Detection of Schizophrenia Based on Brain Structural Analysis, Using Machine Learning over Different Combinations of Multi-slice Magnetic Resonance Images

Schizophrenia is a mental disease with a great number of clinical manifestations that makes diagnosis a great challenge. Until a correct diagnosis is attained, the patient experiments mental suffering that can lead to social conflicts, involuntary accidents, and suicides. Early diagnosis, despite the clinical complexity, is therefore of utmost importance, and several recent studies focus on analyzing structural brain modifications that have been correlated to schizophrenia, and that can be detected in anatomical magnetic resonance images. Previous research applying machine learning to such images presented promising results. However, the scope was limited to analyzing only one or few slices of the brain while not using recent algorithms at the core of the classifiers. This can lead to information loss due to sub-optimal features extraction. In this study we created machine learning models based on Convolutional Neural Networks, and evaluated the best training parameters based on a data set of magnetic resonance images from persons with schizophrenia and from a control group. We analyzed the performance of the classifiers, first trained with individual slices of the brain and later with different combinations of multiple magnetic resonance slices. Our results suggest that it is possible to increase the performance metrics such as accuracy, sensibility, and precision of a classifier from approximately 50% when trained with a single slice, to over 80% when trained with a properly selected combination of slices.