Session 5: Automation and Adaptive Planning - Lombardy/Umbria
le 8 juin 2024 from 11h00 CDT to 12h10 CDT
Scientific Session 5 – Automation and Adaptive Planning
Saturday, June 8, 2024, 11:00-13:10
Scientific Session 5: Automation and Adaptive Planning – Presentation 1
Leveraging IVIM for HCC Hypoxia Assessment for an Online Adaptive MR-Linac System
Ryan Kuhn, Jean-Pierre Bissonnette, Teodor Stanescu, Catherine Coolens, Laura Dawson, Michael Maddalena
University of Toronto,
Purpose: Liver cancer hypoxia may explain poor radiotherapy (RT) outcomes since it increases both radio-resistance and metastatic potential. However, RT planning CT only provides anatomical information. Fortunately, functional MR-based techniques like, intravoxel incoherent motion (IVIM), can create hypoxia maps. By integrating IVIM maps into the online adaptive RT pipeline, we can integrate a boost dose to overcome hypoxia. Using HIF-1a expression as a hypoxia marker helps generate thresholds; however, IVIM parameters specific to liver HCC need to be determined.
Methods: Four patients with hepatocellular carcinoma were imaged using an IVIM sequence and treated on an MR Linac (Unity, Elekta, Stockholm). Contours from the clinical plan were extracted and IVIM data analyzed, creating maps indicating hypoxic potential. IVIM parameters were found by fitting an equation to the signal intensity of varying b-values. Thresholds were calculated to define clusters of high hypoxia scores. Contours were created delineating these clusters and subsequently exported back to the treatment pipeline.
Results: 50% of tumour volumes had hypoxic clusters and 7.6+/-2.6% of those volumes were hypoxic. By adjusting the threshold in accordance with HIF-1a expression, the percent of tumours with hypoxic clusters was unchanged but 14.9+/-4.9% of tumour volumes were found to be hypoxic.
Conclusions: Incorporating IVIM imaging to identify potentially hypoxic regions is feasible and can be integrated into existing adaptive pipelines. Efforts to determine IVIM thresholding specific to primary and metastatic liver cancer are ongoing, offering novel treatment approaches leading ultimately to improved patient outcomes.
Scientific Session 5: Automation and Adaptive Planning – Presentation 2
Evaluation of an Auto-Segmentation Software after 2+ years of Clinical Us.
Derek Liu, Laurie Rumpel, Cody Crewson, Nick Chng, Matthew Sauder, Joshua Giambattista,
Cross Cancer Institute, Allan Blair Cancer Centre, Saskatoon Cancer Centre, BC Cancer Agency Centre for the North
Purpose: We performed an observational study to evaluate an auto-segmentation software after 2+ years of clinical use. Agreement with software contours and workflow task duration times were compared against baseline metrics prior to implementation.
Methods: 1600 patients treated with standard radical external beam treatments between July 2018 and Jan 2023 for selected treatment sites: H&N, breast, lung, and prostate. The auto-segmented structures were compared against clinical contours used during treatment, which were either independently user-generated (before software implementation) or user-modified (after implementation). Metrics included the Dice Similarity Coefficient (DSC) and 95% Distance to Agreement (DTA). In addition, contouring task durations were extracted from the patient management system.
Results: Noticeable increase in agreement with auto-segmented contours was observed across all sites after software implementation. In particular, the mean DSC for the brachial plexus increased from 0.46 to 0.98. A chronological case-by-case review showed an initial transient period where users scrutinize the auto-generated contours but eventually accepted it with minimal modifications. However, the spinal cord agreement remained relatively constant over time for H&N and lung (DSC ~ 0.75, DTA ~ 2 mm) while increased for breast plans involving supraclavicular fields (DSC ~ 0.94, DTA ~ 0.6 mm). The oncologist contouring time for H&N and the dosimetrist time for prostate decreased approximately by 50% on average.
Conclusions: User adaption to the software was observed over time, with increased overall agreement with the auto-generated contours and continued user-edits of high priority structures. Contouring times were significantly reduced for H&N and prostate.
Scientific Session 5: Automation and Adaptive Planning – Presentation 3
Pulling the Trigger: Transitioning from Triggered Imaging with Auto Beam-Hold to Fiducial.
Amanda Cherpak, Jennifer DeGiobbi, R Lee MacDonald, Alex Riopel, James L Robar, Kenny Zhan, Hannah Dahn
QEII Cancer Centre
Purpose: To analyze motion for prostate SBRT patients based on triggered imaging, TI, with auto-beam hold to inform IGRT for adaptive treatment on Ethos with HyperSight.
