An Unsupervised Domain Adaptation Method to Enhance Diagnostic Model Resilience on Heterogeneous Medical Imaging Data

Deep learning models trained on a single dataset often experience performance degradation when applied to external datasets due to data shifts. Data shifts undermine robustness, generalizability, and trustworthiness of clinical AI systems. In this study, we propose VIC-Adaptor, a novel unsupervised domain adaptation (UDA) method that enhances medical imaging diagnostic model's generalizability on heterogeneous datasets and thus the resilience to data complexity in real-world medical applications. VIC-Adaptor introduces a new Variance-Invariance-Covariance Feature Alignment Loss (VIC-FAL), which promotes cross-domain feature alignment, preserves representational diversity, and minimizes redundancy without requiring labels of the data from the target domain during model training. We conducted experiments on two breast imaging datasets (i.e., contrast-enhanced mammography) with varying cohorts, each serving alternately as the source and target domain. Results showed that VIC-Adaptor consistently improves AUC values in the target domain, outperforming several compared existing UDA methods. Our proposed method can contribute to enhancing medical diagnostic model's resilience to heterogeneous imaging datasets, generalizability across the data from different sources/institutions, and increase model users' trust to AI model's clinical utilities in real-world environments.