Enabling the Assessment of Fetal Brain Development and Degeneration with Machine Learning - Project Summary
Diffusion-weighted magnetic resonance imaging (dMRI) is the most promising tool for studying brain
microstructure. However, the application of dMRI to the assessment of fetal brain in-utero is challenged by
unpredictable motion, low signal-to-noise ratio, low spatial resolution, and imaging artifacts. While much effort
has been spent on improving image acquisition and motion compensation techniques, data processing and
analysis methods have remained largely unchanged. Existing biomarker estimation methods in fetal dMRI suffer
from low accuracy and low reproducibility. Moreover, cross-subject and population studies require delineation of
white matter (WM) tracts, which currently can only be performed via highly subjective and time-consuming
manual segmentation. These shortcomings have significantly limited our ability to study the brain at this critical
stage and to detect subtle changes in brain microstructure due to disorders. This proposed project will develop
and validate a new generation of methods for analysis of fetal dMRI data. Unlike existing methods, which are
based on biophysical models of the diffusion signal and mathematical model fitting, the new methods will rely on
data-driven and machine learning techniques. Building on our pioneering works that have shown the potential of
these methods, we will develop deep learning techniques for estimating microstructural biomarkers such as
fractional anisotropy, neurite orientation dispersion, and fiber orientation distribution. The new methods will be
based on two-stage transformer networks, which will be trained using dMRI data from preterm infants and
fetuses. Moreover, we will develop methods that work with undersampled scans and provide a calibrated
measure of estimation uncertainty. We will develop convolutional neural networks to segment WM tracts in the
fetal brain based on the local fiber orientations. To address the noise in the input and target labels, we will build
on our prior works on segmentation with noisy data and labels, shape-aware segmentation, and use of
uncertainty to improve segmentation accuracy. The new technique will generate tracts automatically, with tracts
that are indistinguishable from those created by the best human experts. We will evaluate the new methods
using test-retest and bootstrapping methods and via assessment by experts in fetal brain microstructure and
with histological knowledge of transient fetal fiber pathways. The new methods will enable assessment of fetal
brain microstructure and the impact of neurodevelopmental disorders on tract-specific microstructure with a level
of accuracy, detail, and reproducibility that is currently beyond reach. To definitively demonstrate the value and
significance of the new methods, we will use them to assess the alterations in WM micro-structure due to
congenital heart disease (CHD), which is the most common birth defect. In the process, we will produce the most
comprehensive and detailed picture of the impact of CHD on the fetal brain microstructure ever attempted.
