Explainable depth-wise and channel-wise fusion models for multi-class skin lesion classification

The clinical adoption of deep learning in dermatology requires models that are not only highly accurate but also transparent and trustworthy. To address this dual challenge, this study presents a systematic investigation into deep feature fusion, exploring how to effectively combine complementary representations from diverse neural network architectures. We design and rigorously evaluate six distinct fusion models, first investigating depth-wise and channel-wise strategies for integrating features from powerful Convolutional Neural Network (CNN) backbones, and subsequently incorporating the global contextual awareness of Vision Transformers (ViTs). Evaluated on the challenging 7-class HAM10000 dataset, our optimized architecture achieves a weighted average Precision, Recall, and F1 score of 90%, demonstrating superior diagnostic performance. Crucially, our comprehensive explainable AI (XAI) analysis using Grad-CAM and SHAP reveals that the fusion strategy directly dictates the model's clinical interpretability; our most effective models learn to base their predictions on salient dermatological features, such as border irregularity and color variegation, in a manner that aligns with expert reasoning. This work provides a robust framework and valuable architectural insights for developing the next generation of high-performing, clinically reliable, and transparent AI-powered diagnostic tools.