Enhancing Skin Lesion Classification via Uncertainty-Guided Adaptive Gated Fusion of Multiple Backbones

The accurate diagnosis of skin lesions remains challenging due to the class imbalance and heterogeneous appearance of the lesions. This diversity can destabilize multi-backbone fusion, as different backbones may exhibit miscalibrated confidence and correlated errors on hard or rare cases. A key limitation is that common fusion rules assume that all backbones are equally reliable for every sample, although their confidence may vary across inputs. To address this issue, an Uncertainty-Guided Adaptive Gated Fusion (UAGF) method is proposed. It is an instance-wise fusion mechanism that assigns backbone weights based on per-input reliability. UAGF first normalizes branch embeddings. Then it derives calibrated uncertainty features from temperature-scaled predictions, including entropy, maximum class probability, and the top-2 margin. These cues are used to compute fusion weights that emphasize the most reliable backbone for each input. UAGF is evaluated on ISIC2018 and ISIC2019, achieving accuracies of 95.39% and 94.49% and F1-scores of 94.50% and 91.77% averaged over five random seeds, respectively, indicating that reliability-aware, adaptive weighting is more effective than static fusion rules under class imbalance and heterogeneous lesion appearances. Keywords: skin lesion classification; uncertainty features; adaptive fusion; multi-backbone feature learning