An Ultra-Lightweight Cross-scale Attention Mamba Network for Accurate Skin Lesion Segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Accurate skin lesion segmentation is crucial for early skin cancer detection and management.
• Current deep learning models face a trade-off between computational cost and accuracy, limiting their use in resource-constrained settings.
• Advanced segmentation models often introduce complexity, hindering practical application.

Purpose of the Study:

• To develop a computationally efficient yet accurate deep learning model for skin lesion segmentation.
• To address the limitations of existing models in terms of computational requirements and practical deployment.
• To introduce the Cross-scale Attention Mamba Network (UCA-MNet) for improved medical image segmentation.

Main Methods:

• Developed the Cross-scale Attention Mamba Network (UCA-MNet), a hierarchical encoder-decoder architecture.
• Integrated a Multi-Scale Module (MSM) utilizing a bidirectional Mamba (Bi-Mamba) for linear-complexity long-range dependency modeling.
• Employed Cross-Scale Attention (CSA) and Pyramidal Squeeze Attention (PSA) to capture both local textures and global context.

Main Results:

• UCA-MNet achieved high performance on ISIC-2017, ISIC-2018, and PH² benchmarks with F1 scores of 0.8254, 0.8814, and 0.9202, respectively.
• The model demonstrated significantly reduced computational requirements, using only 0.33 million parameters and 4.3 GFLOPs.
• UCA-MNet is approximately 83 times smaller than the leading VM-UNET model while maintaining competitive accuracy.

Conclusions:

• UCA-MNet effectively segments skin lesions with significantly lower computational demands.
• The model's efficiency supports deployment in resource-limited environments for medical image analysis.
• This research facilitates the practical application of deep learning for early skin cancer detection.