Efficient Synthesis of 3D sMRI for Schizophrenia Classification with Generative Adversarial Networks

1385 

Abstract Submission 

Alexandra Reichenbach¹, Sebastian King¹, Yasmin Hollenbenders¹

¹Heilbronn University, Center for Machine Learning, Heilbronn, Baden-Württemberg

Alexandra Reichenbach  
Heilbronn University, Center for Machine Learning
Heilbronn, Baden-Württemberg

Sebastian King  
Heilbronn University, Center for Machine Learning
Heilbronn, Baden-Württemberg

Yasmin Hollenbenders  
Heilbronn University, Center for Machine Learning
Heilbronn, Baden-Württemberg

Schizophrenia (SCZ) is a heterogeneous psychiatric disease lacking reliable biomarkers (Mohammadi, Rashidi, & Amooeian, 2018) despite genetic, blood, and brain alterations being linked to it. 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 (Gollub et al., 2013). Despite rapid developments in the field of generative models, synthesis of 3D MRI brain data for psychiatric diagnosis support has not yet been demonstrated.

Four GAN architectures based on a deep convolutional GAN (DC-GAN) are adapted to address the technical challenges arising from the specific use case and evaluated for their image synthesis capabilities (Fig. 1). Spectral normalization regularization (SN-GAN) deals with the vanishing gradients problem that often occurs for small sample sizes (Miyato, Kataoka, Koyama, & Yoshida, 2018). Incorporating an encoder (α-SN-GAN) helps to alleviate mode collapse (Kwon, Han, & Kim, 2019). Both problems are also addressed by applying data augmentation during training (DiffAugment) (Zhao, Liu, Lin, Zhu, & Han, 2020). Additionally, a hierarchical approach is adapted (HA-GAN) to reduce the computational cost of the training (Sun et al., 2022) and also combined with the α-SN-GAN to join their advantages (α-HA-GAN). Subsequently, three conditioning approaches are employed for creating the two clinical groups (SZC/HC). Finally, the images are "diagnosed" using a 3D convolutional neural network (3D-CNN). Multiple training datasets consisting of different sizes and ratios of real and synthetic images are evaluated.

Regularization combined with incorporating an encoder (α-SN-GAN) yields synthetic images of high fidelity and diversity shown with qualitative and quantitative evaluation (Fig. 2 left and middle). 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 images for the diagnostic classification task. Increasing the training dataset size with synthetic images 6-fold results in 18% improvement of classifier performance from 61% to 79% accuracy (Fig. 2 right, large dataset).

This work demonstrates the synthesis of high-quality 3D brain sMRI data for two clinical groups from a small dataset. A diagnostic classifier separating real sMRI data from SCZ patients and HC can be trained successfully with the synthetic data and increasing the amount of synthetic training data increases the performance of the classifier by nearly 20%. This increase suggests that the synthetic data is capable of making the algorithm more robust for classifying the real data.
The systematic comparison of GAN architectures for basic training as well as for conditioning the data on the clinical group demonstrates that the architectural choices for the GAN are essential and the resulting data always needs to be evaluated carefully. 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 underlying structural differences between the two clinical groups and might therefore also aid in the research for SCZ biomarkers.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ²

Classification and Predictive Modeling ¹

Other Methods

Informatics Other

Computational Neuroscience

Data analysis

Machine Learning

MRI

Multivariate

Psychiatric Disorders

Schizophrenia

STRUCTURAL MRI

Other - generative adversarial network (GAN); data augmentation