Ground-truth effects in learning-based fiber orientation distribution estimation in neonatal brains

TL;DR

This study investigates how ground-truth choice affects learning-based FOD estimation in neonatal brains by comparing MSMT-CSD and SS3T-CSD as GTs. It trains a U-Net that maps SH representations of diffusion signals to FOD SH coefficients, using GTs derived from both methods, and assesses GT-consistency, input-direction ablations, and age-domain shifts. Results show SS3T-CSD improves crossing-fiber detection and overall FOD accuracy, with robust performance across age groups, especially when more $b_{1000}$ input measurements are used. These findings indicate SS3T-CSD–based GT can enable more accurate and generalizable neonatal white matter mapping and tractography.

Abstract

Diffusion Magnetic Resonance Imaging (dMRI) is a non-invasive method for depicting brain microstructure in vivo. Fiber orientation distributions (FODs) are mathematical representations extensively used to map white matter fiber configurations. Recently, FOD estimation with deep neural networks has seen growing success, in particular, those of neonates estimated with fewer diffusion measurements. These methods are mostly trained on target FODs reconstructed with multi-shell multi-tissue constrained spherical deconvolution (MSMT-CSD), which might not be the ideal ground truth for developing brains. Here, we investigate this hypothesis by training a state-of-the-art model based on the U-Net architecture on both MSMT-CSD and single-shell three-tissue constrained spherical deconvolution (SS3T-CSD). Our results suggest that SS3T-CSD might be more suited for neonatal brains, given that the ratio between single and multiple fiber-estimated voxels with SS3T-CSD is more realistic compared to MSMT-CSD. Additionally, increasing the number of input gradient directions significantly improves performance with SS3T-CSD over MSMT-CSD. Finally, in an age domain-shift setting, SS3T-CSD maintains robust performance across age groups, indicating its potential for more accurate neonatal brain imaging.

Figures (4)

• Figure 1: Workflow overview. Single-shell b1000 dMRI images are sampled from the full series, normalized by b0 and projected into SH space. These images are inputs to a deep learning model predicting GT FODs, using either MSMT-CSD (300 meas.) or SS3T-CSD (108 meas.). The experiments assess: (i) consistency of GT algorithms, (ii) impact of input quantity on model performance, and (iii) model effectiveness across different neonatal developmental stages. Illustrative brain images are from dHCP.
• Figure 2: Qualitative comparison of coronal slices of the FODs used as GT for the model, reconstructed with MSMT-CSD and SS3T-CSD, respectively, from the dMRI scan of a subject at 40 weeks PMA. Both gray matter and white matter compartments are displayed on the fractional anisotropy (FA) image computed from dMRI data. FOD estimation and visualization were performed with MRtrix tournier2012mrtrix.
• Figure 3: Comparison of performance of the models based on incrementing numbers of input directions, using GT: (a) MSMT-CSD; (b) SS3T-CSD. AR and AE under different fiber number configurations and AFD Error are depicted.
• Figure 4: Comparison of performance metrics for FOD estimation models trained and tested on early and late developmental stages using MSMT-CSD and SS3T-CSD ground truths. The agreement rate, angular error, and AFD error are depicted across models trained on early ($M_{\text{early}}$) and late ($M_{\text{late}}$) age groups.