FOD Estimation with Spherical CNNs

Fiber Orientation Distribution (FOD) Estimation enables accurate white matter tractography and microstructural characterization. The fedi_dmri_fod tool implements a rotationally equivariant Spherical Convolutional Neural Network (sCNN) framework optimized for neonatal dMRI data.

For full methodological and validation details, refer to:

Snoussi and Karimi, 2025 – Equivariant spherical CNNs for accurate fiber orientation distribution estimation in neonatal diffusion MRI with reduced acquisition time

Overview

The fedi_dmri_fod tool accurately estimates FODs from data acquired with reduced gradient directions (30% of full protocol), enabling faster acquisitions while maintaining high-quality results.

Key Features

  • Reduced Acquisition Time: Accurate FOD estimation using only 30% of the full gradient direction protocol

  • Rotational Equivariance: Respects the rotational symmetry of diffusion signals

  • Neonatal-Optimized: Trained on 43 neonatal dMRI datasets from the Developing Human Connectome Project (dHCP)

  • Pretrained Model: Automatically downloaded from Hugging Face Hub

  • Multi-Shell Support: Works with multi-shell dMRI data (b-values: 400, 1000, 2600 s/mm²)

Methodology

The tool converts multi-shell dMRI signals to spherical harmonic (SH) coefficients, applies rotationally equivariant convolutions, and outputs SH coefficients representing the FOD.

Performance and Validation

The sCNN significantly outperforms Multi-Layer Perceptron (MLP) and produces FODs and tractography that are quantitatively comparable and qualitatively highly similar to Hybrid-CSD ground truth, despite using only 30% of the full acquisition data.

Results and Visualizations

../_images/Figure1.jpg

Figure 1. Representative FODs from a test subject. (Left column) FODs estimated by the MLP using the full dHCP dataset. (Middle column) FODs estimated by the sCNN using 30% of the diffusion directions. (Right column) Ground truth FODs estimated using Hybrid-CSD with the full dHCP dataset. The sCNN produces FODs that are visually much more similar to the ground truth than the MLP.

../_images/Figure2.jpg

Figure 2. Representative tractography results. (Left) Tractogram generated using MLP-predicted FODs. (Middle) Tractogram generated using sCNN-predicted FODs. (Right) Tractogram generated using ground truth FODs (Hybrid-CSD).

Accessing the FOD Estimation Code

The FOD estimation functionality is implemented in fedi_dmri_fod (see fedi_dmri_fod), with the model in FEDI/models/pytorch/models.py (SphericalCNN_FOD_Neonatal class). The pretrained model is automatically downloaded from Hugging Face Hub (feditoolbox/scnn_neonatal_fod_estimation).

Integration with HAITCH Pipeline

The fedi_dmri_fod tool can be used as part of the HAITCH pipeline (STEP 10) for FOD estimation after motion and distortion correction. For more information, see haitch_starts.