MorphoNet: An Interpretable Hierarchical Deep Learning Framework for Multi-Class Skin Lesion Classification Using...

Accurate and interpretable classification of dermoscopic skin lesions remains challenging because lesion subtypes frequently exhibit overlapping visual characteristics, public datasets are highly imbalanced, and images acquired from different sources may present substantial domain variation. This study proposes MorphoNet, an interpretable two-stage deep learning framework for hierarchical skin lesion classification. A total of 16,845 dermoscopic images from HAM10000, ISIC 2019, and DERM12345 were integrated into eight diagnostic categories. In Stage 1, an EfficientNet-B4-based classifier distinguishes benign from malignant lesions. In Stage 2, separate classifiers identify four malignant and four benign subtypes. Quantitative asymmetry, border, color, and diameter descriptors are incorporated through an attention-guided feature fusion module, while correlation alignment regularization is used to reduce discrepancies in feature distribution across the three data sources. Experiments were repeated using three random seeds. MorphoNet achieved accuracies of 94.01 ± 0.50%, 89.05 ± 0.51%, and 91.84 ± 0.41%, with corresponding F1-scores of 93.70 ± 0.55%, 84.00 ± 1.00%, and 87.97 ± 1.13% for benign-malignant screening, malignant subtype classification, and benign subtype classification, respectively. Ablation analysis showed that the complete framework improved F1-score by 2.52-4.71 percentage points over the visual-only hierarchical EfficientNet-B4 model. MorphoNet also achieved higher numerical performance than ResNet-50, DenseNet-121, and image-only EfficientNet-B4 baselines trained under the same protocol. The learned attention profiles were consistent with clinically relevant morphological characteristics. These findings demonstrate the complementary value of hierarchical classification, ABCD-guided feature fusion, and domain alignment for interpretable multi-source skin lesion analysis, although independent multicenter clinical validation remains necessary.