Uncertainty-guided pancreatic tumor auto-segmentation with Tversky ensemble.

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