MSA-TransUNet: A Multi-Scale Attention Enhanced Transformer-UNet Architecture for Accurate Vessel Segmentation and Visualization in Medical Imaging

Abstract

Accurate segmentation of vascular structures is critical for computer-aided diagnosis, surgical planning, and quantitative vascular analysis. Nevertheless, vessel segmentation remains a formidable challenge due to low contrast, high noise, intricate topology, and the pervasive presence of thin and tortuous vascular branches. To overcome these obstacles, we propose MSA-TransUNet, a novel hybrid architecture that combines convolutional neural networks (CNNs) with Transformer-based global modeling, augmented by a Multi-Scale Attention Module (MSAM) and a Vessel Enhancement Module (VEM) to enhance vessel representation. The MSAM selectively emphasizes vascular features across multiple receptive fields and strengthens both channel-wise and spatial responses, while the VEM introduces a learnable vascular enhancement mechanism inspired by traditional vesselness filtering techniques. In addition, a connectivity-aware loss function (clDice) is incorporated to preserve delicate vascular topology and minimize branch discontinuities. Experiments conducted on three publicly available vascular datasets, including DRIVE, STARE, and CHASE_DB1, demonstrate that MSA-TransUNet outperforms state-of-the-art segmentation models such as UNet, Attention-UNet, and TransUNet. The proposed approach achieves notable improvements in Dice coefficient, connectivity accuracy, and small-vessel recall. These results indicate that MSA-TransUNet offers a robust and effective solution for medical vascular segmentation and has potential to support clinical vessel visualization and diagnosis.