Uncertainty-Aware Ensemble Deep Learning Framework Achieves 96% Accuracy in Skin Lesion Classification
A novel deep learning approach for classifying multi-class skin lesions integrates a vision transformer (MaxViT-Tiny) with CNN architectures (ConvNeXt-Tiny and EfficientNetV2-B0) through deep ensemble learning techniques. It employs Monte Carlo Dropout to estimate predictive uncertainty and Grad-CAM++ for providing visual interpretations. Tested on the HAM10000 dataset, the framework achieves an impressive 96% accuracy and a 99% ROC-AUC with uncertainty-aware filtering (entropy < 1.0, confidence >= 0.7), alongside macro-average scores of 94% for precision, 95% for recall, and 94% for F1. This study, published on arXiv (ID: 2608.11280), aims to improve reliability and transparency in clinical decision-making within AI-assisted dermatology, highlighting the importance of uncertainty quantification and explainability. Further validation across varied datasets is required.
Key facts
- Framework combines MaxViT-Tiny, ConvNeXt-Tiny, and EfficientNetV2-B0 via deep ensemble learning
- Monte Carlo Dropout estimates predictive uncertainty
- Grad-CAM++ provides visual explanations
- Evaluated on HAM10000 dataset
- Achieves 96% accuracy and 99% ROC-AUC under uncertainty-aware filtering
- Macro-average precision, recall, and F1-score are 94%, 95%, and 94%
- Addresses intra-class variability, inter-class similarity, class imbalance, and interpretability
- Announced on arXiv with ID 2608.11280
Entities
—