Automated Skin Lesion Detection and Classification Using a Five-Fold Ensemble of Efficientnet-B4 with Adaptive Metadata Gating

Abstract

Skin cancer incidence is rising globally, and early accurate classification of dermoscopic lesions is critical for improving patient outcomes, particularly for melanoma where delayed detection drastically worsens prognosis. This work presents a comprehensive framework for multiclass skin lesion classification on the ISIC 2019 benchmark, which comprises eight diagnostic categories and a ninth unknown out-of-distribution (OOD) class in the test set. Our system addresses three interlinked challenges: extreme class imbalance, robust integration of patient metadata, and reliable OOD rejection. The architecture utilises an EfficientNet-B4 backbone along with a novel Adaptive Metadata Gating (AMG) module that learns imageconditioned gates bounded by an explicit ceiling in order to prevent shortcut reliance on metadata. Training incorporates an asymmetric focal loss with a double penalty for malignant false negatives, a hierarchically annealed auxiliary malignancy loss, class-aware Mixup/CutMix augmentation, Shades-of-Gray colour constancy, Exponential Moving Averaging, and Stochastic Weight Averaging. HAM10000 data supplements minority classes. Five independently trained EMA-checkpointed models are combined via softmax-temperature weighted logit-space averaging with fiveview test-time augmentation. An entropy-plus-gate composite uncertainty score drives calibrated per-fold threshold tuning for OOD rejection. On the official ISIC 2019 test set, the ensemble achieves balanced accuracy of 57.2%, macro AUC of 0.924, and macro specificity of 95.4%. On a fresh 20% stratified hold-out of ISIC training data, balanced accuracy rises to 80.3% and macro AUC to 0.975, confirming strong in-distribution discriminative capacity, with the test-set gap attributable to domain shift and the presence of unlabelled OOD samples.