Efficient synthesis of 3D MR images for schizophrenia diagnosis classification with generative adversarial networks

Abstract

Schizophrenia and other psychiatric disorders can greatly benefit from objective decision support in diagnosis and therapy. Machine learning approaches based on neuroimaging, e.g. magnetic resonance imaging (MRI), have the potential to serve this purpose. However, the medical data sets these algorithms can be trained on are often rather small, leading to overfit, and the resulting models can therewith not be transferred into a clinical setting. The generation of synthetic images from real data is a promising approach to overcome this shortcoming. Due to the small data set size and the size and complexity of medical images, i.e. their three-dimensional nature, those algorithms are challenged on several levels. We develop four generative adversarial network (GAN) architectures that tackle these challenges and evaluate them systematically with a data set of 193 MR images of schizophrenia patients and healthy controls. The best architecture, a GAN with spectral normalization regulation and an additional encoder (α-SN-GAN), is then extended with an auxiliary classifier into an ensemble of networks capable of generating distinct image sets for the two diagnostic categories. The synthetic images increase the accuracy of a diagnostic classifier from a baseline accuracy of around 61 % to 79 %. This novel end-to-end pipeline for schizophrenia diagnosis demonstrates a data and memory efficient approach to support clinical decision-making that can also be transferred to support other psychiatric disorders.