Young Investigators Symposium - Salle de bal et Foyer

June 23, 2022 from 12:50pm EST to 2:30pm EST

Scientific Session 3 – Young Investigators Symposium
Thursday, June 23, 2022, 12:50-14:30

Scientific Session YIS – Presentation 1

Development and testing of an NLP and radiomics-based predictive model of pain for patients with thoracic spinal bone metastases

Hossein Naseri, Sonia Skamene, Marwan Tolba, Mame Daro Faye, Julia Khriguian, Paul Ramia, John Kildea
McGill University Health Center, Division of  Radiation Oncology - McGill University Health Center, Radiation Oncology Department - Jewish General Hospital, Medical Physics Unit - McGill University Health Centre

Purpose: To evaluate the feasibility of building a radiomics–based machine learning (ML) pipeline for differentiating painless and painful bone metastases (BM) lesions in planning-CT images using lesion-canter-based regions of interest (ROIs).

Materials and Methods: A total of 216 pairs of consultation notes and corresponding planning-CT images of 179 patients (mean age (standard deviation), 68 (士12) y; 91 male) who previously received palliative radiotherapy for T-spine BM between 2016 and 2019 were used in this retrospective study. 631 BM centers were manually identified in the CT images. Physician-reported pain scores were automatically extracted from consultation notes using our in-house natural language processing (NLP) pipeline. Spherical ROIs with various diameters were automatically delineated around expert-identified BM centers, and 107 radiomics features were extracted for each. Data were divided into 70-30% train-test sets. The SMOTE re-sampling and LASSO feature selection methods were applied to the training set. Gaussian process regression (GPR) ML classifiers were trained and evaluated on the test set using precision, sensitivity, F1-score, and area under the receiver operating characteristic curve (AUC).

Results: In the training set, the average AUC, precision, sensitivity, and specificity of our best performing pipeline (on ensemble ROI) from 5-fold cross-validation were 0.91, 0.84, 0.95, and 0.81, respectively. The AUC, precision, sensitivity, and specificity in the test set were 0.81, 0.61, 0.74, and 0.89, respectively.

Conclusion: Our lesion-center-based radiomics pipeline proved capable of differentiating between painful and painless BM lesions.


Scientific Session YIS – Presentation 2

Using deep learning generated CBCT contours for on-line dose assessment of prostate SABR treatments

Conor Smith, Isabelle Gagne, Karl Otto, Carter Kolbeck, Joshua Giambattista, Abraham Alexander, Sonja Murchison, Andrew Pritchard, Erika Chin
University of Victoria, BC Cancer, limbusAI Inc

Purpose: Assuring precision treatment for prostate SABR patients using CBCT images is challenging due to the lack of accurate tissue structure localization and dose estimation.  A common compromise is to use decision trees (DT) based on qualitative visualization of CBCT images.  In this work we developed a system for evaluating patient position that is based on dose that would be received to AI segmented OARs. We then evaluated the current DT method using the new system.

Methods: Pre-treatment CBCTs for ten prostate SABR patients were segmented using Limbus Contour. Models were validated by comparing Limbus Dice Similarity Coefficients (DSC) and Hausdorff Distances (HD) to Expert Interobserver Variability (IOV). On-line registration data was used to superimpose planned dose onto CBCT images (validated with 2%/2mm gamma analysis) and combined with Limbus contours to obtain DVHs. We then compared the DVHs to the planning goals for each fraction.

Results: Relative differences in DSC and HD values for Limbus contours as compared to Expert IOV are ≤ 3.2% and are comparable to expert contours. Over the course of treatment, 8/10 patients did not meet at least one of the mandatory rectum goals, and two patients had a major violation of rectum dose-volume metrics in every fraction treated. When a bladder mandatory goal was not met, the volume was smaller compared to planning CT in all fractions.

Conclusions: Considering the frequency and systematic nature of the mandatory goal violations, we recommend incorporating on-line dose assessment using AI contours into clinical practice.


Scientific Session YIS – Presntation 3

Robust planning of mixed electron-photon radiation therapy for soft tissue sarcoma at standard SAD

Veng Jean Heng, Monica Serban, Marc-André Renaud, Jan Seuntjens
McGill University, Health Centre - McGill University, Gray Oncology Solutions, Princess Margaret Cancer Centre

Purpose: Mixed electron-photon beam radiation therapy (MBRT) utilizes external electron and photon beams to minimize doses to normal tissues by exploiting the electron’s sharp dose falloff. Due to significant in-air scatter, photon-MLC (pMLC) collimated electron beams have necessitated shortened SSD to achieve satisfactory conformality with MBRT. This study demonstrates the feasibility of using pMLC-collimated electron beams at standard SAD as part of MBRT plans.

