Structural MRI (sMRI) is the study of the structure of different parts of the brain and making predictions by comparing the MRI scans of patients and control subjects. By comparing the scans, ML algorithms can be trained to classify patients with and without SZ. Leonard et al. [26] was one of the first to use discriminant function analysis (DFA) to correctly classify the subjects (77% accuracy) from the structural brain scans. The bulk of the work in other sMRI techniques focus on analyzing and comparing Grey Matter (GM) and White Matter (WM), and their corresponding size or density. Other groups used DFA and its variants to classify and detect patients with SZ by considering other Region-of-Interest (ROI) in the brain and were able to achieve similar or better prediction rates by performing DFA on sMRI scans. Through the various studies, we have noticed that researchers tend to make the same conclusion—the risk of SZ may depend on the total amount of neural deviance rather than on anomalies in a single structure or circuit.

Another popular method used in classifying SZ is the use of support vector machine (SVM) classifiers, including non-linear SVM and its variants. SVM forms the majority of the analysis from detection using sMRI images. Customary in most predictive analysis, the SVM models were constructed from one set of subjects (training set) and the model was then applied to a different set of subjects (test set) to cross-validate the model. Many groups also used SVM to compare at-risk mental state (ARMS) SZ individuals with healthy controls (HC). In particular, in the work of Koutsouleris et al. [27], non-linear SVM with multivarite neuroanatomical pattern classification was performed on the sMRI data of individuals with ARMS (early and late) and HC. The accuracy of the method was then evaluated by categorizing the baseline imaging data of individuals with transition to psychosis as compared to those without transition and HC after 4 years of clinical follow-up. The 3-group, cross-validated classification accuracies of the first analysis were 86% in discriminating HC, 91% in discriminating early ARMS, and 86% in discriminating late ARMS. The accuracies in the second analysis were 90% in discriminating HC, 88% in discriminating individuals with transition, and 86% in discriminating individuals without transition. Independent HC were correctly classified in 96% (first analysis) and 93% (second analysis) of cases. Notably, there were several studies that point to better prediction accuracies when combining multiple features than simply employing single-modal features in SVM [28,29,30].

Other ML methods notably include the regression model used by Csernansky et al. [31] to predict SZ among subjects who were similar in age, gender and parental socioeconomic status, with 75% prediction rate. However, it was unable to predict the severity of the condition using the same model. Other notable methods employed include the high-dimensional non-linear pattern classification used by Davatzikos et al. [32] to quantify the degree of separation between patients and control, achieving 81.1% mean classification accuracy. An overview of the work, sample size and accuracy from utilizing machine learning techniques on structural magnetic resonance imaging data is compiled in Table 1.

Summary of work and predictions relating to the detection of SZ using data from structural MRI scans via various artificial intelligence techniques and machine learning algorithms.

Note: The content above has been extracted from a research article, so it may not display correctly.

Please log in to submit your questions online.
Your question will be posted on the Bio-101 website. We will send your questions to the authors of this protocol and Bio-protocol community members who are experienced with this method. you will be informed using the email address associated with your Bio-protocol account.

We use cookies on this site to enhance your user experience. By using our website, you are agreeing to allow the storage of cookies on your computer.