Four Directions, One Solution: Enabling Rapid Diffusion Tensor MRI for Ultra-Low Field Using Deep Learning

This study revisits the tetrahedral encoding strategy originally proposed to accelerate Diffusion Tensor Magnetic Resonance Imaging (DT-MRI) by reducing the requisite number of diffusion-weighted measurements to four. We examine its practical limitations and explore how artificial intelligence (AI) can extend its utility. Specifically, we employ deep learning (DL) to estimate diffusion tensor parameters from four tetrahedrally arranged measurements rather than the conventional six or more, enabling substantially shorter scan durations. This approach is particularly relevant for low-field (LF) and ultra-low-field (ULF) MRI, where long acquisitions are needed to offset low SNR, and for non-compliant populations where extended scan times are impractical.

To overcome the numerical instabilities of traditional tetrahedral encoding, we developed DL models to predict axial and radial diffusivities and the principal eigenvector from four measurements. Synthetic training data spanned a wide range of diffusion tensors with uniformly distributed eigenvalues and orientations. Models were evaluated on digital phantoms and in vivo datasets acquired at 3T and 64 mT.

The DL-based approach improved accuracy in estimating diffusivities, fractional anisotropy, and orientation compared to conventional tetrahedral methods, particularly under low-SNR conditions. Residual errors persisted when the principal eigenvector aligned with scanner axes, reflecting inherent geometric constraints.

By revisiting and refining tetrahedral encoding through AI-driven strategies, this work demonstrates the feasibility of rapid DT-MRI using only four directions. These findings highlight opportunities for clinically viable diffusion imaging in time-constrained or resource-limited settings, while identifying key limitations for future research.