GCMR-Net: A Global Context-Enhanced Multi-scale Residual Network for medical image segmentation

Medical image segmentation is extensively applied in clinical diagnosis, providing physicians with accurate, rapid, and visual analysis tools. With the vigorous development of deep learning, the emergence of U-shaped networks has achieved outstanding performance in medical image segmentation. However, they cannot extract sufficient boundary features due to the continuous downsampling and focus primarily on local features without taking full advantage of global context information. In this paper, we present a Global Context-Enhanced Multi-Scale Residual Network (GCMR-Net) to tackle these problems. Specifically, the first component of GCMR-Net is a Multi-scale residual pooling (MSRP) module, which seamlessly integrates depth-wise separable convolution and atrous convolution to achieve more efficient feature extraction and information transmission. Second, based on the MSRP module, we propose a convolution encoder named MSRPE that progressively extracts abstract and higher-level features through cascaded convolutional layers. The features generated by MSRPE are efficiently reused in the corresponding decoder layers to facilitate the restoration of intricate image details. The third essential component is the global channel–spatial dual attention (GCSDA) mechanism, which enhances the global perception of the model and captures the contextual information to locate targets precisely. Furthermore, we design a feature pyramid compensation and refinement (FPCR) module, which mitigates feature loss in the decoder through various scales of convolution and upsampling operation, generating feature maps ranging from coarse to fine. Extensive comparative experiments are conducted on three biomedical image datasets and the results demonstrate that GCMR-Net exhibits competitive performance compared to state-of-the-art methods.