Causal Uncertainty-Decomposed Ensemble Learning for Confounding-Aware Skin Cancer Detection in Dermoscopic Images

Background/Objectives: Automated dermoscopic image analysis can assist research on early skin cancer detection, but deep learning models may learn acquisition-related shortcuts, including terminal hairs, illumination gradients, gel bubbles, shadows and measurement markings. This study evaluates a causal uncertainty-decomposed ensemble (CUDE) for confounding-aware skin-lesion classification within the ISIC 2019 benchmark setting. Methods: CUDE uses a structured variational autoencoder (SVAE) to impose separate lesion-relevant and acquisition-related nuisance latent heads. Three stochastic latent-space experts are trained on complementary streams and are integrated by a Dirichlet-based fusion module trained with nuisance sampling. The causal graph is used as a modeling prior rather than proof of causal identification. Results: Using a stratified internal split of the labeled ISIC 2019 training collection, CUDE achieved a balanced accuracy of 89.7% and an AUROC of 0.972. Under the predefined synthetic artifact protocol, CUDE showed a relative performance drop of 5.0%, compared with 9.6% for the deep ensemble. Deferring the 10% most uncertain cases increased non-deferred accuracy from 89.7% to 93.2% and reduced false-negative rates for melanoma, BCC and SCC in the non-deferred subset. Conclusions: Within the evaluated ISIC 2019 split and controlled synthetic corruption protocol, structured latent separation, stochastic expert diversity and Dirichlet fusion were associated with improved benchmark robustness and calibration. These results should not be interpreted as evidence of broad real-world robustness or clinical deployment readiness; external multi-center, device-diverse and skin-tone-diverse validation remains required.