Optimizing multi-site schizophrenia differentiation: MRI harmonization with 20 controls per scanner in a study of...

Although structural abnormalities has been reported in schizophrenia, generalizability across MRI scanners and protocols remains a major limitation for clinical application. Our previous study demonstrated that general linear model (GLM)-based harmonization can effectively distinguish patients with schizophrenia from healthy controls (HC) across MRI scanners. In this method, regions of interest (ROIs) showing volume reduction in schizophrenia were pre-defined, and age, sex, and total intracranial volume were included as dependent variables in the scanner and protocol specific GLM (spsGLM). The residuals (ε) of the spsGLM, the difference between estimated and measured ROI volume, were used as an indicator of schizophrenia. In the present study, we assessed required number of HC to apply this method, and adapted it to a larger dataset. We analyzed data from 1179 schizophrenia patients and 2381 HC across 15 MRI scanners. The minimum number of HC required was estimated to be 20. To avoid sampling bias, 20 HC were randomly selected 1000 times, and spsGLM model fitting was implemented for each set. The coefficients of spsGLM were calculated by averaging the results of 1000 trials, and ε was computed. Receiver operating characteristic (ROC) analyses were performed to evaluate ε. Results indicated that the area under the curve (AUC) from ROC analysis ranged from 0.66 to 0.83. ROC analysis using full sample showed an AUC of 0.74. These results were comparable to those obtained using ComBat harmonization or a Random Forrest classifier. In conclusion, scanning 20 HC enables our GLM-based harmonization method to generalize across scanners.