Explainable ensemble transfer learning for skin lesion classification with multi-method explainability validation

Abstract

Introduction Skin cancer is among the most prevalent and life-threatening malignancies worldwide. Early and accurate detection significantly improves therapeutic outcomes. Automated classification of dermoscopic skin lesions remains challenging due to class imbalance, inter-class visual similarity, and lack of interpretable predictions. This study proposes an accurate and interpretable weighted soft voting ensemble for multi-class skin lesion classification. Methods A weighted soft voting ensemble of three heterogeneous deep learning architectures, EfficientNet-B3, ResNet101, and DenseNet201, pretrained on ImageNet and fine-tuned on the HAM10000 dataset comprising 10,015 dermoscopic images across seven diagnostic categories (AKIEC, BCC, BKL, DF, MEL, NV, and VASC), was developed. The framework integrates complementary feature extraction based on compound scaling, residual learning, and dense connectivity, along with class-specific augmentation using differentiated transformation parameters for the minority classes AKIEC, DF, and VASC. Explainability was evaluated using three complementary XAI methods, Grad-CAM, LIME, and Occlusion Sensitivity Analysis (OSA), using the Trustworthiness Metric for AI (TMAI) to quantitatively assess explanation quality. Generalizability was further assessed on the ISIC 2019 benchmark dataset using a separately trained model comprising 25,331 images across eight categories. Results The proposed framework achieved an accuracy of 96.37%, precision of 96.25%, recall of 96.37%, F1-score of 96.23%, and macro-average AUC of 0.983 on HAM10000. The model achieved perfect classification for BCC (100%, 35/35; 95% CI: 90.1%–100.0%) and VASC (100%, 13/13; 95% CI: 77.2%–100.0%), and achieved 99.77% accuracy for MEL. The proposed ensemble surpassed the best individual model, EfficientNet-B3 (95.38%), by 0.99 percentage points and exceeded the clinically recommended AUC threshold of 0.95. Three-fold cross-validation confirmed the consistency of the reported results, achieving a mean accuracy of 95.56% ± 0.32% across all folds. On the ISIC 2019 dataset, the separately trained model achieved an accuracy of 91.47%, F1-score of 91.28%, and AUC of 0.9818 without any architectural modifications. Discussion The proposed heterogeneous weighted ensemble demonstrates high classification performance and consistent generalizability across independent dermoscopic datasets. Quantitative evaluation using TMAI and complementary XAI methods provides additional evidence that the model's predictions are grounded in clinically relevant lesion regions rather than background artefacts, supporting the potential of the framework for accurate and interpretable automated skin lesion classification.