DSGH-Net: Medical Image Segmentation via Dual-Statistical Dynamic Context and Graph-Convolutional Heterogeneous...

Accurate segmentation of medical images is crucial for assisting clinical diagnosis and treatment planning. However, automatic segmentation remains highly challenging due to significant variations in the size and morphology of lesions, as well as their ambiguous boundaries. Although CNN- and Transformer-based architectures have made remarkable progress, there remains room for improvement in context-adaptive modeling and hierarchical feature decoding. First, standard convolutions typically rely on static receptive fields, limiting their context-adaptive capability for complex lesion regions, which may lead to inadequate extraction of multi-scale lesion features. Second, existing decoders predominantly employ relatively homogeneous feature recovery strategies, making it difficult to simultaneously balance deep semantic topological modeling and shallow edge detail recovery. To address these issues, we propose DSGH-Net, a medical image segmentation network designed for multi-scale contextual modeling and hierarchical feature decoding. To overcome the context modeling issue, we design the Dual-Statistical Context Modulation Block (DCM-Block), which integrates Global Average Pooling (GAP) and Global Max Pooling (GMP) to generate content-aware dynamic weights, thereby achieving dynamic multi-scale feature fusion at the bottleneck layer. To tackle feature recovery in the decoding stage, we construct the Heterogeneous Stage Decoder (HSD): in the deep stage, the Topological Dynamic Context Refinement Block (TDCR-Block), integrated with Semantic Calibration Graph Convolution (SC-GCN), is employed to enhance global structural relationship modeling; in the shallow stage, it switches to a lightweight CNN to recover edge details. Furthermore, the Deep Weighted Fusion Attention (DWFA) module is introduced at skip connections to enhance valid skip features and suppress noise interference. Experimental results demonstrate that DSGH-Net achieves competitive performance on four medical image segmentation datasets, including Kvasir-SEG and ISIC2018, outperforming the compared methods across most core metrics. Specifically, on the Kvasir-SEG dataset, DSGH-Net achieves a Dice score of 92.48% and an IoU of 87.80%. Cross-dataset zero-shot generalization experiments further show that, when trained solely on Kvasir-SEG and directly transferred to CVC-ClinicDB for testing, DSGH-Net still achieves a Dice score of 87.63%, with most core metrics outperforming the compared methods, suggesting its preliminary robustness across public polyp segmentation datasets.