage for both the prostate (Scheduled: 52.55%; IQR = 41.75%-64.25% Adapted: 97.3%; IQR=97.1%-97.5%) and nodal chain (Scheduled: 80.15%; IQR = 71.18%-88.08% Adapted: 99.7%; IQR=99.7%-99.7%). Bladder V60Gy increased for adapted plans (6.3%; IQR=5.8%-7.0%) due to the improved coverage of the prostate PTV. A noticeable finding is a sharp increase in bladder V60Gy for bladder filling below 200 cc. For bladder filling above 200 cc, there is a small downward trend of V60Gy. Finally, rectal V70Gy had less variability for adapted plans.
Conclusion: This study represents a patient treated over 28 fractions with an analysis of single dosimetric parameters. While bladder filling remains important, volumes greater than 200 cc showed a lessened dosimetric impact. An optimal volume for patient comfort and adaptation may exist. Further studies with an emulator and dosimetry to the bladder and rectal wall may be more beneficial in patient-related outcomes and toxicity.
Prospective monitoring of standardized radiotherapy compliance
Jeremy Braun, Sarah Quirk, Hali Morrison, Kundan Thind, Lukas Van Dyke
Tom Baker Cancer Centre - University of Calgary, Henry Ford Health Systems
Purpose: To evaluate the usage of standardized radiotherapy protocol compliance software and its ability to assess adherence to institutional protocols in prospective use.
Methods: An automated protocol compliance software tool based on the AAPM Task Group 275 and 263 recommendations has been previously developed, validated, and published  for a prostate cancer patient cohort and is used prospectively. This work demonstrates compliance adherence prospectively compared to a published cohort.
Results: Since February 2020, 112 plans underwent prospective use with the automated protocol compliance software, with a mean pass rate (standard deviation) of 92.3% (6.5%) and mean fail rate (standard deviation) of 6.7% (6.0%), compared to 92.3% (6.1%) and 6.0% (5.8%) in a previous cohort (n = 58). Patient Assessment, Simulation, and Treatment Planning accounted for 8.7%, 0.4%, and 91.0% of failures, respectively, compared to the previous cohort at 2.6%, 0.0%, 97.4%. Contouring checks predominated Treatment Planning failures for both cohorts at 33.3% and 39.3%. Iterative use of this tool demonstrated a mean pass rate improvement of 0.8% (61 plans underwent multiple iterations, with mean pass rates of 94.3% (4.5%) and 95.1% (4.1%) for first and final iterations).
Conclusion: The prospective use of the protocol compliance framework shows sustained ability to maintain protocol compliance adherence. The persistence of high failure rates in contouring can be primarily attributed to non-compliance with TG 263 nomenclature. Continued education and tighter feedback loops with clinicians can improve further compliance.
Impact of compression on heart motion for stereotactic arrhythmia radioablation
Daniel Cecchi, Hali Morrison, Nicolas Ploquin
Univeristy of Calgary, Alberta Health Services
Purpose: The objective of this study is to compare the motion of the heart and the left ventricle (LV) during compression vs. free breathing for patients undergoing ventricular tachycardia radioablation.
Methods: 4DCT scans of two patients with a ventricular tachycardia condition were analyzed to quantify the motion of the heart and LV. Two 4DCTs were acquired: one with a compression paddle, followed by another without (free breathing). The scans were then imported into MIM Maestro (MIM Software Inc.) and the heart and LV were contoured over all 10 breathing phases. Centre of mass (COM) motion from both structures were obtained in the superior/inferior (SI), left/right (LR), and anterior/posterior (AP) directions.
Results: SI, LR, and AP COM maximum displacements for the heart (compression; free breathing) were (3.98mm; 6.60mm), (1.07mm; 0.54mm), (0.76mm; 1.06mm) for patient 1, and (7.59mm; 6.14mm), (1.89mm; 1.14mm), (1.99mm; 1.38mm) for patient 2. SI, LR, and AP COM maximum displacements for the LV were (4.83mm; 8.7mm), (0.8mm; 2.74mm), (1.14mm; 0.83mm) for patient 1, and (6.54mm; 6.69mm), (1.01mm; 2.52mm), (2.56mm; 1.68mm) for patient 2.
Conclusion: The effect of compression on the COM motion of the heart and LV was variable for the two patients indicating that not all patients may have reduced motion with compression. Further patient analysis is warranted to determine optimal immobilization strategies.
Predicting feeding tube insertion in oropharyngeal cancer with a radiomics-based machine learning model
Tricia Chinnery, Pencilla Lang, Anthony Nichols, Sarah Mattonen
Purpose: Patients with oropharyngeal cancer (OPC) treated with chemoradiation suffer treatment-related toxicities which can lead to nutritional deficiencies and weight loss. We aimed to develop a machine learning classifier to predict feeding tube insertion in patients with OPC.
Methods: Primary tumour volumes were contoured on pre-treatment planning CT images for patients with OPC (n=342) treated with chemoradiation. PyRadiomics was used to compute radiomic features from these volumes on the original and filtered images. The dataset was split into independent training (n=239) and testing (n=103) datasets. Feature selection was applied to select the optimal features to predict feeding tube insertion. Multiple machine learning classifiers were built using the selected radiomic features and clinical features on the training dataset. The combined models’ performances were assessed in the testing dataset based on the AUC.
Results: A total of 115 patients (34%) required a feeding tube. Through feature selection, eight predictive features were selected. This included one filtered first-order and seven filtered textural features. Clinical features included body mass index, tumour stage, and nodal stage. The Naïve Bayes model achieved an AUC of 0.69 [95% CI: 0.57-0.79] in the testing dataset, whereas a model comprised of clinical features alone achieved an AUC of 0.60 [95% CI: 0.49-0.72].
Conclusion: This is the first study to use radiomics to predict feeding tube insertion. Once refined and validated, this model could assist physicians in identifying patients who may benefit from prophylactic feeding tube insertion, and ultimately improve quality of life for patients with OPC.
Saving time and reducing errors by automating the generation of treatment planning structures
Nick Chng, Kim Lawyer, Taran Braich, Lindsey Baker
Purpose: OptiMate is a Varian Eclipse script that automates the creation of planning and optimization structures like PTVs, PRVs, and overlap regions. The primary goal of this project is to improve planning efficiency and adherence to planning protocols by minimizing human error. OptiMate is free to use and modify.
Methods: OptiMate is launched from Eclipse once primary contours (CTVs, OARs) have been completed. Templates are defined for sites, which list the structures to create, and the operations necessary to derive them from primary contours. Once a template is loaded, patient-specific adjustments can be made if needed. Supported instructions include any combination of margin creation, Boolean operations, and cropping. Accuracy and timing tests were conducted on representative data for five clinical sites.
Results: Head & Neck, Prostate, Brain, Lung, and Gyne templates were created for two centres within our institution. All automatically generated structures were within 1% of the volume of the corresponding structure in the clinical benchmark, differences being primarily associated with differences in the use of low- vs high-resolution segments. OptiMate was measured to take 5-15 seconds, depending on the number of generated structures. Based on a survey of dosimetrists, the script results in a reduction in planning time of approximately 15-30 minutes per case for an experienced planner.
Conclusion: Automation of this task represented a significant time-savings in our clinical workflows. Future research will look at whether the tool reduces errors in settings like clinical trials in which stringent but unfamiliar contouring requirements may lead to manual errors.
Evaluation of the motion and registration accuracy of the Elekta Symmetry 4D cone-beam mode
Eric Christiansen, Cathy Neath, Aaron Vandermeer
Purpose: To evaluate the accuracy of a 4D cone-beam CT (CBCT) image acquisition mode in capturing the entire range of respiratory tumour motion and in correcting for errors in patient setup.
Methods: The Quasar phantom was used to simulate respiratory motion. A 3D-printed tissue-density insert representing the tumour was placed inside a low-density cedar cylinder representing the lung, all driven by an electric motor.
Motion amplitude, frequency, and imaging parameters were varied to evaluate the effect of each on the resulting 4D-CBCT images. Reference 4D-CT images were also acquired.
The motion amplitude observed in the 4D-CBCT was compared to the amplitude set on the motor. Registration accuracy was determined using high-density surface markers visible in the reference and CBCT images.
Results: The 4D-CBCT acquired over 3 min and using the S20 FOV gave the smallest average (1.3 mm) and upper quartile (1.1 mm) amplitude error.
The range of average registration errors for the different imaging parameter combinations was 0.2-0.4 mm. The average error was 0.3 mm for the 3 min acquisition and S20 FOV and 0.2mm for 3D-CBCT images.
4D-CBCT images acquired with motion amplitudes of 5 and 10 mm at 6 bpm featured severe reconstruction artifacts. These artifacts persisted for the 5 mm amplitude up to 12 bpm.
Conclusion: Given the presence of reconstruction artifacts for 6 bpm, it is recommended not to use the 4D-CBCT image acquisition mode at or below this respiratory frequency. The optimal set of imaging parameters is a 3 min acquisition and S20 FOV.
Prediction of pathologic lymphovascular invasion in non-small cell lung cancer using multi-modality tumour and peri-tumoural radiomics
Jaryd Christie, Perrin Romine, Viswam Nair, Sarah Mattonen
Western University, University of Washington School of Medicine
Purpose: To develop a model using tumour and peri-tumoural radiomic features extracted from CT and PET imaging to predict lymphovascular invasion (LVI) in a cohort of surgically resected non-small cell lung cancer (NSCLC) patients.
Methods: The dataset used in this study consisted of 130 patients with NSCLC. The tumour and 1cm peri-tumoural volumes on both the diagnostic CT and PET images were delineated. Radiomic features were extracted from the volumes. Clinical features were also investigated to predict LVI. Two models were created: a clinical and tumour model, and a clinical, tumour and peri-tumoural model. A nested five-fold cross-validation model implementing logistic regression was used for feature selection and model building. Model performance for LVI prediction was assessed using the AUC.
Results: The model consisting of clinical, tumour and peri-tumoural features outperformed the model with only clinical and tumour features, achieving AUCs of 0.72 (SD: 0.087) and 0.65 (SD: 0.13) respectively. The top performing radiomic features determined by feature selection were wavelet features, both first order and texture, from CT and PET imaging. Peri-tumoural features from both modalities were chosen to be prognostic and augmented tumour features in LVI prediction.
Conclusion: The model developed in this study using features from the tumour and peri-tumoural regions on CT and PET, in combination with clinical features, demonstrated an improvement in prognostic value when compared to a model with tumour and clinical features alone. This model may assist clinicians in identifying patients who may be at a high risk of treatment failure before surgery.
Improved efficiency and precision in the treatment of small cranial targets
Cody Church, David Parsons, R. Lee MacDonald, Alasdair Syme
Dalhousie University, UT Southwestern, Department of Radiation Oncology & Department of Physics & Atmospheric Science
Purpose: Intrafractional motion is an unavoidable reality of radiation oncology when treating small targets. This work proposes two complimentary methodologies to minimize the dosimetric impact of motion: 1) megavoltage (MV) imaging to detect and correct motion with couch translations. 2) Dynamic couch trajectories which improve treatment efficiency by reducing treatment distances during delivery.
Methods: In the first investigation, the dosimetric impact of simulated intrafractional motion is quantified when treating small targets with virtual cones. In the second investigation MV region-of-interest (ROI) imaging plans are generated to verify the detection of simulated motions with a 3D-printed skull. In the third investigation, clinical SRS plans are converted to a delivery at a shortened, virtual isocentre by means of dynamic couch motions; for each plan, plan quality metrics are quantified, and a dose verification is performed.
Results: The dosimetric impact of motion is a function of target and aperture size; smaller apertures demonstrated larger relative dosimetric detriments compared with larger apertures for a given target size and motion trace. MV ROI imaging apertures can be used to achieve sub-mm registration errors while reducing imaging dose. Shortened virtual isocentre deliveries have the potential to reduce treatment times by up to 21.0% through MU-reductions without a significant loss to plan quality.
Conclusion: The magnitude and frequency of SRS magnitudes has been shown to have a clinically relevant dosimetric impact in this work. Two subsequent solutions that independently address motion by means of positional corrections and increased treatment efficiency have been demonstrated.
Iterative Image Reconstruction Methods in Optical CT Radiochromic Gel Dosimetry
Stephen Collins, Andrew Ogilvy, Maria Guenter, Andrew Jirasek, Warren Hare, Michelle Hilts
University of British Columbia – Okanagan, BC Cancer Kelowna
Purpose: To investigate image reconstruction methods and the resulting image quality of optical CT imaged radiochromic gel dosimeters, for use in radiation therapy treatment plan quality assurance.
Methods: Ray tracing simulations were developed to quantify the effects of light refraction and reflection through a custom-built optical CT system. The simulated optical CT data was then utilized to account for refraction and reflection during iterative image reconstruction. Several iterative image reconstruction algorithms were tested and compared in terms of image quality parameters that pertain to the accurate extraction of dosimetric information from a radiochromic gel dosimeter.
Results: By accounting for light refraction and reflection during iterative image reconstruction, the resulting image quality was shown to have drastically improved spatial resolution, contrast resolution, Gamma map pass percentage, spatial-non-uniformity and signal-to-noise ratio, when compared to traditional filtered backprojection (FBP) methods. The fast iterative shrinkage/thresholding algorithm modified with a total variation penalty term (FISTA-TV) was found to obtain the highest image quality when compared to several other iterative algorithms.
Conclusion: The improved image reconstruction techniques allow for precise and accurate extraction of dosimetric information from a radiochromic gel dosimeter making optical CT radiochromic gel dosimetry a viable option for radiation therapy treatment plan quality assurance.
An Evaluation of Dosimetry Optimization for Low-Dose Rate Prostate Brachytherapy
Susan Dang, Alexander Abraham, Saibishkumar Elantholi Parameswaran, Sonja Murchison, Isabelle Gagne, Manuel Rodriguez
BC Cancer - University of Victoria, Parameswaran
Purpose: Despite the several steps taken to ensure the precise placement of the seeds in low-dose rate prostate treatments, uncertainties during the process may contribute to a discrepancy in the planned dose distribution. The purpose of this work is to find correlations between pre-implant and post-implant plan parameters that might lead to prediction of clinical outcome in LDR prostate brachytherapy using machine learning concepts.
Methods: Eighty-five prostate patients previously treated with LDR brachytherapy were evaluated. Calculation of dose distribution generated by the seeds were based on AAPM TG43 formalism and in-house BrachyVIC software. Pre-implant dosimetry objectives, PTV V100, Prostate V100, probabilities of meeting dosimetry objectives (obtained using BrachyVIC), prescription (110Gy/144Gy) and number of voodoos (extra-seeds implanted) were used as independent parameters of a multilinear regression model to predict the post-implant prostate D90 and/or V100 clinical goals.
Results: A linear correlation between post-implant prostate D90 and V100 dosimetry objectives was observed (R2=0.84). There is also a relationship between the pre-implant parameters and post-dosimetry objective, prostate D90. Pre-implant parameters, PTV V100 and prostate V100 probability showed stronger correlations for post-implant clinical goals (coef=2.23 and 0.52, respectively) while other parameters showed less of an influence.
Conclusion: This study shows that there is a correlation between pre-implant and post-implant plan parameters which can be used to build a machine learning algorithm. Training LDR plan databases with the proper classifiers may lead to the implementation of a predictor of post-implant clinical goals and potentially improve the outcome of this treatment modality.
Skin Toxicity during Prostate Bed Radiotherapy (RT) due to Bilateral Hip Prosthesis Avoidance
Melanie Davidson, David Mak, Danny Vesprini
Odette Cancer Centre - University of Toronto Department of Radiation Oncology
Purpose: Limiting entrance dose through hip prostheses to improve dosimetric accuracy can result in unfavorable skin dose and toxicity. We propose a volumetric modulated arc therapy solution that strikes a better balance between dose accuracy and skin dosimetry.
Methods: Our current planning strategy limits entrance dose through hip prostheses using stringent optimization objectives on an avoidance structure. Avoidance efficiency is evaluated by recalculating the plan with prosthesis density set to 20g/cc, and evaluating loss of target coverage from increased attenuation through artificially dense prostheses. We require this loss be within 5% of the original values. This approach has resulted in an uncommon skin toxicity for a prostate bed patient with bilateral hip prostheses, thus we re-optimized with reductions in prosthesis avoidance required to achieve skin Dmax doses 30-50Gy, in 5 Gy increments.
Results: Skin dose increases and target dose coverage and conformity decrease as prosthesis avoidance is prioritized. The large degradation in target coverage at prosthesis density of 20g/cc for plans with the lowest skin dose (30-35Gy) indicate that a significant proportion of dose arises from beams entering the prostheses. A middle range of plans (skin Dmax 40-50Gy) provide a better compromise between skin dose and prosthesis entrance dose, with a 10-20% reduction in target coverage at the enhanced prosthesis density.
Conclusion: Despite recommendations and strategies available for RT planning in the presence of hip prosthesis, there is a fine balance between allowing minimal dose through the prosthesis to optimize target and OAR doses and mitigating its associated dosimetric uncertainties.
Use of the Monte Carlo Method to Relate GAFCHROMIC® EBT3 Film Response to Absorbed Dose for Alpha Particle Dosimetry
Victor Daniel Diaz Martinez, Mélodie Cyr, Devic Slobodan, Nada Tomic, David F. Lewis, Shirin A. Enger
Medical Physics Unit - Department of Oncology - Faculty of Medicine - McGill University
Purpose: The aim of this study was to develop a Geant4-based Monte Carlo software to relate radiochromic film response to absorbed dose for alpha particle dosimetry.
Methods: A Monte Carlo-based software built on the Geant4 simulation toolkit was developed to model and simulate a 241Am sealed source (31.28 kBq, 5 mm diameter) and unlaminated EBT3 film (active layer: 28 ????m) with accurate dimensions and material description. 2 million 241Am atoms were randomly sampled inside the active source volume. The absorbed dose per voxel as well as the total dose inside the active layer of the film (63.5×50.8×0.028 mm3) was scored. The volume was divided into 0.105×0.084×0.028 mm3 voxels. The radioactive decay was handled through explicit simulation of nuclear decay.
Results: Due to the short range of alpha particles, there was no energy deposition in the film voxels outside the active region of the source. The total dose rate in the voxels covering the source was 29.42 Gy/hr.
Conclusion: The developed MC software allowed us to accurately recreate our experimental setup for radiochromic film alpha dosimetry. The obtained dose rate can be used to relate radiochromic film response to absorbed dose.
