DDTracking: A diffusion model-based deep generative framework with local-global spatiotemporal modeling for diffusion...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Diffusion MRI (dMRI) tractography maps brain fiber pathways in vivo.
• Traditional methods are computationally intensive and sensitive to noise.
• Deep learning offers improved efficiency and accuracy but often focuses on local or global information.

Purpose of the Study:

• Introduce DDTracking, a novel deep generative framework for dMRI tractography.
• Reformulate streamline propagation as a conditional denoising diffusion process.
• Apply diffusion models for the first time in fiber tracking.

Main Methods:

• Develop a dual-pathway encoding scheme for local spatial and global temporal features.
• Implement a conditional diffusion model integrating spatiotemporal features for orientation prediction.
• Train all components end-to-end without pretraining for joint optimization.

Main Results:

• DDTracking achieves superior tracking accuracy and computational efficiency compared to traditional and deep learning methods.
• Demonstrate high generalizability across diverse datasets, including synthetic and clinical data.
• Highlight the ability to capture fine-scale details and ensure long-range streamline consistency.

Conclusions:

• DDTracking represents a significant advancement in dMRI tractography using diffusion models.
• The framework offers improved performance and robustness across various imaging conditions.
• The approach enables more precise in vivo mapping of human brain connectomes.