ESTRO 2025 - Abstract Book

S2507

Physics - Autosegmentation

ESTRO 2025

Conclusion: Our study reveals two types of differences among automatic solutions for heart substructures: one about the adopted definitions and another concerning model performance. It also highlights the sensitivity to the presence/absence of CE, depending on the solution. This plebiscites for an ESTRO group to (i) establish a consensus on heart substructure definitions and CT protocols, and (ii) develop an appropriate scoring of auto-contouring solutions, to support future studies on dose-effect relationships.

Keywords: heart substructures, AI, comparison of solutions

References: Jacob et al 2019, Radiat Oncol. Is mean heart dose a relevant surrogate parameter of left ventricle and coronary arteries exposure during breast cancer radiotherapy: a dosimetric evaluation based on individually-determined radiation dose (BACCARAT study). DOI:10.1186/s13014-019-1234-z Jones et al 2024, J Med Imaging Radiat Oncol. Moving beyond mean heart dose: The importance of cardiac substructures in radiation therapy toxicity. DOI:10.1111/1754-9485.13737

3091

Mini-Oral Interpreting convolutional neural network explainability for head-and-neck cancer radiotherapy organ-at risk segmentation Victor I.J. Strijbis 1,2 , Oliver J. Gurney-Champion 3,4 , Dragos I. Grama 1,2 , Berend J. Slotman 1,2 , Wilko F.A.R. Verbakel 1,2,5 1 Radiation Oncology, Amsterdam UMC, Amsterdam, Netherlands. 2 Cancer Treatment and Quality of Life, Cancer Center Amsterdam, Amsterdam, Netherlands. 3 Radiology and Nuclear Medicine, Amsterdam UMC, Amsterdam, Netherlands. 4 Imaging and Biomarkers, Cancer Center Amsterdam, Amsterdam, Netherlands. 5 Varian Medical Systems, a Siemens Healthineers Company, Siemens, Palo Alto, USA Purpose/Objective: Convolutional neural networks (CNNs) have emerged to reduce clinical resources and standardize auto-contouring of organs-at-risk (OARs). Although CNNs perform adequately for most patients, understanding when the CNN might fail is critical for effective and safe clinical deployment. However, the limitations of CNNs are poorly understood because of their black-box nature. Explainable artificial intelligence (XAI) can expose CNNs’ inner mechanisms for classification . Here, we aim to expose essential mechanisms of CNNs for adequate clinical segmentation and explore a novel, computational approach to a-priori flag potentially insufficient contours. Material/Methods: First, 3D UNets were trained in three parotid gland (PG) segmentation situations using (1) synthetic cases; (2) 1925 clinical computed tomography scans with typical and (3) more consistent contours curated through a previously validated auto-curation step 1 . Then, we generated attribution maps for four XAI methods, and qualitatively assessed them for congruency between synthetic and clinical contours, and how much XAI agreed with expert reasoning. To objectify observations, we explored persistent homology 2 intensity filtrations to capture the essential topological characteristics of XAI attributions. Principal component (PC) eigenvalues of Euler characteristic profiles 3 were correlated with spatial agreement (Dice-Sørensen similarity coefficient; DSC). Evaluation was done using area under receiver operating characteristic (AUROC) curve on AAPM data 4 , where, as proof-of-principle, we regard the lowest 15% DSC as insufficient. Results: PatternNet attributions (PNet-A) focused on soft-tissue structures (Fig. 1; green), whereas guided backpropagation (GBP) highlighted both soft-tissue (green) and high-density structures (orange; e.g., mandible bone), which was