Estimating Neural Orientation Distribution Fields on High Resolution Diffusion MRI Scans

Mohammed Munzer Dwedari, William Consagra, Philip Müller, Özgün Turgut, Daniel Rueckert, Yogesh Rathi

Introduction

Diffusion MRI (dMRI) is a critical tool for mapping brain connectivity by estimating the Orientation Distribution Function (ODF), which characterizes the directional diffusion of water molecules along nerve fibers. While Implicit Neural Representations (INRs) have recently emerged as a powerful continuous modeling approach for ODFs, existing methods—such as those based on SIREN architectures—face significant challenges when applied to ultra-high-resolution scans. These global coordinate-based networks often suffer from the "curse of scale," leading to excessive training times, high computational costs for inference, and a tendency to over-smooth fine structural details, blurring the very features high-resolution scans are intended to capture.

Methodology: The HashEnc Framework

To overcome these limitations, we introduce HashEnc, a framework that replaces global Multi-Layer Perceptrons (MLPs) with a multi-resolution grid-hash encoding strategy. This approach, inspired by recent advances in computer graphics, shifts the learning burden from a single large network to a set of local feature embeddings.

Overview of the HashEnc architecture, featuring multi-resolution hash-encoding of input coordinates followed by ODF coefficient prediction via a lightweight MLP.

As shown in the architecture overview, HashEnc utilizes nested grids at varying resolutions to store local features. These coordinates are hashed into fixed-size lookup tables to maintain a low memory footprint. For any 3D coordinate $v$, the system retrieves and interpolates local embeddings, which are then processed by a compact MLP head to predict ODF coefficients $c(v)$. This hybrid architecture allows for high-frequency spatial information to be captured locally, preventing the blurring associated with global coordinate-based models. Furthermore, the framework incorporates a Bayesian linear layer ($W$) for uncertainty quantification, enabling the estimation of confidence intervals for fiber orientations across whole-brain volumes.

Experimental Setup

The framework was evaluated using the Wang et al. Connectome Dataset, which provides ultra-high-resolution (760 µm³ isotropic) human brain dMRI scans. To establish a high-signal-to-noise ratio (SNR) reference, a 6-session average (420 total gradient directions) was used as the ground truth. The models were trained on single-session data (70 gradient directions) to assess their performance in recovering complex structures from noisy, low-dose signals.

Results and Discussion

HashEnc demonstrates significant improvements over state-of-the-art SIREN-based models in both efficiency and reconstruction quality.

Performance and Efficiency

Quantitatively, HashEnc achieved a 10% enhancement in the Feature Similarity Index (FSIM) for General Fractional Anisotropy (GFA) maps. The computational gains are even more pronounced: training is 3x faster, and whole-brain inference is reduced from 173 seconds (SIREN) to just 4 seconds on a CPU. This efficiency stems from the model's parameter sparsity; during inference for a single voxel, HashEnc utilizes fewer than 13,000 parameters—less than 0.1% of its total capacity.

Convergence comparison between SIREN and HashEnc. HashEnc achieves superior detail in 13 minutes compared to SIREN after 24 hours of training.

Qualitative Analysis

Visual comparisons highlight HashEnc's ability to preserve sharp edges and intricate textures, particularly in complex regions like the cerebellum and white-to-gray matter transitions. While SIREN models often require thousands of epochs and many hours to resolve basic structures, HashEnc recovers thin fiber tracts and high-contrast boundaries within minutes, as demonstrated in the temporal progression of training results.

Conclusion and Future Work

HashEnc provides a scalable and efficient solution for modeling neural orientation fields in high-resolution dMRI. By leveraging local embeddings, it avoids the over-smoothing typical of global INRs while drastically reducing computational overhead. Future research will focus on relaxing the assumption of spatially constant noise and exploring adaptive grid resolutions to further optimize the representation of complex cortical regions.

Summary updated

BibTeX
@inproceedings{dwedari2024estimating,
  author = {Mohammed Munzer Dwedari and William Consagra and Philip Müller and Özgün Turgut and Daniel Rueckert and Yogesh Rathi},
  title = {{Estimating Neural Orientation Distribution Fields on High Resolution Diffusion MRI Scans}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2024},
  doi = {10.1007/978-3-031-72104-5\_30},
}

Loading related content...

Figure