MSARAE: Multiscale adversarial regularized autoencoders for cortical network classification | JoVE Visualize

Area of Science:

Neuroscience
Artificial Intelligence
Data Science

Background:

Limited data availability and privacy regulations pose significant challenges in cerebral cortex research.
Deep learning models require substantial training data for optimal performance and generalization, with small sample sizes leading to overfitting.
Augmenting data is crucial for enhancing the capabilities of deep learning models in neuroscience.

Purpose of the Study:

To propose a novel data augmentation method, the Multi-Scale Adversarial Regularized Autoencoder (MSARAE), to address data limitations in cortical structural connectivity analysis.
To improve the performance and generalizability of deep learning models for classifying brain conditions using augmented data.
To enhance the capture of topological and multi-scale features in cortical networks.

Main Methods:

Data preprocessing and construction of cortical structural connectivity networks.
Leveraging Laplacian eigenvectors to enrich topological information within the networks.
Utilizing variational autoencoders with multi-scale graph convolutional layers for feature extraction.
Implementing an adversarial regularization mechanism to minimize latent space distribution discrepancies and improve representational capacity.

Main Results:

The MSARAE model demonstrated superior performance in augmenting and classifying cortical structural connectivity compared to existing methods.
Experiments on major depression disorder (MDD), Human Connectome Project (HCP), and Alzheimer's Disease Neuroimaging Initiative (ADNI) datasets validated the model's effectiveness.
The adversarial regularization successfully aligned latent representations with real data distributions, enhancing model generalizability.

Conclusions:

The MSARAE provides an effective solution for data augmentation in brain imaging research, particularly for studies with limited sample sizes.
The proposed method enhances the ability to analyze and classify neurological and psychiatric disorders by improving the quality and quantity of training data.
This approach holds significant potential for advancing deep learning applications in neuroscience and clinical diagnostics.