ESTRO 2025 - Abstract Book

S2520

Physics - Autosegmentation

ESTRO 2025

Fig.2 Performance monitoring using DSC, HD95 (a) and SDSC (b) for bladder (Vendor A Vs. Vendor B in red, Vendor B Vs. clinical contours in light blue)

Conclusion: The retrained DLAS model has improved segmentation consistency for prostate, bladder, rectum, femur_head_l and femur_head_r, with reduced discrepancies between AI-generated and clinical contours. SPC-based monitoring effectively detected auto-segmentation variations and outliers, providing a reliable and easy framework for quality assurance and early detection of monthly degradation. Prospective data is recommended for establishing DLAS performance benchmarks in future clinical implementations.

Keywords: Deep learning, Auto-segmentation, AI monitoring

3311

Digital Poster AI-generated contouring: software updates and image quality impact in daily practice Jimmy Fontaine 1 , Maud Suszko 1 , Emilie Bonnet 2 , Mathieu Bosset 2 , Sebastien Clippe 2 , Carole Sola 1 , Francesca Di Franco 3 , Jean-Baptiste Guy 2 1 Medical physics Unit, Centre Marie Curie, Valence, France. 2 Radiation oncology Unit, Centre Marie Curie, Valence, France. 3 Medical physics, IRUDIGI SARL, Bayonne, France Purpose/Objective: Artificial-intelligence (AI)-based automatic contouring software streamlines radiotherapy workflows by reducing contouring time. However, clinical implementation demands rigorous validation, particularly due to frequent software updates that affect contouring accuracy. Moreover, the impact of image acquisition and reconstruction parameters on AI performance remains underexplored. This study evaluated the effect of software updates on contouring accuracy and assessed how image quality variations influence AI performance to support seamless clinical integration. Material/Methods: Two patient cohorts were analysed. The software updates cohort (40 CT scans: 20 thorax, 10 male pelvis, 10 head and neck (H&N)) compared six versions of a commercial AI contouring software (Limbus AI, Radformation, USA). Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (HD95%) were calculated for AI-generated contours against