ESTRO 2025 - Abstract Book

S2421

Physics - Autosegmentation

ESTRO 2025

340

Digital Poster Has AI Changed How We Contour? A Comparative Study of Pre- and Post-AI Contouring Practices Ciaran Malone 1 , Samantha Ryan 1 , Jill Nicholson 1,2 , Pierre Thirion 1,3 , Aodh MacGairbhith 1 , Sinead Brennan 1,2 , Orla McArdle 1 , Frances Duane 1,2 , Gerard G Hanna 2,1 , Ruth Woods 1 , Brendan McClean 1,4 1 Department of Radiation Oncology, St.Luke's Radiation Oncology Network, Dublin, Ireland. 2 Applied Radiation Therapy Trinity, Discipline of Radiation Therapy, Trinity St.James’s Cancer Institute, Trinity College Dublin, Dublin, Ireland. 3 School of Medicine, Trinity St.James’s Cancer Institute, Trinity College Dublin, Trinity College Dublin, Dublin, Ireland. 4 Department of Physics, UCD, Dublin, Ireland Purpose/Objective: The integration of artificial intelligence (AI) in radiation oncology is transforming clinical workflows, particularly in automated contouring, with opportunities for increased efficiency and consistency. This study evaluates whether AI has changed how clinicians contour in our clinic by comparing contouring volumes pre- and post-AI introduction and assessing the need for ongoing AI performance monitoring. Specifically, we assess the impact of AI on organ-at risk (OAR) volumes, inter-observer variation, and the potential dosimetric implications of bias introduced by AI. Material/Methods: We analyzed pre- and post-AI contours approved by a Senior Radiation Oncologist across three anatomical sites: Head and Neck (H&N), thorax, and prostate. The pre-AI group included 52 H&N cases, 43 thorax cases, and 40 Prostate cases, totalling 2043 contours. The post-AI group included 44 H&N cases, 42 thorax cases, and 40 Prostate cases, totalling 2248 contours. Comparisons were made between pre-AI manual contours, post-AI amended contours, and raw AI-generated contours in terms of volumetric differences and inter-observer variability. An analysis was conducted to assess the dosimetric impact of using raw AI-generated structures, focusing on whether the “worst-case scenario” of structure bias posed a high risk of clinically significant dosimetric consequences. Results: Significant shifts in mean OAR volume were observed after the introduction of AI contouring. Specifically, differences were noted in 14 of 22 H&N structures, 3 of 14 thorax structures, and 6 of 12 prostate structures, with discrepancies consistently being in the direction of the raw AI-generated contours. The use of AI reduced inter observer variation in contouring for 33 out of 47 structures, while the remaining 14 structures showed no significant change, indicating increased consistency overall. Dosimetric analysis between manual and AI-generated contours showed all evaluated structures had a Mean Absolute Error (MAE) <1.1Gy and Root Mean Square Error (RMSE) <2.6Gy compared to institutional dose volume constraints (DVCs). Twenty structures showed minimal differences (MAE <0.5Gy, RMSE = 0.9Gy). Structures with significant bias had an increased MAE of 1.9Gy (RMSE = 3.7Gy) relative to manual contours, with the oesophagus, brachial plexus and rectum showing differences exceeding 10Gy, underscoring the need for careful review. Conclusion: AI contouring shows promise for improving efficiency and reducing variability. However, it can introduce biases potentially leading to significant dosimetric impacts. Multidisciplinary teams (MDTs) must collaborate to monitor AI performance, understand it’s limitations, and address human factors. A robust QA and feedback loop is crucial to ensure AI integration enhances clinical outcomes without unintended consequences.

Keywords: Bias, Human Factors, AI