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

TL;DR

This work tackles the challenge of diffusion imaging with portable ultra-low-field MRI by introducing a nine-direction, single-shell ULF DTI sequence and two novel post-processing tools: a Bayesian, angularly aware bias-field correction (Beta-DSW) and a joint spatio-angular superresolution method (DiffSR). DiffSR operates in spherical-harmonic space with icosahedral projections and graph-based angular modeling, leveraging aggressive SH-space augmentations and a U-Net with global attention to restore diffusion contrast and directionality under constrained SNR and angular sampling. Across synthetic degradation tests, matched HF-UFL comparisons, and ADNI-based tract analyses, the approach improves FA/ADC accuracy, V1 coherence, and tractography fidelity, and enhances the detection of AD-related white-matter changes, demonstrating potential clinical utility for portable DTI and sequence harmonization. The work also provides code for DiffSR, supports cross-site applicability, and discusses limitations such as b-value mismatches and the need for multi-shell generalization to further close the HF-ULF gap.

Abstract

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 $\href{https://github.com/markolchanyi/DiffSR}{public \space use}$.

Figures (10)

• Figure 1: Direction-specific bias fields encountered in ULF DTI. Application of our nine-direction ULF DTI sequence to a spherical 40$\%$ Polyvinylpyrrolidone phantom displays smooth intensity fields that vary with diffusion encoding gradient direction. The top row displays the forward (i.e., standard) ULF DTI acquisition with diffusion weighting (excluding the interleaved low-b acquisitions), while the bottom row displays the same ULF DTI sequence but with diffusion encoding gradient directions acquired in reverse order.
• Figure 1: DiffSR reconstruction accuracy under synthetic spatial and angular downsampling for Connectom HCP data. DiffSR reconstruction accuracy is shown with respect to standard trilinear upsampling for the b=3000$\frac{s}{mm^2}$ (top) and b=5000$\frac{s}{mm^2}$ (bottom) shells for synthetically downsampled HF DTI data from the MGH HCP dataset. The raw (i.e., gradient directions and low-b’s) HF DTI data were spatially downsampled with trilinear interpolation between 2mm and 4mm isotropic spatial resolutions at 0.25mm intervals. The data was also angularly downsampled by choosing random gradient direction subsets at ratios of 0.15 to 0.55 with respect to the original gradient number in the respective shell. Of note, the b=1000$\frac{s}{mm^2}$ and b=3000$\frac{s}{mm^2}$ shells contain 64 diffusion encoding gradient directions, while the b=5000$\frac{s}{mm^2}$ shell contains 128 diffusion encoding gradient directions, leading to the same gradient subset ratio containing twice as many diffusion encoding gradient directions in the b=5000$\frac{s}{mm^2}$ shell as compared to other shells. Shown are the MAE and LNCC for the SH coefficient channels (channel-wise average), FA and ADC reconstructions. Also shown is the voxel-wise mean angular error for the V1 reconstructions.
• Figure 2: Beta and DSW parametric priors for bias correction. Shown in the left two columns are the HCP priors for beta distribution coefficients separated by tissue class: white matter, grey matter and cerebrospinal fluid. Shown in the right column are the directional ($\mathbf{v1}_{\mu}$) and attenuation ($\kappa$) priors for the DSW distribution, as well as the hard tissue segmentations in MNI space used for $\alpha_L$ and $\beta_L$ tissue class separation.
• Figure 2: Per-tract unnormalized fractional anisotropy and apparent diffusion coefficient measurements across ULF DTI reconstruction variants. Shown (A) are non-z-scored (i.e., un-normalized) distributions of tract-averaged FA (top) and ADC (bottom) for our 18-subject cohort with matched conventional HF DTI and ULF DTI sequences across a subset of white matter tracts segmented with Tracula. The native ULF DTI sequence with standard preprocessing (orange) is compared with Beta-DSW bias-corrected ULF DTI without (green) and with superresolution with DiffSR (purple) in terms of overall agreement with matched HF DTI measurements (blue).
• Figure 3: Overview of the DiffSR forward pipeline. The pipeline consists of an icosahedral projection of the spherical harmonic coefficients derived from the raw DTI signal, followed by two layers of graph-weighted convolutions to capture proximal angular dependence. The filtered projected amplitudes are then re-projected back to spherical harmonic space via matrix inversion, which we call a graph-weighted convolution block. The transformed spherical harmonic coefficients are then propagated through a U-Net CNN with an eight global token attention block at the bottleneck. Finally, the propagated SH coefficients are filtered through a second graph-weighted convolution block to generate the superresolved DTI volume.