ESTRO 2024 - Abstract Book

S3739

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

Purpose/Objective:

The ESTRO 2023 Physics Workshop hosted the Fully Automated Radiotherapy Treatment Planning (Auto-RTP) Challenge. The challenge was designed to take a snapshot and address the question: How close are we to fully automated radiotherapy treatment planning? To answer this question, the challenge provided participants with CT images from 16 prostate cancer patients (mix of prostate only, prostate + nodes, post-operative prostate bed + nodes) across 3 challenge stages with the goal of automatically generating treatment plans with minimal user intervention. The purpose of this report is to describe our team’s (UABRO-KillingCancerWithCode) winning approach which was developed to rapidly adapt to contouring guidelines and treatment prescriptions that differ significantly from those used in our clinical practice.

Material/Methods:

Our fully-automated planning pipeline was composed of two major components: 1) auto-contouring and 2) auto planning engines. Both engines were developed in-house and triggered via advanced DICOM transfer operations. The auto-contouring engine utilized a dataset of 600 prostate cancer patients previously treated in our clinic to train 3D U-Net models for normal tissues (seminal vesicles, bladder, rectum, bowels, and bowel bag), as well as, prostate, nodal, and surgical bed clinical target volumes. The auto-planning engine, built using the Eclipse Scripting Application Programming Interface (Varian Medical Systems), automatically defined target volumes, field geometry, and all relevant planning parameters, prior to automated optimization and dose calculation. RapidPlan models were developed using the RapidCompare tool previously developed by our group 1 . An advanced optimization scheme was implemented to generate a training plan dataset (n=150 cases, 50 for each treatment type) for RapidPlan model generation. Elements of this optimization scheme were used to iteratively re-train and refine plan quality to meet challenge DVH objectives and goals. We report overall (0-100, where 100 is a perfect score, overall = OAR + target scores) , organs-at-risk (OAR, 0-50, where 50 is a perfect score), and target (0-50, where 50 is a perfect score) scores from the automatically-generated plans on our own contours (used for planning) and “consensus” contours (which were used to score challenge submission leaderboards) across all three stages of the challenge.

Results:

Our team placed 1st in all three stages of the competition. Our leaderboard scores were 79.9, 77.3, and 78.5, on the Data Format Confirmation, Web Contest, and Onsite Event stages, respectively. On our own contours, average (+/- std dev) OAR and target scores were 43.2 (+/- 7.1) and 41.6 (+/- 3.0), 42.4 (+/- 10.1) and 41.2 (+/- 1.7), and 42.7 (+/- 5.4) and 40.9 (+/- 4.2) for the Data Format Confirmation, Web Contest, and Onsite Event stages, respectively. On the consensus contours, average (+/- std dev) OAR and target scores were 41.5 (+/- 6.5) and 38.4 (+/- 10.1), 41.4 (+/- 4.6) and 35.9 (+/- 6.6), and 40.6 (+/- 7.1) and 37.9 (+/- 10.8) for the Data Format Confirmation, Web Contest, and Onsite Event stages, respectively. Highest scores were observed for prostate only cases with an average overall score greater than 90 for all cases.

Conclusion:

We developed an automated planning pipeline for prostate cancer radiotherapy planning. Our approach was able to adapt well to new contouring and planning guidelines. Future studies are needed to confirm clinical acceptability of automatically generated plans and their readiness for clinical integration.