DermaMamba: A Dual-Branch Vision Mamba Architecture with Linear Complexity for Efficient Skin Lesion Classification

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Image Analysis
• Computer Vision

Background:

• Accurate skin lesion classification is vital for early detection of malignant melanoma.
• Convolutional Neural Networks (CNNs) excel at local features but lack global context.
• Vision Transformers (ViTs) offer global context but have high computational cost (O(n^2)).

Purpose of the Study:

• To develop DermaMamba, an efficient dual-branch architecture for skin lesion classification.
• To integrate CNNs for local features and Vision Mamba (VMamba) for efficient global context modeling (O(n)).
• To improve diagnostic accuracy and computational efficiency for clinical deployment.

Main Methods:

• Proposed DermaMamba: a dual-branch fusion architecture combining CNNs and Vision Mamba (VMamba).
• Introduced a state space fusion mechanism with adaptive weighting for dynamic feature balancing.
• Incorporated medical domain knowledge via multi-directional scanning and ABCDE rule integration.
• Evaluated on the ISIC dataset.

Main Results:

• Achieved 92.1% accuracy, 91.7% precision, 91.3% recall, and 91.5% mac-F1 score.
• Outperformed the best baseline by 2.0% accuracy.
• Demonstrated 2.3× inference speedup and 40% memory reduction.
• Showed statistically significant improvements (p < 0.001, Cohen's d > 0.8) and maintained performance on challenging cases.

Conclusions:

• DermaMamba effectively balances local and global features for superior skin lesion classification.
• The model offers significant diagnostic accuracy and computational efficiency.
• DermaMamba presents a promising solution for clinical deployment in resource-constrained environments.