starU‐Net: An enhanced U‐Net architecture with star operation and multi‐view fusion for improved vessel segmentation

Abstract Accurate vessel segmentation is crucial for diagnosing vascular diseases and supporting research in developmental biology. However, existing methods struggle to preserve fine capillaries and the topological continuity of vascular structures. In this study, we aim to develop a novel deep learning architecture that specializes in thin vessel segmentation for improved continuity. We propose starU‐Net, a four‐layer encoder‒decoder framework for vessel segmentation. It integrates three components in a problem‐driven design: (1) an enhanced feature extraction module leveraging a “star operation” to enable multiplicative feature interaction, (2) a shallow network design to minimize spatial degradation, and (3) a multi‐view feature fusion module using dynamic snake convolution to capture continuous tubular structures from multiple orientations. The model was trained and evaluated on three public retinal datasets and a novel, self‐constructed chick embryo dataset. Against CNN‐based, transformer‐based, and foundation‐model baselines, starU‐Net obtained the highest sensitivity, F 1 , area under the curve, and centreline Dice on all three public retinal datasets, at the cost of slightly lower specificity. The improvement in centreline Dice provides quantitative support for the reduction in thin vessel discontinuity. On the chick embryo dataset, whose vessels are markedly wider, the proposed starU‐Net was competitive but not leading. In summary, starU‐Net is an architecture that combines star operation‐based feature extraction with multi‐view tubular feature modeling to improve vascular segmentation. The proposed framework demonstrates strong performance in both clinical retinal imaging datasets and self‐constructed chick embryo angiogenesis data, serving as a promising computational tool for clinical decision support, biomedical research, and large‐scale vascular phenotyping in biomedical applications.