Research And Application Of Functional Connectivity Method Based On Brain Magnetic Resonance Imaging Data

How to yield accurate diagnosis of mental disorders is currently the most important research topic in the neuroscience field.Since mental disorders(such as schizophrenia(SZ),schizoaffective disorder(SAD)and bipolar disorder(BPP))have many overlapping clinical symptoms,subjective diagnosis relying on symptoms tends to result in misdiagnosis of mental disorders to other similar disorders and affects the treatment of mental disorder.Using brain imaging(such as brain magnetic resonance imaging(MRI))to explore the mechanism and objective indicators of disorders as well as applying data mining to refine the disorder categories based on neuroimaging measures are promising methods to promote accurate diagnosis.Regarding the two aspects,in the thesis we propose supervised and unsupervised machine learning methods to study brain functional MRI(f MRI)data of psychoses.Our study finds the commonality and specificity of several psychoses with similar clinical symptoms.In addition,our study provides new insights for potential biotypes of related disorders from a neuroimaging view.Due to the inaccuracy of traditional diagnosis of mental disorders,many studies applied brain MRI data to identify objective indicators.SZ,SAD and BPP have many overlapping clinical symptoms,so it is difficult to distinguish them in clinical practice.In this thesis,we applied a classification framework using dynamic functional connectivity as features to distinguish them,in order to explore their common and unique aspects in brain functional connectivity.In this study,we used large-sample f MRI data of 623 subjects including SZ patients,SAD patients,BPP patients and healthy controls.First,the study used a sliding window method to calculate the whole-brain dynamic functional connectivity,and then extracted the connectivity states of each subject by independent component analysis.Next,we performed four-group(SZ,SAD,BPP and healthy controls)classification and pair-group(e.g.SZ vs.SAD)classification using support vector machine within a 10-fold cross-validation framework,and then summarized the common and unique connectivity changes among the disorders.The results show that our method achieved an average accuracy of 69% in the four-group classification and accuracy of greater than 80% in the pair group classifications.By summarizing the functional connectivity features that have a significant impact on the classification,we found that the important connectivity changes are mainly between the thalamus and cerebellum,between the central posterior gyrus and thalamus in the three types of disorders.Our research shows that dynamic functional connectivity may provide biological evidence for the diagnosis of mental disorders.Since the diagnosis labels of mental illnesses may be inaccurate,how to use unsupervised machine learning methods to obtain reliable disease categories based on brain images is an important issue in the neuroscience field.SZ and autism spectrum disorder(ASD)have much similarity in clinical symptoms.In this study,using the static brain functional connectivity estimated by f MRI,an unsupervised clustering method based on the graph kernel was proposed to explore the image-driven biotypes of the two disorders.First,we calculated the static functional connectivity of the whole brain,and then used the graph-based substructure mining algorithm to mine frequent subgraphs,then calculated graph kernel similarity matrix and performed clustering.Through our method,we got the potential biotypes across the two disorders.We found significant differences in functional connectivity between biotypes.Our research shows that the unsupervised clustering method based on graph kernel helps us discover potential biotypes of mental disorders,and can provide a new insight for the diagnosis of psychoses.