ESTRO 2025 - Abstract Book

S3420

Physics - Machine learning models and clinical applications

ESTRO 2025

3304

Digital Poster Outcomes for Human versus Machine Learning Prostate-Only Radiation Therapy Treatment Planning Jeff D Winter 1,2 , Leigh Conroy 1,2 , Matthew Ramotar 1 , Anna T Santiago 3 , Charles Catton 1,2 , Peter Chung 1,2 , Chris Mcintosh 4 , Thomas G Purdie 1,2 , Ale Berlin 1,2 1 Radiation Medicine Program, Princess Margaret Cancer Centre, Toronto, Canada. 2 Department of Radiation Oncology, University of Toronto, Toronto, Canada. 3 Department of Biostatistics, University Health Network, Toronto, Canada. 4 Department of Medical Biophysics, University of Toronto, Toronto, Canada Purpose/Objective: Machine learning (ML) automation in radiotherapy (RT) treatment planning offers potential improvements in efficiency and standardization. Previous ML planning performance clinical evaluation focused on comparing ML generated and human-generated RT plans using quantitative dosimetric evaluation and physician RT plan selection in a controlled study 1 . However, to date there has been no assessment of clinical outcomes in a real-world RT workflow. Our objective here is to compare genitourinary (GU) and gastrointestinal (GI) toxicity differences between ML- and human-generated RT plans in a prospective patient cohort with prostate cancer to further support wide scale ML planning adoption. Material/Methods: We prospectively evaluated ML-generated RT plans for standard-of-care 60 Gy in 20 fractions prostate-only patients treated between August 2020 and September 2022. ML planning was performed with a clinically validated ML dose prediction model combined with a dose optimization step to produce clinically deliverable RT plans. ML planning without any user intervention was the default RT planning approach for all cases. The treating radiation oncologist had the option to either select the ML-generated plan for patient treatment or to request a human-generated RT plan for comparison before selecting either the ML- or human-generated RT plan for treatment. To assess outcomes, we extracted physician-reported GU and GI toxicities from the electronic medical record with a minimum 180-day follow-up. To compare clinical outcomes between ML- and human-generated plans, we performed a toxicity-free survival Kaplan-Meier analysis for grade 2+ GI and GU toxicities and assessed significance via the log rank test. Results: The study cohort included 93 patients with a median follow-up of 34 (range 6 – 51) months. Radiation oncologists selected the ML-generated RT plan for treatment in 76 of 93 (82%) of cases, which is greater than the previously reported 61% of ML-generated RT plans selected for patient treatment using the same ML model in a different clinical workflow 1 . There were no treatment-related grade 2+ GI toxicities in the study cohort. Figure 1 illustrates the toxicity-free survival for GU grade 2+ toxicities, with the log-rank test showing no significant differences between ML- and human-generated RT plans (p = 0.46).