PaperlayerView exact revision on arXiv

Source-linked AI summary

Prior-Guided Implicit Neural Representations for Single-Subject Diffusion MRI Super-Resolution

Abdulkader Ghandoura, Marsil Zakour, William Consagra, Yogesh Rathi

arXiv:2609.00981v1eess.IVcs.CV

TL;DR

High-resolution dMRI is costly to acquire, while low-resolution scans make fine white-matter structure difficult to recover and existing INR methods can be slow or lack anatomical priors. The paper pre-trains an INR on a high-resolution template, registers it to each subject, and fine-tunes it on low-resolution data. On HCP data, the framework trains faster and achieves higher-quality reconstructions than competing INR-based methods at 4× through-plane super-resolution.

  • Problem

    Long acquisitions drive clinical protocols toward low-resolution dMRI, while existing INR methods can be slow and lack anatomical priors for constraining ill-posed super-resolution.

  • Method

    The framework pre-trains an INR on a high-resolution template, maps it to a new subject through deformable registration, and fine-tunes it for low-resolution data consistency.

  • Results

    The framework achieves faster training and higher-quality reconstructions than two INR-based baselines on HCP diffusion data, recovering fine anatomical structures at 4× through-plane super-resolution.

  • Takeaways & Limitations

    Template-guided transfer learning provides a subject-specific dMRI super-resolution approach that supports arbitrary target voxel sizes while incorporating anatomical priors.

Abstract

from arXiv · show

Resolving complex fiber geometries in brain white matter requires high-resolution diffusion MRI at the cost of long acquisition times. This leads many clinical protocols to opt for low-resolution scans, making downstream microstructure estimation and tractography challenging. Implicit neural representations (INRs) can model the diffusion signal continuously, enabling native single-subject super-resolution by querying the network at arbitrary spatial coordinates, yet existing methods often suffer from long training times and lack a mechanism to incorporate anatomical priors to regularize super-resolution by constraining the space of plausible reconstructions. To address these limitations, we propose a novel transfer-learning framework that pre-trains an INR on a high-resolution template and then adapts it to subject-specific scans via registration and fine-tuning. For $4\times$ through-plane super-resolution from 5 mm to 1.25 mm on Human Connectome Project (HCP) data, our method reduces NRMSE by 36-49% and increases FSIM by 24-43% over a recent baseline with $6\times$ faster training, outperforming competing INR-based methods across both image quality and domain-specific metrics. Code is available on the project page at https://abdulkaderghandoura.github.io/research/msc-thesis/ .

1 Introduction

Diffusion MRI provides information about brain microstructure and fiber pathways, but the long acquisition times required for high spatial resolution force many clinical protocols to use thick slices. The proposed INR framework incorporates a high-resolution anatomical prior through template pre-training and registration-guided fine-tuning, improving reconstruction quality and training speed.

  • Motivation: Diffusion MRI probes brain tissue structure in vivo and supports characterization of microstructure and large-scale white matter pathways.These capabilities motivate its use in studies of neurodegenerative, neuropsychiatric, and cognitive disorders.
  • Motivation: High-spatial-resolution dMRI requires long acquisitions, encouraging thick-slice clinical protocols that sacrifice through-plane resolution and downstream anatomical detail.The lost detail affects analyses including tractography and microstructure estimation.
  • Approach: The proposed framework reconstructs high-spatial-resolution dMRI signals from low-resolution clinical-grade scans using a continuous, orientation-consistent INR.Because the representation is continuous over the image domain, it supports super-resolution at arbitrary scales.
  • Approach: The method uses a high-resolution template as an anatomical prior through registration-guided fine-tuning, while template pre-training amortizes computational cost.This design estimates subject-specific INRs from low-resolution data and targets the ill-posed loss of detail from very thick slices.
  • Results: 6× faster training and superior perceptual quality are reported against recent INR approaches for 4× through-plane super-resolution.The comparison is made on the stated super-resolution setting.

2 Related Work

INRs represent continuous fields and naturally support arbitrary-coordinate super-resolution, but dMRI super-resolution methods have largely focused on parameter maps, angular resolution, or supervised settings. The paper positions its template-guided approach within this gap by using continuous spatial representations and anatomical priors for subject-specific reconstruction.

