ESTRO 2025 - Abstract Book

S102

Invited Speaker

ESTRO 2025

importance of data science competitions has become increasingly evident, as the specialty relies heavily on image driven clinical decision-making. Data science competitions in radiation oncology have broadly advanced AI methodologies for critical radiotherapy (RT) tasks, including image processing, contouring of tumors and organs at risk, dose optimization, and outcome prediction. Notably, auto-contouring represents a clinically impactful area where the advantages of data science competitions have been exceptionally evident. Illustrating this trend, the recently completed Head and Neck Tumor Segmentation for MR-Guided Applications (HNTS-MRG) 2024 Challenge specifically focused on multi-timepoint MRI-guided tumor auto-contouring in patients with head and neck cancer (HNC). Responding to the critical shortage of large-scale, publicly available MRI datasets for RT planning and adaptive treatment, the HNTS-MRG 2024 Challenge utilized a comprehensive dataset of 202 HNC patients with pre RT and mid-RT MRI scans accompanied by expert-generated annotations for primary and nodal gross tumor volumes. The challenge attracted extensive international participation, with 19 teams from around the world submitting models utilizing diverse AI methodologies. Analysis of submitted models yielded several important insights. Pre-RT contouring tasks saw AI models consistently surpassing human expert interobserver variability, demonstrating that state-of-the-art deep learning methods can potentially achieve clinical-grade contouring performance in conventional imaging scenarios. Conversely, mid-RT contouring, reflecting the adaptive RT context characterized by dynamic, treatment-induced anatomical changes, proved to be substantially more difficult. Participant performance variability in this adaptive scenario underscored significant remaining gaps in current methodologies, indicating that capturing HNC tumor evolution throughout RT represents an important frontier for future AI research. Beyond contouring accuracy, the forthcoming full public release of the meticulously annotated HNTS-MRG dataset will serve as a valuable resource for addressing additional related research areas, such as AI uncertainty quantification. Overall, the experiences and outcomes from the HNTS-MRG Challenge highlight how structured data science competitions effectively foster collaborative AI advancements, driving potential for clinical progress in radiation oncology.

4818

Speaker Abstracts The physicist role in the clinical reirradiation workflow Charles Mayo Radiation Oncology, University of Michigan, Ann Arbor, USA

Abstract:

Reirradiation (reRT) accounts for a substantial fraction of patient volume clinics with photon (10-20%) or with proton (30-50%) therapy. While there is substantial multi-institutional evidence for de-novo treatments to guide optimizing tumor control vs risk of serious toxicity there is are significant gaps in evidence for patients treated with reRT. To fill these gaps, physicists play a central role in developing cohesive approaches to assure safe treatments and clear reporting on a per patient basis while also enabling systematic aggregation of evidence to model outcomes and guide treatments for future patients. This presentation will detail clinical workflow and applications our physics team has developed in treating over 3000 patients with reRT, to smoothly integrate with physician and dosimetry teams. Clinical examples in spine, liver, brain and head and neck that push the boundaries of our understanding will be used to illustrate advantages of the workflow. Statistics on evaluation completion times will be used to examine throughput.