The human cerebral cortex, as the outer layer of gray matter in the brain, plays a vital role in both neuroscience and clinical medicine (Agirman et al., 2017). It accounts for more than half of the brain’s volume and is responsible for regulating high-level cognitive functions such as perception, thought, language, attention, and motor control (Van Essen et al., 2018). Morphological analysis of the cerebral cortex is essential for advancing our understanding of human perception and cognitive mechanisms. Such analyses are often conducted using structural magnetic resonance imaging (sMRI), a technique widely applied in clinical settings. sMRI offers the advantage of accurately visualizing brain structures, including the location and extent of atrophy and the detailed anatomy of various brain regions. This imaging modality is one of the most commonly used tools for diagnosing psychiatric disorders (Feng et al., 2021). By observing the structure of the cerebral cortex, doctors can more accurately assess the patient's cognitive function and rule out cognitive dysfunction caused by other brain diseases (Kong et al., 2023; Zhu et al., 2021).

In recent years, clinical brain imaging research on the cerebral cortex has made some progress (Drevets et al., 2008; Guo et al., 2015). Some researchers use characteristics such as cortical thickness and surface area as objective indicators to study (Frangou et al., 2022). Including the structural characteristics of the cerebral cortex in the research scope is expected to provide more specific information for more in-depth cognitive research. However, this method ignores the connectivity between different areas of the brain. Numerous studies have demonstrated that structural connectivity reflects inter-regional communication and is closely linked to cognitive function changes. By constructing and studying cortical structural connectivity networks, we can more accurately and comprehensively understand the role of the cerebral cortex in cognitive processes (Fang et al., 2012; Kim et al., 2021). This will deepen the understanding of neuroscience and provide a more precise approach to the diagnosis and treatment of mental illnesses.

Existing studies have shown that imaging-based research on the cerebral cortex often lacks sufficient statistical power due to insufficient sample size, which reduces the ability to scientifically explain disease mechanisms and changes in neurological function (Razzak et al., 2018). In addition, too small a sample size not only affects the repeatability of clinical research, but also limits the verification of research results in a larger group. In the context of deep learning being widely used in brain network analysis, insufficient data has become a key factor restricting model performance. Since deep learning methods rely on a large amount of high-quality data to achieve good performance, limited samples can easily lead to problems such as overfitting. Therefore, improving the adaptability of the model to different data under the premise of limited data has become one of the important challenges facing current cerebral cortex classification research. To address these issues, we decided to introduce data augmentation to improve the reliability and generalization of classification studies. Data augmentation is a method that generates new samples by transforming and extending data to increase the size of the dataset (Mumuni and Mumuni, 2022; Zhao et al., 2021). Variational autoencoder (Han et al., 2019; Pu et al., 2016) is a common data augmentation method that maps data into a continuous latent space so that it can generate new samples similar to the training data. In addition, it can be used in combination with other generative models such as generative adversarial networks (GANs) (Aggarwal et al., 2021; Goodfellow et al., 2014) to further improve the quality and diversity of generated samples.

Specifically, we propose a multi-scale adversarial regularized variational autoencoder. Our model includes a topological enhancement module to further utilize the topological structural information in spatial feature extraction. Topological information is incorporated into the input of the autoencoder, and the Laplacian matrix is used to enhance the autoencoder's ability to perceive edges (Merris, 1994). The model mainly uses graph variational autoencoders to generate cortical structural connectivity. In order to capture structural connectivity at different spatial scales (Guo et al., 2022), we use graph neural networks (GNNs) of different scales as encoders. To help the autoencoder better learn the distribution of real data, we introduce an adversarial training module into the multi-scale graph variational autoencoder. The introduction of the adversarial training module can be used to identify the potential representation of the autoencoder, thereby forcing the potential representation to match the prior distribution. Through continuous iterative training, the accuracy of multi-scale graph variational autoencoder generation is improved. Finally, the generated structural connectivity and original data are input into the classification model.

We conduct comprehensive tests on classification tasks on two datasets: gender detection by sMRI in the Human Connectome Project dataset and disease detection by sMRI in the major depressive disorder dataset. In addition, we also tested the model's diagnostic performance for Alzheimer's disease on the Alzheimer's Disease Neuroimaging Initiative dataset. The MSARAE we proposed can effectively perform data augmentation on the dataset. The enhanced classification results significantly outperform other methods without data augmentation, and compared with the state-of-the-art baseline classification methods, our method improves the accuracy by 2.03 %, 3.59 %, and 3.19 % on the HCP, MDD, and ADNI datasets, respectively. Our method also performs better than other data augmentation methods. Experiments have shown the important contribution of each component module to model performance, including the topological enhancement module, the multi-scale variational autoencoder, the adversarial training module, etc. In summary, the main contributions of our work are as follows:

A new autoencoder structure for data augmentation and classification of cortical structural connectivity.

A new topological enhancement module that uses the Laplacian matrix to extract topological information on cortical structural connectivity.

The encoder structure in the autoencoder is modified to better capture cortical structural connectivity at different spatial scales.

An adversarial regularization module is introduced to improve the accuracy of multi-scale graph variational autoencoder generation.

Moreover, the remainder of this paper is organized as follows. Section 2 provides a detailed analysis of related work, including both traditional and deep learning approaches, to demonstrate the necessity of our proposed method. Section 3 introduces the proposed MSARAE model in detail, covering data preprocessing, model architecture, loss functions, and training strategies. Section 4 presents the datasets, experimental results, and comparative analyses. Finally, Section 5 concludes the paper, and Section 6 discusses potential future directions.