  • Implicit Neural Representations: INRs parameterize continuous functions that map spatial or spatiotemporal coordinates to field values, with limited prior adoption in dMRI.Existing dMRI examples include SIREN-based signal modeling and multi-shell extensions.
  • Proposed Framework: The proposed framework combines template pre-training, subject-template registration, and low-resolution subject fine-tuning.Figure 1 presents these as three sequential stages.
  • Diffusion MRI Super Resolution: dMRI is represented as a 4D spatial-and-diffusion array, so super-resolution can target spatial resolution, angular resolution, or both.The diffusion volumes correspond to gradient directions and diffusion-weighting values.
  • Diffusion MRI Super Resolution: Most prior work targets spatially super-resolved parameter maps or angularly super-resolved raw signals, whereas arbitrary-scale spatial signal super-resolution benefits downstream tractography.Continuous spatial querying permits streamline propagation at a chosen resolution.
  • Diffusion MRI Super Resolution: Existing INR approaches for spatial dMRI super-resolution have been limited to supervised encoder-decoder settings, which may be problematic under clinical pathology-related distribution shifts.The proposed framework addresses this motivation with a template-based anatomical prior.

3 Methodology

The method combines a continuous spherical-harmonic INR with anatomical registration and transfer learning for single-subject diffusion MRI super-resolution. A high-resolution template prior is adapted to low-resolution subject data through deformation-aware fine-tuning and regularization.

  • Prior-guided transfer learning: The framework fits an INR to a high-resolution dMRI scan, registers subject coordinates to the template, and fine-tunes the prior on low-resolution subject data.The three stages are template-space pretraining, non-rigid registration, and subject-specific adaptation.
  • Angular representation: Spherical harmonics provide the angular basis, restricted to even orders because diffusion signals have antipodal symmetry.The basis functions are real-symmetric and relate the orientation distribution function to the diffusion signal through the Funk-Radon transform.
  • Implicit neural representation: The NODF model represents diffusion continuously over spatial location v ∈ R3 and orientation p ∈ S2 using spherical-harmonic coefficients parameterized by a SIREN-based INR.This representation supports evaluating the diffusion signal at arbitrary spatial locations and orientations.
  • Angular regularization: Training combines voxelwise data fidelity with an angular smoothness prior whose strength is controlled by λc and whose precision is encoded by Rγ.The angular prior uses the spectral density of a spherical Matérn prior and harmonic order.
  • Subject-specific fine-tuning: The registered template prior approximates subject-voxel downsampling by Monte Carlo integration over transformed voxel footprints, while fine-tuning uses Adam and spatial total-variation regularization.Uniform slice averaging is the experimental special case, and arbitrary-scale spatio-angular super-resolution is obtained from the fine-tuned network.

Experimental Setup

The study uses HCP diffusion data to evaluate single-subject super-resolution against two INR-based baselines using image-quality, microstructural, orientational, and tractography metrics.

  • Data: The HCP-YA data contain 1.253 mm isotropic voxels, three diffusion shells, 90 directions per shell, and 18 non-diffusion-weighted images.Experiments use only the b = 1,000 s/mm2 shell, with one subject training the template INR and 10 randomly selected subjects used for testing.
  • Competing Methods: The proposed method is compared with NODF and NODF-HashEnc, which support arbitrary-scale super-resolution through continuous spatial querying.Supervised CNN methods are excluded because arbitrary-scale dMRI upsampling remains essential for tractography at chosen resolutions.
  • Evaluation Metrics: Evaluation measures FA and GFA image quality with NRMSE, FSIM, PSNR, and SSIM, plus ODF-L2 and principal-eigenvector angular errors for orientation content.Tractography from super-resolved fODFs is also compared qualitatively with ground-truth tractography.
  • Implementation Details: The template-space INR uses a 10-layer SIREN with width 1,024, batch size 128, learning rate 10^-6, and subject fine-tuning with total variation regularization.All models are trained on an NVIDIA GeForce GTX 1080 Ti GPU with 11 GB VRAM.
  • Evaluation Protocol: Table 1 reports quantitative results across training times on 10 randomly selected HCP subjects using mean ± standard deviation.The table emphasizes NRMSE, FSIM, PSNR, and SSIM comparisons across methods and training budgets.

