Abstract

Multi-contrast magnetic resonance imaging (MRI) plays a vital role in braintumor segmentation and diagnosis by leveraging complementary information fromdifferent contrasts. Each contrast highlights specific tumor characteristics,enabling a comprehensive understanding of tumor morphology, edema, andpathological heterogeneity. However, existing methods still face the challengesof multi-level specificity perception across different contrasts, especiallywith limited annotations. These challenges include data heterogeneity,granularity differences, and interference from redundant information. Toaddress these limitations, we propose a Task-oriented Uncertainty CollaborativeLearning (TUCL) framework for multi-contrast MRI segmentation. TUCL introducesa task-oriented prompt attention (TPA) module with intra-prompt andcross-prompt attention mechanisms to dynamically model feature interactionsacross contrasts and tasks. Additionally, a cyclic process is designed to mapthe predictions back to the prompt to ensure that the prompts are effectivelyutilized. In the decoding stage, the TUCL framework proposes a dual-pathuncertainty refinement (DUR) strategy which ensures robust segmentation byrefining predictions iteratively. Extensive experimental results on limitedlabeled data demonstrate that TUCL significantly improves segmentation accuracy(88.2\% in Dice and 10.853 mm in HD95). It shows that TUCL has the potential toextract multi-contrast information and reduce the reliance on extensiveannotations. The code is available at:https://github.com/Zhenxuan-Zhang/TUCL_BrainSeg.