Methods: Seventeen patients with implanted gold seeds received 36.25 Gy/5 fx (PTV margin = 4 mm) on a TrueBeam STx linac. Images were acquired every 5 s during treatment and auto beam-hold was activated for seed motion ≥3 mm. Two additional patients received 60 Gy/20 fx fiducial-free adaptive treatment (PTV margin = 7 mm) on Ethos with HyperSight. Treatment times and patient shift data were reviewed in preparation of adaptive SBRT dose delivery.
Results: Five fractions were excluded from TI analysis due to setup issues. Average 3D shift for the remaining 80 fx’s was 3.0 mm ± 1.9 mm, IQR = 1.8 mm. Four shifts resulted in a vector ≥4 mm. For these treatments, every interruption resulted in a shift to improve seed position. Ignoring shifts < 1 mm, 66 fx’s (95%) had ≤2 beam interruptions, 3 fx’s had 3, and 1 fx had 4. Average elapsed time following a CBCT before beam-hold was 6.1 min ± 2.3 min, IQR = 2.6 min. On HyperSight CBCTs, patient position shifted by 0.11 mm/min (patient 1, R2 = 0.75) and 0.16 mm/min (patient 2, R2 = 0.81) during the adaptive process.
Conclusions: Fiducial-based TI resulted in a maximum of 2 repeat CBCTs for almost all patients. Based on predicted beam delivery time, inter-field HyperSight CBCT could be scheduled considering patient shift trends and margins.
Scientific Session 5: Automation and Adaptive Planning – Presentation 4
An updated implementation of 4π radiotherapy visualization in Eclipse using the ESAPI environment.
Brian Little, Christopher Thomas, R. Lee MacDonald, Alasdair Syme, Christopher G. Thomas
Nova Scotia Health
Purpose: Utilize the Eclipse® ESAPI programming interface to develop a 4π radiotherapy system, encompassing features such as collision detection, providing beams-eye-view (BEV) analysis, ray tracing BEV overlap with OARs, and 3D visualizations.
Methods: The research team enhanced and re-implemented previous 4π radiotherapy methods using C# (Microsoft Corporation, Redmond WA, USA), incorporating libraries for geometric calculations and mesh manipulation (geometry3Sharp), along with visualization tools (HelixToolkit). A collision model was created by deforming a cylinder around the patient scan and using a 3D CAD model of the gantry and couch, enabling the detection of gantry/couch and gantry/patient collisions. Raytracing was employed through PTV/OAR overlap in the BEV to assign costs to the objective function for selecting apertures for optimized trajectories.
Results:The ESAPI 4π radiotherapy system allows users to evaluate/visualize/optimize cranial and extra-cranial treatment plans within the TPS. Due to the heavily optimized data structures used within this implementation, the calculation total time for collision detection and ray tracing has decreased from approximately 2 hours to under 15 minutes. Further investigation is needed to assess the scalability of this speed-up with variations in ray resolutions and the reduction of vertices and faces used on CAD meshes within the collision model.
Conclusion: An ESAPI 4π radiotherapy software solution was implemented using C# and provides users with the capability to evaluate, visualize, and optimize treatment plans for both cranial and extra-cranial cases directly within the treatment planning system (TPS).
Scientific Session 5: Automation and Adaptive Planning– Presentation 5
Substantial Reduction in Irradiated Normal Brain Volume Using a Small-Margin Weekly Adaptive Protocol on the 1.5T MR-Linac.
Thomas Mann, James Stewart, Amir Safavi, Jay Detsky, Chia-Lin Tseng, Deepak Dinakaran, Hany Soliman, Sten Myrehaug, Hanbo Chen, John Hudson, Shawn Binda, Arjun Sahgal, Mark Ruschin
Sunnybrook Odette Cancer Centre,
Purpose: The standard-of-care for glioblastomas includes radiotherapy using a 1.5cm-2.0cm clinical target volume (CTV) margin around the gross tumor volume (GTV). The UNITED adaptive MR-Linac trial used a 5mm CTV margin (plus discretionary FLAIR signal) with weekly adaptation of target volumes. This study evaluates the dosimetric impact of small margin adaptive versus conventional margin radiotherapy.
Methods: Ten patients treated under the UNITED protocol were re-planned using a conventional 1.5cm CTV margin contoured according to EORTC guidelines. EORTC and UNITED plans used the same dose prescription of 60Gy/30 fractions and clinical objectives. Plans were created using the Monaco treatment planning system, Unity 7FFF beam model, and IMRT delivery. Relevant dose-volume metrics for organs-at-risk (OAR) were compared between planning methods using Wilcoxon signed-rank tests (α=0.007, multiple comparisons). Coverage of weekly dynamic target volumes with the static EORTC plan was also assessed.