5 Results

The proposed framework improves quantitative and qualitative dMRI super-resolution over INR baselines, including on clinical data, while ablations show benefits from fine-tuning and explicit thick-slice modeling.

  • 5.1 Quantitative and Qualitative Evaluation: ODF L2 norm error is 0.08–0.09 for the proposed method versus 0.11–0.23 for NODF and 0.12–0.13 for HashEnc.GFA absolute error is 0.023–0.026 for the proposed method versus 0.04–0.08 for the baselines, while angular error is 17–20° versus 28–47° for NODF and 19–52° for HashEnc.
  • 5.1 Quantitative and Qualitative Evaluation: The proposed method significantly outperforms both baselines on all seven primary metrics at every training budget (α = 0.05).FA PSNR and FA SSIM are likewise statistically significant across the 10 subjects.
  • 5.1 Quantitative and Qualitative Evaluation: After 5 minutes of fine-tuning, the proposed method produces DTI, fODF, and tractography reconstructions closer to ground truth than baselines trained for 1 hour.NODF shows blur and NODF-HashEnc shows blocking artifacts, whereas the proposed method better preserves fine structures and pathways.
  • 5.2 Generalization to Clinical Data: The clinical evaluation uses a 0.9375×0.9375×6 mm acquisition with a non-integer 4.8× through-plane upsampling factor and realistic acquisition noise.Because isotropic ground truth is unavailable for real clinical scans, this evaluation is qualitative.
  • 5.2 Generalization to Clinical Data: Fine-tuning recovers the right-anterior lesion in the clinical case, whereas the registered template without fine-tuning omits it.The reconstruction is driven by subject data consistency, with the healthy template serving as initialization and regularization; this remains a single qualitative case.
  • 5.3 Ablation Study: Removing the downsampling operator increases GFA NRMSE by up to 8.6% and decreases FSIM by up to 4.5% at 30 minutes.This supports explicitly modeling each thick slice’s spatial extent during reconstruction.

6 Discussion

The discussion attributes the method’s advantage under extreme anisotropy to globally coherent spatial representations combined with an anatomical prior. It also argues that domain-specific diffusion MRI metrics better reflect reconstruction quality than conventional image metrics.

  • Global spatial coherence and the proposed prior enable effective super-resolution under extreme anisotropy.NODF’s SIREN basis captures spatial correlations across the full volume, unlike locally interpolated hash-grid embeddings.
  • Hash-grid encoding limits recovery of missing through-plane details across wide interslice gaps.Off-grid interpolation lacks global spatial relationships, diminishing anisotropy and directional information and degrading tractography.
  • PSNR and SSIM can favor oversmoothed reconstructions and fluctuate during training despite poorer perceptual quality.These natural-image metrics systematically favor blur and correlate poorly with expert radiologist assessments.
  • The proposed method produces sharper, more anatomically detailed outputs than NODF after one hour of training.Qualitative inspection finds NODF remains blurry at that point, whereas the proposed reconstruction preserves more detail.
  • On ODF reconstruction error, eigenvector angular error, and tractography fidelity, the method outperforms both baselines across all training times.

7 Conclusion

The paper concludes that template-guided transfer learning provides a fast, high-quality INR framework for diffusion MRI super-resolution. On HCP data, it recovers fine anatomical structures at 4× through-plane super-resolution that neither baseline resolves.

  • The framework pre-trains an INR on a high-resolution template, maps it to new subjects through deformable registration, and fine-tunes subject-specific scans.
  • The method achieves significantly faster training and higher-quality reconstructions than two recent INR-based methods on HCP diffusion data.
  • At 4× through-plane super-resolution, the framework recovers fine anatomical structures that neither baseline resolves.
  • Future work includes dMRI-specific registration, multiple b-values, angular super-resolution, measured slice-excitation profiles, and validation on pathological cohorts.
Loading 2609.00981v1…