ESTRO 2025 - Abstract Book

S2448

Physics - Autosegmentation

ESTRO 2025

1832

Mini-Oral Developing a Deep Learning Segmentation QA Dashboard Guus B. Spenkelink, Xander Staal, Jean-Paul J.E. Kleijnen Medical Physics, Haaglanden Medical Center, Leidschendam, Netherlands

Purpose/Objective: Deep Learning Segmentation (DLS) is commercially available from several vendors, enabling rapid automatic contouring. However, AI tools carry risks due to the black-box nature of the technology and may fail unexpectedly, potentially resulting in segmentation mistakes with adverse consequences for the patient. To address these risks, new regulations and guidelines like the EU Artificial Intelligence Act and the Netherlands’ Guideline for High Quality AI in Healthcare recommend continuous monitoring by deployers. A DLS QA dashboard has been developed at our clinic to aid monitoring DLS quality over time and identify trends or biases related to imaging, patient, and plan parameters. Material/Methods: The dashboard evaluates DLS performance by monitoring the extent of manual adjustments to the DLS contours. It consists of two parts: a data processing part and a data exploration part. Data processing During treatment plan preparation, original and manually adjusted DLS segmentations are automatically exported from the TPS. An external script calculates contour comparison metrics, and collects relevant data on the autosegmentation, imaging, patient, and treatment plan. The output is automatically added to the DLS dashboard database. In addition to standard contour metrics, Surface DSC and Added Path Length are of particular interest, as they correlate most closely with the time required for manual contour corrections [1],[2] , allowing for indirect monitoring of the time savings provided by DLS. Data exploration The QA dashboard provides an interactive view of DLS performance over time, enabling monitoring across various metrics and all imaging, patient, or plan parameters. Smoothed trendlines are included to help highlight trends over time. Users have full control over the dashboard by selecting X-axis, Y-axis, and colour variables via dropdown menus. Results: As an example of clinical results in the dashboard, a scatterplot of Dice coefficient versus time is shown, with colour representing the organ at risk.