ESTRO 2025 - Abstract Book

S2517

Physics - Autosegmentation

ESTRO 2025

3237

Digital Poster Evaluating the consistency of deep learning-based autocontouring between different treatment planning system versions Afxentia Karousiou 1 , Mark Gooding 2,3 , Djamal Boukerroui 2 , Constantinos Zamboglou 4,5 , Konstantinos Ferentinos 4 , Paul J Doolan 6 1 Medical Physics, University of Surrey, Surrey, United Kingdom. 2 Inpictura, Ltd, Abingdon, United Kingdom. 3 Division of Cancer Sciences, University of Manchester, Manchester, United Kingdom. 4 Radiation Oncology, German Oncology Center, Limassol, Cyprus. 5 Radiation Oncology, University of Freiberg, Freiberg, Germany. 6 Medical Physics, German Oncology Center, Limassol, Cyprus Purpose/Objective: Accurate contouring is essential for radiotherapy treatment planning, where defining the tumour and organs at risk (OARs) impacts the effectiveness of treatment. Deep learning autocontouring tools offer potential improvements in efficiency and consistency over manual contouring [1,2] . According to the recent RCR guidelines, one of the primary goals of post-implementation monitoring is to detect any change in performance [3] and thus it is important to assess updates between software versions. The aim of this study is to evaluate the performance of a deep learning based autocontouring tool across software versions. Material/Methods: Autocontours were generated in RayStation (RaySearch Laboratories, Stockholm, Sweden) versions 11B, 12A, 2024A and 2024B for 80 clinical cases (20 prostate, 20 head-and-neck, 20 breast and 20 lung). The autocontours were compared to a manually-drawn gold-standard contour set from a recent study [2] . AIQUALIS (Inpictura Ltd, Abingdon, UK) was used to calculate the geometric similarity between manual and automatic contours. In this work, we evaluated differences using the surface Dice Similarity Coefficient (sDSC) [2,4] , 3D spatial plots and spider plots.

Results: No notable differences were observed between versions 11B and 12A for prostate structures as seen in Figure 1.