Improving the level of autism discrimination with augmented data by GraphRNN

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Deep learning for autism research relies heavily on datasets, but current public datasets like ABIDE are limited by small sample sizes and heterogeneity.
• Existing methods to improve recognition accuracy focus on feature selection and data augmentation, with limited success.

Purpose of the Study:

• To address the limitations of existing autism datasets by generating synthetic brain network data.
• To improve the effectiveness of deep learning models in autism disease research through enhanced data.

Main Methods:

• Utilized a graph recurrent neural network (GraphRNN) to learn the edge distribution of real brain networks.
• Generated synthetic brain network data to augment existing datasets.
• Evaluated the impact of synthetic data on the performance of discriminant models.

Main Results:

• The generated synthetic data significantly improved the classification ability of subsequent classifiers.
• Classification accuracy of a 50-layer ResNet was improved by up to 30% with the use of synthetic data compared to models trained without it.

Conclusions:

• GraphRNN-generated synthetic data is a promising approach to overcome data limitations in autism research.
• This method offers a substantial improvement in diagnostic accuracy for autism spectrum disorder using deep learning.