CodeX Radiation Therapy Treatment Data (RTTD): Smarter Data in the Fight Against Cancer
Purpose: Sharing of radiation therapy (RT) treatment details between patients’ care providers is critical for care coordination. Typically this is achieved by manually transcribing information from radiation oncology information systems (ROIS) into documents for distribution. But transcription is burdensome and error prone and seldom standardized, which in turn stymies data reuse. Therefore, a standardized solution that can replace manual transcription has the potential to improve care coordination and facilitate data reuse. The CodeX Radiation Therapy Treatment Data (RTTD) project is such a solution.
Methods: The RTTD project is using Health Level Seven (HL7) Fast Healthcare Interoperability Resources (FHIR) to:
- Develop, test, and deploy open data standards and APIs for interoperable, multi-purpose exchange of RT data.
- Enable ROIS to generate/share patients’ RT data.
- Make RT data readily available and displayable.
Results: Radiotherapy FHIR profiles have been modeled and defined within the minimal Common Oncology Data Elements (mCODE) Implementation Guide and RTTD Implementation Guide. Both implementation guides are being leveraged to support the transmission of RT treatment summaries. The RTTD project aims to standardize all relevant RT related data, to improve care coordination and facilitate the reuse of RT data.
Conclusion: The RTTD team aims to have RT data modeled in HL7 FHIR profiles and implemented in vendor systems to test real-time, end-to-end transfer of RT data in clinical practice. The project is implementing a multi-phase pilot for testing. This pilot will include end-of-treatment, in-progress, and prescription level details. Radiation oncology centers and vendors are invited to participate.
Finite element method analysis of parallel-plate glass vessels for use in a water calorimeter under clinical photon and electron beams
Mark D'Souza, James Renaud, Arman Sarfehnia
Ryerson University, National Research Council Canada, Sunnybrook Health Sciences Centre
Purpose: Water calorimeters (WCs) are used to determine absolute dose-to-water by measuring the radiation-induced temperature rise. The point of measurement is contained within a glass vessel (GV) filled with high-purity water to minimize endo/exothermic reactions. The heat transfer correction (kht) accounts for heat loss/gain at the point of measurement due to thermal conduction and convection. The aim was to study the effect of the GVs dimensions/position on kₕₜ in clinical applications to develop a framework to build more optimized GV designs in the future.
Methods: Finite element method software was used to simulate a parallel-plate GV inside a WC operated vertically. The top/bottom glass thicknesses along with the position of the vessel with respect to the measurement point were varied for several energies (6MV, 6MeV, 9MeV, 18MeV). kₕₜ was determined by comparing simulation results in the presence of heat transfer to an idealized simulation in the absence of heat transfer.
Results: The effect of GV dimensions on kₕₜ was minimized when the point of measurement was more than 8mm away from the front surface of the GV. Beyond 8mm, there was less variation on kₕₜ for the 6MV, 9MeV, and 18MeV beams as a function of position when front/back thicknesses were identical. For a 6MeV beam, kₕₜ was most stable when the top glass thickness was 0.7mm and the back thickness was 1.3mm.
Conclusion: Our study enables GV designs tailored to specific applications that result in a stable kₕₜ and reduces the overall uncertainty on WC-measured dose.
Did You Read the Policy? A System for Document Management and Staff Learning
Horizon Health Network
Purpose: Tracking policy and procedure documents and disseminating this information to staff can be challenging. A software application, Oncology Learning Management System (OLMS), was developed to address this challenge, with the goals to improve communication and accountability, organize staff training, and track new and revised policies within the oncology department.
Methods: OLMS is a desktop application with a backend MSSQL database. Users register through the software and select to associate themselves from a list of staff groups, site groups, committees, and/or credentials. Similarly, documents are uploaded to the database and are configured to specify which groups are required to review the document, and at what frequency. Once an uploaded document is approved, each staff required for review is notified automatically via email. Users must login to OLMS to record their reviews in the database. To facilitate training, some documents are configured with a quiz which must be completed before recording the review. A mechanism for staff to provide feedback is also integrated and is linked to the document to assist in future policy revisions. Monthly reports are sent to each user and the department managers, which include the quiz results and a dashboard highlighting results from the previous month.
Results: To date, over 500 documents have been uploaded to OLMS, resulting in 12000 document reviews by 59 staff members.
Conclusion: A survey conducted before and after the implementation of OLMS demonstrates a significant improvement in staff satisfaction with the organization and management of policy and procedure documents.
Absolute Dose Determination in Synchrotron Produced X-ray Beams: Commissioning of an Aluminum Calorimeter
Islam El Gamal, Jean Dessureault, Malcolm McEwen
Carleton University, National Research Council Canada
Purpose: The performance of a novel aluminum-based calorimeter design to measure the dose rate of synchrotron produced X-ray beams has been assessed.
Methods: The dose rate of the 65-140 keV monochromatic X-ray beams produced by the Canadian Light Source will be evaluated. The beamline produces an X-ray beam with a non-uniform field size on the order of 1 mm and a peak dose rate of 1 Gy/s. Given the small beam size, a vacuum-based calorimeter with an aluminum core was designed, matching the beam profile, and a prototype built. Measurements in 5 different monochromatic beams were used to assess the signal to noise performance of the calorimeter in addition to an upper bound for the uncertainty on model derived heat loss corrections.
Results: Thermal modeling predicts conductive heat-losses < 3 %. Preliminary measurements have shown reproducibility at the 0.01 % - 0.25 % level and agreement with our thermal model at the 0.7 % level. Additionally, the absence of non-aluminum material in the beam path results in a Monte Carlo derived attenuation/scatter correction of 1.025.
Conclusion: The results demonstrate the reproducibility of the radiation induced temperature rise for 5 different beam energies in addition to agreement with the thermal models used. The use of this calorimeter as part of a new protocol to determine the dose rate for a synchrotron beam is expected to result in a significant reduction in uncertainty relative to current methods while providing metrological traceability.
Using Random Forests to Model Nasogastric Tube Requirement Using Patient-Reported Outcomes and Treatment Factors
Jonathan Elbaz, Owen Paetkau, Demetra Yannitsos, Lisa Barbera, Wendy Smith
University of Calgary, Tom Baker Cancer Centre
Purpose: The goal of this study was to predict nasogastric (NG) tube requirement using patient-reported outcome measure (PROM) responses and pre-treatment factors.
Methods: Data from 35 patients receiving radiotherapy for head and neck cancer with a curative intent were retrospectively analysed. Data contained responses to a PROM survey given at consultation, consisting of modified MD Anderson Symptom Inventory for Head and Neck Cancer and Edmonton Symptom Assessment System Revised surveys. Other data included pre-treatment factors and planned radiation doses to organs at risk. Correlations between these metrics and NG tube requirement were considered using Pearson correlation coefficients. Random forest classification models were created using stratified k-fold cross validation with the correlated metrics. Hyperparameter optimisation was performed with random search.
Results: 5 patients required an NG tube during treatment. The highest correlation values found through Pearson coefficient analysis were age (-0.395), a PROM question asking to rate tiredness from 0 (least severe) to 10 (most severe) (0.442), and a PROM question asking to rate the problem of mucus in the mouth and throat from 0 (least severe) to 10 (most severe) (0.413). These results were statistically significant. A random forest model using these three metrics produced a model with a training accuracy of 0.79±0.03 and a cross-validation accuracy score of 0.77±0.07.
Conclusion: Patient responses at consultation to PROM questions regarding tiredness and mucus in the mouth and throat along with age were found to be prognostic of eventual NG tube requirement during treatment with a random forest model.
Effects of cell and nucleus size on microdosimetry of gold nanoparticle enhanced radiation therapy in a realistic tissue model
Elizabeth Fletcher, Martin Martinov, Rowan Thomson
Purpose: To assess sensitivity of patterns of energy deposition to cell and nucleus radii within realistic cellular targets using multiscale tissue models for gold nanoparticle dose-enhanced radiation therapy (GNPT).
Methods: A realistic multiscale tissue model is developed for efficient Monte Carlo simulations of energy deposition in subcellular targets. Approximately 550 cells with randomly assigned cell and nucleus radii are embedded in a macroscopic tissue phantom and irradiated with 20-100 keV photons. Each cell contains thousands of 25 nm radius gold nanoparticles (GNPs) in one of two configurations: in a perinuclear shell or in four spherical endosomes.
Results: At a constant macroscopic dose and photon energy, the specific energy imparted to a cell compartment (energy per mass deposited in a nucleus or cytoplasm) is largely random at doses below approximately 0.2 Gy but is correlated with cell and nucleus radius at higher doses. When GNPs are in the perinuclear configuration, specific energy deposition to the nucleus and cytoplasm decrease by factors of 2 and 3 respectively when nucleus radius and cytoplasm thickness increase by factors of 3. When GNPs are in endosomes, both nucleus and cytoplasm specific energy decrease by a factor of 4 when cell radius increases by a factor of 3.
Conclusion: At macroscopic dose levels greater than 0.2 Gy, specific energy imparted in cell compartments is strongly correlated with cell and nucleus radii. This is an important consideration for GNPT treatment planning, as all cells cannot be assumed to experience the mean population specific energy.
Commissioning a novel flattening filter free beam total body irradiation technique
Rebecca Frederick, Alana Hudson, Greg Pierce
University of Calgary - Tom Baker Cancer Centre
Purpose: To develop and implement a total body irradiation (TBI) technique using flattening filter free (FFF) beams.
Methods: FIRE-TBI (Flattening filter free IRradiation at an Extended distance for Total Body Irradiation) is an extended source-to-skin distance (SSD), anteroposterior-posteroanterior, volumetric modulated arc therapy technique. Our automated treatment planning uses the Eclipse Scripting Application Programming Interface (ESAPI) for beam placement. The algorithm uses body contour height and width at the umbilicus for field isocentre placement and arc length. In addition, standard multileaf collimator (MLC) templates are applied depending on patient dimensions.
We completed a full commissioning process to verify treatment planning system (TPS) accuracy at an extended SSD. Measurements included absolute and relative dose verifications. Patient-specific quality assurance (PSQA) procedures were tested for six standard MLC templates and two plans post-optimization. Lastly, we did an end-to-end test by creating and delivering a 200cGy FIRE-TBI plan to an anthropomorphic phantom. Internal and surface doses were monitored.
Results: FIRE-TBI automated treatment planning has solicited positive feedback from dosimetrists at our centre. Absolute and relative dose measurements for commissioning were within expected uncertainty for a beam model generated from data at 100cm SSD. All PSQA passed at a minimum of 96.1% using gamma criteria of 2%/2mm. The anthropomorphic phantom average internal dose was 206±15cGy, and surface dose was >180cGy (90% of prescription) at all measured locations.
Conclusion: FIRE-TBI is a feasible technique and implementable using a standard TPS beam modeling process. Clinical implementation is on-going as part of a prospective comparative study at our institution.
Automatic vertical couch compensation using Varian’s Developer Mode for extended field treatments
Rebecca Frederick, Devin Van Elburg, Alana Hudson, Greg Pierce
University of Calgary - Tom Baker Cancer Centre
Purpose: To dosimetrically test if Varian’s Developer Mode can compensate for six degrees of freedom (6DoF) couch vertical deflections that occur during extended field treatments like total body irradiation.
Methods: We performed measurements on one Varian PerfectPitch 6DoF couch. Three solid water blocks (33.2lbs) were set up at longitudinal positions of 83cm, 115cm, and 160cm. We delivered 150MU to an ionization chamber in each block with no deliberate vertical shift between them. This was repeated for three additional couch loads (maximum 250.4lbs). Percent dose changes relative to 160cm were compared with inverse square law (ISL) predicted changes due to 6DoF automatic couch compensation. Lastly, we delivered 150MU fields to two solid water blocks (83 and 160cm) dynamically using Developer Mode. A vertical shift was added to compensate for couch vertical deflections during the longitudinal shift between fields. These measurements were repeated for three additional loads. Dose changes over the longitudinal range were compared between dynamic deliveries with and without compensation.
Results: The maximum couch vertical deflection was 3.70mm, occurring for the lightest load. Maximum dose changes in this case were 0.70±0.02% for static and 0.85±0.04% for dynamic measurements. Agreement between ISL predicted and measured dose changes was typically <0.10% (maximum 0.14%). For all couch loads, introducing a downwards vertical motion when extending the couch improved the relative dose change to 0.01-0.21% from 0.53-0.85% uncompensated.
Conclusion: Vertical couch deflections due to automatic couch compensation are systematic and dosimetrically measurable. We found Developer Mode a viable tool to correct for these deflections.
Implementation of a stand-alone precalculated Monte Carlo code to calculate the dose deposited by alpha and beta-emitters from the Ra-224 decay chain
Mojtaba Hoseini-Ghahfarokhi, Yuji Kamio, Keyvan Jabbari, Jean-FranÃ§ois Carrier
CRCHUM, CHUM, Champlain Valley Physicians Hospital
Purpose: To develop a standalone platform to calculate the deposited energy by alpha and beta particles emitted by the Ra-224 decay chain using the precalculated Monte Carlo (PMC) method. PMC code can overcome the time-consuming process of general-purpose MC codes like GEANT4 using a precalculated database.
Methods: The energy spectra of alpha and beta particles were extracted from GAMOS 6.2.0 outputs and used to define the primary sources emitting mono-directionally in GEANT4. To generate the databases, main track information such as particle position, step number, deposited energy and track length were recorded in a text file. The precalculated database was integrated in a Matlab-based PMC code to calculate the deposited energy in a voxelized phantom. The results were benchmarked with corresponding GEANT4 standard MC calculations. Some key parameters like StepMax were optimized to yield the best agreement.
Results: Implementing proper filters on deposited energy by various types of secondary particles shows that the primary alpha particles contribute >99.9% of the cumulative dose of alpha decays. The comparison between calculated normalized deposited energy exhibited a good consistency between PMC and GEANT4 codes with differences <2% for all types of sources.
Conclusion: The good agreement between PMC and GEANT4 calculations shows the potential of the PMC method to be implemented in clinical TPS for current and future alpha and beta radionuclide applications.
How the Bladder filling Affects Dosimetry During Prostate Treatment
Amjad Hussain, Ryan Rivest, Eric VanUytven
Purpose: This study was designed to find a correlation between bladder filling and the relative dose received.
Methods: Prostate SBRT and Hypofractionation is the standard treatment at our institute. In this study, a total of 382 patients were analyzed, including 55 SBRT Prostate, 162 hypofractionated prostate ± seminal vesicles (SV), and 165 hypofractionated prostate + pelvis patients. SBRT prescription was: 35Gy, 36.25Gy, or 37.5Gy in 5 fractions. Hypofractionated prostate/SV were delivered as 60Gy in 20 fractions and prostate/pelvis as 70Gy in 28 fractions. Bladder dose constraints for prostate SBRT are: V18.1Gy<40%, V28Gy<15%, V37Gy<10cc and V40Gy<1.5%. For 60Gy/20 fractions, dose constraints are: V60Gy<5%, V56Gy<15% and V46Gy<35%, whereas for 70Gy/28 fractions, bladder constraints are: V70Gy<10%, V60Gy<15% and V50Gy<35%. Bladder doses were compared against bladder volumes.
Results: For the whole cohort the mean bladder volume was 172cc, whereas 10th and 90th percentile bladder volumes were 120cc and 468cc respectively. None of the bladder dose constraints (VD(%)) exceed the tolerance limits for prostate SBRT treatments. For Prostate/SV hypo (60Gy/20#) 3.7% (6 out of 162) patients received doses higher than the set tolerances. For prostate/pelvis hypo (70Gy/28#) dose tolerances exceeded for 13.3% (22 out of 165) patients. There was a strong correlation between the bladder volume and VD(%). The bladder volume was < 200cc for 90% (25 out 28) patients. For all the 382 patients a lower VD(%) was observed with increasing bladder volume.
Conclusion: There was a strong correlation between the bladder volume and the dose received.
Towards evaluation of motion using a radiochromic gel dosimeter in a Quasar motion phantom
Kevin Jordan, Derek Gillies, Brandon Lefevre, Stewart Gaede
London Regional Cancer Program - London Health Sciences Centre - Dept of Medical Biophysics - Western University
Purpose: To develop vessels and methods for evaluating the impact of motion management techniques using 3D gel dosimetry and a Quasar pRESP motion phantom (ModusQA).
Methods: Custom, 7.9 cm diameter by 12 cm length cylinders were prepared by cutting 0.025cm thick polyethylene terephthalate sheets, folding and welding the seam. Flat, circular ends were machined from 0.1 cm thick PETG sheets and welded to the sleeves. A 0.02 cm thick sheet of polycarbonate (PC) was wrapped around the vessel to protect the wall from abrasion while moving inside the Quasar phantom and provide coupling to the motion motor via a custom bracket. A mount was also developed for coupling the vessels to the Vista16 optical cone beam CT scanner. The gel formulation contained: 5% gelatin, 60 mM sulfuric acid, 0.5 mM tannic acid, 0.3 mM ferrous ammonium sulfate and 0.05 mM xylenol orange. The gel was placed near the linac isocentre for a single, 4x4 cm field size, x-ray beam delivery perpendicular to the direction of motion. Individual samples were irradiated for the static and sinusoidal motion cases
Results: Static dose distributions agreed with the treatment planning system <2% along the central axis and cross profiles showed good agreement. The optical CT reconstructions for the static and motion irradiations recorded the impact of motion on the 3D dose distribution and demonstrated dose smearing.
Conclusion: This work shows it is possible to assess the impact of motion during radiotherapy of a 3D dose distribution.
Near-surface measurements of orthovoltage x-ray beam using a thin water tank
Kevin Jordan, Derek Gillies, Homeira Mosalaei
London Regional Cancer Program - London Health Sciences Centre - Dept of Medical Biophysics - Western University
Purpose: Comparison of 100 kVp x-ray, 0-5 mm depth doses measured with a thin water tank or a set of polystyrene sheets placed above a parallel plate Markus ion chamber.
Methods: Scanning water tank measurements for depths 5-100 mm were obtained with a cylindrical IBA CC13 ion chamber. A Markus chamber with a waterproof cap was used in water for 1-20 mm depths. For 0-5 mm readings, either a set of 0.48 mm thick, polystyrene sheets or a thin water tank was placed on the bare Markus chamber mounted flush in a thick PMMA plate. The thin water tank, dimensions 100 x 100 x 20 mm, was constructed by folding a 0.25 mm thick polycarbonate sheet. Soap was added to lower water surface tension to obtain uniform layers > 0.5 mm thick.
Results: Data sets were normalized at 5 mm depth. Reproducible measurements were obtained until the waterproof cap was within 1 mm of the water surface, effectively limiting measurements to 2 mm depth. While the polystyrene sheet data was distinctly different, the thin tank data agreed with the extrapolated CC13 and Markus in-water data.
