Fig. 7.

The DL model trained on HCP cohort could capture the differences of FNs between healthy controls (HC) and schizophrenia (SCZ) patients. (a) Personalized FNs of two HCs (the second and third rows) and one SCZ patient (the fourth row), identified by the DL model trained on HCP cohort, isolines of the FNs at a value of 0.15 in different colors (the fifth row), and their corresponding average FNs of all testing HCP subjects (the first row). (b) Personalized FNs computed using the DL model had significantly higher functional within-network homogeneity than those computed using the spatially-regularized NMF (bottom right, $p<{10}^{-5}$, Wilcoxon signed rank test). (c) Receiver operating characteristic (ROC) curves of classification models built on the personalized FNs of one run of the 2-fold cross-validation. (d, e) Classification rates and AUC values of 100 runs of 2-fold cross-validation, and the null distribution of classification rates and AUC values from permutation test (1000 runs of 2-fold cross-validation using permuted data).