Deep Learning applications for discriminating diagnosis of schizophrenia and bipolar disorder based on anatomical MR imaging

Abstract

Mental illnesses such as bipolar disorder and schizophrenia break the proper mental health equilibrium and disrupt the interaction of oneself with the environment. The diagnosis of these two mental disorders is limited to the appearance of clinical manifestations. An early categorization of psychiatric disorders is problematic due to the scarce of prolongated symptoms, episodes, and outbreak in initial stages of the mental illnesses. The present bachelor’s thesis aims to develop an automated diagnosis pipeline to discriminate patients with schizophrenia or bipolar disorder from healthy subjects, based on the analysis of brain subcortical structures. Brain T1 MRI data was considered to develop the comparison of patients and control brain morphology. A first convolutional neural network was trained with adult brain atlases developed by A. Hammers, which includes 30 subjects and 95 subcortical regions. Additionally, an analogous model was trained to segment 24 brain structures. Dice coefficient demonstrated de functionality of the networks, respectively obtaining 0.89 and 0.99 score. A second classification network was trained with 125 healthy control individuals, 50 schizophrenia patients and 49 bipolar disorder patients, from UCLA Consortium for Neuropsychiatric Phenomics database. Control health against psychiatric patients obtained 0.96 accuracy in test set, whereas schizophrenia against bipolar disorder approach exhibited 0.94 accuracy in test set. This project’s purpose is to demonstrate the possibility of developing alternative diagnosis tools to differentiate schizophrenia and bipolar disorder patients from healthy controls, based on morphology analysis of brain MR images.