Generating Realistic Brain MRIs via a Conditional Diffusion Probabilistic Model.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Acquiring Magnetic Resonance Imaging (MRI) data is costly, limiting sample sizes for neuroscience deep learning models.
• Generative Adversarial Networks (GANs) are popular for MRI synthesis but suffer from instability and limited data diversity.
• Diffusion Probabilistic Models (DPMs) offer stability but demand significant computational resources.

Purpose of the Study:

• To develop a computationally efficient method for synthesizing high-quality, diverse 3D brain MRIs.
• To address the limitations of existing GANs and DPMs in MRI synthesis for neuroscience research.
• To enable robust training of deep learning models with limited MRI datasets.

Main Methods:

• Proposed a conditional DPM (cDPM) with a memory-efficient, fine-grained training strategy.
• Trained a 2D cDPM to generate MRI subvolumes conditioned on spatially distant MRI slices.
• Utilized an attention network to learn interdependencies between MRI slices for anatomy-consistent 3D MRI generation.

Main Results:

• Generated realistic-looking 3D brain MRIs with high quality.
• Ensured synthesized MRIs share a similar distribution to real MRI data.
• Successfully diversified training datasets for deep learning models.

Conclusions:

• The proposed cDPM offers a stable and computationally efficient solution for 3D brain MRI synthesis.
• This method can significantly aid neuroscience research by overcoming MRI data acquisition limitations.
• The approach facilitates the creation of diverse and high-quality synthetic MRI datasets for enhanced model training.