Streamline tractography of the fetal brain in utero with machine learning

TL;DR

The paper addresses the challenging problem of fetal brain tractography, where low dMRI quality and rapid developmental changes hinder accurate reconstruction. It introduces a transformer-augmented neural network that fuses diffusion orientation from tensor fits, streamline propagation history, global brain location cues, tissue segmentation, and a spatio-temporal fixel atlas to predict the next propagation direction. The approach is trained on millions of manually refined fetal tractograms and evaluated on 11 independent scans, achieving superior Dice, precision, and recall across nine tracts compared with conventional FACT and RNN baselines, and receiving favorable qualitative expert assessments. This work substantially advances in utero diffusion MRI analysis by enabling more accurate, reproducible tract reconstruction and enabling comprehensive structural connectivity analyses during fetal development.

Abstract

Diffusion-weighted magnetic resonance imaging (dMRI) is the only non-invasive tool for studying white matter tracts and structural connectivity of the brain. These assessments rely heavily on tractography techniques, which reconstruct virtual streamlines representing white matter fibers. Much effort has been devoted to improving tractography methodology for adult brains, while tractography of the fetal brain has been largely neglected. Fetal tractography faces unique difficulties due to low dMRI signal quality, immature and rapidly developing brain structures, and paucity of reference data. This work presents the first machine learning model for fetal tractography. The model input consists of five sources of information: (1) Fiber orientation, inferred from a diffusion tensor fit to the dMRI signal; (2) Directions of recent propagation steps; (3) Global spatial information, encoded as distances to keypoints in the brain cortex; (4) Tissue segmentation information; and (5) Prior information about the expected local fiber orientations supplied with an atlas. In order to mitigate the local tensor estimation error, a large spatial context around the current point in the diffusion tensor image is encoded using convolutional and attention neural network modules. Moreover, the diffusion tensor information at a hypothetical next point is included in the model input. Filtering rules based on anatomically constrained tractography are applied to prune implausible streamlines. We trained the model on manually-refined whole-brain fetal tractograms and validated the trained model on an independent set of 11 test scans with gestational ages between 23 and 36 weeks. Results show that our proposed method achieves superior performance across all evaluated tracts. The new method can significantly advance the capabilities of dMRI for studying normal and abnormal brain development in utero.

Figures (11)

• Figure 1: Example diffusion tensor and color fractional anisotropy images for fetal brains scanned at 24, 28, 32, and 36 gestational weeks.
• Figure 2: The proposed fetal tractography method. The method encodes the information in the 3D volume of diffusion orientation distribution using transformer and convolutional blocks, generating feature maps at three different scales. The spatial location of the current streamline propagation point is used to interpolate these feature maps. A similar procedure is followed to encode the information in the tissue segmentation map and the fixel atlas image registered to the subject, although using a light-weight fully-convolutional network. These are combined with prior streamline propagation directions and with "position vector" features that represent the global location of the current point in the brain. A set of fully-connected layers fuse these features to predict the next streamline propagation direction.
• Figure 3: Our proposed scheme for encoding the position of the current tractography propagation point in the brain. We encode this information as the normalized distance with respect to the centers of mass of five non-coplanar cortical parcellation regions. This figure shows the axial, sagittal, and coronal views depicting the cortical parcellation regions that are visible in the shown slices, spheres denoting the centers of mass of those regions, and 3D views that show the lines connecting these centers to an arbitrary point within the brain mask. The fetuses shown in this figure are 24 and 31 weeks of gestational age.
• Figure 4: Axial, coronal, and sagittal views of the spatio-temporal fixel atlas for 25, 30, and 35 gestational weeks. In order to better view the atlases for lower gestational ages, all atlases have been displayed to the same size.
• Figure 5: We launch a streamline from the center of each gray matter voxel that has at least one white matter voxel neighbor. The direction of the first step is selected to be the direction of the line connecting the center of the gray matter voxel to the center of the neighboring white matter voxel. In this figure, the red voxels are gray matter and the green voxels are white matter. We have selected three arbitrary gray matter voxels and have shown the direction of the first step for the streamlines launched from those seed points with yellow arrows. To enhance the tractogram diversity, random jittering is applied to the location of the seed point and the direction of the first step as explained in the text. These augmentations are not portrayed in this figure.