Machine-Learning Enhanced Diffusion Tensor Imaging with Four Encoding Directions

Purpose: This study aims to reduce Diffusion Tensor MRI (DT-MRI) scan time by minimizing diffusion-weighted measurements. Using machine learning, DT-MRI parameters are accurately estimated with just four tetrahedrally-arranged diffusion-encoded measurements, instead of the usual six or more. This significantly shortens scan duration and is particularly useful in ultra-low field (ULF) MRI studies and for non-compliant populations (e.g., children, the elderly, or those with movement disorders) where long scan times are impractical. Methods: To improve upon a previous tetrahedral encoding approach, this study used a deep learning (DL) model to predict parallel and radial diffusivities and the principal eigenvector of the diffusion tensor with four tetrahedrally-arranged diffusion-weighted measurements. Synthetic data were generated for model training, covering a range of diffusion tensors with uniformly distributed eigenvectors and eigenvalues. Separate DL models were trained to predict diffusivities and principal eigenvectors, then evaluated on a digital phantom and in vivo data collected at 64 mT. Results: The DL models outperformed the previous tetrahedral encoding method in estimating diffusivities, fractional anisotropy, and principal eigenvectors, with significant improvements in ULF experiments, confirming the DL approach's feasibility in low SNR scenarios. However, the models had limitations when the tensor's principal eigenvector aligned with the scanner's axes Conclusion: The study demonstrates the potential of using DL to perform DT-MRI with only four directions in ULF environments, effectively reducing scan durations and addressing numerical instability seen in previous methods. These findings open new possibilities for ULF DT-MRI applications in research and clinical settings, particularly in pediatric neuroimaging