Fast learning of fiber orientation distribution function for MR tractography using convolutional neural network.

Area of Science:

• Neuroimaging
• Diffusion-weighted magnetic resonance imaging (DW-MRI)
• Computational neuroscience

Background:

• Fiber orientation distribution functions (fODFs) are crucial for brain tractography in DW-MRI.
• Current methods like multi-shell multi-tissue constrained spherical deconvolution (MSMT-CSD) require extensive diffusion measurements, limiting clinical feasibility due to time and motion artifacts.
• Accelerating DW-MRI acquisition is essential for practical clinical applications.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (CNN) method for reconstructing fODFs from downsampled diffusion-weighted images (DWIs).
• To accelerate DW-MRI acquisition while maintaining accurate fODF reconstruction.
• To leverage spatial continuity across voxels to compensate for reduced diffusion measurements.

Main Methods:

• A CNN model was designed to process spherical harmonics (SH)-represented DWI signals and output fODF coefficients.
• The network incorporates surrounding voxel information to exploit spatial continuity, compensating for reduced gradient directions.
• Performance was evaluated on simulated data and in vivo Human Connectome Project (HCP) datasets, comparing against MSMT-CSD using angular correlation coefficient (ACC) and mean angular error (MAE).

Main Results:

• The CNN method demonstrated superior performance over MSMT-CSD on simulated and in vivo data, especially with fewer DWIs.
• With DWIs reduced from 95 to 25, median ACC for CNN was 0.91 vs. 0.77 for MSMT-CSD.
• The CNN method achieved lower mean angular error (MAE), particularly in multi-fiber regions, and enabled effective tractography with only 25 DWIs.

Conclusions:

• The proposed CNN-based method successfully reconstructs fODFs from significantly reduced DWIs (up to 11-fold reduction).
• This approach offers a streamlined reconstruction procedure, showing great potential for accelerating DW-MRI acquisition.
• The method provides accurate fODF reconstruction, facilitating improved brain connectivity analysis and neurological dysfunction exploration.