Learning Sequential Information in Task-based fMRI for Synthetic Data Augmentation.

Area of Science:

• Medical Image Analysis
• Neuroimaging
• Machine Learning

Background:

• Training data scarcity is a major challenge in medical image analysis, particularly for task-based functional magnetic resonance imaging (fMRI).
• Task-based fMRI data, characterized by spatio-temporal dynamics from cognitive tasks, requires extensive datasets for robust model training.
• Existing datasets are often insufficient for developing accurate diagnostic or analytical tools.

Purpose of the Study:

• To develop a novel approach for generating synthetic, high-resolution, task-specific fMRI sequences.
• To create augmented training datasets for downstream machine learning tasks using the generated synthetic fMRI data.
• To evaluate the efficacy of synthetic fMRI data augmentation for a specific clinical application.

Main Methods:

• Adaptation of the alpha-Generative Adversarial Network (α-GAN) architecture, combining strengths of GANs and variational autoencoders.
• Implementation of novel strategies for aggregating temporal information within the synthetic fMRI generation process.
• Evaluation of synthetic fMRI data through visual inspection and performance assessment on an autism spectrum disorder (ASD) classification task.

Main Results:

• The proposed method successfully generates high-resolution, task-specific synthetic fMRI sequences.
• Visualizations confirm the quality and realism of the generated spatio-temporal fMRI data.
• Synthetic fMRI data demonstrated significant effectiveness in augmenting training datasets for ASD classification, improving learning outcomes.

Conclusions:

• The developed synthetic fMRI generation approach effectively overcomes training data limitations in medical image analysis.
• The α-GAN adaptation provides a powerful tool for creating realistic and task-specific neuroimaging data.
• Synthetic fMRI data augmentation shows promise for enhancing the performance of machine learning models in clinical applications like ASD diagnosis.