S1970
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
Digital Poster 3967
Robustness evaluation of Deep Learning auto- planning models for breast cancer on out-of distribution (OOD) patients Michele Zeverino 1 , Anastasiia Barysheva 1 , Maud Marguet 1 , Wendy Jeanneret-Sozzi 2 , Fernanda Herrera 2 , Emil Schueler 3 , Raphael Moeckli 1 1 Institute of Radiation Physics, Lausanne University Hospital, Lausanne, Switzerland. 2 Radiation Oncology Department, Lausanne University Hospital, Lausanne, Switzerland. 3 Department of Radiation Physics, MDACC, Houston, USA Purpose/Objective: This study aimed to evaluate the robustness of a deep learning (DL)-based auto-planning model developed for left-sided breast cancer and its adapted version for right-sided cases1. The assessment focused on the quality of automated treatment plans generated for an unseen cohort of patients with larger breast volumes—representing cases outside the original training distribution—compared with two control cohorts. Material/Methods: Thirteen patients from an external centre formed the out-of-distribution (OOD) cohort (EXT_L), characterized by larger breast volumes and planning target volumes (PTVs). Two internal control cohorts, each including 13 patients, represented smaller (INT_S) and medium (INT_M) breast volumes. Model parameters could be modified if needed to improve plan quality. Performance evaluation included fulfilment of clinical goals (see Figure 1) and a blinded qualitative assessment using a four-point scale (lower score = better quality) performed by a medical physicist. Statistical differences across cohorts were tested using the Mann–Whitney U test (p < 0.05).
Conclusion: A Bayesian model that can be shared and updated between institutions whilst preserving privacy has been demonstrated. This approach has the advantage of allowing collaborations for treatment sites with limited available data. References: [1] Siva S, Bressel M, et al FASTRACK II Investigator Group. Stereotactic ablative body radiotherapy for primary kidney cancer (TROG 15.03 FASTRACK II): a non-randomised phase 2 trial. Lancet Oncol. 2024 Mar;25(3):308-316. doi: 10.1016/S1470-2045(24)00020- 2. Keywords: SABR, kidney, Bayesian iterative predictive model
Made with FlippingBook - Share PDF online