MHS U-Net: Multi-scale hybrid subtraction network for medical image segmentation

Abstract

Medical image segmentation plays a critical role in modern clinical diagnosis. However, existing methods face challenges such as insufficient feature extraction, limited spatial modeling capabilities, and restricted generalization ability with low computational cost. To address these challenges, we propose a Multi-scale Hybrid Subtraction Network (MHS U-Net) for Medical Image Segmentation. First, a pretrained PVTv2-B2 is integrated as the encoder to enhance the model's adaptability and feature extraction capability for complex multi-modal medical images. Second, a Multi-Layer Shift Perception Attention (MSPA) mechanism is designed at the bottleneck to capture fine-grained high-level features by deepening the network structure, while effectively suppressing the surge in computational cost through shift operations. In the decoder, a Multi-Scale Hybrid Convolution Subtraction Decoder (MSHCSD) is developed, to improve the modeling of spatial relationships within and around lesions and significantly enhance the model's generalization ability through integrating group convolution, gating mechanisms, and deep convolutional blocks. Additionally, to address the insufficient utilization of multi-scale feature interactions, a Multi-Scale Subtraction Module (MSSM) is proposed to strengthen cross-scale feature fusion through differential information extraction and complementary feature enhancement, thereby achieving the precise localization and segmentation of lesion regions. Experimental results on 14 public datasets across five imaging modalities demonstrate that MHS U-Net consistently outperforms state-of-the-art methods in metrics and visual results. Moreover, MHS U-Net requires only 5.48G FLOPs and 11.59M parameters, significantly lower than most existing models. Overall, MHS U-Net offers an excellent balance between model performance and size in multi-modal medical image segmentation tasks.