Interpretable Skin Lesion Classification via Attention-Augmented Ensemble With Two-Stage Training

Abstract

Skin cancer, particularly, melanoma, remains a major global health challenge, with early detection being critical for effective treatment and improved patient outcomes. Traditional diagnostic approaches, which rely on clinician expertise, are susceptible to variability and have limited scalability. Deep learning models often struggle with extreme class imbalance and poor interpretability in this context. We propose an interpretable ensemble framework comprising three CNN backbones—EfficientNet-B0, DenseNet-169, and ResNet-50—each evaluated with and without a custom channel attention mechanism; the final ensemble combines EfficientNet-B0 + CA, DenseNet-169 + CA, and ResNet-50. To enhance feature discriminability for minority classes, we introduce a specialized two-stage training strategy: initial stabilization on balanced subsets followed by global fine-tuning on the full distribution. The weighted ensemble achieved 90.81% accuracy and an 86.68% F₁-score on the HAM10000 dataset. Grad-CAM visualizations provide post hoc evidence, confirming the model's focus on clinically relevant pathological markers. This study demonstrates that architectural refinements combined with strategic learning protocols significantly improve multiclass diagnostic performance in imbalanced medical datasets.