Conclusion: Without these measurements, surface dose estimates required linear extrapolations of 3% and 13% were required from the Markus and CC13 in-water measurements, respectively. The thin water tank is a simple and effective phantom for near surface dosimetry in the horizontal plane.
Considerations in left anterior descending artery (LAD) toxicity in large volume breast radiotherapy
Tania Karan, Dr. Louise Wade, Dr. Alanah M Bergman, Tara Menna, Dr. Cheryl Duzenli
BC Cancer - Vancouver Centre
Purpose: VMAT can provide an alternative to standard 4-Field breast radiotherapy. A Carbon-fibre Adjustable and Re-usable Accessory (CARA) is being investigated to support large breast volume patients. This VMAT planning study examines dosimetry to the left anterior descending artery (LAD) for breasts treated with and without support.
Methods: The planning study included five patients previously treated with the 4Field technique using the CARA device. Four of the patients were planned on DIBH scans and one was planed on a free-breathing scan. Every patient had a CT scan with and without CARA per study protocol. VMAT plans were generated on both CARA and non-CARA scans. VMAT optimisation volumes were generated based on the original 4-Field beam-edges. The LAD was contoured following peer-reviewed cardiac atlas guidelines. Plans were optimised with constraints on the heart only, or specifically to reduce LAD doses. LAD V15Gy<10%, V30Gy<2% and V40Gy<1% constraints were assessed, based on radiation toxicity described in literature.
Results: In 4-Field plans, LAD constraints were met in the CARA and non-CARA arms for DIBH patients. No VMAT plan exceeded the V40Gy constraint. Without explicit optimisation objectives, the V15Gy was exceeded in all cases and V30Gy was exceeded in 4/5 patients planned with CARA. Once the LAD was introduced into the optimisation, V15Gy and V30Gy were successfully reduced to meet constraints in all plans.
Conclusion: For VMAT breast planning, exit dose can result in LAD doses exceeding dosimetric constraints for toxicity. Care should be taken to contour and optimise the LAD to minimise dose.
Therapeutic applications of 2.5 MV photon beam in radiation therapy
Navid Khaledi, Bonnie Yao, Rao Khan, James L. Gräfe
Department of Physics - Ryerson University, Howard University
Purpose: Dose enhancement with high-Z nanoparticles in megavoltage radiation therapy is difficult due to the energy dependence of the photoelectric effect. In this study we investigated the possibility to achieve clinically relevant treatment plans for a 2.5 MV photon beam. The 2.5 MV beam has energy bridging the gap between kV and MV beams.
Methods: An open-source research treatment planning system, matRad, was commissioned using measured 2.5 MV beam data. Treatment plans were produced for prostate, liver, nasal cavity, Glioblastoma Multiforme (GBM), and orbit cases. The doses to planning target volumes (PTVs) and normal tissues as well as integral doses were computed. The results were compared with a conventional 6 MV photon beam.
Results: The difference between the dose to 98% of PTV volume (D98%) for 2.5 MV and 6 MV beams was less than 3% in all cases. For some shallow tumors, such as GBM, the dose to healthy tissues such as the optic nerve and brain was 19% and 7.5% less than 6 MV, respectively. For the nasal cavity and orbit, the doses of optic chiasm and brainstem were lower than 6 MV plans. However, the integral dose for deep seated tumors, such as the prostate and liver, was 12.1% and 11.4% higher than 6 MV, respectively. For the remaining cases the integral dose was 6 to 19% lower than the 6 MV.
Conclusion: The results show that 2.5 MV beam can provide an acceptable coverage to the target while keeping normal tissues below the critical limits. The results of this study motivate future work investigating dose enhancement with high-Z nanoparticles.
Comparison of in-house and RayStation robust optimization methods for skin-flash in breast cases
Rafael Khatchadourian, Caroline Duchesne
CIUSSS de l'Est de l'île de Montréal
Purpose: The skin-flash method is used to account for anatomical changes such as swelling during treatment in order to preserve effective PTV coverage. We compared plans obtained with an in-house developed skin-flash method to those obtained with the robust optimization tool of the RayStation (RS) treatment planning system (TPS).
Methods: Twelve breast patients were planned using both approaches. The in-house method consists of extending the patient body up to 2 cm beyond the skin contour around the tumour and overriding it with a lung density during optimization. MLCs thus conform to and irradiate the area of possible tumour extension during multi-fraction treatments. Robust optimization simulates a set of deformed images, each corresponding to the motion of the PTV in the anterior or ipsilateral direction and with an amplitude of 2 cm. The optimizer will then seek a robust solution for all scenarios.
Results: On the planning CT, robust plans were comparable to flashed plans with the marked difference that contralateral breast in the former was systematically hotter by at most 4Gy. Robust plans had also a larger hot-spot volume (107% of prescription) and a higher mean dose to heart by an average of 0.5Gy for left breasts. Comparison of flashed and robust plans on deformed CT (simulating swelling) did not reveal notable differences except for the former having generally higher max dose values.
Conclusion: Given the absence of a significant dosimetric advantage in robust plans and their higher computationally intensive nature, we decided to preserve our in-house flash method for breast cases.
Independent Assessment of a Commercial Automated Contouring Software for Lung SBRT
Charles Kirkby, Hong-Wei Liu, Abhijit Ghose, Petra Grendarova, Stephanie Seberg, Xinzhou Li, Conor Shaw
University of Calgary - Jack Ady Cancer Centre, Central Alberta Cancer Centre, Grande Prairie Cancer Center
Purpose: Automated contouring of CT images for radiation therapy planning has the potential to save time, increase patient throughput, reduce repetitive tasks, and reduce errors. In this work we independently assess one commercial convolution neural network-based solution against contours from physician experts from multiple clinics for lung SBRT patients, the goal: to discern whether the AI software generates contours consistently and within inter-expert variability.
Methods: A set of 10 CT data sets were anonymized and fed into a trial version of the LIMBUS Contour (ver. 1.4, Limbus AI Inc., Regina, SK), which generated a contour set (aorta, brachial plexus, esophagus, heart, kidneys, liver, lungs, bronchial tree, trachea and spinal canal). Three radiation oncologists from community oncology centers in Alberta independently drew those same contours on copies of the CT data sets, blind to the AI results, under typical clinical conditions (Varian Contouring ver. 15.6, Varian Oncology Systems, Palo Alto, CA, USA). Dice Similarity Coefficient (DSC), 50% Hausdorff Distance (HD) and 95%HD quantified contour overlap with reference curves set as either the AI or an expert. Paired t-tests with Bonferroni correction were used to identify statistical significance
Results: The AI performed well. Worst case differences in means of DSC, 50%HD, and 95%HD were 6.2%, 0.9 mm, and 2.2 mm respectively. While statistically significant differences arose, differences were not considered clinically significant.
Conclusion: Our assessment indicates that LIMBUS Contour generated contours within a clinically acceptable margin inter-operator variability and can serve as a baseline for OAR contour generation for SBRT lung patients.
Automatic catheter modelling in 3D transrectal ultrasound images from high-dose-rate prostate brachytherapy using a deep learning and feature extraction pipeline
Nicole Kitner, Jessica R. Rodgers, Tamas Ungi, Martin Korzeniowski, Tim Olding, Chandra P. Joshi, Parvin Mousavi, Gabor Fichtinger
Queen's University, Kingston Health Sciences Centre
Purpose: Manually segmenting catheters for ultrasound-based prostate high-dose-rate brachytherapy (HDR-BT) is a laborious process that is subject to human error and variability, impacting dosimetric outcomes. A tool that automatically tracks catheters in 3D TRUS images would provide a secondary check to the manually tracked catheter positions.. We propose a novel two-step pipeline involving a deep learning approach followed by a feature extraction technique.
Methods: The three-dimensional (3D) transrectal ultrasound (TRUS) images and their corresponding ground truth catheter positions were obtained for 97 prostate HDR-BT patients treated at Kingston Health Sciences Centre between 2017 and 2021. The ground truth catheters were manually tracked by medical physicists during procedures. After pre-processing, 77 images were exported for training with the 3D U-Net deep learning architecture to learn to automatically identify catheters in the 3D TRUS images. The resulting predictions from the deep learning model were then transformed using the 3D Hough transform algorithm for line detection.
Results: The proposed pipeline correctly identified (within 3 mm) the positions of 321 of the 343 catheters while identifying 21 false positives. The average distance between the ground truth and predicted positions was 1.56 ± 0.16 mm, with the average maximum distance being 2.01 ± 0.07 mm.
Conclusion: The proposed deep learning model in combination with the 3D Hough transform provides clinicians with a tool to reduce time spent on verification of initial manual catheter trackings, and equips clinicians with a method of reducing uncertainties and improving clinical workflow during procedures.
Validation of a Treatment Planning System Model for Total Body Irradiation
Anastasia Kolokotronis, Malik Brunet-Benkhoucha, Étienne Roussin
Purpose: To implement a step and shoot lateral-field technique for total body irradiation (TBI) in our clinic, validation of RayStation’s treatment planning system’s (TPS) ability to accurately compute dose at extended source-to-surface distance (SSD) was required.
Methods: Percentage depth dose (PDD), in-plane, cross-line, and diagonal profile measurements were performed using a CC13 ionisation chamber and a microdiamond detector. Dose-to-surface and build-up regions were measured with a parallel-plate chamber. Acquisition measurements were performed using an 18 MV photon beam laterally directed towards a Blue Phantom 2 IBA water tank and various solid water configurations at an extended SSD of 365 cm. A beam spoiler was positioned at various distances to evaluate the optimal distance to achieve ideal dose-to-surface.
Results: The existing 18 MV photon beam model developed under standard SSD conditions accurately predicted extended SSD data. The CC13 PDD and profile measurements fit with less than a 2% difference for fields of various sizes in comparing to TPS results. Diagonal measurements were added in the model, providing a more accurate representation of the beam profile. An optimal beam spoiler distance of 10 cm from the phantom resulted in a surface dose of 83.8%. The beam spoiler was added as an ROI in the TPS, which was also found to give the best model to measure agreement.
Conclusion: The TPS model for an 18 MV beam under normal operating SSD conditions can be used to perform dosimetry plans for patients at extended SSD, up to 365 cm.
Validation of Calypso Surface Beacons for DIBH Radiation Therapy
Sashika Kumaragamage, Boyd McCurdy, David Sasaki
University of Manitoba, CancerCare Manitoba
Deep Inspiration Breath Hold (DIBH) is a method used in radiation therapy (RT) to deliver adequate radiation dose to superficial targets such as the breast while limiting cardiac dose and resulting cardiotoxic effects. In this study, we explore the use of the Calypso electromagnetic array and small transponders known as surface beacons to track the chest wall in 3D and in real time. The study consists of three major components: a phantom study in which the tracking capabilities of the system are tested with precisely controlled mobile phantoms, the effects of electromagnetic interference were assessed using multiple metallic objects near the surface beacon, and a volunteer study where five individuals are coached through twenty DIBH cycles with surface beacons fastened to their chest wall. Total errors below 1 mm were attained in the phantom study when the beacon was moved through step-function and sinusoidal motion. Eddy current interference was not found to produce statistically significant changes in tracking accuracy despite being in the vicinity of small steel spheres, copper plats, or a prosthetic hip. The volunteer study showed wildly varying accuracy in achieving DIBH target displacements and in consistency between sessions; however overall errors remained below 5 mm across all volunteers. As such the Calypso surface beacon shows promise as a potential safe method for tracking DIBH for RT.
An Automated Treatment Plan Alert System to Safeguard Cancer Treatments in Radiation Therapy
Paul Kump, Junyi Xia, Erwei Bai, Sridhar Yaddanapudi, Keith Wachowicz, Nawaid Usmani, Don Yee, Jordan Wong, Gino Fallone, Jihyun Yun
SUNY Maritime College, Infondrian LLC, University of Iowa, University of Alberta, BC Cancer
Purpose: Severe patient harm can occur in radiation oncology due to treatment errors, for example, when delivering treatment using corrupted treatment plan data. To safeguard radiotherapy treatment, a novel machine learning (ML)-based treatment alert system is proposed to automatically check if a given treatment plan is consistent with its intended use for prostate, head and neck, breast cancer treatments.
Methods: For each treatment plan, a heat map is computed by utilizing plan parameters such as plan apertures and planning dose. Then, the heat map, which is characterized by a high-dimensional vector, is reduced to a much lower dimension using an ML-based dimension-reduction technique. The extracted information from this two-step algorithm is used to train and test several ML classifiers with real clinical treatment plans from a hospital in the United States. Random forest, softmax regression, k-nearest neighbors, shallow neural network, and support vector machines—each with several different tunings/architectures—are each evaluated under a 5-fold cross-validation scheme.
Results: Dimensionally reduced vectors are so well clustered among similar intended uses that classifiers can predict the intended use with an accuracy of about 91%, demonstrating the efficacy of the proposed method for extracting treatment plan information. Among all the tested classifiers, the random forest classifier makes the most accurate predictions.
Conclusion: Each plan’s predicted use, once it is returned by the classifier, can then be compared with its actual documented intended use, and a warning flag raised if the predicted and actual uses are inconsistent with each other, possibly reducing treatment errors.
Varian HyperArc® Delivery QA: Using 3D Polymer Gel Dosimetry (PGD) with CBCT readout to quantify inter-LINAC reproducibility
Tenzin Kunkyab, Derek Hyde, Michelle Hilts, Andrew Jirasek
The University of British Columbia Okanagan, BC Cancer Kelowna
Purpose: 3D polymer gel dosimetry (PGD) is a promising tool for dose verification for highly complex radiation therapy techniques. The aim of the current study is to illustrate the potential and reproducibility of a novel PGD technique using CBCT read-out for clinical 3D dosimetry by investigating the precision of HyperArc® (Varian) SRS delivery on two matched Truebeam units using a common SRS treatment plan.
Methods: Two PGDs were manufactured using previously established protocols. Three targets were defined to mimic an SRS treatment plan. The PGD experiments (irradiation and CBCT imaging for dose read-out) were conducted using two LINACS (Truebeam 1 and 2). A calibration curve obtained from Truebeam 1 was used to calibrate both sets of gel CBCT images. 3D gamma analysis was used to compare planned (Eclipse®, Varian) and resulting CBCT PGD measured dose distributions using dose criteria of 5% 1mm, 3% 1mm and 2% 1mm.
Results: 3D gamma were >95% for all measured criterion: 99.1%, 97.1%, 95.1% were obtained for Truebeam 2 and 99.9%, 98.5%, and 96.6% for Truebeam 1 (self-calibrated results) for 5% 1mm, 3% 1mm and 2% 1mm respectively, using a dose threshold of 30%. The dose profile along the x, y and diagonal direction through the SRS targets shows good matching against the Eclipse® planned dose.
Conclusion: The results from the current study showed that, when manufacturing conditions were carefully controlled, PGD can be used to quantify the reproducibility of an SRS delivered 3D dose distribution delivered on matching LINAC treatment units.
Evaluation of Deep Learning Algorithms to Segment Non-Small Cell-Lung Cancer (NSCLC)
Tenzin Kunkyab, Derek Hyde, Zheng Liu, Zhila Bahrami
The University of British Columbia Okanagan, BC Cancer Kelowna
Purpose: Manual contouring by a Radiation Oncologist remains the gold standard for the segmentation of gross tumour volume (GTV). An accurate and efficient segmentation algorithm could be beneficial, as it would save time and reduce observer variability. The current study aims to compare four different deep learning algorithms to segment GTV on CT images of NSCLC patients. The application of these deep learning algorithms is twofold: 1) Use them as a part of an automated radiomics workflow; and 2) potentially use them for computer-aided detection.
Methods: 563 NSCLC patients from The Cancer Imaging Archive (TCIA) website were used for training and validating the deep learning models. This research examined a number of different deep learning model architectures, including U-Net, Attention U-Net, UNETR, and CoTr. The deep learning models were trained (500 epochs) and validated using 563 NSCLC patients from The Cancer Imaging Archive (TCIA) website. As an independent test set, a dataset of 91 patients from the clinical database was used.
Results: This study showed that the UNETR had the best dice correlation coefficient compared to the other three algorithms tested: U-Net (0.66), Attention U-Net (0.68), UNETR (0.78), and CoTr (0.72).
Conclusion: The deep learning segmentation algorithm UNETR is capable of producing accurate segmentation for an automated radiomics workflow. Further testing is required for implementation as computer-aided detection.
Multi-step correction-based algorithm for conversion of EPID images into water-equivalent images
Ivan Kutuzov, Boyd McCurdy
Purpose: EPID is an instrument that can be used in clinical dosimetry. There have been attempts to use EPID as an absolute water-equivalent detector. EPID signal is generally proportional to the incident x-ray fluence. However, it is influenced by several factors such as differential off-axis response, individual pixel sensitivity variation, and noise. The purpose of this study was to identify contributions of these factors and remove them from the image, to transform EPID images into water-equivalent images.
Methods: Pixel sensitivity matrix and differential off-axis detector response were evaluated. Series of rectangular fields were delivered with incremental lateral displacement of the EPID detector, and their images were acquired. Then the ratios of pixel signals were calculated to assess individual pixel sensitivity variation. Differential off-axis detector response function was calculated using ratio of the EPID open field image with pixel sensitivity difference removed, and the open field image measured using a reference water-equivalent ion chamber array.
Series of ten modulated IMRT fields were used to assess performance of the developed conversion algorithm. The images from the test fields were acquired using both EPID and ion chamber array. Then EPID images were processed using previously found pixel sensitivity and off-axis correction functions and compared against the reference images using chi-comparison.
Results: Chi-comparison using 2%/2mm passing criterion resulted in at least 95% passing rate for all test images.
Conclusion: The algorithm that converts EPID images into water-equivalent images was developed and validated and can potentially be used in clinical dosimetry or for research purposes.
Update from AAPM TG263U1
Renée Larouche, Charles Mayo, Laurent Tantot, Ying Xiao, Elizabeth Covington
Centre hosptialier de l'Université de Montréal (CHUM), University of Michigan, Centre intégré universitaire de santé et de services sociaux de l'Est-de-l'Île-de-Montréal - Hôpital Maisonneuve-Rosemont, University of Pennsylvania
Purpose: Standardized nomenclature for structure names and dosimetric data allow for easier data aggregation, improvements in efficiency and practice quality, and enable big data efforts in radiation oncology. The American Association of Physicists in Medicine (AAPM) Task Group 263 published nomenclature guidelines to assist with these goals in January 2018 and an update is currently underway to further expand the radiation oncology nomenclature standard.