Methods: A retrospective cohort of 14 soft tissue sarcoma (STS) patients treated using VMAT were replanned with MBRT on an in-house treatment planning system. Both VMAT and MBRT plans were planned for the Varian TrueBeam linac. MBRT plans consisted of an electron component, including up to 5 energies, and a 6 MV photon component. Both modalities were collimated with only pMLCs and planned at standard SAD of 100cm as step-and-shoot apertures. Robust optimization was performed on the CTV by accounting for 5mm setup error scenarios using a column generation algorithm. The final dose distributions were recalculated by Monte Carlo using EGSnrc.

Results: Compared to the clinical VMAT plan, all MBRT plans achieved superior organs-at-risk sparing without the use of bolus. Coverage of the CTV by 95% of the prescription dose was found to be more robust to setup errors with MBRT than VMAT. Wider electron penumbras at standard SAD were compensated by the optimizer by favoring sharp penumbra photons.

Conclusion: For STS, MBRT plans using electrons at standard SAD is dosimetrically equivalent to shortened SSD. MBRT deliveries can thus be automatically performed without intra-fraction couch repositioning.


Scientific Session YIS – Presentation 4

Design and implementation of a prototype radiotherapy menu in a patient portal to reduce patient anxiety and facilitate data sharing

Kayla O'Sullivan-Steben, Luc Galarneau, Susie Judd, Andréa Laizner, Tristan Williams, John Kildea
McGill University

Purpose: To improve patient understanding of and access to their radiotherapy data through the use of a novel Radiotherapy menu in a patient portal incorporating the mCODE data standard.

Methods: A prototype Radiotherapy menu was developed in our institutional patient portal following a participatory stakeholder co-design methodology. Customizable page templates were designed to render key radiotherapy data in the portal’s patient-facing mobile phone app. DICOM-RT data were used to provide patients with relevant treatment parameters and generate pre-treatment 3D visualizations of planned treatment beams, while the mCODE data standard was used to provide post-treatment summaries of the delivered treatments. A focus group was conducted to gather initial patient feedback on the menu.   

Results: Pre-treatment, the Radiotherapy menu provides patients with a personalized treatment plan overview, including a personalized explanation of their treatment, along with an interactive 3D rendering of their body and treatment beams for visualization. Post-treatment, a summary of the delivered radiotherapy is provided, allowing patients to retain a concise personal record of their treatment that can easily be shared with future healthcare providers. Focus group feedback was overwhelmingly positive. Patients highlighted how the intuitive presentation of their complex radiotherapy data would better prepare them for their radiation treatments.

Conclusion: We successfully designed a prototype Radiotherapy menu in our institution’s patient portal that improves patient access to and understanding of their radiotherapy data. We used the mCODE data standard to generate post-treatment summaries in a way that is easily shareable and interoperable.


Scientific Session YIS – Presentation 5

Expiration CT Texture-Based Radiomics: A New Biomarker for Assessing the “Quiet” Zone in COPD?

Meghan Koo, Ryan Au, Wan Tan, Jim Hogg, Jean Bourbeau, Miranda Kirby
Ryerson University, Western University, Centre for Heart Lung Innovation, Montreal Chest Institute of the Royal Victoria Hospital

Purpose: To investigate the association between texture-based radiomics features on expiration computed tomography (CT) images with established pulmonary function measures reflecting lung hyperinflation, and compare these emerging features to existing measures of CT gas trapping. 

Materials and Methods: Participants from the longitudinal CanCOLD cohort performed full-expiration CT imaging and plethysmography for measurement of the residual volume to total lung capacity (RV/TLC) at baseline and 3-year follow-up.  CT gas trapping measurements included the low attenuation below -856HU (LAA856) and disease probability map of functional small airway disease (DPMfSAD). A total of 86 features from five texture-radiomics feature sets were extracted. The features were used for principal component analysis (PCA) to create a new representative variable. Multiple linear regression models (MLR) were used to determine association for baseline and annualized change in RV/TLC with the extracted PCA measurement, LAA856 and DPMfSAD.

Results: A total of 1111 participants were investigated (n=234 never-smoker, n=325 at-risk, n=314 mild COPD, n=238 moderate-severe COPD). In separate MLR models to predict RV/TLC at baseline, all three CT measurements were independently significant (p<0.001), however the PCA component had the highest R2 value (PCA: R2=0.76; LAA856: R2=0.51; DPMfSAD: R2=0.66). Among the three CT measurements for predicting 3-year ΔRV/TLC, the PCA component had the highest R2 value (PCA: R2=0.20, p<0.001; LAA856: R2=0.19, p=0.032; DPMfSAD: R2=0.19, p>0.05).

