EFEMNet: An efficient feature extraction multi-attention convolutional neural network for skin lesion segmentation

Abstract

Melanoma is the skin tumor with the highest mortality rate, and timely diagnosis based on dermoscopic images is an essential task in melanoma prevention and treatment. However, complex dermoscopic image morphology and unclear image edges affect the accurate diagnosis of melanoma. This study presents an encoder–decoder architecture (EFEMNet) to segment dermatologic lesions. First, residual concatenation is introduced in the encoder part to enhance the feature retention capabilities. Second, Coordinate Attention is utilized to identify and localize the target region more accurately for dermoscopic images with different morphologies. Third, an efficient feature extraction module (EFEM) is designed to improve up-sampling operations and to extract and fuse features efficiently. Finally, Global Attention Module (GAM) Attention is added to the output layer to integrate the dimensions in space and channels to solve the problem of unclear edges in dermoscopy images. The suggested method is evaluated on various datasets, such as ISIC 2018 and EFEMNet, and segmented skin lesions more accurately than some state-of-the-art network models, achieving 92.52% Dice, 86.81% IoU, 96.01% Accuracy, 93.86% Recall, and 92.92% Precision. The proposed method is shown to outperform other methods across all evaluation indices, and the effectiveness of the functional module is validated through a series of ablation experiments.