Diffusion MRI GAN synthesizing fibre orientation distribution data using generative adversarial networks.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Neuroimaging

Background:

• Machine learning requires extensive training data, which is often unavailable for rare diseases.
• Generative neural networks can create synthetic medical data, but complex data types like Fibre Orientation Distributions (FODs) remain challenging to generate.
• FODs model diffusion in the brain using spherical harmonics, requiring multiple 3D volumes for representation.

Purpose of the Study:

• To develop a novel method for generating synthetic Fibre Orientation Distributions (FODs) using advanced generative models.
• To address the challenge of limited training data for machine learning in rare pathologies.
• To validate the anatomical accuracy and properties of the generated synthetic FODs.

Main Methods:

• Training an alpha-Wasserstein Generative Adversarial Network (α-WGAN) model.
• Utilizing the Human Connectome Project (HCP) dataset for training.
• Comparing generated synthetic FODs against a validation dataset for anatomical accuracy and property matching.

Main Results:

• Successful generation of synthetic FODs using the α-WGAN model.
• Generated FODs demonstrated anatomically accurate fibre bundles and connectomes.
• Properties of synthetic FODs closely matched those of the validation dataset.

Conclusions:

• The developed α-WGAN approach effectively generates complex synthetic medical imaging data, specifically FODs.
• This method holds significant potential for augmenting limited clinical datasets, particularly for rare diseases.
• The approach is adaptable for generating other complex medical imaging data types, advancing machine learning applications in medicine.