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

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Dermatology

Background:

• Automated dermoscopic image analysis aids early skin cancer detection.
• Deep learning models may exploit acquisition artifacts (hairs, shadows) instead of lesion features.
• Confounding factors in image data can hinder diagnostic accuracy.

Purpose of the Study:

• Evaluate a causal uncertainty-decomposed ensemble (CUDE) for confounding-aware skin lesion classification.
• Assess CUDE's robustness against acquisition-related shortcuts in dermoscopic images.
• Improve the reliability of AI models for early skin cancer detection.

Main Methods:

• CUDE employs a structured variational autoencoder (SVAE) with separate latent heads for lesion and nuisance factors.
• Three stochastic latent-space experts are trained and fused using a Dirichlet-based module.
• Nuisance sampling and a causal graph prior are utilized for improved model training.

Main Results:

• CUDE achieved 89.7% balanced accuracy and 0.972 AUROC on the ISIC 2019 benchmark.
• CUDE demonstrated a 5.0% relative performance drop with synthetic artifacts, outperforming deep ensembles (9.6%).
• Deferring uncertain cases improved accuracy to 93.2% and reduced false negatives for melanoma, BCC, and SCC.

Conclusions:

• Structured latent separation, expert diversity, and Dirichlet fusion enhance robustness and calibration in benchmark settings.
• CUDE shows promise for more reliable AI-based skin lesion classification.
• Further validation across diverse clinical settings, devices, and skin tones is essential for real-world deployment.