EffiCAT: A synergistic approach to skin disease classification through multi-dataset fusion and attention mechanisms

Abstract

Early and accurate diagnosis of skin diseases is essential for their efficient treatment and effective management. Conventional approaches typically depend on the use of a single dataset, which can introduce biases and limit the generalizability of the models due to dataset-specific idiosyncrasies. This study presents a novel hybrid model, named EffiCAT (EfficientNet Concatenation Attention Technology), for the categorization of skin diseases, specifically focusing on four classes named Actinic Keratosis (ACK), Basal Cell Carcinoma (BCC), Melanoma (MEL), and Melanocytic Nevus (NEV). EffiCAT enhances traditional approaches by integrating features from two different convolutional neural networks, EfficientNet B0 and EfficientNet B4, through feature concatenation. This is followed by applying advanced attention modules, specifically a Dual Channel Attention Layer applied twice and a Convolutional Block Attention Module (CBAM), to refine feature representation and focus on relevant patterns more effectively. Our method is evaluated on a combined dataset composed of HAM10000 and PAD-UFES-20, which enhances the diversity and volume of training samples to improve generalization across various skin types and conditions. The inclusion of multiple datasets helps mitigate the biases associated with single-dataset training and enhances the robustness of the model. EffiCAT attained a test accuracy of 94.48%, with precision, recall, and F1 score all closely aligned at 94.48%. These metrics not only illustrate the efficacy of our method but also underscore its superiority in handling varied and complex skin disease presentations through refined attention-driven feature concatenation. Additionally, external validation was performed on the ISIC 2018 dataset, where the model achieved a test accuracy of 92.08%, with precision of 92.45%, recall of 92.08%, and an F1 score of 92.15%, further confirming its robustness and generalizability. The model’s architecture efficiently leverages concatenated features enriched with attention mechanisms, setting a new standard for image-based diagnostic models.