Conclusions: Our results show that CT texture-based radiomics features may be able to better quantify the heterogeneous structural changes that occur due to small airways dysfunction than existing CT gas trapping measurements in COPD.


Scientific Session YIS – Presentation 6 

Validation of a daily adaptive radiation therapy pipeline incorporating machine learning automated treatment planning on CBCT with dose accumulation

Aly Khalifa, Maryam Golshan, Inmaculada Navarro, Jason Xie, Chris McIntosh, Thomas G. Purdie, Victor Malkov, Tony Tadic, Jeff Winter
University of Toronto, Princess Margaret Cancer Centre

Purpose: To demonstrate and validate a machine learning (ML) treatment planning pipeline for prostate adaptive radiation therapy (ART) on daily iterative-reconstructed cone beam computed tomography (CBCT) imaging.

Methods: For CBCT dose computation, we generated a CBCT-to-density table by performing deformable image registration (DIR) between CT and CBCT images collected from 5 patients with prostate cancer and determining the correlation between CT and CBCT voxel histograms. We simulated daily 4270 cGy in 7 fraction CBCT-based ART plans for 10 patients with prostate cancer using a ML model trained on CT imaging only. We validated the CBCT-to-CT DIR using Dice coefficients between DIR-propagated and expert manual contours and performed dose accumulation across all fractions. We compared accumulated dose for daily ART with the reference plan using clinically relevant dose-volume objectives.

Results: Dose calculation on CBCT images demonstrated excellent accuracy, with 99.7 ± 0.8% of all irradiated voxels within 1%/1mm agreement with CT, averaged over all patients. Mean Dice coefficients between manual and DIR-propagated CTV, bladder and rectum were ≥ 0.97. Comparing the reference plan to the accumulated adapted plans, the reference planning target volume (PTV) V95% was 99.7±0.1% and adapted clinical target volume (CTV) V95% was 100±0.001%. Rectum V90% and V75% were 6.6±1.9% and 11.4±2.6% on the reference plan, and 3.4±1.3% and 7.6±1.9% for adapted plans. 

Conclusions: We successfully validated an end-to-end pipeline for assessing cumulative dose for daily ART using ML planning on CBCT for prostate treatment and dose accumulation. Adapted plans showed improved rectum sparing while maintaining target coverage.


Scientific Session YIS – Presentation 7

Low-cost, kilovoltage, isocentric radiotherapy (kVIRT) machine design optimisation and treatment characterisation

Jericho O'Connell, Michael Weil, Magdalena Bazalova-Carter
University of Victoria, Sirius Medicine LLC

Purpose: To optimise, design, and benchmark a low-cost, kilovoltage, isocentric radiotherapy (kVIRT) machine for deployment in underserved populations globally.

Methods: A novel, robust, low-cost treatment machine design has been proposed featuring a kV x-ray tube with the intent of delivering 18 Gy to a 10 cm deep lesion with a target-to-skin dose ratio (SDR) of 4.5 within a 30 minute treatment. EGSnrc Monte Carlo (MC) simulations of x-ray tubes with peak voltages of 225, 300, and 320 kV, respectively, were used to ascertain whether the SDR, spatial dose distributions, and beam currents were adequate to meet the dose and dose rate constraints. Dose distributions were assessed with the TOPAS MC code for 360 degree arc treatments of cylindrical water phantoms, an XCAT head phantom, and a lung patient to investigate the additional benefits of gold nano-particles (GNPs) and non-coplanar beams.

Results: The 225 and 320 kV x-ray simulations demonstrated doses exceeding 18 Gy for treatment lengths of 20 minutes at depths of 10 cm with SDRs of 3.2 and 4.0 for a 360 degree arc. With the addition of two non-coplanar beams, SDRs increased to 9.2 and 11.3, respectively. The addition of GNPs resulted in significantly decreased OAR DVH values.

Conclusion: We demonstrate a novel, low-cost, kilovoltage, isocentric radiotherapy system meeting our initial treatment targets. Design optimisation will need to be followed by manufacturing and experimental characterisation ahead of utilisation of the system in low-income countries.


Scientific Session YIS – Presentation 8

Development of a hybrid alanine-calorimetry absorbed dose standard for linac electron beams

Rodi Surensoy, Bryan Muir
National Research Council of Canada - Carleton University

Purpose: To develop an electron beam absorbed-dose standard by calibrating alanine dosimeters traceable to calorimetry primary standards in high-energy beams and thereafter using alanine to establish absorbed dose in low-energy beams.

