Extracting diffusion tensor fractional anisotropy and mean diffusivity from 3-direction DWI scans using deep...

Area of Science:

• Medical Imaging
• Machine Learning
• Neuroscience

Background:

• Diffusion MRI (dMRI) is crucial for neuroimaging.
• Reconstructing diffusion tensor imaging (DTI) metrics like fractional anisotropy (FA) and mean diffusivity (MD) from limited dMRI data is challenging.
• Developing efficient reconstruction methods is vital for broader clinical application.

Purpose of the Study:

• To develop and evaluate machine learning (ML) methods for reconstructing FA and MD values from 3-direction dMRI.
• To compare the performance of ML models against conventional linear-least-squares (LLS) methods.

Main Methods:

• Two ML models, a multilayer perceptron (MLP) and a convolutional neural network (CNN) U-Net, were implemented.
• Models were trained on dMRI brain scans (N=46) to map undersampled dMRI signals to FA/MD maps.
• Reconstruction accuracy was evaluated using an independent cohort (N=20) and compared to 6-direction LLS reconstructions.

Main Results:

• The 3-direction U-Net model achieved significantly lower absolute FA error compared to 3-direction MLP and 6-direction LLS.
• FA error for U-Net was 0.06 ± 0.01, compared to 0.08 ± 0.01 for MLP and 0.09 ± 0.03 for LLS.
• MD reconstruction errors were not significantly different across the 3-direction MLP, 3-direction U-Net, and 6-direction LLS methods.

Conclusions:

• The U-Net model demonstrates superior accuracy in reconstructing FA from 3-direction dMRI compared to MLP and LLS methods.
• MD reconstruction accuracy was comparable across all tested methods.
• These findings suggest the potential of U-Net for efficient and accurate dMRI data reconstruction.