ESTRO 2025 - Abstract Book

S2548

Physics - Autosegmentation

ESTRO 2025

Conclusion: Federated learning with geographically-dispersed real-world radiotherapy data is a feasible method for training lung tumour segmentation deep learning models, with comparable results to centralized training on curated data. Uniform procedures for data pre-processing and model evaluation, and stable network connectivity were essential for federated learning. Wide heterogeneity of imaging settings and varying clinician delineations remain significant challenges when utilizing real-world data to train and validate artificial intelligence models.

Keywords: deep learning, lung tumour, automated segmentation

References: [1] SP Primakov et al. Nat Comm. 2022 Jun 14;13(1):3423. 4013

Digital Poster Cross-institutional validation of prostate tumor auto-segmentation using multiparametric MRI Ruben Bosschaert 1 , Josiah Simeth 2 , Eduardo H Pais Pooch 3 , Mar Fernandez Salamanca 1 , Ivo G Schoots 3 , Neelam Tyagi 2 , Uulke A van der Heide 1 , Harini Veeraraghavan 2 , Tomas M Janssen 1 1 Radiation Oncology, The Netherlands Cancer Institute, Amsterdam, Netherlands. 2 Medical Physics, Memorial Sloan Kettering Cancer Centre, New York, USA. 3 Radiology, The Netherlands Cancer Institute, Amsterdam, Netherlands Purpose/Objective: Delivering a focal boost to the gross tumor volume (GTV) in prostate cancer radiotherapy has shown improved outcome compared to whole gland prostate radiotherapy. This treatment requires accurate GTV delineation on multiparametric MRI (mpMRI), which is time consuming, subject to inter-observer variability and prone to errors.