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

Purpose To develop and evaluate machine-learning methods that reconstruct fractional anisotropy (FA) values and mean diffusivities (MD) from 3-direction diffusion MRI (dMRI) acquisitions. Methods Two machine-learning models were implemented to map undersampled dMRI signals with high-quality FA and MD maps that were reconstructed from fully sampled DTI scans. The first model was a previously described multilayer perceptron (MLP), which maps signals and FA/MD values from a single voxel. The second was a convolutional neural network U-Net model, which maps dMRI slices to full FA/MD maps. Each method was trained on dMRI brain scans (N = 46), and reconstruction accuracies were compared with conventional linear-least-squares (LLS) reconstructions. Results In an independent testing cohort (N = 20), 3-direction U-Net reconstructions had significantly lower absolute FA error than both 3-direction MLP (U-Net(3-dir): 0.06 +/- 0.01 vs. MLP3-dir: 0.08 +/- 0.01,P< 1 x 10(-5)) and 6-direction LLS (LLS6-dir: 0.09 +/- 0.03,P= 1 x 10(-5)). The MD errors were not significantly different among 3-direction MLP (0.06 +/- 0.01 x 10(-3)mm(2)/s), 3-direction U-Net (0.06 +/- 0.01 x 10(-3)mm(2)/s), and 6-direction LLS (0.07 +/- 0.02 x 10(-3)mm(2)/s,P> .1). Conclusion The proposed U-Net model reconstructed FA from 3-direction dMRI scans with improved accuracy compared with both a previously described MLP approach and LLS fitting from 6-direction scans. The MD reconstruction accuracies did not differ significantly between reconstructions.

Automated summary

The study aimed to develop and evaluate machine-learning methods for reconstructing FA and MD values from 3-direction diffusion MRI acquisitions. Two models were compared, with the U-Net model showing improved accuracy in FA reconstruction compared to MLP and LLS methods, while there was no significant difference in MD reconstruction accuracy.