Methods: AAPM Task Group 263U1 began in February 2021 with the following charges: (1) solicit feedback about implementation of the initial report; (2) update the TG-263 worksheet to concepts not covered in the initial report; (3) provide translations in French and Spanish and; (4) inform how standardized nomenclature can mitigate errors using failure modes and effects analysis (FMEA). Members from the French and Spanish working groups have worked to translate approximately 700 structure names and ensure consistency with the initial report. Surveys regarding usage of TG-263 were sent to all AAPM, American Society for Radiation Oncology (ASTRO), and American Association of Medical Dosimetrists (AAMD) members.
Results: A survey conducted in February 2022 resulted in 692 responses. Most notable, 42.4% of respondents worked in clinics that had not adopted TG-263 nomenclature guidelines. Discussions are underway on proposed changes and additions to the proposed nomenclature guidelines. Translations in French and Spanish are complete. Work on the FMEA is ongoing.
Conclusion: Implementing TG-263 guidelines now will aid future efforts in the field. Efforts are underway to increase knowledge of TG-263 guidelines, as well as the future TG263U1 guidelines.
Optimization of polymer-embedded fiber Bragg gratings dosimeter for radiotherapy
Marie-Anne Lebel-Cormier, Tommy Boilard, Martin Bernier, Luc Beaulieu
CHU de Québec - Université Laval
Purpose: In order to minimize the noise and maximize the dose response of our polymer-embedded fiber Bragg gratings (FBGs) dosimeter, we set out to characterize the impact of the gratings writing parameters and the UV glue holding the fiber in its polymer envelop.
Methods: The standard deviation over a 10 sec acquisition (10,000 values) for FBGs having lengths varying between 0.5 and 7.0 mm for a transmission intensity relatively constant (between 1.16 and 1.80 dB) was characterized. We also measured the standard variation for a 10.0 mm FBG having a transmission intensity ranging from 0.26 to 3.17 dB and a 1.0 mm FBG having a transmission intensity ranging from 0.16 to 1.91 dB. To characterize the impact of the UV glue, 4 detectors using different UV glue (DSM-Desotech 950-200, OVATION 1.54-2K, NORLAND OPTICAL ADHESIVE 68) and Krazy Glue. The previous version of the detector used DSM-Desotech glue.
Results: The gratings writing parameters that minimize the noise are 2 or 3 mm long FBG having a transmission intensity around 2 or 3 dB. An increase of over 20 % is also obtained for the detector response using OVATION compared to DSM Desotech UV glue. Similar dose responses are obtained for DSM Desotech UV glue and Krazy Glue.
Conclusion: Using the OVATION glue will allow a 20% increase signal compared to the previous version of the dosimeter. Furthermore, the previous version already had writing parameters similar to the optimized ones obtained here. Hence, no improvement in signal to noise ratio is expected.
Novel Dosimetric Method for Dosimetric Analysis of the Multiple Nucleus-Targeted AuPd103 Radioactive Nanoparticles for a Radionuclide Treatment
Jason Lee, Carl Kumaradas, Stephen McMahon
Ryerson university - Odette Cancer Center, Queen's University Belfast
Purpose: Develop a dosimetric method to simulate the dose distribution and the biological effect of nucleus targeted AuPd¹⁰³ nanoparticles using a local effective model (LEM) and Voxel Dose Kernel (VDK) convolution.
Methods: The VDK of electrons, photons from AuPd¹⁰³NP decays, and gold shell ionization were calculated using a Monte-Carlo simulation(MC). A Cubic water phantom with 10µm sidelength was created and 1000 AuPd¹⁰³NPs were randomly distributed. The total number of gold shell ionization events and AuPd¹⁰³NP decays in a phantom for a target dose were calculated and randomly assigned to NPs in a phantom. VDKs were convoluted to each NP location to create a three-dimensional dose and isodose map. For validation purpose, an isodose map and three-dimensional dose of the same geometry was produced using the direct MC simulation method. Using the same method, a three-dimensional dose map of the cell with either targeted or nontargeted AuPd¹⁰³NPs was calculated. Dose-response was calculated using a LEM. RBE of NPs was calculated compared to the Pd103 photons.
Results: The difference in an isodose map and the average dose of the phantom was less than 5%. The computation time for MC simulation and the convolution method was 30 hrs and 40 minutes, respectively. RBE of targeted and nontargeted AuPd¹⁰³NP were 4.59 and 3.47.
Conclusion: We showed three-dimensional VDK convolution combined with a LEM can predict the biological effects of multiple radionuclide nanoparticles with an acceptable uncertainty. The nucleus targeted AuPd¹⁰³NP demonstrated the potential to be a great candidate for radionuclide brachytherapy source.
A Step Towards Personalized CT Dose Profiles
Ronan Lefol, Yannick Lemarechal, Philippe Despres
Purpose: A limitation to patient-specific Computed Tomography (CT) dose profiling has been the lack of computational tools permitting automated organ segmentation and personalized dose quantification in a timely manner. The continuous development of GPU-enabled algorithms opens the door to rapid, personalized dose calculation in diagnostic imaging. This study aims to demonstrate the feasibility of a GPU-centered workflow combining a convolution neural network (CNN) for multi-organ segmentation and a fast Monte-Carlo tool (GPUMCD) for CT dose calculation.
Methods: A CNN based on UNet architecture was adapted for multi-organ segmentation. The 2017 Lung CT Segmentation Challenge (LCTSC) dataset, which contains 60 thoracic, professionally segmented, DICOM images was utilized as a primary development source. A five times cross validation was used over the dataset for training. The Dice score (DSC) was utilized as a final scoring metric; A DSC of 1 representing a perfect overlap. Additionally, a set of partial image predictions was performed in order to improve the segmentation quality.
Results: The mean DSC were obtained as 0.75 (spinal cord), 0.95 (right lung), 0.91 (left lung), 0.91 (heart), and 0.6 (esophagus) for the LCTSC dataset. Additional scoring was performed with the same network architecture with mean scores of 0.85 (Aorta), 0.9 (Liver), 0.82 (Bladder).
Conclusion: This work demonstrates the feasibility of using automated contouring as a component of a personalized CT dose calculation tool. Combined with a fast Monte Carlo dose calculation tool such as GPUMCD, this segmentation pipeline can generate patient-specific dose-to-organ profiles to fuel epidemiological studies in diagnostic imaging.
A digital infrastructure to foster research data management good practices in medical imaging
Yannick Lemarechal, Samuel Ouellet, François Pelletier, Ronan Lefol, Luc Beaulieu, Philippe Després
Purpose: Sane research data management (RDM) is increasingly promoted and required by funding agencies and international initiatives. The Cancer Imaging Archive (TCIA), for instance, enforces the use of DICOM for data submission, and some projects hosted by TCIA are starting to fully exploit the potential of this standard. Unfortunately, tools and platforms supporting RDM are not widely available or fine-tuned for specific purposes. To address this in medical imaging, a robust digital infrastructure was built to securely manage and analyze data with full traceability and FAIR principles in mind.
Methods: The infrastructure, relying on open-source assets, offers DICOM-compliant storage services (Orthanc), an authentication and authorization frontend (Kheops), a S3-compatible storage volume (RedHat Ceph), a job dispatcher, an event bus (RabbitMQ), and a cluster to execute high-performance computing (HPC) tasks. The entire software stack is deployable as containers (Docker) managed by an orchestrator (OpenShift), and hosted in an institutional Tier 3 data center.
Results: The digital infrastructure was successfully used to implement an automated Monte Carlo dose recalculation pipeline, where images are ingested, dose calculated and results stored back as DICOM objects. The entire data lifecycle is preserved in DICOM objects (e.g. simulation parameters, code version used) for full traceability.
Conclusion: A digital platform was developed to support RDM in medical imaging, and successfully used to automate a Monte Carlo pipeline. Other use-cases, including ML-based segmentation pipelines, are being implemented with the same philosophy: enforcing FAIR principles through purposely designed digital infrastructures and tools.
Sensitivity Analysis of Non-Coplanar Trajectory Radiotherapy to Intrafractional Motion
John Lincoln, R. Lee MacDonald, Alasdair Syme, Christopher G. Thomas
Dalhousie University, Nova Scotia Health
Purpose: To quantify the sensitivity of 4Pi non-coplanar trajectory radiotherapy to translational intrafractional patient motion using a dose-surrogate cost function methodology.
Methods: An example patient undergoing APBI (left-sided) was chosen for this study. A depth-based cost function was performed for every couch and gantry angle combination, yielding dose-surrogate maps for PTV and two OARs (heart and ipsilateral lung). Translational motions in three dimensions were simulated by applying lateral, longitudinal, and vertical shifts (±5 mm, ±10 mm, and ±20 mm). These shifts were applied to the patient’s isocenter. Dosimetric impact of the motions were quantified by calculating the difference between a cost function map before applying any motion (unshifted) and a new cost function map with an applied user-specified translation (shifted). This was repeated for all translations, always relative to the unshifted case.
Results: Dose-surrogate differences ranged from 0.1% to 35.2%. Lateral shift analysis estimated differences of 1.2% per mm for the PTV. Longitudinal and vertical shift analysis estimated differences of 0.1% per mm and 1.5% per mm, respectively, for the PTV. When considering the OARs, lateral, longitudinal, and vertical shift analysis estimated dose differences of 1.2% per mm, 0.03% per mm, and 1.7% per mm.
Conclusion: This methodology offers a way to quantify the sensitivity of 4Pi non-coplanar trajectory radiotherapy to positional uncertainties due to translational intrafractional motion.
Automated thyroid nodule detection and segmentation in ultrasound images using contrastive unsupervised representations learning
Ningtao Liu, Claire Park, Jaron Chong, Aaron Fenster, Shuiping Gou
Robarts Research Institute - Western Uinversity, Department of Medical Imaging, Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education - School of Artificial Intelligence
Purpose: Thyroid cancer is one of the most common cancers with rapidly increasing incidence. Currently, diagnosis of thyroid nodules using ultrasound (US) images is time-consuming, labor-intensive, and dependent on the experience of the radiologist. Therefore, deep learning-based computer-aided diagnosis (CAD) methods have been implemented to improve diagnostic evaluation of thyroid cancer. However, these methods require large-scale datasets with medical image annotation, which relies on the experience of radiologists for interpretation. We propose a novel patch-scale augmented self-supervised method (PASS) trained on unlabeled US images to improve the performance and convergence speed of thyroid nodule detection and segmentation.
Methods: The method is based on contrastive representation learning with a patch shuffling image augmentation method. The representation of the thyroid US images, learned from the self-supervised model, were used as pre-trained weights, then transferred to the baseline network.
Results: We evaluated the model on half of the MICCAI TN-SCUI2020 dataset. Compared with the randomly initialized baseline detection model (Faster R-CNN1), our method improves the mean average precision from 57.63 to 61.91 (7.43% improvement), which improves the performance of the baseline model trained on the full training set (mean average precision=61.84). See the supplementary for detailed performance comparison. Furthermore, the model initialized with the pre-trained weights converged faster (~10000 iterations less).
Conclusion: We developed the novel PASS to represent the features of US images. Our method shows the capability to improve thyroid nodule detection and segmentation performance, and the convergence speed during training, showing potential utility to improve diagnostic efficiency in clinical workflow.
Dose Reconstruction of Control-Point Specific Intrafraction Motion during Linac-Based Stereotactic Radiotherapy (SRT)/Radiosurgery (SRS)
Lee MacDonald, Mark Ruschin
Nova Scotia Health, University of Toronto
Purpose: Stereotactic radiotherapy/radiosurgery (SRT/SRS) employs highly conformal dose distributions, requiring a high degree of accuracy to be safely delivered. Recent publications have described the degree of intrafraction motion using modern radiosurgery immobilization . The purpose of the present study is to develop and evaluate a workflow for quantifying the dosimetric sensitivity to intrafraction motion for linac-based SRT/SRS plans.
Methods: Multi-target SRS patients were planned with two separate techniques: per standard clinical workflow with VMAT, and an intraarc binary collimation (iABC) conformal technique. Synthetic motion traces were used to simulate patient motion, informed by previously published translationally modelled characterization of 3D intra-fraction motion of patients immobilized using thermoplastic masks. For seven cases, with two techniques, one hundred synthetic traces were incorporated into the distribution and the dosimetric change relative to the plan without motion was quantified.
Results: The average loss in target coverage with intrafraction motion incorporated was 1.94 ± 0.94% (range -0.19 to 7.83%) and 0.64 ± 0.55% (range 0.20 to 3.23%) for VMAT and iABC, respectively. iABC had significantly less coverage loss with intra-fraction motion as compared to VMAT in 80% (20 of 24) of targets simulated. The average maximum OAR dose difference was 0.92Gy and 1.24Gy for VMAT and iABC, respectively.
Conclusion: By incorporating intra-fraction motion into dose distributions for linac-based plans, changes in plan quality can be quantified. The degree of modulation showed a strong correlation with coverage loss. This workflow can help inform optimization of dose distribution with dose fidelity with plan modulation prior to treatment.
Comparison of AAA and AXB Dose Calculation Algorithms for Clinical Thoracic SABR Spine Treatment Plans
Nicholas Majtenyi, Ernest K. Osei, Daniel N. Glick, Johnson Darko
Grand River Regional Cancer Center
Purpose: The purpose of this work was to evaluate the dosimetric impact of AAA and AXB clinical treatment planning algorithms on the computation of dose in thoracic SABR spine treatment plans.
Methods: The CT dataset of nine patients who were previously treated with SABR (24 Gy in 2 fractions, 10 MV FFF, 2400 MU/min) for thoracic spinal cancer were retroactively selected for this study. Treatment plans were created in the Eclipse TPS using both AAA and AXB and the same optimization criteria. Generated plans met our institutional guidelines for CTV and PTV coverage and normal organ dose constraints. Dosimetric parameters and DVH characteristics (e.g. maximum and minimum dose, D5, D95, V100) were extracted for target volumes and organs-at-risk and were statistically compared using a paired, two-tailed t-test.
Results: Maximum dose, mean dose, D2, and D5 for each tumour volume were significantly different (p ≤ 0.05) between AAA and AXB plans. Additionally, the CTV and PTV V99 and V100 values from AXB plans were 2.6-4.9% higher (p ≤ 0.01). The maximum dose to the spinal cord PRV was nearly equivalent between AAA and AXB, while the maximum dose to esophagus was significantly higher for the AXB plans but met tolerance criteria.
Conclusion: A significant difference in dosimetric parameters of tumour volumes was observed between AAA and AXB in SABR spine treatment plans. Thus, careful consideration should be taken into account regarding the choice of algorithm to be used for the creation of SABR spine treatment plans.
Variability in Beam Quality Measurements on Mammography Units in a Screening Study
Aili Maki, James Mainprize, Martin Yaffe
Sunnybrook Research Institute
Purpose: Our group developed a physics QC program for the mammography and tomosynthesis equipment used for a large breast cancer screening study. HVL and tube output measurements from the physicists’ equipment surveys are collected in a database for use in estimating radiation doses. To understand variation in dose estimates, we examined the statistics of these measurements.
Methods: Physicists’ survey reports were obtained for all mammography equipment being used. HVL and tube-output measurement information from these reports is tabulated in a database, and includes make and model of the mammography equipment and dosimeter, kV and target-filter combination. An analysis of variance was performed on the measurements made at the most commonly measured target-filter-kV combinations. Population and group means and standard deviations were calculated and tested for statistically significant differences (p value < 0.05).
Results: To date, there are 5,128 HVL and 6,235 output measurements, made on 280 mammography machines, (7 models) using 5 makes of dosimeter. Some statistically significant differences were observed in HVL (mm Al) and tube output (mR/mAs) values between dosimeters makes.
Conclusion: Many factors may contribute to the observed variability in the measurements. These include differences in measurement technique and dosimeter technology, as well as actual variability in x-ray beam characteristics between individual machines. Preliminary analysis suggests that some of this variability may be due to dosimeter make. Care should be taken when using measured values to estimate dose. Further work is required to determine how much of the variability is due to differences in beam quality.
Estimating the relative biological effectiveness of neutron radiation for inducing clustered DNA damage via Monte Carlo simulation of direct and indirect action
James Manalad, Logan Montgomery, John Kildea
Research Institute of the McGill University Health Centre, Cancer Centre of Southeastern Ontario, Gerald Bronfman Department of Oncology
Purpose: Exposure to ionizing radiation can induce stochastic biological effects in the human body. For neutrons, the risk of these biological effects is energy dependent and previous Monte Carlo studies have linked this dependence with the relative biological effectiveness (RBE) of neutrons to induce difficult-to-repair clusters of DNA damage (CDDs). However, these studies have only modeled direct radiation action, excluding the potentially influential effects of indirect action. In this project, we investigated the effects of indirect action on CDDs and estimated the CDD-induction RBE of neutrons due to the combined effects of direct and indirect action.
Methods: An existing Monte Carlo track-structure pipeline of our group was updated to include indirect action. Following this update, we simulated the irradiation of our custom nuclear DNA model (built using TOPAS-nBio) by monoenergetic neutrons and reference X-rays. The resulting CDDs were analyzed and compared to determine energy-dependent neutron RBE.
Results: With indirect action included, we found a significant increase in CDD yields, cluster length, and lesion count per cluster. Our estimated energy-dependent neutron RBE followed the same trend as previous findings, but is lower in magnitude because the relative damage impact of indirect action was greater for X-rays than for neutrons.
Conclusion: We have demonstrated the significant influence of indirect action in radiation-induced CDDs. Our findings suggest that the energy-dependent risk of neutron-induced stochastic effects is related to, but not completely explained by, the induction of CDDs. Thus, the investigation of factors such as DNA damage repair and non-targeted radiation effects is warranted.
Stereotactic Optimized Automated Radiotherapy (SOAR): a novel automated planning solution for multi-metastatic SRS compared to HyperArc™
Thomas Mann, Amanda Swan, Kundan Thind, Nicolas Ploquin
University of Calgary - Tom Baker Cancer Center, Henry Ford Health Systems
Purpose: Stereotactic Radiosurgery (SRS) planning for multi-metastatic patients is a complex process. The commercial HyperArc™ solution is limited to specific beam angles and immobilization. This study compares plan quality between the novel SOAR solution and HyperArc™.
Methods: Twenty-five multi-metastatic SRS plans treated within the last 4 years were re-planned using SOAR and HyperArc™. Plans used the same number of isocenters and beam angles. SOAR used patient-specific characteristics, collision prediction, beam angle optimization, and automation with Eclipse Scripting to design and build treatment plans as described in our published works. Dose tuning structures, treatment fields, and optimization objectives were all added automatically. The automatic lower dose objective (ALDO) and SRS Normal Tissue Objective were used for HyperArc™ plans. All plans were normalized to have 100% or greater target coverage of all targets with the prescription isodose line. Multiple dose-volume metrics and SRS specific plan quality metrics were compared between HyperArc™ and automated plans using double-sided Wilcoxon signed-rank tests (α=0.05).
