The Dependence of MRI Parameters on the Accuracy of Program Binary Classification of Schizophrenia Based on Resting State fMRI Data

The article explores the dependence of the accuracy of the binary classification of subjects based on the presence of schizophrenia pathology using fMRI data on two parameters: scanning duration and spatial smoothing. The data set was acquired on a Siemens Magnetom Verio 3 T MRI scanner and included 36 subjects undergoing treatment at the State budgetary institution of health care “Psychiatric Hospital no. 1 Named after N.A. Alexeev” as well as 36 subjects from the control group. The total scanning duration for each subject was 900 time points (648 s) with a resolution of 2 × 2 ×  2 mm3. To simulate different scan durations, the data was divided into 3 sets of 300 time points, each of which was independently used for classification. To analyze the effect of spatial smoothing, Gaussian blur with convolution kernels of 4, 6, 8, 10, 12 mm was used. For classification 38 machine learning methods from the scikit-learn software library were used. To generate sets of feature vectors for classification, well-known regional homogeneity (ReHo) methods and correlation matrices calculated using the structural and functional atlas of the CONN software package were used.