Active Domain Adaptation Based on Probabilistic Fuzzy C-Means Clustering for Pancreatic Tumor Segmentation

Abstract

Pancreatic cancer is a highly lethal disease, for which mortality closely parallels incidence. Precise delineation of the Gross Tumor Volume (GTV) plays a pivotal role in ensuring effective radiotherapy for pancreatic cancer. Although recent methods have achieved promising results in GTV segmentation, it remains challenging to generalize from one domain to another, especially when the target domain is a different modality with severe domain shifts. To address this challenge, we propose a novel active domain adaptation framework to enhance domain adaptation for GTV segmentation. Specifically, we refer to the latent feature space of the synthesized target domain to select domain-specific representative samples from a specific target domain for annotation and model fine-tuning. To suppress noise and enhance the edge information, we decouple the network into an additional edge regression task that is used to further mine the contextual information of the edge pixels. Experiments on our self-collected pancreas tumor dataset and a public dataset show that our method outperforms state-of-the-art methods by a significant margin, achieving an average Dice score improvement of 2.16% and 1.86% in the two target domains on the pancreas tumor dataset, respectively.