Results: Differences in dose to the brainstem, cochlea, optic nerves and chiasm were statistically significant favoring SOAR. The median V12Gy volume was 14.2cc for automated planning and 11.7cc for HyperArc™ plans (p=0.4). The median conformity index was not found to be statistically different but median gradient index favored HyperArc™.
Conclusion: The novel SOAR solution had better OAR sparing while maintaining similar target coverage to HyperArc™. SOAR is also able to simultaneously create multi-isocenter plans, is not limited to a specific immobilization device, and provides more freedom in beam angle selection.
Automated plan quality checks for complex stereotactic radiosurgery plans
Thomas Mann, Kundan Thind, Nicolas Ploquin
University of Calgary - Tom Baker Cancer Center, Henry Ford Health Systems
Purpose: Plan quality metrics in radiation therapy assist radiation oncologists, physicists, and planners in evaluating the clinical suitability of a treatment plan. Calculating these metrics manually can be time-consuming and prone to errors due to the repetitive nature of the task. The efficiency of an automated plan quality check application for Stereotactic Radiosurgery (SRS) was assessed and compared to a manual calculation method.
Methods: Eclipse scripting allows the automatic addition of dose structures using the assigned prescription dose and Boolean operators were used to create prescription dose structures and half the prescription dose structures. Volumes from the appropriate structures were used to calculate the RTOG Conformity Index, Paddick Conformity Index, and Gradient Index. Results were color-coded based on the target coverage with the prescription dose. To determine the efficiency of automated and manual methods plan quality metrics were calculated for five patients, with the number of targets ranging from one to fifteen metastases, and calculation times were compared.
Results: The average time for automated calculation of plan quality metrics was 2.9 minutes (range: 0.4 to 7.7 minutes. Manual calculation took an average of 22.2 minutes (range: 7.7 minutes to 51.8 minutes). The time saved by using automated checks ranged from 7 minutes to 44 minutes for the most complicated case.
Conclusion: Automated plan quality checks were approximately 8 times faster than manually calculating plan quality metrics. Automation reduces the time required for plan evaluation, particularly for multi-metastatic cases.
A new approach for characterization of patient-specific 3D dosimetric data using Haralick texture analysis
Lymad Mansour, Luc Beaulieu, Eric Vigneault, Rowan Thomson
Carleton University, Université Laval
Purpose: To introduce a new approach for quantitative analysis of the spatial distribution of dose deposited within clinically-relevant datasets.
Methods: We apply Haralick texture analysis to 3D dose distributions, illustrating the new approach using example datasets for 125I permanent implant prostate brachytherapy. Monte Carlo simulations under “TG43” (water, no interseed attenuation) and “TG186” (realistic tissues/seeds simulated) conditions are used to generate dosimetric datasets. We consider varying intraprostatic calcification (IC; 0-3% by volume) in idealized (8mm3 calcifications scattered throughout prostate with uniformly-spaced seeds) and patient-specific (DICOM-CT, RTPLAN) models. Haralick features are computed to characterize each dose distribution, focusing on homogeneity, contrast, entropy, correlation, and local homogeneity.
Results: Textural features are observed to quantify and reflect changes within dosimetric distributions due to simulation conditions, idealized versus realistic models, and patient/treatment-specific components. For the same patient (3% IC), texture measure differences range from 1 to 80% for TG43 versus TG186 doses. For TG186 datasets, texture measures vary with %IC, e.g., homogeneity decreases as IC increases from 0 to 3% corresponding to high doses associated with calcifications in random positions. For patient-specific models, homogeneity generally increases with %IC corresponding to dosimetric patterns due to more localized, contiguous IC regions in these CT-derived models.
Conclusion: A novel approach for quantitative analysis of 3D dosimetric distributions using Haralick textural analysis has been successfully applied to clinical data. Textural features quantify heterogeneities and patterns in dose distributions, with trends interpretable using radiation dosimetry. Future work include will explore connections with radiobiological modelling and outcomes.
Monte Carlo modeling to investigate the suitability of using single-cell DNA sequencing of irradiated cells in the mutation signature analysis of low- and high-LET radiation
Felix Mathew, James Manalad, John Kildea
Purpose: Behjati et al. (2016) has shown that radiotherapy-associated second cancers exhibit mutation signatures (MS) specific to ionizing radiation. We also expect distinct MS for low- and high-LET radiations but stratifying radiotherapy-associated cancers by LET is an arduous task. We hypothesize that cells irradiated in vitro can serve as an alternative to tumor cells to identify radiation-induced MS by performing DNA sequencing at the single-cell level. With this in mind, we used Monte Carlo modeling to investigate if low- and high-LET radiation introduces distinguishable DNA damage patterns in cells’ genomes.
Methods: Our single-cell geometric nuclear DNA model, developed in-house using the TOPAS-nBio toolkit, was exposed to both low-LET photons and high-LET neutrons of various doses. Both direct and indirect radiation action were tracked for all secondary particles generated, including reactive oxygen species. Various types of DNA damage were scored with repeated simulations.
Results: We observed a significant increase in the number of radiation-induced DNA damage sites with dose for both radiation qualities that should be experimentally measurable using single-cell DNA sequencing. On comparing, we found that the number of base lesions was significantly higher for low-LET photons than for neutrons, whereas the number of clustered DNA damages was higher for neutrons than photons.
Conclusion: Our simulations show that the DNA damage induced by low- and high-LET radiation should be distinguishable in a cell population. Therefore, human cells exposed to radiation in vitro can be used as an alternative for tumor cells to study MS if examined individually using single-cell sequencing.
IMRT planning of a challenging case of right chest wall plus bilateral IMC, axilla and supra-clavicular lymph nodes
Xiangyang Mei, Gary Bracken, Andrew Kerr
Cancer Centre of Southeastern Ontario - Kingston Health Science Centre
Purpose: We demonstrate an IMRT planning case of very challenging right chest wall with bilateral IMC, axilla and supra-clavicular (SC) lymph nodes, when conventional 3D conformal and VMAT plans are not acceptable.
Methods: A post bilateral mastectomy patient was planned for radiation therapy with a prescription dose of 4250cGy in 16 fractions to her right chest wall and bilateral IMC, axilla and SC nodes with a large PTV spanning 39.5cm laterally and 25cm in the Sup-Inf direction. Our efforts with conventional 3D conformal and VMAT techniques failed to achieve an acceptable plan due to the extensive PTV volume and close proximity to the heart. A fixed gantry 10-field sliding window IMRT plan was generated, and optimized to achieve acceptable PTV coverage and low dose to the heart, both lungs, spinal canal, and liver. The right chest wall region PTV was extended 5mm outside the body surface, and 1.0cm bolus was used for plan optimization to account for breathing motion, but 0.5cm bolus was used for the final plan calculation and actual treatment.
Results: A clinically acceptable plan was achieved using the 10-field sliding window IMRT technique. For clinically acceptable and comparable target coverage the IMRT plan achieved adequate OAR sparring whereas the VMAT plan exceeded heart and lung dose tolerances. The IMRT plan was deliverable in 10 minutes on a Varian Truebeam linac at 600 MU/min.
Conclusion: IMRT technique can achieve clinically acceptable plan for very challenging case of Radiation therapy to chest wall plus bilateral lymph nodes.
Dynamic myocardial PET image denoising via unsupervised learning
Mersede Mokri, Sanaz Kaviani, Daniel Juneau, Claire Cohalan, Jean-Francois Carrier
University of Montreal
Purpose: 13N-ammonia myocardial perfusion positron emission tomography (PET) is a dynamic imaging modality that evaluates coronary artery disease (CAD). In dynamic imaging modality, a high noise level affects quantitative accuracy. Conventional denoising methods such as filtering decrease image resolution, and deep learning supervised algorithms require preparing vast data sets for high accuracy. Our study aims to address previous denoising issues and enhance dynamic image quality by using unsupervised deep learning.
Methods: In this study, ten myocardial 20 minutes dynamic 13N-ammonia PET/CT images acquired in. Ten-second frame reconstructed images were considered noisy images, while 60-sec frame reconstruction was ground truth. We designed a 3D U-Net consisting of an encoder and decoder with skip connections. We used a static image and patient CT as input to the network. The performance of the network was evaluated by comparing peak signal-to-noise ratio (PSNR) and Structural Similarity Index (SSIM) to Gaussian and NLM filtering methods.
Results: The Mean PSNRs are 18.35 and 18.03(dB) for NLM, Gaussian filtering, while this value is 20.29 and 20.25 (dB) for the network using the static image and network using CT, respectively. The calculated mean SSIM for our algorithms is 0.713and 0.764 for static and CT priors, while these values are 0.48 and 0.47 for Gaussian and NLM. SSIM in the network using static and network using CT increase 51% and 62% compared to NLM.
Conclusion: The proposed 3D-UNet unsupervised learning method enhances the quality of dynamic PET/CT imaging through reducing noise levels without requiring extensive data set.
Intracellular Acidification Monitored by Chemical Exchange Saturation Transfer MRI: Effect of Cariporide on Rat C6 Brain Tumor
Maryam Mozaffari, Nivin Nystrom, Alex Li Robarts, Miranda Bellyou, Timothy Scholl1, Robert Bartha
Robarts Research Institute- Western University
Purpose: Intracellular pH (pHᵢ) is a hallmark feature of the tumor microenvironment that can be altered by drug treatment. NHE1, a membrane transport protein, is involved in pHi regulation. Blockage of NHE1 by cariporide as a therapeutic strategy produces tumor acidification. We have shown that cariporide selectively acidifies U87MG gliomas in mice. However, the effect of cariporide on different tumor models is unknown. Tumor pHi can be monitored in vivo using a chemical exchange saturation transfer (CEST) MRI technique called AACID. The AACID value is inversely related to pHᵢ. This study aimed to determine whether cariporide could selectively acidify C6 gliomas in rats monitored by AACID-CEST-MRI.
Methods: Glioma cells were injected into the brain of six rats. Rats were scanned on day 7 and day 14 following implantation. On day 14, rats received cariporide inside a 9.4T MRI scanner, and AACID-CEST images were acquired through the tumor before injection and 160 minutes afterward.
Results: Twenty minutes after cariporide injection, the average AACID value in the tumor significantly increased. A second maximum occurred 100 minutes post-injection, but the AACID value in the tumor was not significantly increased. Although an increase in the AACID value in the contralateral region was also observed, the prompt effect of cariporide on tumor tissue was more noticeable.
Conclusion: Selective tumor acidification was not observed following cariporide injection. Instead, acidification occurred in the tumor and contralateral tissues. This discrepancy in comparison to the U87MG tumor model in mice may be related to differences in tissue vasculature which causes cariporide to infiltrate the healthy tissue.
Deformable Image Registration Quality Assurance using Deep Learning Segmentation
Joel Mullins, Byron Wilson
The Ottawa Hospital
Purpose: To develop a quantitative approach to deformable image registration quality assurance (DIR QA) using deep learning (DL) structure segmentation that is streamlined and human-interpretable.
Methods: A head-and-neck deep learning structure segmentation model was previously trained and validated in the RayStation treatment planning system, and applied to treatment planning CT scans for 20 patients that required a replan over their treatment course. To evaluate the DIR relating the original (reference) and rescan (target) study sets, the reference structures were mapped using the DIR to the target study set and analytically compared to the corresponding target structures using the mean distance-to-agreement (DTA). The detectability of the DIR QA procedure was assessed by inducing registration errors through deliberately misaligned points-of-interest (POI) for four patient cases.
Results: The statistical distribution of the mean DTA for each structure was used to establish tolerance and action levels for the evaluation of future DIRs. Visual inspection of structures that exceeded the tolerance/action levels revealed the cause to be either registration errors or deep learning structure segmentation inconsistencies across the two study sets. Applying the DIR QA procedure to the DIRs with POI-induced errors demonstrated an increase in the structures that exceeded the tolerance/action thresholds for 25/32 cases compared to the baseline DIR.
Conclusion: A DIR QA procedure was developed involving the quantitative comparison of DL-segmented structures on reference and target study sets applied to 20 patient cases. The DIR QA procedure provides an expedient evaluation of both DIR quality and DL-segmentation.
Investigating the suitability of existing facilities for a new 177Lu prostate-specific membrane antigen radionuclide therapy program
Nathan Murtha, Hali Morrison, Michael Roumeliotis, Sarah Quirk, Wendy Smith, Adam Blais
Tom Baker Cancer Centre
Purpose: A recent phase 3 trial determined that 177Lu prostate-specific membrane antigen (PSMA) targeted therapy prolonged survival for specific prostate cancer patients. This study determines if it is possible to accommodate a 177Lu-PSMA therapy program within existing facilities.
Methods: At the new Calgary Cancer Centre in Calgary, AB, existing facilities designed for 131I treatment, PET/CT-sim uptake, PET/SPECT injection, and orthovoltage therapy were investigated. Each room was examined for its capability to provide shielding and to accommodate the workflow of 177Lu-PSMA therapy. Two methods of shielding calculation were performed: (i) National Council on Radiation Protection and Measurements report 151, with workload defined in terms of the activity of 177Lu administered, and (ii) using the online RadPro shielding calculator. The workflow requirements related to 177Lu-PSMA therapy were also considered.
Results: Of the rooms examined, only the PET/CT-sim room did not meet radiation safety goals. The 131I treatment rooms are best suited, meeting shielding goals and accommodating workflow requirements. The most challenging workflow requirement to meet were private hot washrooms. The lack of private hot washrooms in the PET/CT-sim and PET/SPECT facilities is a limiting factor against their use in a 177Lu therapy program. The orthovoltage room would require retrofitting of an existing “cold” washroom.
Conclusion: It is possible to accommodate a 177Lu-PSMA therapy program in three of the four types of existing facilities. The existing 131I treatment rooms are most appropriate. This is primarily due to private hot washrooms inside each room, which shield the patient and contain contamination.
Implementation of HyperArc for Single Isocenter Frameless Stereotactic Radiosurgery (SRS) of Multiple Brain Metastases
Theodore Mutanga, Yizhen Wang, Julia Giovinazzo, Heather Fineberg, Raxa Sankreacha, Anthony Brade, Luluel Khan
Trillium Health Partners Credit Valley Hospital
Purpose: To report on our implementation and initial experience with HyperArc for frameless single-isocenter stereotactic radiosurgery treatments of multiple brain metastases on a linear accelerator in a community hospital setting.
Methods: HyperArc (Varian Medical Systems, Palo Alto, CA) is a dedicated hardware and software solution that delivers SRS (or fractionated SRS (SRT) to multiple brain metastases using a single isocenter. The dosimetric accuracy of HyperArc plans was assessed using measurements with a small volume ionization chamber and radiochromic film (EBT-XD, Ashland Inc., Wayne, NJ). Plan quality was assessed using the RTOG and Paddick conformity indices (CI_RTOG and CI_Paddick), Gradient Index (GI_Paddick), ICRU Homogeneity Index (HI), as well as the normal brain V12Gy. Clinical benefit was assessed using the number of metastases treated per session and the reduction in overall treatment times.
Results: For PTV diameters ranging from 0.4cm-3.7cm, the average difference between measured dose and planned dose was -3.9% +-1.7% with a gamma (3%/1mm) passing rate of 98%+-2%. Twelve patients with 103 targets (median 8, range 5-19) have been treated using HyperArc in 12 treatment sessions (39.2 min/session). Compared to our single-isocenter-single-target approach, this represents machine time savings of 69.4 hours. CI_RTOG, CI_Paddick, GI_Paddick, and HI were 1.14-+0.32, 0.87+-0.10, 4.19+-1.15 and 1.25+-0.04 respectively. Median V12Gy was 3.6 cc (range 0.7 – 74.1).
The Encompass mask system was well tolerated with pre and post treatment shifts < 0.5mm and rotations < 0.3 degrees.
Conclusion: HyperArc is an excellent option for delivering linac based frameless SRS treatments safely in significantly shorter times with excellent plan quality.
Clinical validation of a power Doppler-based needle visualization method in high-dose-rate prostate brachytherapy procedures
Nathan Orlando, Jonatan Snir, Kevin Barker, David D'Souza, Vikram Velker, Lucas Mendez, Aaron Fenster, Douglas Hoover
Western University, Robarts Research Institute, London Health Sciences Centre
Purpose: Needle tip identification is critical for safe and effective high-dose-rate brachytherapy (HDR-BT) treatment. Using standard brightness (B)-mode ultrasound (US) imaging, needle tip visualization can be limited due to image artifacts. We propose a power Doppler (PD) US approach for improving needle visualization intraoperatively, validated in clinical HDR-BT procedures.
Methods: Needle perturbation was done using a custom-built wireless mechanical oscillator, designed specifically for brachytherapy applications. The device utilizes a cylindrical end-piece to fit overtop standard brachytherapy needle mandrins allowing for two points of contact and thus improved PD signal consistency. Our method was evaluated in five prostate cancer patients who received standard HDR-BT treatment as part of a prospective feasibility clinical trial. Needle tip identification accuracy was compared between B-mode and PD US, including a breakdown of performance based on needle visibility in B-mode US.
Results: For all 63 needles across five HDR-BT patients, B-mode US and PD US offered mean tip error of 0.9 ± 0.7 mm and 0.8 ± 0.5 mm, respectively, with PD US reducing variation. PD US demonstrated improved performance as needle visibility decreased, showing the ability to reduce tip identification error compared to B-mode US for outlier needles.
Conclusion: Our PD US needle visualization method is cost-effective and easy-to-implement, requiring no modifications to the clinical equipment or additions in the operating room. This method has the potential to improve needle visibility in HDR-BT procedures, potentially improving treatment accuracy. It may be extended to other needle-based procedures providing the possibility for widespread impact.
SmartAdapt® results for propagating head and neck structures from diagnostic CT to radiotherapy CT images
Siobhan / Shevonne Ozard
*** NA ***
Purpose: To investigate the performance of the deformable image registration (DIR) application SmartAdapt® for propagating head and neck organ at risk contours from diagnostic to radiotherapy (RT) CT images.
Methods: The CT images and structures were from a publicly available example data set. To test SmartAdapt® the contours on the diagnostic CT scans were propagated to the RT CT scans and compared to the published RT contours using the Dice Similarity Coefficient (DSC).
Results: For the head and neck organs at risk the median DSC was: brainstem 0.78±0.04, larynx 0.67±0.03, parotid gland 0.74±0.01, submandibular gland 0.72±0.01, mandible 0.74±0.03, and esophagus 0.64±0.05.
