Biomarker-Aware Mixture-of-Experts for Device-Invariant Personalized Skin Diagnosis

Abstract

Smartphone-based skin analysis is limited by device heterogeneity, uncontrolled illumination, and subject-specific variability, yet existing methods treat calibration, domain adaptation, and multi-task prediction in isolation without capturing the causal structure underlying biomarker interactions. We propose an end-to-end framework for device-invariant skin diagnosis built on three principles: (i) hierarchical hybrid normalization formulated as a causal intervention to disentangle diagnostic signals from device-induced nuisance, (ii) a keypoint-guided Skin Mixture-of-Experts that organizes experts around dermatological biomarkers with conditional routing, and (iii) a causal expert deliberation module that enables experts to exchange evidence through clinically motivated dependency constraints, moving beyond simple expert averaging to structured causal reasoning over multi-biomarker outputs. Experiments on multiple public dermatology benchmarks demonstrate consistent improvements over strong 2024--2025 vision and medical foundation baselines.