ESTRO 2025 - Abstract Book

S2469

Physics - Autosegmentation

ESTRO 2025

Purpose/Objective: Deep learning-based autocontouring is now widely used in radiotherapy, due to its consistency and time savings [1] . Recent ESTRO/AAPM guidelines [2] recommended ongoing surveillance after clinical implementation. Scheduled re evaluation of test cases for system upgrades, or as a periodic audit, is insufficient for detecting automation bias or review fatigue. This work investigated how patient autocontours are edited over time using a novel tool for monitoring. Material/Methods: Using a prototype version of the software AIQUALIS (Inpictura, Abingdon, UK), the unedited autocontours from RayStation treatment planning system (Raysearch Labs, Stockholm, Sweden) and final clinical contours were compared for 234 individual patients (across all anatomical sites), treated between January 2023 and August 2024. During this period, four different radiation therapists reviewed and corrected the autocontours. The geometric similarity between the autocontours and clinical contours was quantified using the surface Dice Similarity Coefficient (sDSC) [3] (with a tolerance of 1 mm), and other measures. The software also provided a 3D map highlighting the location of the edits per structure over a cohort of patients. Comparisons were made with commissioning, using data from a previous study [1] . An intervention was made to try and reduce automation bias and review fatigue, by showing the radiation therapists intermediate results in a presentation in July 2024. Results: The difference between autocontours and clinical contours was (mostly) greater during commissioning, as shown in Figure 1. The average sDSC at commissioning/during monitoring was: brain 0.70/0.74; eyes 0.76/0.97; kidneys 0.35/0.95; lacrimal glands 0.88/0.95; lenses 0.91/0.93; lips 0.71/0.78; mandible 0.79/0.95; optic nerves 0.77/0.74; pituitary 0.79/0.85; prostate 0.35/0.38; submandibular glands 0.92/0.95.

There were notable differences in editing between radiation therapists – an example is shown in Figure 2 for the brain. After the intervention at the start of July, there was early evidence of increased editing. For example, the average sDSC for the brain structure for radiation therapist 4 decreased from 0.95 before the intervention, to 0.66 after the intervention.