PPubMed10 May 2026

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

Joshua Mawuli Ametepe, James Gholam, Leandro Beltrachini, Mara Cercignani, Derek Kenton Jones

Abstract

1## Introduction Diffusion-weighted imaging (DWI) is sensitive to the microscopic motion of water molecules within tissue, thus providing a unique contrast which complements T1 and T2-weighted MRI . As diffusion is anisotropic in brain white matter , typically three images with orthogonal diffusion directions are acquired and averaged to provide rotationally-invariant mean apparent diffusion coefficient (ADC) maps . Clinically, DWI has become the standard diagnostic scan for acute stroke and is increasingly acquired in other conditions such as multiple sclerosis, brain tumors, and dementia . G

A figure from Four Directions, One Solution: Enabling Rapid Diffusion Tensor MRI for Ultra‐Low Field Using Deep Learning — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

Deep learning enables accurate diffusion tensor MRI from just four directions, dramatically speeding scans especially at low and ultra‑low field strengths.

What’s new

i.Developed deep‑learning models to predict axial/radial diffusivities and the principal eigenvector from four tetrahedral diffusion measurements
ii.Showed the DL approach improves accuracy over the standard tetrahedral method under low‑SNR conditions in simulations and in‑vivo data at 3T and 64 mT
iii.Analyzed the ‘axis problem’ where tensors aligned with scanner axes cause errors and demonstrated the DL model reduces but does not eliminate this error
iv.Provided a multi‑layer perceptron architecture and training pipeline using synthetic tensors with uniformly distributed eigenvalues and orientations

How they did it

Synthetic training data spanning a wide range of diffusion tensors (uniform eigenvalues and orientations) were generated with Rician noise (SNR = 5). Two multi‑layer perceptron models (diffusivity and eigenvector) were trained in TensorFlow using the Adam optimizer (learning rate 0.001) for 16 epochs. The models were evaluated on digital phantoms and in‑vivo datasets acquired at 3 T and 64 mT.

What they found

DL model achieved lower FA estimation variance than the standard tetrahedral method across both fixed and rotated gradient conditions (Figure 9)
In Monte Carlo simulations (2000 iterations) across SNR levels, the DL model showed greater robustness to noise than the standard approach (Figure 3)
Using synthetic data with SNR = 5 (typical for 64 mT), the DL model outperformed the standard tetrahedral method in diffusivity and orientation accuracy
DL model reduced angular error of the principal eigenvector compared with the standard method (Figure 5)

Caveats

✕Residual errors remain when the principal eigenvector aligns with the scanner’s x, y, or z axes (axis problem)
✕Method is not yet suitable for advanced applications such as whole‑brain tractography
✕Performance degrades with participant motion that rotates the tissue relative to the gradient vectors

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

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

The paper, in five minutes

Ask a question about this paper

Similar papers in all four archives

Similar papers in all four archives

The paper, in five minutes