Deep Learning for Automated Dermatoscopic Skin Lesion Classification: Addressing Class Imbalance and Visual Explainability
Cutaneous melanoma and non-melanoma skin cancers represent a major global public health challenge where early dermatoscopic triage significantly improves patient survival. However, deploying deep Convolutional Neural Networks (CNNs) on clinical dermatoscopic benchmarks such as HAM10000 faces three primary biomedical engineering challenges: (1) severe class imbalance exceeding 60:1 between benign melanocytic nevi and rare pathologies, (2) patient- and lesion-level data leakage across random train-test splits, and (3) opaque "black-box" decision-making that limits clinical trust. In this work, we present an end-to-end, leakage-free classification and visual explainability pipeline utilizing an ImageNet-pretrained EfficientNet-B0 backbone optimized with a smoothed inverse-frequency class-weighted cross-entropy loss. Enforcing strict lesion_id-grouped stratified splitting across N = 10,015 dermatoscopic images (7,995 train, 1,028 validation, 992 held-out test), our framework achieves a Macro ROC-AUC of 0.9654, a Weighted ROC-AUC of 0.9662, a Macro Specificity of 96.19%, and an overall diagnostic accuracy of 83.97%. Notably, the model attains a 0.9850 ROC-AUC on basal cell carcinoma and 0.9414 ROC-AUC on malignant melanoma. Integrating Gradient-weighted Class Activation Mapping (Grad-CAM) confirms that predictions localize along irregular pigment networks and lesion borders rather than optical imaging artifacts. Open-Source Code & Weights: https://github.com/Dvka-c/dermatoscopic-lesion-cnnInteractive Web Demo: https://huggingface.co/spaces/Dvka-c/skin-lesion-gradcam
Authors
- Daulet Dastan
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24198556
- Primary Topic
- Cutaneous Melanoma Detection and Management
- Type
- preprint