Comparative Analysis of Pre-trained CNN Models for Neurobiological Disease Classification

Machine learning has gained enormous popularity in the diagnosis of neurobiological diseases due to its pattern recognition and diagnostic prediction ability. Numerous studies have explored machine learning with neuroimaging data to separate schizophrenic patients from healthy controls but only a small number of them have included bipolar patients, allowing the most clinically pertinent discrimination between two psychotic groups. Also, it has been proven by previous studies that machine learning-based methods are not that effective in drawing out complex patterns from neuroimaging data. In order to address these issues, we have implemented three different pre-trained CNN models (VGG-16, MobileNet, and Xception) to classify schizophrenia (SZ), bipolar disorder (BD), and healthy controls (HC) automatically based on the gray matter volume images. In our study, 66 bipolar patients, 66 schizophrenia patients, and 66 healthy controls were taken into account. Experimental findings show that among the three classifiers, Xception shows the best performance with an accuracy of 90.21%.