gms | German Medical Science

Introduction: Schizophrenia (SCZ) is a heterogeneous neurological disease lacking reliable biomarkers [1] despite genetic, blood, and brain alterations being linked to the disease. Automated decision support for diagnosing psychiatric diseases using deep learning (DL) classifiers based on structural magnetic resonance imaging (sMRI) of the brain is currently investigated. However, these classifiers typically require large datasets for training, which are not available for SCZ patients. To overcome this obstacle, we synthesize artificial data for SCZ patients and a healthy control (HC) group using generative adversarial networks (GAN) based on 193 3D sMRI images of SCZ patients and HC from the MCIC dataset.

Methods: Four GAN architectures based on a deep convolutional GAN (DC-GAN) are adapted to fit the problem and evaluated for their image synthesis capabilities. Spectral normalization regularization (SN-GAN) deals with the vanishing gradients problem that often occurs for small sample sizes [2]. Incorporating an encoder (α-SN-GAN) helps to alleviate mode collapse [3]. Both problems are also addressed by applying data augmentation during training (DiffAugment) [4]. Additionally, a hierarchical approach is adapted (HA-GAN) to reduce the computational cost of the training [5], and also combined with the α-SN-GAN to join their advantages (α-HA-GAN). Subsequently, three conditioning approaches – one classifier per class, auxiliary classifier, projection discriminator - are employed for creating the two clinical groups (SZC / HC). Finally, the images are “diagnosed” using a 3D convolutional neural network (3D-CNN). Multiple datasets consisting of different ratios of real and synthetic images are evaluated.

Results: Regularization combined with incorporating an encoder (α-SN-GAN) yields synthetic images of high fidelity and diversity shown with both qualitative and quantitative evaluation. The hierarchical approaches as well as data augmentation for training produces data of lesser quality. Furthermore, we demonstrate that the α-SN-GAN conditioned with an auxiliary classifier produces synthetic images that trains the 3D-CNN equally well as the real ones for the classification task. Increasing the training dataset size for synthetic images resulted in 15% improvement of classifier performance from 65% to 80% accuracy.

Discussion: This work demonstrates the synthesis of high-quality brain sMRI data capable of training a diagnostic DL classifier for SCZ. Four GAN architectures designed to handle the problems arising from training GANs on small datasets are systematically compared. The winner architecture is subsequently extended by three conditional approaches that are systematically compared as well. This approach can be adapted to bolster other imaging modalities such as fMRI for training multimodal classifiers that have shown promise for SCZ diagnosis. Furthermore, the auxiliary α-SN-GAN has the potential to reveal the underlying structural differences between the two clinical groups and might therefore aid in the research for SCZ biomarkers.

Conclusion: Generating synthetic (neuro)imaging data is a promising approach, especially for clinical use cases with inherently small dataset sizes. With this work, we demonstrate the ability to train GANs even on a complex, small dataset for a psychiatric disorder that lack objective diagnostic tools. However, the architectural choices for the GAN are essential and the resulting data always needs to be evaluated carefully.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.