Conclusion: This study found that propagating head and neck structure contours from diagnostic CT scans to radiotherapy CT scans with SmartAdapt® DIR within Eclipse version 13.6 was only successful in a minority of cases. Propagating contours from diagnostic scans to RT scans is a challenging task which affected the DSC results.
How much do Pion and Ppol vary between annual output measurements?
Siobhan / Shevonne Ozard
*** NA ***
Purpose: To determine the 2SD variation for Pion and Ppol where SD is the standard deviation of the measurements. Also of interest was control of larger variation and opportunities for improvement of process efficiency.
Methods: The linear accelerators were Varian iX units producing photon and electron beams (6MV, 10MV, 15MV; 6MeV, 9MeV, 12MeV, 16MeV, 20MeV). The number of data points for Pion and Ppol for analysis ranged from three to seven. A cylindrical chamber, PTW 30013 Waterproof Farmer, was used for photons and higher energy electrons (12MeV, 16MeV, 20MeV). A parallel plate chamber, PTW 34045 Advanced Markus, was used for 6MeV and 9MeV electrons and the cross calibration with the cylindrical chamber at 20MeV.
Results: For the cylindrical chamber, the variation of Pion was found to be typically larger compared to the variation of Ppol and this may be due to inconsistent wait times for equilibrium to re-establish after changing the bias voltage. The largest 2SD values for Pion and Ppol were associated with the 20MeV cross-calibration of the parallel-plate chamber, and future work could investigate if this variation can be further reduced. Ppol for photons with the cylindrical chamber exhibited 2SD variation less than 0.001. Similarly, Pion for electrons for the parallel plate chamber showed 2SD variation less than 0.002.
Conclusion: This study led to standardization of wait times after changing the bias voltage. Improved process efficiency could be realized by using previously measured values for Ppol for photons with the cylindrical chamber and Pion for electrons for the parallel plate chamber.
Pharyngeal constrictor dose constraints prognostic of patient-reported dysphagia
Owen Paetkau, Sarah Weppler, Jaime Kwok, Harvey Quon, Wendy Smith
University of Calgary, Tom Baker Cancer Center
Purpose: The goal of the study was to identify pharyngeal constrictor dose constraints prognostic of patient reported dysphagia to improve treatment planning methods in head and neck radiotherapy.
Methods: A 66-oropharynx patient retrospective cohort, treated with 70 Gy in 33 fractions, completed the MD Anderson Dysphagia Inventory survey at a follow-up time of greater than 12 months to evaluate dysphagia incidence. 34 patients reported moderate or severe results on the survey and were classified as exhibiting late patient-reported dysphagia, while the remaining 32 patients were considered asymptomatic. Four pharyngeal constrictor muscle substructures were contoured by a single author: cricopharyngeal constrictor muscle and the superior, middle, and inferior pharyngeal constrictor muscles. Dose volume histogram metrics were identified in D1% intervals and V50cGy intervals for each organ. A decision tree classifier model was developed for each DVH metric with 1000 randomly sampled training sets to classify patient symptoms.
Results: The accuracy and sensitivity of each decision tree classifier model were evaluated. Principal component analysis was used for feature selection to reduce the number of proposed models. The cricopharyngeal constrictor V42Gy at a threshold of 35.9±6.9% and inferior pharyngeal constrictor V37Gy at a threshold of 61.4±9.9% were identified as models of interest due to accuracy (0.71±0.10, 0.67±0.10) and sensitivity (0.72±0.13, 0.78±0.13) scores.
Conclusion: Current pharyngeal constrictor dose constraints consider the mean dose to the entire organ. Adding cricopharyngeal constrictor and inferior pharyngeal constrictor dose constraints prognostic of patient-reported dysphagia to the treatment planning process improve outcomes in head and neck radiotherapy.
2D Transfer learning to predict late patient-reported dysphagia
Owen Paetkau, Sarah Weppler, Fletcher Barrett, Lingyue Sun, Alex Leakos, Deylin Yiao, Graydon Hall, Jared Kraus, Roberto Medeiros de Souza, Wendy Smith
University of Calgary, Tom Baker Cancer Center
Purpose: The goal of this study was to use the deep learning technique, transfer learning, to classify patients at risk of patient-reported dysphagia in head and neck radiotherapy.
Methods: This retrospective study included 133 head and neck cancer patients treated with curative radiotherapy. Patients completed the MD Anderson Dysphagia Inventory (MDADI) survey more than 6 months after completion of treatment. Those with an MDADI summary score >40 were identified as patients exhibiting patient-reported dysphagia. A transfer learning procedure was designed to predict patient-reported dysphagia based on a coronal, sagittal and axial CT and dose image-set. For each patient, five central image-sets of the pharyngeal constrictors were used for a total of 665 image-sets. Further data augmentation was performed by applying vertical and horizontal flips, translations, and rotations. The architecture was the EfficientNet model with the pretrained ImageNet weights. Pre-treatment factors were input as features to the model. A 70%/15%/15% split was used for training, validation, and testing of the model. Ninety-six models were trained with different data augmentation steps.
Results: A total of 73 patients exhibited dysphagia. The best model used average pooling and data augmentation of ±1.25 radian rotations and ±10% translations. The model’s training accuracy and validation accuracies were 80% and 75% respectively. The test accuracy of the model was 60%.
Conclusion: The relatively high training and validation accuracies indicate the potential of transfer learning to predict late patient-reported dysphagia. Ongoing work continues to examine model architecture, data augmentation, and parameter selection to improve generalizability to testing data.
Scratching the (dose) surface: demonstrating the power and potential of dose-surface maps to investigate spatial effects of treatment planning parameters on delivered dose to the rectum
Haley Patrick, John Kildea
Purpose: To demonstrate how dose-surface maps (DSMs) can be used to investigate the influence of PTV margins and fractionation schedules on the level of agreement between planned and delivered dose to the rectum during prostate radiotherapy.
Methods: Three retrospective prostate cancer cohorts treated with daily IGRT were used for this study: 20 patients prescribed 36.25 Gy in 5 fractions with 5 mm PTV margins (5fx-5mm), 20 patients prescribed 60 Gy in 20 fractions with 7 mm margins (20fx-7mm), and the 20fx-7mm cohort replanned with 5 mm margins (20fx-5mm). The influence of margins was investigated using the 20fx-7mm and 20fx-5mm cohorts, and the influence of fractionation using the 5fx-5mm and 20fx-5mm cohorts. Delivered doses were determined by contouring daily CBCTs, registering planned beams to them, and calculating and summing rectum DSMs. Spatial agreement between planned and delivered DSMs was assessed using multiple-comparisons permutation testing.
Results: Both the 20fx-7mm and the 20fx-5mm cohorts were found to have significantly lower delivered doses than planned in similar subregions of the posterior rectum, indicating margin choice did not influence how planned and delivered doses differed. In contrast, the 5fx-5mm cohort had no significant differences between planned and delivered doses. While this suggests improved dose delivery with increased hypofractionation, an assessment of confounding factors is underway to draw firm conclusions.
Conclusion: DSM analysis revealed reduction of prostate PTV margins does not alter how planned and accumulated rectal wall doses vary spatially from each other. Further investigation is required to understand the influence of fractionation schedules.
Characterization of an eye model including a uveal melanoma for Monte Carlo dosimetric simulations in TOPAS
Audran Poher, Francisco Berumen, Joseph Perl, Yunzhi Ma, Luc Beaulieu
Université Laval, SLAC National Accelerator Laboratory
Purpose: Brachytherapy is a treatment method commonly used for uveal melanoma and Monte Carlo simulations are the best way to model this method. The inclusion of an eye model in TOPAS allows to fully simulate brachytherapy eye treatments. This work has been done to characterize a TOPAS eye model containing a uveal melanoma.
Methods: The eye model was created based on the study by Lesperance et al (Med Phys 41, 2014). Brachytherapy simulations involving the eye with a 16mm in diameter eye plaque loaded with iodine-125 LDR sources were performed. The eye plaque and the sources were modelled and validated in TOPAS. Multiple cases were studied: one following the TG-43 recommendations (the whole geometry is made of water and no interseed effects are considered) and another with the “true” geometry considered. In each case, doses were scored in (0.5 mm)3 voxels. Dose values of specific locations, maxima, minima and dose averages of structures of interest were compared with reported data from the study by Lesperance et al.
Results: Measured doses with TOPAS from specific locations such as the center of the lens, the center of the optical disk or the fovea show relative differences of 6% on average with data from Lesperance et al in the TG-43 and complete geometry cases. Noticeable differences were calculated in the maxima and minima of dose distributions of the structures of interest.
Conclusion: Specific dose parameters generally agree with the reference data showing potential for this eye model in medical physics contexts such as Monte Carlo brachytherapy simulations.
Managing PSQA workload after long history of measurement-based analysis
Alejandra Rangel Baltazar, Iram Munawar, Michael Jensen
Trillium Health Partners
Purpose: To implement a strategy that allows reduction of the PSQA workload while maintaining treatment quality.
Methods: The PSQA program at CFRCC was re-visited from different perspectives including a review of PSQA historical data, evaluation of the existing setting/tools at the clinic and mitigation of potential risks. LINAC QA program frequency and tests were also reviewed. A log file and second MU calculation software was commissioned, tested and compared against 2 measurement-based methods currently available at CFRCC. Thresholds for each type of check were defined aided by a sensitivity analysis.
Results: A new process was lay out in May 2020 for single phase prostate, prostate bed and prostate boost plans. The new process includes the addition of a VMAT secondary MU calculation at physics chart check and moves PSQA verification after the 1st treatment fraction, utilizing the treatment log file without the need of a QC specific plan delivery session. Documentation was created to guide the course of action when a case might fail. Since 2020, Secondary MU calculation agreement has shown to be within 3%. Control charts for the selected site has shown PSQA pass rates ranging from 97% to 100%; all cases under control limits. A quarterly reproducibility test ensures the integrity of the new method. With the new process, approximately 5 VMAT cases per week are redirected to a measurement free, after 1st fraction analysis.
Conclusion: It is possible to selectively reduce the time spent for PSQA. This time can now be directed to focus on complex PSQA and machine QA.
Comparison of three commercial methods of cone-beam computed tomography (CBCT) based dosimetric analysis of head and neck patients with weight-loss
Satyapal Rathee, Amr Heikal, Ben Burke
University of Alberta
Purpose: This investigation compares three commercial methods of cone beam computed tomography (CBCT) based dosimetric analysis to a method using repeat computed tomography (CT).
Methods: Seventeen head and neck patients treated in 2020, and with a repeat CT, were included in the analyses. The planning CT was deformed to anatomy in repeat CT to generate a reference plan. Two of the CBCT based methods generated test plans by deforming the planning CT to CBCT of fraction N using VelocityAI™ and SmartAdapt®. The third method compared directly calculated doses on the CBCT for fraction 1 and fraction N, using PerFraction™. Maximum dose to spinal cord (Cord_dmax) and dose to 95% volume (D95) of planning target volumes (PTVs) were used to assess “need to replan” criteria.
Results: The VelocityAI™ method provided results that most accurately matched the reference plan in “need to replan” criteria using either Cord_dmax or PTV D95. SmartAdapt® method overestimated the change in Cord_dmax (6.77% vs 3.85%, p <0.01), change in cord volume ((9.56% vs 0.67%, p < 0.01) resulting in increased false positives in “need to replan” criteria, and performed similarly to VelocityAI™ for D95, but yielded more false negatives. PerFraction™ method underestimated Cord_dmax, did not perform any volume deformation, and missed all “need to replan” cases using cord dose. It also yielded high false negatives using the D95 PTV criteria.
Conclusion: The VelocityAI™ based method using fraction N CBCT is most similar to the reference plan using repeat CT; the other two methods had significant differences.
Evaluating beam characteristics and stability for ECG gated SBRT of ventricular tachycardia
Cristiano Reis, James Robar, Greg Berryhill, Scott Karnas, Stewart Gaede
London Health Sciences Centre, Dalhousie University - Nova Scotia Health
Purpose: To evaluate LINAC beam characteristics and stability when gated under typical frequencies of the cardiac cycle for stereotactic body radiation therapy (SBRT) of ventricular tachycardia (VT).
Methods: A 6 MV beam from a Clinac iX linear accelerator was gated using an Elegoo UNO R3 Arduino microcontroller board to provide a logical signal to the linac. Gating windows were defined to evaluate beam output under high frequency gating delivery mode within typical human cardiac cycle periods between 500 ms and 1000 ms.
Results: A maximum difference of 2.7% is found in chamber response per MU relative to non-gated mode and decreases with increasing the duty cycle. Comparison of beam profiles relative to non-gated beam delivery shows that differences are within 0.4% for all gating windows investigated. Differences in flatness relative to the non-gated delivery decrease as the beam-on time increases relative to beam off time. For all cardiac cycles investigated, most differences are less than 5% for beam-on time greater than equal 200 ms. A similar behavior was observed for crossplane and inplane symmetry of the profiles for the gated beams which were found to converge to the value obtained for non-gated delivery as the beam-on time was increased above 200 ms
Conclusion: Our results demonstrate that dose linearity, beam flatness and symmetry are not strongly affected when gating the beam at high frequencies within typical human cardiac cycles. Deviation of beam flatness and symmetry are reduced for beam-on time greater than or equal to 200 ms.
Contaminant electron origins on a 0.5T inline linac-MR
Michael Reynolds, Patricia A. K. Oliver, Tania Wood, Keith Wachowicz, Ben Burke, B. Gino Fallone
Alberta Health Services
Purpose: Herein we investigate a Linac-MR device with an inline magnetic field with respect to the radiation CAX. The nature of the Lorentz force in this orientation captures electron contaminants in the magnetic field, and directs them towards the machine isocenter. The purpose of this research is to identify the origin and magnitude of the contaminant electrons.
Methods: The Linac-MR was modelled in the EGSnrc Monte Carlo BEAMnrc code in the presence of the main magnetic field, which was generated in Opera3D. PDDs and dose profiles were then generated in DOSXYZ within a 30x30x30 cm3 water phantom. These data sets were compared to corresponding measurements taken in a water tank using an IC10 cylindrical chamber. Once benchmarked, a modified ZLAST option within the BEAMnrc code was enabled to score the X, Y, and Z origin of all electrons that reach the isocenter of the Linac-MR.
Results: Simulated PDDs matched measurements with a 100% gamma pass rate (criteria: 2%, 2mm), with notable differences at the very surface where measurement uncertainty is correspondingly high. Beam profile measurements at various depths had a gamma passing rate of >90% using the same criteria. It was found that the contaminant electrons present at isocenter are generated downstream of the region within the proximal magnetic pole plate, where we see an increasing fluence of electrons closer to the phantom surface.
Conclusion: The origins of contaminant electrons within the Linac-MR device are outside of the linac head, increasing in number with decreasing distance to the phantom surface.
Classification of Chronic Obstructive Pulmonary Disease using CT Images and Pre-trained Convolutional Neural Networks
Sara Rezvanjou, Amir Moslemi, Wan-Cheng Tan, James C Hogg, Jean Bourbeau, Miranda Kirby
Department of Physics - Ryerson University, Centre for Heart Lung Innovation - University of British Columbia, Respiratory Epidemiology and Clinical Research Unit - Research Institute of the McGill University Health Centre
Purpose: Computed tomography (CT) imaging of lungs can be used to classify patients with chronic obstructive pulmonary disease (COPD) using 2D convolutional networks (CNNs). However, it is unknown which CT slice should be selected, and whether pre-trained CNN models outperform the naïve model. The objectives of this study were to: 1) compared the accuracy of the naïve model for COPD classification using three different CT slices, and, 2) compare the performance of four established pre-trained CNN models for COPD classification.
Methods: CT images were evaluated from the CanCOLD study. COPD classification was determine using GOLD criteria. A total of 80%/20% of subjects were used for training/testing. A 4-layer CNN was considered the naïve model. Three CT slices (upper, middle, lower) were evaluated for classification accuracy. Using the optimal slice, classification accuracy of the naïve model was compared to four pre-trained models (DenseNet201, VGG16, ResNet50, and InceptionV3).
Results: A total of n=608 no COPD and n=593 COPD subjects were evaluated. Using the naïve model, the CT middle slice had the highest accuracy (68%) compared to lower (66%) and upper (66%) slices for COPD classification. Classification accuracy of pre-trained models InceptionV3, Densenet201, VGG16, and Resnet50, were 71%, 67%, 65%, and 66%, respectively.
Conclusion: The slice selected for 2D-CNN model training impacts COPD classification accuracy, and the middle slice provided the highest accuracy. Among all models, the pre-trained Inception201 had the best performance. Therefore, even with a relatively small dataset, properly tuned pre-trained models can classify COPD with moderate accuracy.
3D-printed modification of a daily QA phantom for all-in-one evaluation of 6DOF couch positioning with surface and kV imaging localization
Andrew KH Robertson, Bradford Gill, Robin JN Coope, Samantha AM Lloyd
BC Cancer, Canada’s Michael Smith Genome Sciences Centre - BC Cancer
Purpose: To adapt the established daily test of radiographic image guidance to include performance evaluation of an optical surface guidance imaging system and a 6DOF couch.
Methods: The established daily QA test checks OBI performance using the Varian Geometric Phantom. Held by a jig with multiple mounting positions and indexed to the couch using an index bar, the block is aligned to lasers before remounting in a second position. This introduces known translations to the block’s position from isocentre that are then compared to couch shifts computed by the OBI match software. To incorporate 6DOF couch functionality, a new jig was 3D-printed in ABS plastic that introduces translations and rotations to the phantom position. Masking tape was applied to the block surface to make it visible to an AlignRT system. A comparison was then made between known block displacements and those measured via surface and CBCT imaging.
Results: Longitudinal data shows that radiographic imaging, the 6DOF couch, and the optical surface guidance system compute, perform, and detect the same translations and rotations within 1 mm and 0.3°. Time to complete daily warmup and QA was not impacted.
Conclusion: An existing daily QA procedure was successfully modified to incorporate monitoring of both 6DOF couch and optical surface guidance system performance, at low cost and with minimal impact to workflow during morning QA. Design files for the 6DOF jig are available upon request.
Markerless Dynamic Tumour Tracking (MDTT) using soft tissue anatomical surrogates for liver cancer
Maryam Rostamzadeh, Marie-Laure Camborde, Roy Ma, Tania Karan, Mitchell Liu, Ante Mestrovic, Isaac Tai, Allison Tammark, Alanah Bergman
Purpose: To demonstrate pre-clinical feasibility for Markerless Dynamic Tumor Tracking (MDTT) using dome-of-liver/diaphragm interface as a motion-surrogate for liver radiotherapy.
