Image-encoded biological and non-biological variables may be used as shortcuts in deep learning models trained on multisite neuroimaging data

OBJECTIVE: This work investigates if deep learning (DL) models can classify originating site locations directly from magnetic resonance imaging (MRI) scans with and without correction for intensity differences. MATERIAL AND METHODS: A large database of 1880 T1-weighted MRI scans collected across 41 sites originally for Parkinson’s disease (PD) classification was used to classify sites in this study. Forty-six percent of the datasets are from PD patients, while 54% are from healthy participants. After preprocessing the T1-weighted scans, 2 additional data types were generated: intensity-harmonized T1-weighted scans and log-Jacobian deformation maps resulting from nonlinear atlas registration. Corresponding DL models were trained to classify sites for each data type. Additionally, logistic regression models were used to investigate the contribution of biological (age, sex, disease status) and non-biological (scanner type) variables to the models’ decision. RESULTS: A comparison of the 3 different types of data revealed that DL models trained using T1-weighted and intensity-harmonized T1-weighted scans can classify sites with an accuracy of 85%, while the model using log-Jacobian deformation maps achieved a site classification accuracy of 54%. Disease status and scanner type were found to be significant confounders. DISCUSSION: Our results demonstrate that MRI scans encode relevant site-specific information that models could use as shortcuts that cannot be removed using simple intensity harmonization methods. CONCLUSION: The ability of DL models to exploit site-specific biases as shortcuts raises concerns about their reliability, generalization, and deployability in clinical settings.