Multi-confound regression adversarial network for deep learning-based diagnosis on highly heterogenous clinical data

Automated disease detection in medical images using deep learning holds promise to improve the diagnostic ability of radiologists, but routinely collected clinical data frequently contains technical and demographic confounding factors that differ between hospitals, negatively affecting the robustness of diagnostic deep learning models. Thus, there is a critical need for deep learning models that can train on imbalanced datasets without overfitting to site-specific confounding factors. In this work, we developed a novel deep learning architecture, MUCRAN (Multi-Confound Regression Adversarial Network), to train a deep learning model on highly heterogeneous clinical data while regressing demographic and technical confounding factors. We trained MUCRAN using 16,821 clinical T1 Axial brain MRIs collected from Massachusetts General Hospital before 2019 and tested it using post-2019 data to distinguish Alzheimer's disease (AD) patients, identified using both prescriptions of AD drugs and ICD codes, from a non-medicated control group. In external validation tests using MRI data from other hospitals, the model showed a robust performance of over 90% accuracy on newly collected data. This work shows the feasibility of deep learning-based diagnosis in real-world clinical data.