Generating Novel Brain Morphology by Deforming Learned Templates.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Generative models for 3D structural brain MRI are crucial for research.
• Existing methods synthesizing images directly may struggle with intricate morphological details.
• Attribute-specific generation (age, sex, disease state) is a key challenge.

Purpose of the Study:

• To develop a novel 3D brain MRI generation method using latent diffusion models (LDMs).
• To improve the synthesis of morphologically-plausible and attribute-specific brain MRI samples.
• To address limitations of direct image synthesis approaches.

Main Methods:

• Proposed MorphLDM, a 3D brain MRI generation method based on state-of-the-art latent diffusion models (LDMs).
• Utilized a learned template and synthesized deformation fields instead of direct image synthesis.
• Employed a specialized encoder-decoder architecture and minimized a registration loss between original and deformed templates.

Main Results:

• MorphLDM outperformed existing generative baselines in empirical evaluations.
• Achieved superior performance in image diversity, adherence to input conditions, and voxel-based morphometry.
• Demonstrated the ability to generate high-quality, attribute-specific 3D brain MRI samples.

Conclusions:

• MorphLDM offers a promising new approach for generating realistic 3D brain MRI data.
• The method effectively captures intricate morphological details and attribute specificity.
• This work advances generative modeling for neuroimaging applications.