DERMIS: An explainable dual-branch ensemble framework for robust melanoma detection

Accurate melanoma lesion classification is inherently linked with early diagnosis and improved prognosis. However, class imbalance, non-interpretable deep models, and poor feature extraction abilities restrict their clinical use. To address these limitations, Dual-branch Ensemble for Robust Melanoma Identification System or DERMIS is proposed, which integrates DenseNet201 and MobileNetV3-Large with an attention-based fusion approach to surpass state-of-the-art models in all evaluation measures. Synthetic malignant images are crafted by Generative Adversarial Networks (GAN)-based image augmentation with balanced class representation and improved model generalizability. Aggressive data preprocessing operators like normalization, random rotation, flipping, zoom, and shearing improve data quality and model robustness. Explainable AI tools, i.e., LIME, SHAP, and Grad-CAM, facilitate interpretable model decision-making with comprehensive interpretability analysis. DERMIS superiority over benchmark models is proven through exhaustive comparative analysis with 95.11% accuracy, 96.66% precision, 93.13% recall, and 94.86% F1-score, outperforming EfficientNetB3, BeIt, InceptionResNetV2, DenseNet201, ResNet101, CoaTNet, and MobileNetV3 architectures significantly. The model was trained and validated against dataset containing 10,605 RGB images with binary melanoma classification using an 80:20 training–testing split. Exhaustive ablation studies confirm the synergistic benefits of the dual-branch fusion strategy over single baseline models. Grad-CAM heatmaps have the best clinical interpretability among LIME and SHAP, accurately localizing pathologically important features like asymmetric morphology, irregular borders, and abnormal pigmentation patterns. DERMIS is an important breakthrough in melanoma detection with automated methods, offering best trade-off between diagnostic accuracy, computational efficiency, and clinical interpretability.