Methods:  Alanine dosimeters are irradiated in a water phantom with 4 MeV to 22 MeV linac electron beams.  A horizontal beam geometry is used with an SSD of 100 cm and a 10 × 10 cm2 field size shaped by a clinical applicator. A Bruker EPR spectrometer is used to read out the intensity of the alanine signal.  In order to use alanine to determine absorbed-dose in low-energy electron beams, the alanine signal from the spectrometer must be independent of electron-beam energy. This variation is quantified using Monte Carlo calculations of the absorbed dose to water and the absorbed dose to alanine for the beam energies investigated. Five cylindrical and parallel-plate ionization chambers are calibrated against this absorbed dose standard. Results are corrected for alanine absorbed-dose energy-dependence and holder (wall) effects.

Results: Monte Carlo simulations show that absorbed-dose energy-dependence of alanine is ±0.2 % for the electron beam energies investigated. Beam quality conversion factors for the ionization chambers investigated here are in good agreement with literature data considering combined systematic uncertainties (0.8 %, k=1) for the results of this work.

Conclusions: The low absorbed-dose energy-dependence of alanine allows the use of alanine as an absorbed-dose standard in low-energy electron beams. The agreement of results measured here with literature data demonstrates the accuracy of this hybrid alanine-calorimetry standard.


Scientific Session YIS – Presentation 9

Xenon-enhanced Dual-energy Chest Tomosynthesis for Functional Imaging of Respiratory Disease

Fateen Basharat, Jesse Tanguay
Ryerson University

Purpose: To investigate the feasibility of xenon-enhanced dual energy (XeDE) chest tomosynthesis for functional imaging of lung using an experimental phantom study.

Methods: We performed experiments using a custom-built x-ray imaging cabinet and a chest phantom consisting of a xenon-filled chamber, solid acrylic slabs and Al slats. The fractional molar concentration of xenon inside the chamber was set to 0.5. A sealed air-filled cavity was positioned inside the vacuum chamber simulating a ventilation defect. Low-energy and high-energy projection images were acquired with an angular range of ±30^o  in increments of 1^o. The low-energy and high-energy tube voltages were fixed at 60 kV, 140 kV with 1.1 mm added Cu filter on the HE spectrum, respectively. Images were acquired with Rad-92 X-ray tube and a CsI/CMOS energy integrating X-ray detector (XINEOS-3030HS) with 151.8 μm pixel pitch. The total entrance air kerma was 7 mGy. Low-energy and high-energy coronal slices were reconstructed using filtered backprojection with a coronal of 5 mm. From the reconstructed low-energy and high-energy slices, a soft-tissue suppressed XeDE tomosynthesis image was produced.

Results: The XeDE tomosynthesis slices permitted visualization of the ventilation defect by providing contrast between xenon background and unventilated region of the phantom. The acrylic boundary of the defect is suppressed. The relative contrast of the ventilation defect in XeDE tomosynthesis slice was 57%.

Conclusion: Xenon-enhanced DE tomosynthesis may enable functional image of respiratory disease and should be further investigated as an alternative to approaches to functional imaging based on computed tomography and magnetic resonance imaging.


Scientific Session YIS – Presentation 10

Accurate, on-demand neural networks for respiratory motion forecasting

Neil Johnson, Keith Wachowicz, Satyapal Rathee, Gino Fallone, Jihyun Yun
University of Alberta - Cross Cancer Institute

Purpose: Developing accurate, on-demand neural networks for respiratory motion forecasting for non-invasive intra-fractional tumour-tracked radiotherapy (nifteRT) using a hybrid linear accelerator/magnetic resonance imaging system (Linac-MR).

Methods: Long short-term memory recurrent neural networks (LSTM-RNNs) are trained on 120 seconds of 3D tumour motion data to predict a target’s future coordinates based on its previously observed positions. Accuracy is evaluated as the root mean square error (RMSE) over 300 second simulated treatments (42 patients, 158 fractions) for a range of imaging rates and forecasting times.

Super-convergence regularization prevents overfitting and greatly accelerates training, and homogeneous network ensembles mitigate solution instability without extending training time. This allows us to propose a novel network adaptation strategy in which networks are intermittently fully retrained during treatment. We then suggest a clinically feasible strategy for nifteRT on a Linac-MR using on-demand, fraction-specific neural network-based motion forecasting.

Results: We report a mean RMSE of 0.54 mm (range 0.04 – 1.68, SD 0.33) for abdominothoracic tumours with a mean motion amplitude of 4.92 mm (range 0.32 – 14.51, SD 3.08) at an acquisition rate of 240 ms and a forecasting time of 280 ms. RMSE is found to be independent of acquisition rate between 120 and 280 ms, and to roughly double with tripled forecasting time between 120 and 520 ms. Intermittent retraining is found to yield 18% better results than traditional online learning.

Conclusions: Our novel approach to network training results in improved accuracy as well as a practical method for producing on-demand neural networks for respiratory motion forecasting.