Ensemble AI framework hits 96% accuracy in skin lesion classification
MaxViT-Tiny + ConvNeXt-Tiny + EfficientNetV2-B0 ensemble flags unreliable predictions with MC Dropout...
Skin cancer diagnosis from dermoscopic images remains difficult due to high intra-class variability, inter-class similarity, and class imbalance. A new paper by Rofiqul Islam and Lilatul Ferdouse introduces an ensemble deep learning framework that addresses these challenges while adding interpretability. The framework combines a vision transformer (MaxViT-Tiny) with two CNN-based models (ConvNeXt-Tiny and EfficientNetV2-B0) via deep ensemble learning. Monte Carlo (MC) Dropout estimates predictive uncertainty, allowing the system to flag low-confidence predictions for human review rather than blindly outputting a diagnosis.
Evaluated on the HAM10000 dataset, the framework achieves 96% accuracy and 99% ROC-AUC under uncertainty-aware filtering (entropy < 1.0, confidence >= 0.7). Macro-average precision, recall, and F1-score reached 94%, 95%, and 95%, with weighted-average scores of 96% across all metrics. Grad-CAM++ visual explanations highlight the specific lesion regions influencing each decision, making the model's reasoning transparent to clinicians. The combination of uncertainty estimation and explainable AI represents a meaningful step toward trustworthy computer-aided diagnosis, reducing the risk of silent misclassification in critical medical settings.
- Ensemble combines MaxViT-Tiny, ConvNeXt-Tiny, and EfficientNetV2-B0 for multi-class skin lesion classification
- 96% accuracy and 99% ROC-AUC on HAM10000 with uncertainty-aware filtering (entropy < 1.0, confidence >= 0.7)
- MC Dropout flags unreliable predictions while Grad-CAM++ provides visual explanations for clinician trust
Why It Matters
Clinicians can trust AI diagnoses only when uncertainty is quantified; this framework sets a new bar for explainable, safety-conscious medical ML.