LeaNet: Lightweight U-shaped architecture for high-performance skin cancer image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate skin cancer diagnosis relies heavily on image segmentation.
• Medical devices require lightweight algorithms due to limited computing power.
• Existing lightweight models struggle to capture complete lesion edge information, leading to pixel loss.

Purpose of the Study:

• To develop a high-performance, lightweight network for skin cancer image segmentation.
• To address the limitations of current models in capturing local and global feature information of lesion edges.
• To improve the accuracy and efficiency of skin cancer segmentation on resource-constrained medical devices.

Main Methods:

• Proposed LeaNet, a novel U-shaped network incorporating multiple attention blocks.
• Utilized a dilated efficient channel attention (DECA) module for global/local contour information.
• Implemented an inverted external attention (IEA) module for enhanced data sample correlation and an attention bridge (AB) for multi-level feature extraction.

Main Results:

• LeaNet demonstrated superior performance on ISIC2017 and ISIC2018 datasets compared to large and lightweight models.
• Achieved improvements in Accuracy (ACC), Sensitivity (SEN), and Specificity (SPEC) metrics.
• Significantly reduced model parameter count and computational complexity, outperforming existing methods.

Conclusions:

• LeaNet provides a highly effective and efficient solution for skin cancer image segmentation.
• The network's design overcomes limitations in capturing lesion edge details in lightweight models.
• LeaNet is suitable for deployment on medical devices with limited computational resources, enhancing diagnostic capabilities.