Interpretable Skin Cancer Identification Using a Hybrid Deep Learning and XAI Framework on HAM10000

Deep learning-based automated classification of dermatoscopic skin lesions has exhibited promising potential in diagnostics. However, two prominent issues need to be addressed before achieving high-quality diagnostic tools: inconsistent performance in the case of imbalanced classes and poor clinical interpretability of models. Even though some studies have attempted to leverage both deep and shallow learning by combining pretrained convolutional neural networks (CNNs)-based feature extraction with classical machine learning (ML) models, very few of them systematically explore several model combinations based on various clinically important metrics, such as F1-score, precision, recall, accuracy, etc., and utilize decision threshold calibration techniques. In this research, we present an evaluation of a systematic framework with threshold calibration for the comparison of several hybrid models on seven-class skin lesion classification (multi-class) on the HAM10000 dataset. In particular, we used deep features extracted from three pretrained CNN architectures, i.e., DenseNet201, InceptionV3 and EfficientNet-B4. These deep features were used as inputs for six different classical classifiers. As a result, we obtained 18 comparable hybrid models that were then systematically compared by multiple clinically relevant metrics: accuracy, macro-precision, macro-recall, macro-F1, ROC-AUC, Precision-Recall-AUC, and log loss. Also, fold-wise optimization of decision thresholds was performed, which was based on the maximization of the macro-F1 score. Finally, we found out that DenseNet201 with an SVM-RBF classifier yielded the highest performance among all 18 tested models, showing 90.88% accuracy, 90.7% macro-precision, and 0.921 ROC-AUC. To analyze the clinical plausibility, top-performing models were further explained with explainable artificial intelligence (XAI) techniques: Grad-CAM, LIME and Occlusion Sensitivity. Results show that the most successful models concentrated mostly on lesion-specific areas. Overall, this study contributes a reproducible hybrid-XAI model-selection framework rather than a single black-box classifier, supporting more transparent and clinically meaningful skin lesion diagnosis.