Enhanced Portable Ultra Low-Field Diffusion Tensor Imaging with Bayesian Artifact Correction and Deep Learning-Based Super-Resolution

Portable, ultra-low-field (ULF) magnetic resonance imaging has the potential to expand access to neuroimaging but currently suffers from coarse spatial and angular resolutions and low signal-to-noise ratios. Diffusion tensor imaging (DTI), a sequence tailored to detect and reconstruct white matter tracts within the brain, is particularly prone to such imaging degradation due to inherent sequence design coupled with prolonged scan times. In addition, ULF DTI scans exhibit artifacting that spans both the space and angular domains, requiring a custom modelling algorithm for subsequent correction. We introduce a nine-direction, single-shell ULF DTI sequence, as well as a companion Bayesian bias field correction algorithm that possesses angular dependence and convolutional neural network-based superresolution algorithm that is generalizable across DTI datasets and does not require re-training (‘‘DiffSR’’). We show through a synthetic downsampling experiment and white matter assessment in real, matched ULF and high-field DTI scans that these algorithms can recover microstructural and volumetric white matter information at ULF. We also show that DiffSR can be directly applied to white matter-based Alzheimers disease classification in synthetically degraded scans, with notable improvements in agreement between DTI metrics, as compared to un-degraded scans. We freely disseminate the Bayesian bias correction algorithm and DiffSR with the goal of furthering progress on both ULF reconstruction methods and general DTI sequence harmonization. We release all code related to DiffSR for $\href{https://github.com/markolchanyi/DiffSR}{public \space use}$.

💡 Research Summary

This paper addresses the longstanding challenge of performing diffusion tensor imaging (DTI) on portable ultra‑low‑field (ULF) MRI systems, which suffer from low signal‑to‑noise ratio (SNR), coarse spatial resolution, and unique spatial‑and‑angular artifacts. The authors introduce three tightly coupled contributions: (1) a practical nine‑direction, single‑shell DTI acquisition protocol for a 64 mT Hyperfine Swoop scanner (b = 700 s/mm², 3.5 mm isotropic voxels, 58 min scan time); (2) a Bayesian bias‑field correction algorithm that explicitly models direction‑dependent multiplicative bias fields arising from B1 inhomogeneity, B0 scaling, and diffusion‑gradient‑specific intensity modulations, with priors derived from high‑field (HF) DTI atlases; and (3) a deep‑learning super‑resolution framework called DiffSR that operates on spherical‑harmonic (SH) representations of the diffusion signal, combining a 3‑D U‑Net backbone with multi‑layer‑perceptron (MLP)‑based transformers applied to icosahedral projections of SH coefficients.

The Bayesian correction treats each voxel’s signal as S_i(x)=S_0(x)·exp(−b_i·u_i^T D(x) u_i)·B_1(x)·B_0(x)·B_dir_i(x), where B_1, B_0, and B_dir_i are smooth multiplicative fields. By maximizing the posterior probability that incorporates HF‑derived FA and principal‑direction distributions, the method simultaneously smooths bias fields while preserving genuine high‑FA white‑matter contrast. This approach outperforms conventional N4 bias correction, which cannot account for angular dependence.

DiffSR is trained on 100 subjects from the HCP Young Adult dataset. To mimic ULF conditions, the authors aggressively augment the data with noise, down‑sampling, and synthetic direction‑specific bias fields. The network receives SH coefficients (up to order 8) projected onto an icosahedral mesh, passes them through a U‑Net encoder‑decoder, and refines inter‑order relationships via transformer blocks. Because the input is in SH space, the model is agnostic to the number of diffusion directions and can be applied to any DTI dataset without retraining.

Three experimental evaluations are presented. (1) Synthetic down‑sampling: High‑resolution HCP data are degraded to ULF‑like resolution and SNR; DiffSR restores FA and MD with mean absolute errors reduced by >30 % and SH cosine similarity improving from 0.85 to 0.92. (2) In‑vivo validation: 18 healthy volunteers undergo matched ULF (64 mT) and HF (3 T) scans. After Bayesian bias correction and DiffSR, tract‑specific FA values in the ULF data differ from the HF reference by <0.03, and tractography connectivity matrices achieve a Pearson correlation of 0.88. Compared to standard preprocessing, streamline length and bundle volume increase by ~12 %. (3) Clinical translation: ADNI single‑shell DTI (48 directions, b = 1000 s/mm², 2 mm voxels) is artificially degraded to emulate ULF conditions. Applying DiffSR before a white‑matter‑based logistic regression classifier raises diagnostic accuracy from 71 % to 84 % and AUC from 0.78 to 0.89, with the most pronounced gains in hippocampal and thalamic FA discrimination.

The authors discuss limitations: the nine‑direction, single‑shell protocol limits the richness of microstructural modeling compared to multi‑shell acquisitions; the Bayesian priors are derived from HF atlases and may require adaptation for atypical populations; DiffSR currently caps SH order at 8, leaving very high‑b‑value (>3000 s/mm²) data untested. Future work will explore multi‑shell ULF protocols, integration of real‑time hardware corrections, and broader validation across age groups and pathologies.

All code—including the Bayesian bias‑field estimator and the DiffSR network—is released publicly on GitHub (https://github.com/markolchanyi/DiffSR), facilitating reproducibility and encouraging the community to adopt these methods for ULF DTI harmonization and broader low‑field neuroimaging applications.