Schizophrenia versus Healthy Controls Classification based on fMRI 4D Spatiotemporal Data

A wide array of machine learning approaches have been employed for differentiating patients with mental health disorders from healthy controls using neuroimaging data. However, almost all such methods have been applied on inputs based on connectivity matrices or features derived from the neuroimaging data. Only a few papers recently have considered such classification based on the original voxel-based spatiotemporal data. In this paper, we report the performance of a few cutting edge machine learning algorithms on voxel-based fMRI data to classify healthy controls and patients with schizophrenia. The methods that we employed included convolutional neural networks, convolutional recurrent neural networks with long short-term memory and a transfer learning approach for classification based on Wasserstein generative adversarial networks. In order to reduce the computational burden to fit in with available hardware, we had to reduce the original 4-dimensional data to 3-dimensional inputs for almost all architectures. Our results indicate that the relatively simpler architecture based on convolutional neural networks showed reasonable unambiguity in grouping patients from healthy controls. In contrast, the performance of the other two more complex architectures that we employed were comparatively poorer.