FUSCANet: Enhancing Skin Disease Classification Through Feature Fusion and Spatial-Channel Attention Mechanisms

Skin diseases represent a prevalent global health issue that significantly impacts the physical and mental well-being of patients. With the widespread application of computer vision technology in dermatology, automating skin lesion classification through computer algorithms has become a crucial method for improving diagnostic efficiency and reducing the mortality rate due to malignant skin conditions. In response to this need, a novel lightweight neural network model, called Feature fUsion and Spatial-Channel Attention Network (FUSCANet) model, is proposed in this paper, based on the MobileViT framework, aiming at classifying multi-class skin disease images on mobile or embedded devices. Firstly, a newly designed Leaky MobileNetV2 (LMV2) block, integrated into the proposed model, allows to enhance its ability to extract features from skin images. Then, a novel Multi-Scale Feature Aggregation (MSFA) layer is added to enhance feature extraction across various scales, effectively capturing more comprehensive features and fusing feature maps from different scales to improve final feature reuse. Lastly, a newly designed Enhanced Convolutional Block Attention Module (ECBAM) increases the model’s focus on important information, allowing for the final extraction and summarization of features. The proposed FUSCANet model is evaluated on four different datasets (the public PAD-UFES-20, HAM10000, and ISIC 2019 datasets, and a private dataset) containing images of various skin diseases. The obtained experimental results show that FUSCANet outperforms existing models on most evaluation metrics, while maintaining low parameter counts, making it suitable for deployment in resource-constrained environments.