FCA-Net: Fully context-aware feature aggregation network for medical segmentation

Accurate segmentation of lesion based on the Dermatoscopy images and Colonoscopy polyp images is beneficial for subsequent diagnosis and treatment. Although effectiveness has been verified in the fields of skin segmentation and polyp segmentation, there are still some challenges. Skin lesions and polyps often have different sizes and shapes, and a lack of clear boundaries between the lesion area and the background. To address this issue, we propose a new fully context-aware feature aggregation network (FCA-Net). It features three innovative designs: the edge perception module (EPM), the boundary-guided feature aggregation module (BFAM), and the iterative context aggregation module (ICAM). The EPM extracts initial boundary guidance maps from high-level and low-level features simultaneously, the BFAM incorporates boundary information into the segmentation network, enhancing these hierarchical features, better preserving boundary details and repositioning the calibrated objects. The ICAM leverages a fully context-aware approach to better exploit dependencies among features at different scales for more effective feature aggregation. Extensive experiments on two categories of datasets demonstrate that our proposed model outperforms state-of-the-art methods in the segmentation of lesion areas for different diseases including skin lesions and polyps.