Academic

Ahmed Jameel Mohammed a b, Masoud M. Hassan c

https://www.sciencedirect.com/science/article/pii/S2590123026030100?via%3Dihub

Reliable skin cancer detection from dermoscopic images is essential for early diagnosis and improved clinical outcomes. Although deep learning models perform strongly on benchmark datasets, most approaches apply a uniform classifier to all samples and treat uncertainty as a post-hoc measure rather than integrating it into the decision process. We propose an Entropy-Guided Three-Phase Ensemble Framework (EG-3PEF) that embeds uncertainty into classification. A base model estimates predictive entropy and routes each sample, via a validation-derived median entropy threshold, to a Certain phase (L2) for high-confidence cases or an Uncertain phase (L3) for ambiguous cases. High-confidence samples are classified using the proposed Entropy-Filtered Soft Voting (EFSV) strategy, which selects the most confident models per sample, whereas ambiguous cases use a more expressive ensemble. Targeted MixUp–CutMix augmentation is applied only to high-entropy training samples to strengthen decision-boundary learning without altering the global distribution, and uncertainty is further quantified using Monte Carlo Dropout (T = 100). The framework is evaluated on two independent dermoscopic datasets, ISIC 2019 and the Kaggle Skin Cancer. EG-3PEF achieves 97.60% accuracy, 97.57% F1-score, and 0.9969 AUC on ISIC 2019, and 97.42% accuracy, 97.40% F1-score, and 0.9944 AUC on the Skin Cancer dataset, outperforming recent CNN- and ensemble-based methods on both. By treating entropy as an active routing signal, EG-3PEF preserves near-perfect accuracy on confident cases while improving on ambiguous, boundary-level lesions where diagnostic errors most often occur. This cross-dataset consistency indicates that uncertainty-aware, phase-specific design offers a reliable and interpretable direction for trustworthy computer-aided skin cancer screening.