Equivariant Spherical CNNs for Accurate Fiber Orientation Distribution Estimation in Neonatal Diffusion MRI with Reduced Acquisition Time

TL;DR

A rotationally equivariant Spherical Convolutional Neural Network framework tailored for neonatal dMRI, which significantly outperforms a Multi-Layer Perceptron (MLP) baseline across multiple quantitative metrics, including Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Angular Correlation Coefficient (ACC), angular error, and peak match rate, indicating superior FOD estimation accuracy.

Abstract

Early and accurate assessment of brain microstructure using diffusion Magnetic Resonance Imaging (dMRI) is crucial for identifying neurodevelopmental disorders in neonates, but remains challenging due to low signal-to-noise ratio (SNR), motion artifacts, and ongoing myelination. In this study, we propose a rotationally equivariant Spherical Convolutional Neural Network (sCNN) framework tailored for neonatal dMRI. We predict the Fiber Orientation Distribution (FOD) from multi-shell dMRI signals acquired with a reduced set of gradient directions (30% of the full protocol), enabling faster and more cost-effective acquisitions. We train and evaluate the performance of our sCNN using real data from 43 neonatal dMRI datasets provided by the Developing Human Connectome Project (dHCP). Our results demonstrate that the sCNN achieves significantly lower mean squared error (MSE) and higher angular correlation coefficient (ACC) compared to a Multi-Layer Perceptron (MLP) baseline, indicating improved accuracy in FOD estimation. Furthermore, tractography results based on the sCNN-predicted FODs show improved anatomical plausibility, coverage, and coherence compared to those from the MLP. These findings highlight that sCNNs, with their inherent rotational equivariance, offer a promising approach for accurate and clinically efficient dMRI analysis, paving the way for improved diagnostic capabilities and characterization of early brain development.

Figures (6)

• Figure 1: Distribution of the postmenstrual ages for the 43 neonatal dMRI datasets included in the study.
• Figure 2: Sagittal, axial, and coronal views of representative examples data from neonatal dMRI in the dHCP dataset.
• Figure 3: Flowchart illustrates the entire data processing and analysis pipeline, including the use of neonatal dMRI datasets, FOD estimation, data simulations, the sCNN architecture, and the outputs of the sCNN.
• Figure 4: 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 MSMT-CSD with the full dHCP dataset. The sCNN produces FODs that are visually much more similar to the ground truth than the MLP.
• Figure 5: Zoomed-in views of regions of interest (ROIs) with complex fiber configurations, highlighting differences between FODs predicted by MLP, sCNN, and MSMT-CSD (ground truth). The sCNN preserves anatomical structure and closely resembles the ground truth, whereas the MLP exhibits increased noise and reduced structural clarity. These ROIs correspond to those shown in Figure \ref{['FOD_full']}.