Uncertainty-guided pancreatic tumor auto-segmentation with Tversky ensemble

• DOI: 10.1016/j.phro.2025.100740
• Publication Date: 2025-03-08T23:29:04Z
• Authors: Cenji Yu; Skylar S. Gay; Aashish C. Gupta; Rachael M. Martin...; Ethan B. Ludmir; Yao Zhao; Jack Duryea; Xinru Chen; Carlos E. Cardenas; Jinzhong Yang; Albert C. Koong; Tucker J. Netherton; Dong Joo Rhee; Laurence E. Court

Pancreatic gross tumor volume (GTV) delineation is challenging due to their variable morphology and uncertain ground truth. Previous deep learning-based auto-segmentation methods have struggled handle tasks with truth not accommodated stylistic customizations. We aim develop a human-in-the-loop pancreatic GTV segmentation tool using Tversky ensembles by leveraging uncertainty estimation techniques. In this study, we utilized total of 282 patients from the pancreas task Medical Segmentation Decathlon. Thirty were randomly selected form an independent test set, while remaining 252 divided into 80-20 % training-validation split. incorporated loss layer during training train five-member ensemble varying contouring tendencies. The predicted probability maps estimating pixel-level uncertainties. Probability thresholding was employed on resulting generate final contours, which eleven contours extracted for quantitative evaluation against truths, variations in threshold values. Our achieved DSC 0.47, HD95 12.70 mm MSD 3.24 respectively optimal configuration. outperformed Swin-UNETR configuration that state-of-the-art result medical decathlon. study demonstrated effectiveness employing ensemble-based technique segmentation. approach provided clinicians consensus map could be fine-tuned line preferences, generating greater confidence.