Results: The reduction in irradiated normal brain volume was statistically significant with a median V60Gy of 98.4 cm3 for UNITED plans compared to 169.8 cm3 with the EORTC margin (p=0.005). Differences in median max doses to the brainstem, eyes, lenses, optic nerves, and chiasm were not significant between planning methods however a reduction in intermediate doses was observed. Four patients had CTV coverage (V57Gy) with the conventional EORTC plan fall below 95% during their 6 week course due to adapted CTV volumes growing outside the original PTV.
Conclusions: Our adaptive UNITED approach substantially reduced the volume of normal brain tissue being irradiated while ensuring full target coverage and respecting OAR doses.
Scientific Session 5: Automation and Adaptive Planning – Presentation 6
Evaluating the performance of an MR tumor autocontouring algorithm in the context of intra- and inter-observer variabilities.
Gawon Han, Arun Elangovan, Jordan Wong, Asmara Waheed, Keith Wachowicz, Nawaid Usmani, Zsolt Gabos, B. Gino Fallone, Jihyun Yun
University of Alberta, BC Cancer
Purpose: To quantify inherent intra- and inter-observer variations in manual contours from multiple experts on MR images of cancer patients, and evaluate our tumor autocontouring algorithm against the manual contours.
Methods: 30 cancer patients (10 liver, 10 prostate and 10 lung) were scanned using a 3T MRI with a 2D bSSFP sequence at 4 frames/s. Three experts, each in two separate sessions, manually contoured the tumor in 300 dynamic images/patient. For autocontouring, an in-house built U-Net-based autocontouring algorithm was used, whose hyperparameters were optimized with Covariance Matrix Adaptation Evolution Strategy for each patient, expert, and session. 160 images were used for hyperparameter optimization and training of the algorithm, 70 for validation, and 70 for testing. The contours were compared using Dice’s coefficient (DC) and Hausdorff distance (HD).
Results: Autocontours generated by our algorithm (patient-, expert-, and session-specifically trained) were compared against manual contours from the (a) same expert, same session; (b) same expert, different session; and (c) different experts. For (a)-(c), DC=0.91, 0.86, 0.78 and HD=3.1, 4.6, 7.0 mm were achieved, respectively. The evaluation metrics for (b) surpassed the manual intra-observer variations observed between two sessions (DC=0.85, HD=4.9 mm), and those for (c) exceeded the manual inter-observer variations (DC=0.77, HD=7.2 mm).
Conclusions: Our algorithm generates tumor autocontours that faithfully emulate the contouring tendencies of each expert, achieving 91% accuracy to manual contours, but with remarkable efficiency (<55 ms). The consistency between autocontours and manual ones is comparable to the intra- and inter-observer variabilities observed across liver, prostate, and lung cases.
Scientific Session 5: Automation and Adaptive Planning– Presentation 7
Automatic segmentation of Organs at Risk in Head and Neck cancer patients from CT and MRI scans.
Sebastien Quetin, Andrew,] Heschl, Mauricio Murillo, Shirin A. Enger, Farhad, Maleki
McGill University, University of Calgary
Purpose: Deep Learning has been widely explored for Organ at Risk (OAR) segmentation; however, most studies have focused on a single modality, either CT or MRI, not both simultaneously. This study presents a high-performing Deep Learning pipeline for automated segmentation of 30 OARs from MRI and CT scans in Head and Neck (H&N) cancer patients.
Methods: The segmentation challenge provided CT and MRI T1 data from 42 H&N cancer patients alongside annotation for 30 OARs to develop a segmentation pipeline. First, we performed non-rigid registration of CT and MRI volumes, followed by cropping to remove irrelevant regions. We created two versions of the CT volume to represent soft and bone tissues. These were concatenated with the MRI volume and used as input to an nnU-Net pipeline. The trained model was used to predict the segmentation masks for an independent test set of 14 new patients, only available through the challenge website and not directly accessible to the challenge participants. Mean Dice score and Hausdorff distance were used to assess the performance of the trained model.
Results: For the independent test set, the mean Dice score and Hausdorff distance were 0.778±0.107 and 5.717±6.480, respectively. As a comparison, the prior state-of-the-art achieved by other participants resulted in a mean Dice score of 0.768±0.119 and a mean Hausdorff distance of 3.542±1.623.
Conclusion: The proposed pipeline achieved the highest Dice score among all participants and sets a new state-of-the-art for H&N OAR segmentation in cancer patients.