Methods: The Brainlab Vero4DRT linac MDTT module, designed for lung tracking, was applied in a novel way. Phantom: An in-house dome-of-liver shaped insert was built, compatible with the QUASAR motion platform. The “dome” was contoured on 3Dhelical CT images as a “Markerless Tracking Structure” (MTS). A single-field, multi-aperture beam was delivered to the film plane under static and MDTT delivery conditions. Patient Study: A pre-clinical, IRB-approved imaging-only study was offered to ten patients undergoing standard liver RT. The dome-of-liver was contoured on patient “exhale-phase” CTs as the MTS. The MDTT imaging workflow was performed. ExacTrac KV-imaging log files containing detected vs correlation-model-predicted dome positions from each session were analyzed.
Results: Phantom: The average/SD difference between the detected vs predicted dome-of-liver position during MDTT beam delivery was 0.51/0.19 mm, respectively. The film-measured 2D gamma (2%, 2mm, 30% threshold) comparing static vs MDTT deliveries was 98.2%.
Patient Study: The MDTT model building was successful in all 10 patients. With TRACKING enabled during the pre-MV BEAM-ON state, the ExacTrac-reported range of average/SD differences between detected vs predicted dome-of-liver position was 1.11-3.52 mm/0.62-2.67 mm, respectively. The average/RMS error over all 10 patients was 2.30/1.61 mm.
Conclusion: MDTT workflows using dome-of-liver as an MTS was feasible using the Vero4DRT linac. Care must be taken to select good kV-imaging angles for the model-building step. This method could replace invasive implanted marker-based tumour tracking for liver radiotherapy.
Dosimetric evaluation of VMAT treatment plans for patients with stage IIB or III non-small cell lung carcinoma
Amani Shaaer, Johnson Darko, Darin Gopaul, Ernest Osei
Grand River Regional Cancer Centre
Purpose: Volumetric-modulated arc therapy (VMAT) is a well-established an external-beam radiotherapy technique for patients with small cell lung cancer (NSCLC). The main challenge in VMAT planning is the lack of consistency in acceptable target coverage and normal tissue constraints. Our objective is to develop institutional criteria for VMAT treatment plans acceptability based on current experiences and resources.
Methods: The CT dataset of 20 patients with NSCLC were randomly included in this study. An isotropic 5 mm margin was added to the ITV to generate the CTV and an additional 5 mm margin was added to the CTV to generate the PTV. All treatment planning were accomplished with 2-3 VMAT arcs. The dose fractionation scheme used was 60Gy in 30 fractions and was normalized based on RTOG-0617 (V100 = 95%). The planning acceptance criteria for bilateral lung and spinal canal were V20Gy ≤ 37% and Dmax < 83.5% of prescribed dose, respectively.
Results: Adequate target coverage was achieved for all patients reviewed. The mean PTV Dmean was 106 ± 2% (range: 104-110%). The mean Dmax for spinal canal was 55±12% (range: 30– 71%). The mean bilateral lung V20Gy and V5Gy were 20±8% (range: 5-36%) and 51±17% (range: 15-68%), respectively. Renormalizing the PTV coverages such that V95=95% helped to reduce OARs doses.
Conclusion: The results of this study can provide the basis for the development of local criteria for VMAT treatment plans. Predefined dose-volume objectives can be achieved. However, coverage renormalization of targets may be required to reduce some critical organs dose.
4D Monte Carlo dose reconstructions using surface motion measurements: a feasibility study
Meaghen Shiha, Joanna E. Cygler, Robert MacRae, Emily Heath
Carleton University, The Ottawa Hospital
Purpose: To outline a framework for 4D dose reconstruction and assess the feasibility of its implementation for NSCLC patients undergoing VMAT treatments.
Methods: Our 4D dose reconstruction method uses the 4Ddefdosxyznrc/EGSnrc user code to model motion and deformation of patient anatomy during radiation delivery. Exhale phase CT images are registered to inhale phase CT images to generate a patient specific respiratory model. During treatment fractions, the patient surface motion was tracked with the RADPOS 4D dosimetry system and used as a surrogate for tumour respiratory motion. Linac log files were extracted after each treatment fraction and used to synchronize the machine parameters with the patient respiratory motion trace. 4D dose calculations for each fraction were compared to the planned dose calculated on the phase averaged CT and the exhale CT images.
Results: Use of the 4D method is demonstrated here for one patient with a right upper-mid lobe tumour. Respiratory trace measurements and log files were extracted for 5 treatment fractions. An average respiratory amplitude of 10.0±1.2 mm was measured for this patient. There were no significant inter-fraction differences in the reconstructed doses to the CTV and OAR. There were no significant differences between the planned doses and the reconstructed doses.
Conclusion: We have demonstrated the 4D dose reconstruction method using linac log files and patient surface motion measurements as a surrogate for tumour motion. This method can be used to investigate dosimetric impact of respiratory motion on delivered dose.
DLG Parameter Optimization using ESAPI Automation
David Sinn, Brad Warkentin
Alberta Health Services - Cross Cancer Institute
Purpose: To develop a more efficient method of optimizing the Eclipse DLG parameter for both NDS120HD and NDS120 MLCs based on automation using ESAPI-based scripting.
Methods: Verification plan doses were calculated for a large selection of TG119 and clinical patient plans for comparison to measurements using the PTW Octavius 4D phantom and 1000 SRS detector panel. Ten 6FFF, six 10MV, and thirteen 10FFF plans were evaluated. Each dose was re-calculated (Acuros v15.6) for 7-9 DLG values, varied in 0.05 or 0.1 mm increments. ESAPI scripting was used to automate the plan calculation. 3D gamma analysis compared each measured/reconstructed 3D dose with the sets of doses calculated over the DLG space. Gamma passing rates were averaged over all plans for a given energy, and polynomial fits were performed to find the DLG yielding optimal passing rates. Overall, 312 plans were.
Results: Optimized DLGs for the 10MV beams were within 0.05 mm of values determined from measurement alone. For the unflattened beams using the NDS120HD MLCs, the optimized DLG values of 0.58 mm (6FFF) and 1.18 mm (10FFF), were 0.31 mm and 0.75 mm smaller, respectively, than measured values, which improved gamma passing rates (2%/1mm) by 1.7% and 10.8%.
Conclusion: ESAPI and other automation tools may make robust DLG optimization much more feasible than previously possible. This may result in improved accuracy in treatment planning modeling of IMRT dose distributions.
A convolution-superposition fluence model for the Varian HD120 MLC applied to 3D VMAT dose calculations
Cross Cancer Institute
Purpose: To develop a fast-calculating fluence model for the Varian HD120 MLC, and to use this fluence to calculate 3D VMAT dose in a cylindrical phantom. To-date no fluence model for the Varian HD120 MLC has been published.
Methods: The methodology for developing the HD120 MLC fluence model is similar to that developed earlier for the Varian Millennium 120 MLC but with modifications. In this model, parameters for 2D inter-leaf leakage maps, 2D tongue-and-groove maps, and 2D MLC transmission maps were optimized based on aS1200 EPID images acquired from five commissioning test fields and measured DLG values. An extra-focal head model was applied to determine the final 2D fluence for a given field. For fluence calculation on a cylindrical water phantom, a 2-D curvature-correction map was applied, and gantry-dependent VMAT fluences were convolved with a 6MV 3D Monte Carlo-generated dose kernel to calculate 3D doses for VMAT plans.
Results: The fluence model showed an excellent fit with all five EPID commissioning fields, where fluences were convolved with an EPID dose kernel for comparison. 2D gamma analysis (2%/2mm) of the fields showed a 94.7% average pass rate, where fluence was calculated in 25 seconds per field on a standard PC. 3D gamma analysis (3%/3mm) of VMAT doses with dose thresholds of 20% and 80% showed average pass rates of 95% and 92% respectively.
Conclusion: A fast and accurate fluence model for the Varian HD120 MLC has been developed and applied to calculating 3D VMAT dose in cylindrical phantom.
Comparison of Bayesian Networks and Conventional Machine Learning Techniques for Cystitis Prediction in Cervical Cancer Patients
Kailyn Stenhouse, Philip McGeachy, Sofia Spampinato, Kari Tanderup, Sarah Quirk, Kevin Martell, Ina Jürgenliemk-Schulz, Kathrin Kirchheiner, Michael Roumeliotis
University of Calgary - Tom Baker Cancer Centre, Aarhus University Hospital, University Medical Center Utrecht, Medical University of Vienna
Purpose: To develop a data-driven Bayesian Network (BN) for cystitis prediction in cervical cancer patients undergoing brachytherapy and compare performance to conventional machine learning (ML) models.
Methods: Cystitis toxicity patient data (n=1379) from the EMBRACE I trial was processed for predictive modelling. From 44 initial features, ANOVA, Pearson, Mutual Information, and Chi2 correlations were used to identify 21 clinical and dosimetric features. Patients with missing feature data were removed. The remaining patients (n=1056) were used for predicting grade ≥1 cystitis as defined by Common Terminology Criteria for Adverse Events v3.0 (No cystitis=819, cystitis=237). A BN and six ML algorithms (GB: Gradient Boosting, RF: Random Forest, AB: AdaBoost, KNN: K-Nearest Neighbours, LR: Logistic Regression, SVM: Support Vector Machine) were compared. Each model was evaluated over 100 random samplings of training and validation data (90%/10%). Balanced accuracy (mean of sensitivity and specificity), precision, and recall were calculated. For each metric, BN performance is reported with the ML model maximum and minimum performance.
Results: The BN model demonstrated a similar balanced accuracy comparable to the ML (maximum, minimum) models (BN: 61.49±3.72%, RF: 59.00±3.84%, KNN: 54.27±4.46%). For precision, the BN model did not outperform all ML models (BN: 61.45±4.17%, RF: 67.15±7.71%, KNN: 55.99±6.21%), however, it outperformed ML in recall (BN: 62.68±5.14%, SVM: 46.21±12.11%, LR: 33.12±5.94%). These results indicate that the BN model is predicting fewer false negatives with comparable false positives.
Conclusion: BN modelling demonstrated improved toxicity prediction capabilities compared to conventional ML. Future work will incorporate expert input to refine the network structure.
Are dosiomics features robust against variations in dose calculation for prostate external beam radiotherapy plans?
Lingyue Sun, Wendy Smith, Charles Kirkby
University of Calgary - Tom Baker Cancer Centre, University of Calgary - Jack Ady Cancer Centre
Purpose: To assess the stability of dosiomics features against variations in dose calculation algorithm (DCA) type, version, and dose grid size.
Methods: Twenty-seven prostate EBRT patients with two-phase pelvic irradiation were retrospectively included. For each patient, the planned dose distributions were recalculated using two algorithms (AAA and Acuros XB), two versions (13.6 and 15.6), and three dose grids (2, 2.5, and 3mm) in the EclipseTM TPS, resulting in 12 dose distributions. 93 dosiomics features were extracted from each dose distribution using PyRadiomics for these regions-of-interest: high-dose PTV (PTV_High), low-dose PTV (PTV_Low), 1cm rind around PTV_High (PTV_Ring), rectum, and bladder. For each dosiomics feature and each patient, the coefficient of variation (CV) was calculated. Hierarchical clustering was used to group features with high and low variability and to determine a threshold CV value. Features with high variability were further analyzed using three-way repeated-measures ANOVA to investigate the effect of these dose calculation factors on the features.
Results: For the PTV_Low, PTV_Ring, and rectum, all the dosiomics features showed low CV with an average CV≤0.26. For PTV_High, six dosiomics features had CV>0.26 with DCA type and grid size being the major contributing factors for variation. For bladder, one dosiomics feature had an average CV>0.26, although none of the three dose calculation factors led to statistically significant variations.
Conclusion: While most dosiomics features were stable given the variations in dose calculation, some features for the PTV_High and bladder showed significant variations. This may warrant attention when datasets with various dose calculations are included.
Using principal component analysis to explore prognostic power of dose-volume histograms and dosiomics features
Alec Swallow, Lingyue Sun, Wendy Smith
University of Calgary - Tom Baker Cancer Centre
Purpose: To investigate if adding dose-volume histograms (DVHs) and dosiomics information using principal component analysis (PCA) improves biochemical failure-free survival (BFFS) prediction.
Methods: 1774 localized prostate adenocarcinoma patients diagnosed between 2010 to 2016 and treated with curative external beam radiation therapy were retroactively recruited. DVHs for the high-dose PTV and 214 dosiomics features from the high-dose CTV and high-dose PTV were analyzed. PCA was used to derive two principal components (PCs) from the DVHs and 20 PCs from all dosiomics features. Random survival forest (RSF) was used for BFFS prediction. Cox proportional hazards model was used to identify statistically significant variables for BFFS prediction. Three RSF and Cox proportional hazards models were constructed. 1), using clinical parameters (Model A); 2), using clinical parameters and DVH-derived PCs (Model B); 3), using clinical parameters and PCs derived from both DVH and dosiomics features (Model C). RSF performance was evaluated using concordance index (c-index). 1064 patients were used for training while 710 were reserved for validation using a random stratified split.
Results: All Cox proportional hazard-based models found three statistically significant clinical parameters to BFFS prediction. Neither DVH-derived PCs were statistically significant. Four dosiomics-derived PCs were statistically significant. For RSF-based models, Model A had a c-index of 0.65/0.62 for the training/validation set. Model B had a c-index of 0.67/0.61. Model C had a c-index of 0.70/0.61.
Conclusion: Four dosiomics features from both PTV and CTV were significant. DVH features have low prognostic power. The cause of model C overfit must be investigated further.
Feasibility of MapCHECK-based and log file-based quality assurance for mixed electron-photon radiation therapy at standard SAD
Yee Man Tai, Veng Jean Heng, Marc-André Renaud, Monica Serban, Jan Seuntjens
McGill University, Gray Oncology Solutions, McGill University Health Centre, University Health Network
Purpose: To develop a QA protocol for mixed electron-photon beam radiotherapy (MBRT) using MapCHECK and trajectory log files.
Methods: An MBRT plan consisting of step-and-shoot deliveries of a 6MV photon beam and electron beams in 5 energies was robustly optimized for a soft-tissue sarcoma patient. The plan was delivered to MapCHECK using a Varian TrueBeam linac with collapsed gantry angle at 100cm SAD. A MapCHECK Monte Carlo phantom was modelled in detail on EGSnrc and the planned dose to the MapCHECK detectors were scored in a DOSXYZnrc simulation. The agreement between the simulated and measured dose distribution was evaluated using gamma analysis. The axis positions and monitor unit (MU) at each control point of the delivery were retrieved from the trajectory log files to recalculate the delivered dose distribution. The planned and the re-calculated dose to the MapCHECK detectors were compared.
Results: The comparison between the measured and the simulated overall dose distribution to the MapCHECK demonstrated a gamma passing rate of 90.11% with the gamma criteria of 3%/2mm. The log file-recalculated and planned dose to the MapCHECK detectors were compared. More than 97% of the detector doses differed by less than 3% for all energy components.
Conclusion: The agreement among the measured, log file-reconstructed and planned dose distribution demonstrated the potential of MapCHECK and trajectory log files as MBRT QA tools. The measurement using MapCHECK and collection of log files were performed in a single delivery and provided an efficient and comprehensive QA for the complex MBRT plan.
Measuring impact of oxygen on radiotherapy treatment using finite element method on mouse’s vascular network using numerous surviving models compared to the linear quadratic model
Jeremie Tanguay, Michèle Desjardins, Louis Archambault, Corinne Chouinard
Université Laval, Université Laval-Chu de quebec
Purpose: To evaluate oxygen partial pressure in a numerical model in order to model the effect of tumor hypoxia on cell survival to radiation therapy.
Methods: Starting from a vascular network, it is possible to evaluate oxygen partial pressure using the finite element method to solve the diffusion equation. To succeed, we used coupled equations to consider oxygen transport inside the vessels, the advection, while also considering the diffusion outside the vascular network. Since the two mechanisms of transport are mutually dependent, the system is solved iteratively until equilibrium is achieved. Once our oxygen level is stable, we evaluate cell survival by modeling a treatment map for various survival models from literature (Strigari, Wouters, Antonovich). Differing from the linear quadratic model that is most used, those models take oxygen levels into consideration. Our model allows to simulate varying physiological conditions such as modifying oxygen levels to mimic hyperbaric treatment or increasing oxygen consumption to mimic tumor metabolism.
Results: Our first exploration of models that take into account oxygen level reveal its impact on cell survival compared to the linear quadratic model. by combining our oxygen distribution model with different radiobiological calculation approaches that includes the effects of oxygen on cell survival, we've shown that asymmetrical distribution can be correctly represented, which is not possible with the LQ model.
Conclusion: This work suggests that considering oxygen levels when administrating treatment can have a significant impact on cell survival and provides a framework to quantify this impact under various conditions.
Gafchromic film and scintillator detector measurements in phantom with a novel intensity-modulated brachytherapy endorectal shield
Alana Thibodeau-Antonacci, Shirin A. Enger, Hamed Bekerat, Té Vuong
Medical Physics Unit - Department of Oncology - McGill University, Department of Radiation Oncology - Jewish General Hospital
Purpose: To validate Monte Carlo (MC) dosimetry simulations in solid-water/water for a novel MRI-conditional intensity-modulated brachytherapy (IMBT) endorectal shield using GAFCHROMIC® EBT3 Film and a scintillator detector (MedScint, Québec).
Methods: A single-grooved IMBT tungsten shield with a 180° emission window, diameter of 15-mm and length of 80-mm was developed. It contained a 2-mm central source channel to permit the insertion of a 6F catheter. Dosimetric properties of the shield were calculated in water using RapidBrachyMCTPS, a MC-based treatment planning system, and the MicroSelectron-v2 Ir-192 source. Dose was scored by simulating 108 radioactive decays using a 1-mm3 voxel grid. 2D dose distributions at different distances from the source were measured with EBT3 films in solid-water. The global gamma index (2-mm and 2% criteria) was calculated to evaluate the agreement between the MC and measured dose maps. The azimuthal anisotropy plot at 1-cm was determined in water using a scintillator detector.
Results: The gamma passing rate was greater than 90% on the unshielded side for distances up to 13-cm from the source. Better agreement was obtained on the shielded side as the passing rate was greater than 95% for all cases. The measured and simulated anisotropy plots agreed within 15%. The positional uncertainty was estimated to be 1-mm. The measured data was contained within this uncertainty.
Conclusion: Radiochromic film and scintillator detector measurements validated the MC dosimetry calculations for a novel endorectal IMBT shield. Additional measurements will be performed with an improved setup to reduce the positional uncertainty.
