DSHR-Net: A Deep Snake High-Resolution Network for Dorsal Hand Vein Segmentation

Dorsal hand vein segmentation is crucial for medical imaging analysis and biometric identification. However, conventional encoder–decoder networks often struggle with blurry boundaries and indistinct tubular shapes in low-contrast vein images. This stems from standard convolutional kernels’ limitations in capturing complex geometries and the loss of spatial details due to down-sampling. To overcome these issues, we propose the deep snake high-resolution network (DSHR-Net), an enhanced U-Net architecture. DSHR-Net utilizes a high-resolution pathway with multiresolution aggregation modules (MAMs) to preserve fine spatial information. Critically, it integrates dynamic snake convolutions (DSConvs) into the encoder–decoder structure, significantly enhancing its ability to model the deformable, tubular characteristics of veins and improving feature extraction. Furthermore, multilevel semantic supervision and squeeze-and-excitation (SE) attention mechanisms guide the network to focus on refined low-level spatial representations and adaptively prioritize critical channel features. We also introduce a new annotated dorsal hand vein dataset (SDNU). Extensive experiments on both our SDNU dataset and the public GPDS dataset demonstrate DSHR-Net’s superiority over U-Net, achieving Dice improvements of 3.71% (SDNU) and 4.92% (GPDS), with corresponding intersection over union (IoU) gains of 4.84% and 5.82%. These results highlight its exceptional ability to segment fine curved vessels in low-contrast dorsal hand vein images, significantly outperforming existing methods.