Learning Spatially-Continuous Fiber Orientation Functions

Tyler A. Spears, P. Thomas Fletcher

Abstract

Mapping the human connectome relies on diffusion MRI (dMRI) to trace neural pathways. However, dMRI is often limited by low resolution and noise. Traditional tractography algorithms estimate fiber directions between discrete grid points using simple interpolation, which can miss fine-scale details. We present FENRI (Fiber orientations from Explicit Neural RepresentatIons), a deep learning model that generates continuous, high-resolution maps of brain fiber orientations. By treating the brain's fiber architecture as a continuous function, FENRI enables the reconstruction of the connectome with unprecedented detail and spatial smoothness.

Introduction

The human connectome is a complex network of neural pathways. Diffusion MRI (dMRI) measures the movement of water molecules to reveal the direction of nerve fibers. A critical step in analyzing this data is tractography—tracing these pathways through the brain. Tractography requires knowing the fiber direction at every point along a curve. Because MRI data is captured on a discrete 3D grid (voxels), researchers typically rely on trilinear interpolation to estimate directions between grid points. This method is often insufficient for capturing fine-scale changes, especially in noisy or low-resolution data.

Methodology: The FENRI Architecture

FENRI moves beyond traditional fixed-grid upsampling by modeling the brain's fiber architecture as a continuous field. The architecture consists of several key components:

  1. Encoder: A 3D Cascaded Residual Network (CARN) compresses low-resolution input data into a latent feature space, capturing essential geometric information.
  2. Continuous Decoder: Instead of a standard upsampling layer, FENRI uses a multi-layer perceptron (MLP) that takes a specific 3D coordinate as input and predicts the fiber Orientation Density Function (fODF) at that exact point. The fODF is represented by Spherical Harmonic (SH) coefficients.
  3. Local Ensemble: To ensure spatial continuity and prevent "jumps" between grid cells, the model queries features from the eight nearest neighbors of a target point and blends them.
  4. Positional Encoding: Utilizing Fourier positional encoding (similar to NeRF), FENRI learns high-frequency spatial details that are typically lost during standard interpolation.

Experimental Results

The model was evaluated using the HCP Young Adult dataset and a synthetic benchmark, ISMRM-sim.

Quantitative Performance

FENRI was compared against standard trilinear interpolation and a fixed-grid deep learning model (Fixed-Net). Across multiple metrics—Weighted Mean Squared Error (WMSE), Jensen-Shannon Distance (MSJSD), and Weighted Average Angular Error (WAAE)—FENRI consistently outperformed the baselines.

Quantitative comparison of FENRI against baselines across three major error metrics: WMSE, MSJSD, and WAAE. FENRI (top row, blue) shows significantly lower error distributions.

Qualitative Analysis

Visual inspection confirms that FENRI produces more anatomically plausible reconstructions. In complex regions of the brain, FENRI recovers fine-scale structures and fiber orientations that are blurred or lost by other methods.

Qualitative comparison of fiber orientation reconstruction. FENRI's outputs (second column) closely match the Ground Truth (first column) and provide sharper, more accurate details than Fixed-Net or Trilinear interpolation.

Discussion and Conclusion

FENRI demonstrates that spatial continuity is a vital property for high-fidelity tractography. By breaking the "grid bias" of traditional upsamplers, FENRI allows for arbitrary scaling and produces smoother, more accurate streamlines. While standard interpolation remains competitive for large, simple fiber bundles, FENRI's learned priors are particularly effective in resolving ambiguous or fine-scale neural structures. Future work will explore integrating FENRI into probabilistic tractography frameworks to further enhance connectome mapping in clinical settings.

Summary updated

BibTeX
@inproceedings{spears2024learning,
  author = {Tyler A. Spears and P. Thomas Fletcher},
  title = {{Learning Spatially-Continuous Fiber Orientation Functions}},
  booktitle = {ISBI},
  year = {2024},
  doi = {10.1109/isbi56570.2024.10635838},
}
Figure