Computational protocol: Clinical Utility of Machine-Learning Approaches in Schizophrenia: Improving Diagnostic Confidence for Translational Neuroimaging

Similar protocols

Protocol publication

[…] T1 images were resliced (1 mm isotropic) and segmented into gray, white, and CSF tissue using the SPM8 Diffeomorphic Anatomical Registration Through Exponentiated Lie algebra (DARTEL) algorithm (). GM and WM images were separately warped onto a group average template and normalized to MNI space. To correct for variation due to field inhomogeneity, the images were bias field corrected using 60 mm FWHM setting using SPM8 (). To confirm that the higher inhomogeneity in the ultra-high field (7 T) did not affect the integrity of tissue segmentation process, we compared the total GM tissue volume from the 3- to 7-T scans. There was no significant difference [paired t(38) = 0.18, p = 0.9] in the GM volume. Furthermore, the total GM tissue volumes obtained from 3 to 7 T scans were very highly correlated (r = 0.93, p < 0.001), indicating that there were no systematic differences in the tissue segmentation between the 3- and 7-T scans (Figure ). The normalized, modulated, unsmoothed WM, and GM images for the 3- and 7-T datasets were then used as inputs to the separate linear SVM classifiers. In this approach, each subject’s input image is considered as a datapoint in a high-dimensional space of anatomical information (defined by GM or WM volumes). A hyperplane producing the greatest margin between the datapoints of the opposite groups (controls and patients) was identified using the multivariate information from the input images. A linear rather than non-linear kernel matrix was computed as input into the SVM classifier, as this allows the extraction of weight vectors of the high-dimensional data and also reduces the likelihood of overfitting (). The SVM analysis was carried out using Pattern Recognition for Neuroimaging Toolbox (PRoNTo), following the standard manualized descriptions ( approaches such as the one used here utilize a similarity matrix derived from all datapoints when developing classifiers; this obviates the need for explicit dimensionality reduction and optimizes computation efficiency (). To measure the test performance and to validate the classifier, a leave-one-subject-out (LOSO) cross validation approach was employed, where the classifier is trained on all subjects except one, which is used as test data. Balanced accuracy, specificity, sensitivity, and predictive values for each classifier were obtained and statistical significance of these measures was determined by way of permutation testing (n = 1000 permutations with random assignment of patient/control labels to the training data). […]

Pipeline specifications

Software tools SPM, PRoNTo
Application Neuroimaging analysis
Organisms Homo sapiens