UniCurvSeg: Learning Contextual Prompt for Universal Vessel-Like Curvilinear Structure Segmentation

Universal model excels at handling multi-task and multi-modal medical image segmentation, effectively addressing the variations between different tasks. However, most existing universal models fail to adequately consider the interrelations between different structures, overlooking their latent semantic connections. In this paper, we propose a novel prompt-driven model based on in-context learning that aims to achieve 3D universal vessel-like curvilinear structure segmentation (UniCurvSeg) across domains and modalities. In UniCurvSeg, a learnable task prompt is designed to capture the correlations across all tasks, integrating a modality prompt to help the model recognise features from different modalities and a region prompt to accurately localize target regions. These prompts are injected into feature maps through cross-attention, addressing the challenges of data heterogeneity and annotation discrepancies. In addition, we incorporate a dual attention mechanism to learn rich hierarchical representations of vessel-like structures. Extensive experimental results demonstrate that UniCurvSeg outperforms existing methods on 9 datasets (1662 images) across three modalities (CT, MRI, MRA).