EA-Net: Research on skin lesion segmentation method based on U-Net.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Accurate skin lesion segmentation is crucial for clinical diagnosis but is hindered by variations in location, shape, and scale.
• Existing Convolutional Neural Networks (CNNs) struggle to preserve local image features and highlight relevant information, limiting their clinical application.

Purpose of the Study:

• To develop an enhanced CNN model (EA-Net) for more accurate skin lesion segmentation.
• To improve the preservation of local features and the relevance of feature maps during segmentation.

Main Methods:

• Proposed EA-Net, a U-Net based CNN incorporating a pixel-level attention module (PA) in the encoder to preserve local features.
• Integrated a spatial multi-scale attention module (SA) post-decoder to refine feature maps and enhance spatial relevance.
• Evaluated EA-Net on the ISIC 2017 and ISIC 2018 skin lesion datasets.

Main Results:

• EA-Net demonstrated superior performance compared to U-Net on both datasets.
• Achieved average Dice score improvements of 1.94% (ISIC 2017) and 5.38% (ISIC 2018).
• Showed increases in Intersection over Union (IoU) by 2.69% and 8.31%, and decreases in Average Symmetric Surface Distance (ASSD) by 0.3783 and 0.5432 pixels.

Conclusions:

• The proposed EA-Net effectively improves skin lesion segmentation accuracy by integrating attention mechanisms.
• The model excels in segmenting lesions with obscure boundaries and complex conditions, proving the value of attention in medical image analysis.