Curvature-aware selective feature interaction network for skin lesion segmentation.

With the increasing prevalence of dermatological diseases, skin lesion segmentation has gained significant attention in medical image analysis. Despite substantial advancements in deep learning-based segmentation methods, there remain two major challenges in achieving robust and accurate skin lesion segmentation: 1) insufficiency in capturing feature interactions due to the semantic gap in the encoder-decoder architecture. 2) Susceptibility to redundant interaction information, which fails to exploit the shape characteristics of skin lesions. In this paper, we propose a Curvature-Aware Selective Feature Interaction Network (CASFI-Net) to address these issues. Concretely, to enhance the decoder structure, we introduce a Multi-Grain Feature Interaction (MGFI) module that uses attention mechanisms to guide cross-resolution information interaction, effectively integrating low-level detail features from the encoder and high-level semantic features from the decoder, thereby bridging the semantic gap between them. Furthermore, skin lesions typically have smooth and nearly circular edges, making them more sensitive to curvature variations. We present a Curvature-Aware Selective Feature (CASF) module that evaluates the curvature of the feature maps generated by the MGFI module. By employing a fast curvature selection mechanism, this evaluation allows for the selective retention of the most informative feature channels, emphasizing key edge features while reducing redundant information. Experimental results in three skin image datasets demonstrate that CASFI-Net outperforms current state-of-the-art methods in skin lesion segmentation while maintaining a low computational cost.