MEFP-Net: A Dual-Encoding Multi-Scale Edge Feature Perception Network for Skin Lesion Segmentation

Skin lesion segmentation is an indispensable step in the diagnostic process of skin diseases. Using deep learning networks for skin lesion segmentation can enhance the work efficiency of medical personnel. However, skin lesions in dermoscopy images possess characteristics such as uneven region sizes and inconspicuous region edges, making it difficult for existing neural networks to accurately segment them. To address these issues, a Multi-scale Edge Feature Perception Network (MEFP-Net) is proposed in this paper for skin lesion segmentation. In order to allow the network to comprehensively capture the structural features of skin lesions, an additional encoder branch is introduced into it, along with newly designed Global Information Extraction Modules (GIEMs), enabling global contextual information and detailed features to be simultaneously captured. In the decoding part of MEFP-Net, newly designed Multi-scale Adaptive Feature Fusion Modules (MAFFMs) are used to adaptively extract features of different scales within the channels and fuse these features deeply. Additionally, after each MAFFM, a Convolutional Block Attention Module (CBAM) is utilized to enhance the network's perception and utilization of detailed features. Finally, a newly designed Atrous Pooling Dense Perception Module (APDPM) is utilized to enhance the network's representation of boundary features. Additionally, the combined BCE-Dice loss function is used to addresses the issue of data class imbalance. Experiments, conducted on three datasets, demonstrate that MEFP-Net outperforms traditional and state-of-the-art networks in performing skin lesion segmentations, based on the two most widely used evaluation metrics in segmentation tasks, by achieving Intersection over Union (IoU) values of 84.53%, 85.71%, and 65.01%, and Dice similarity coefficient (DSC) values of 90.90%, 91.86%, and 77.59%, respectively. In addition, the proposed MEFP-Net network exhibits higher robustness and generalization ability than traditional networks.