CMDB Net: Multimodal De-Biased Fusion Network for Robust Dermatological Diagnosis

Abstract Effective clinical decision support in dermatology requires AI systems that reliably integrate multimodal patient data—including clinical metadata and lesion images, while mitigating modal bias that compromises diagnostic fairness and robustness across diverse patient populations. We propose the Case Mining-Based De-Biased Fusion Network (CMDB Net) to systematically address modality bias in multimodal dermatological diagnosis, advancing clinical decision support toward fairness and trustworthiness. CMDB Net combines supervised contrastive learning for robust cross-modal case feature mining with a dual-branch cross-modal attention mechanism regularized by Jensen–Shannon divergence to dynamically balance modality contributions and suppress feature bias. Extensive evaluation on PAD-UFES-20 and Derm7pt datasets demonstrates significant improvements in balanced accuracy (4.0% and 10.6% respectively), indicating enhanced diagnostic fairness for under-represented disease classes. CMDB Net achieves 86.2% accuracy on PAD-UFES-20 while maintaining superior and consistent performance across multiple backbone architectures. By systematically mitigating modal bias, CMDB Net improves both diagnostic accuracy and fairness, providing a trustworthy and task-agnostic framework for clinical decision support deployable in diverse settings, from centralized medical centers to resource-limited clinical environments. The code for this study is publicly accessible at https://github.com/lifeng-medical/CMDB-Master.