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

Accurate skin lesion classification is crucial for the early detection of malignant lesions, including melanoma, as well as improved patient outcomes. While convolutional neural networks (CNNs) excel at capturing local morphological features, they struggle with global context modeling essential for comprehensive lesion assessment. Vision transformers address this limitation but suffer from quadratic computational complexity O(n²), hindering deployment in resource-constrained clinical environments. We propose DermaMamba, a novel dual-branch fusion architecture that integrates CNN-based local feature extraction with Vision Mamba (VMamba) for efficient global context modeling with linear complexity O(n). Our approach introduces a state space fusion mechanism with adaptive weighting that dynamically balances local and global features based on lesion characteristics. We incorporate medical domain knowledge through multi-directional scanning strategies and ABCDE (Asymmetry, Border irregularity, Color variation, Diameter, Evolution) rule feature integration. Extensive experiments on the ISIC dataset show that DermaMamba achieves 92.1% accuracy, 91.7% precision, 91.3% recall, and 91.5% mac-F1 score, which outperforms the best baseline by 2.0% accuracy with 2.3× inference speedup and 40% memory reduction. The improvements are statistically significant based on a significance test (<i>p</i> < 0.001, Cohen's d > 0.8), with greater than 79% confidence also preserved on challenging boundary cases. These results establish DermaMamba as an effective solution bridging diagnostic accuracy and computational efficiency for clinical deployment.