Diffusion Bridge Models for 3D Medical Image Translation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Diffusion tensor imaging (DTI) is vital for brain microstructure analysis but is time-consuming to acquire.
• T1-weighted (T1w) magnetic resonance imaging (MRI) is more readily available but offers different information.
• Bridging the gap between these modalities is crucial for enhancing neuroimaging datasets.

Purpose of the Study:

• To develop a novel diffusion bridge model for 3D brain image translation between T1w MRI and DTI modalities.
• To enable high-quality generation of DTI fractional anisotropy (FA) images from T1w images and vice versa.
• To facilitate cross-modality data augmentation and reduce reliance on extensive DTI acquisition.

Main Methods:

• A diffusion bridge model was employed for 3D brain image translation.
• The model was trained to generate DTI FA images from T1w images and T1w images from DTI data.
• Performance was evaluated using perceptual similarity, pixel-level agreement, and distributional consistency metrics.

Main Results:

• The diffusion bridge model successfully generated high-quality synthetic DTI FA images from T1w images.
• The generated images accurately captured anatomical structures and preserved white matter integrity information.
• Synthetic data achieved comparable performance to real data in sex and Alzheimer's disease classification tasks.

Conclusions:

• The diffusion bridge model provides a promising solution for enhancing neuroimaging datasets.
• This approach can reduce the need for lengthy DTI acquisition, improving efficiency.
• The model has significant potential to impact neuroimaging research and clinical decision-making.