Improvements to material density effect corrections in EGSnrc
Sehmimul Hoque, Frédéric Tessier, Ernesto Mainegra-Hing, Reid Townson
National Research Council Canada, University of Waterloo
Purpose: One crucial aspect of EGSnrc is the ability to model arbitrary material composition and density accurately. The effective density in the reference frame of a moving particle depends on its kinetic energy, and energy-dependent tables of corrections to the density for each material are typically calculated in advance of running the simulation. Previously, it had to be calculated using a separate software or web interface. The current work involves integrating the density effect correction calculation from the NIST ESTAR software directly into the EGSnrc toolkit, to be performed as a part of the simulation initialization stage.
Methods: Density effect calculations from ESTAR were translated from Fortran to C++, and integrated into the EGSnrc material initialization section of a simulation. The updated version of the calculation was compared against the ESTAR code using an automated fuzzy testing suite over all elements and many compounds that generated over 10,000 variations of inputs as validation. The existing density correction files distributed with EGSnrc were also compared. The accuracy of one parameterization in the density effect calculation was improved, by replacing an approximation with a root-finding method.
Results: The revised density corrections were shown to agree with ESTAR. The improvements to the accuracy of the algorithm increased the density factor by 1%-3% for most elements and compounds over a wide energy range.
Conclusion: This work enhances the traceability of EGSnrc simulations by including the source code for an essential component of the calculations as a part of the open-source version-controlled toolkit.
Comparison of Machine Learning Models Trained on Synthetic Radiomic and Clinical Data
Lorna Tu, Herve HF Choi, Haley Clark, Samantha AM Lloyd
University of British Columbia - BC Cancer Agency
Purpose: To compare cancer survival prediction quality for machine learning (ML) models trained on a synthetic radiomic and clinical feature dataset.
Methods: Sixty-one radiomic features were extracted from CT-based, manually drawn contours for 132 non-small cell lung cancer (NSCLC) patients. A synthetic training dataset of radiomic and clinical data was generated for 106 examples using CTGAN. Synthetic data quality was assessed using feature distributions and statistical tests. Predictive models of 2-year survival were trained on real or synthetic data using combinations of four feature selection methods (mutual information, ANOVA F-test, recursive feature elimination, random forest (RF) importance weights) and six ML algorithms (RF, k-nearest neighbours, logistic regression, support vector machine, XGBoost, Gaussian Naïve Bayes) and tested on real data. Model performance was compared using balanced accuracy and area under the precision-recall curve (PR-AUC).
Results: Real and synthetic datasets were similar, with an average 1 minus Kolmogorov-Smirnov test statistic of 0.871 for continuous features. Chi-square test confirmed agreement for discrete features (p<.001). The best model was XGBoost using RF importance-based features, with differences in accuracy and PR-AUC between models trained on each dataset of 0.0069 and 0.0105 respectively. Gaussian Naïve Bayes using features selected using ANOVA F-test had the largest differences in accuracy and PR-AUC, 0.3403 and 0.3783 respectively.
Conclusion: RF and XGBoost models trained on real/synthetic datasets performed similarly, suggesting tree-based algorithms may be more robust and suitable for NSCLC data augmentation via dataset synthesis. Gaussian Naïve Bayes models had the largest difference in performance, even with tree-derived feature selection.
Evaluation of a commercially developed thoracic autocontouring deep learning model
Madeleine Van de Kleut, Miller MacPherson, Dal Granville
The Ottawa Hospital
Purpose: To quantitatively evaluate a commercially developed deep learning model’s autocontouring performance through comparisons with expertly contoured organs at risk (OARs) on thoracic radiation treatment planning CT scans.
Methods: 123 segmented and peer-reviewed thoracic CT scans from a single center were used to evaluate the agreement between autosegmented OARs and those contoured manually by a radiation oncologist (RO). Autosegmentation was performed using a commercially developed deep learning model (ADMIRE v3.11, Elekta, AB, Stockholm, Sweden). This model was developed using a residual UNET architecture with training data that was independent from the testing data used in this study. We compared autosegmented esophagus, heart, lung, and spinal canal structures to manually segmented structures using the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD). DSC and HD were evaluated within the superior-inferior range that included both auto and RO contours for the esophagus and spinal canal due to inconsistencies in the manually drawn contours.
Results: The mean DSC was highest for the right and left lungs (equally), followed by the heart, spinal canal, and esophagus, at 0.97 ± 0.01, 0.92 ± 0.03, 0.85 ± 0.04, and 0.77 ± 0.07, respectively. Mean HD ranged from 9 ± 9 mm for the spinal canal, to 21 ± 10 mm for the right lung.
Conclusion: The commercially developed deep learning model investigated can perform fast and accurate autosegmentation of OARs relevant to thoracic radiotherapy. Such models have the potential to reduce the time required to generate radiotherapy treatment plans and reduce inter-observer variability in segmentation.
A comparison of portal dosimetry for the Varian Halcyon 2.0 and multiple detector systems
Kurt Van Delinder, Martin Shim, Ryan Studinski, Roxana Vlad
Juravinski Cancer Centre (JCC), Hamilton Health Sciences
Purpose: To validate clinical treatment plans generated on a dual-multileaf collimator (MLC) Halcyon 2.0 Linac with a predefined beam model. Additionally, to compare the portal dosimetry (PD) tool with film and Mobius-3D/FX used in our institution for patient-specific QA (PSQA).
Methods: A total of 50 clinical VMAT plans consisting mainly of pelvis were evaluated using the gamma analysis index (γ) with Gafchromic film (3%/2mm), ion chamber point measurements, Mobius3D (3%/3mm), MobiusFX (3%/3mm) and PD. Various γ tolerances (3%/3mm to 1%/1mm) were explored to analyze the differences between the PD dose prediction and the dose-map delivered to the EPID.
Results: The average γ pass rates for all 50 plans were above 95% for each method. The mean pass rates (γ ≤ 1) were 99% for film (3%/2mm) and -0.07% for ion chamber (% difference of calculated vs measured). The dose calculated with Mobius 3D (3%/3mm) and the Halcyon’s treatment log delivery files using Mobius MFX (3%/3mm) produced a mean γ of 99% for both modalities. The mean γ pass values were 100 to 98% for the various γ criteria used to analyze the differences between PD dose prediction and delivered dose.
Conclusion: The γ results indicate a good agreement between PD results analyzed with various tolerance criteria, film and Mobius 3D/FX for 50 plans delivered on a Halcyon-Linac. The PD tool is an efficient and fast method that can be potentially used for routine PSQA pending further investigation to determine its sensitivity to different errors in MLC positioning.
Three-dimensional transvaginal ultrasound guidance for gynecologic perineal template interstitial brachytherapy
Devin Van Elburg, Kevin Martell, Tyler Meyer, Sarah Quirk, Robyn Banerjee, Aaron Fenster, Michael Roumeliotis
University of Calgary - Tom Baker Cancer Centre, University of Western Ontario
Purpose: To describe our institutional workflow using three-dimensional transvaginal ultrasound (3DTVUS) for gynecologic template interstitial high-dose-rate brachytherapy (IS-HDRBT).
Methods: Template interstitial IS-HDRBT patients with vaginal tumours are suitable for intraoperative 3DTVUS. The 3DTVUS apparatus is secured to the inferior end of the operating room bed rails and the ultrasound probe is inserted to the top of the patient’s vagina. A 3DTVUS image is acquired via 360-degree probe rotation with sequential sagittal slices acquired to 7cm depth. This image enables clinicians to visualize gross tumour volume or normal tissue (e.g. vaginal wall, rectum) distal to the top ~2cm of vagina. Intended needle positions are identified in software, then the probe rotates to the selected 2D frame for live ultrasound guidance during needle insertion. Intermediate 3DTVUS images can assist in determining required adjustments or needle additions to produce a high-quality implant. Following the implant, the 3DTVUS apparatus is detached from the bed. The perineal template is guided over the free length of the implanted needles, followed by the vaginal cylinder. The standard brachytherapy treatment workflow proceeds following institutional standard for imaging, treatment planning, and delivery.
Results: To date, our institution has successfully used 3DTVUS guidance for perineal template implants in three patients. Eight needles total have been inserted under 3DTVUS guidance, all which clinicians were satisfied with placement relative to the intended target and surrounding normal tissue.
Conclusion: Intraoperative 3DTVUS visualization of implanted needles in a rigid and reproducible coordinate system has the capability to improve target coverage and normal tissue avoidance.
The Canadian experience in commissioning a Halcyon-Linac in a Pinnacle-MOSAIQ based centre
Roxana Vlad, Orest Ostapiak, Ryan Studinski, Kevin Diamond, Martin Shim, Gordon Chan, Robert Hunter
Juravinski Hospital and Cancer Centre at Hamilton Health Sciences
Purpose: This work describes a Pinnacle/MOSAIQ institution's experience in commissioning and first clinical use of two Halcyon 2.0 units.
Methods: Commissioning involved validating the pre-defined beam model and portal dosimetry application through a series of ion chamber and diode water tank scans, radiochromic film measurements and end-to-end tests. Using the water tank on the treatment couch required a custom-made support. Baseline data was acquired for mechanical tests, multi-leaf collimator and imaging systems. A clinical implementation committee designed new clinical processes, new planning protocols, and coordinated staff training. Plans for the first 25 patients treated on each of two units were validated using film, ion-chamber, Mobius3D/FX and Varian’s portal dose tool. Results were compared with those published.
Results: Good agreement was found among measured, calculated and published beam data for the pre-configured Halcyon beam model when using the recommended detectors for measurement. Testing of mechanical subsystems was found to be within specifications which meet or exceed those of TrueBeam units. Planning protocols and workflows for pelvis-based sites were developed by selected staff trained for the Aria-Eclipse environment. The various methods used for patient-specific quality assurance yielded consistent results, however, the portal dose prediction tool requires a gamma pass rate set to >98% at 2%/2mm in order to capture clinically significant errors that were intentionally produced.
Conclusion: Two Halcyon units have been successfully introduced, commissioned, and implemented clinically to treat sites in the pelvis in a centre that previously used only Pinnacle/MOSAIQ.
Applications of Deep Learning for Differential Diagnosis of Lung Cancer
Mitchell Wiebe, Dr. Rasika Rajapakshe
University of British Columbia - Okanagan Campus, BC Cancer Agency
Purpose: Histopathologic tissue evaluation represents the gold standard for differential diagnosis of lung cancer, but requires invasive procedures that risk major complications and potential death. Histologic patterns have been visualized with tomographic imaging of excised lung tissue, demonstrating potential for non-invasive diagnosis. To the best of our knowledge, the effect of spatial resolution (SR) on differential diagnosis of gold standard pathology slides has not been studied.
Methods: Pathology slides from The Cancer Genome Atlas (TCGA) (Adenocarcinoma (LUAD)=823, Squamous-cell carcinoma (LUSC)=753, and Normal=591) were used to train (70%) and test (30%) a convolutional neural network (Inception-v3) as a binary (Tumor/Normal) and three-way classifier (LUAD/LUSC/Normal). Slides were split into 512x512 pixel tiles at 2.5x magnification (SR=4micron/pixel) and reduced SRs of 8,16,32,64, and 128micron/pixel were simulated with Lanczos3 low-pass filters. Slide-level predictions were obtained by averaging constituent tile predictions and performance was evaluated by area under the ROC curve (AUC) and 95% confidence intervals by bootstrapping. An arbitrary performance cutoff was set at AUC=0.95.
Results: For the binary classifier, the minimum SR that was classified at the cutoff was 64micron/pixel (AUC=0.980, CI=0.963-0.992). For the three-way classifier, the minimum SR that was classified at the cutoff was 16microns/pixel (AUC-LUAD=0.940, CI=0.920-0.957), (AUC-LUSC=0.940, CI=0.922-0.957), and (AUC-Normal=0.992, CI=0.984- 0.998).
Conclusion: This research shows a strong ability to differentiate between Tumor/Normal tissue at SR=64microns/pixel and LUAD/LUSC/Normal tissue at SR=16microns/pixel. Further research is required to confirm that similar performance can be achieved through non-invasive imaging.
An open-source, customizable Winston-Lutz-style quality assurance system for multi-target, single isocentre, stereotactic radiosurgery
Patricia Oliver, Tania Wood, Lesley Baldwin
Cross Cancer Institute
Purpose: In cranial stereotactic radiosurgery (SRS), treatment time increases linearly with lesion number, increasing patient discomfort and motion-related errors. Treating multiple intracranial lesions simultaneously with a single isocentre is a practical solution. However, targeting accuracy of small treatment volumes away from the isocentre should be validated before adopting this technique. This work provides software and a phantom design to implement a multi-target Winston-Lutz (MTWL) test.
Methods: Plans for a 3D-printed phantom are provided to accommodate various target positions and beam geometries. The software generates multileaf collimator positions to facilitate plan creation. Target locations are found using the Hough circle detection algorithm modified to be robust in the presence of image noise and artefacts. The MTWL software works in conjunction with the Pylinac Python library.
Results: This software was validated by comparing with a synthetic phantom and with a commercial system. The maximum discrepancy is comparable to the resolution (0.2 mm) of the synthetic data, and discrepancies with the commercial system are < 0.5 mm. These discrepancies are considered reasonable since they are comparable to the variation among repeatability tests performed with our MTWL phantom and the commercial phantom.
Conclusion: This work presents a multi-target Winston-Lutz test. The software and phantom design are open-source and fully customizable. We have validated our software using synthetic data and by comparison with a commercial system. Successful completion of this test provides confidence in the linac’s ability to deliver a multi-target, single isocentre treatment plan with sufficient geometric accuracy according to the chosen tolerance level.
Initial Experience for Treatment Planning System Commissioning of the 0.5 T Inline Rotating Bi-Planar Linac-MR System
Shima Yaghoobpour Tari, Patricia Oliver, David Sinn, Tania Wood, Stephen Steciw, Brad Murray, Amr Heikal, Gino Fallone, Eugene Yip
Cross Cancer Institute, MagnetTx Oncology Solutions
Purpose: In preparation for the clinical use of the first 0.5 T inline hybrid Linac-MR (LMR), the 6 MV FFF radiation beam was characterized to model the beam for the treatment planning system (TPS). Following the TPS commissioning, validation tests were performed.
Methods: The LMR system has a magnetic field aligned with the radiation beam with isocentre of 120 cm and bore opening of 60 x 110 cm2. The LMR has a double focused MLC system to define the field and jaws only in the Y direction. The beam data was acquired using the PTW BEAMSCAN MR water tank with PTW semiflex 3D chamber for field sizes between 3 x 3 and 25 x 25 cm2. The beam was modeled within the Varian’s Eclipse system for Pencil Beam Convolution (PBC) algorithm. Point comparisons were made and PDD and profile validation were done using an in-house Gamma analysis code. Finally, the Delta4+MR phantom was used to evaluate a set of IMRT plans.
Results: Tissue Maximum Ratio and Relative Dose Factor measurements were within 0.8% compared to the calculation. With a 3%|3 mm criteria, PDD Gamma analysis showed 100% of points agreed at depths beyond Dmax, however, the buildup region had Gamma>1. The inline profiles were in agreement with an average of 96.42% and the crossline profiles with an average of 94.40% for 3% and 3 mm. The IMRT plans delivered on Delta4+MR had Gamma passing rates greater than 92.5%.
Conclusion: TPS for the 0.5 T LMR system was commissioned and validated to be used clinically.
A four-dimensional dynamic conformal arc approach for real-time tumour tracking: a treatment planning study
Timothy Yau, Stewart Gaede
Western University - London Regional Cancer Program
Purpose: To investigate a 4D dynamic tumour tracking (DTT) dose optimization using dynamic conformal arc (DCA) for normal tissue dose reduction compared to standard VMAT using an internal target volume (ITV) approach for stereotactic ablative radiotherapy (SABR) of lung cancer.
Methods: Thirty-three free-breathing 4D-CT-based VMAT lung SABR cases were retrospectively analyzed. The VMAT treatments were planned on the untagged average CT. Each DCA plan was created using the same beam geometry as the VMAT plan and normalized on the end-exhale CT. An in-house MATLAB script parsed both plans into each respiratory phase from the 4D-CT. The MLCs in the DCA plan conformed to the GTV in each respiratory phase to simulate DTT. Dose distributions for each plan were accumulated on the end-exhale CT. V5 and V20 for the normal lung and D99, D95, D50, D1, and mean dose for the GTV and critical organs were compared using a two-tailed, paired t-test.
Results: The GTV D95 in the DTT plan increased by an average of 0.18% (p=0.89). Significant decreases in all dose parameters for the normal lung were observed with the DTT plan (p<0.0001). Mean dose to critical organs also decreased by an average 54.5% (p<0.0001). In twelve cases, the spinal cord D1 dose increased by up to 5 Gy, but remained within clinical tolerance.
Conclusion: A conformal arc approach to DTT significantly improved healthy tissue sparing compared to an ITV-optimized VMAT plan while maintaining GTV dose. However, care must be taken when the GTV is near the spinal cord.
Improvement on Cone Beam Computed Tomography in radiation treatment with anti-scatter grid using a deep learning neural network
Yutong Zhao, Kaiming Guo, Richard Lee, Naseer Ahmed, Arbind Dubey, Boyd McCurdy
CancerCare Manitoba, Sunnybrook Health Science Centers
Purpose: Cone-beam computed tomography (CBCT) is a significant procedure in image-guided radiation therapy. The scattering effects bring the unwanted photons as contaminations into CBCT data, resulting in image artifacts and poor contrast. Anti-scatter grid (ASG) was the physical approach to compress this scatter effect. In this work, we developed a post-process method to generate a high-quality scatter estimation dataset with ASG for deep-learning CBCT scatter removal project.
Methods: Monte-Carlo simulation is the “golden rule” in research of radiation transport. Patient’s planning CT (pCT) used as input into EGSnrc software toolkit (EGS_cbct module) to perform the Monte-Carlo simulation (collimated point light source with a calculated 125 keV X-tube spectrum and 1 billion (1E9) photons) to get the scattered and primary CBCT projection image.
Results: The simulated total energy fluence at the imager is used as the input to predict the scatter-only energy fluence as the output, which is compared with ground truth (MC-simulated, scatter-only images). U-Net is the model architecture, which converges after a five-hour training process.
Conclusion: To account for the scatter transmission through ASG, we applied the grid transmission function, so the scatter of the CBCT projection after ASG can be calculated. Performing the volume reconstruction of those corrected CBCT projections, we obtained corrected CBCT (cCBCT) dataset which is used to compare with the planning CT dataset. As preliminary results show that the low frequency scatter are removed as well as CBCT image quality improved. In future, we will optimize our approach more and investigate the comparison HU unit between pCT and cCBCT.