Meta-Classifier Fusion of Deep Ensemble Representations for Enhanced Skin Lesion Diagnosis

Although ensemble learning has proven to be effective for skin lesion classification, existing methods often under-use the complementary information embedded in heterogeneous convolutional neural networks (CNNs). In this work, we propose a meta-classifier framework that leverages high-dimensional feature representations from a frozen, diversified ensemble of CNNs to improve diagnostic robustness. To exploit both decision-level and feature-level diversity, we extract penultimate layer embeddings from each ensemble member, concatenate them into a unified feature vector, and train a meta-classifier on this enriched representation. We evaluated multiple meta-classifier architectures, including lightweight neural networks and kernel-based models, and benchmarked their performance against conventional ensemble averaging strategies. Experiments on the ISIC 2018 dataset (13,000+ dermoscopic images) demonstrate that our approach outperforms both individual CNNs and classical ensemble methods, achieving an accuracy of 91.32% (versus 90.15% in prior work). The meta-classifier’s ability to learn cross-model feature interactions and suppress ensemble uncertainty is validated through ablation studies and robustness tests against noisy input. This work highlights the untapped potential of hierarchical fusion in ensemble-based medical image analysis.