Advances in Fetal Brain Imaging

INTRODUCTION

Sonography has been widely used to assess fetal growth and identify fetal anomalies ¹, and although fetal neuro sonography can analyze brain surface anatomy and maturation², it has limitations in measuring small brain substructures and outlining the transient fetal brain compartments. Conversely, fetal MRI offers better soft tissue contrast than prenatal ultrasound and can visualize both hemispheres with great detail without being disrupted by artifacts such as acoustic shadowing that frequently affect fetal ultrasound, especially at higher gestational ages. Therefore, fetal MRI is a valuable tool for confirming or excluding suspected abnormalities detected by ultrasound, thus optimizing perinatal management and influencing pregnancy outcomes ³.

Fetal brain MRI has proven to be a valuable tool for accurately assessing brain development in utero. With superior soft tissue contrast in T2-weighted images, fetal MRI enables accurate measurement of the shape and size of various brain structures ⁴, which can be useful in studying a broad range of congenital brain disorders ^5–8. Super-resolution image reconstruction methods can reconstruct volumetric images of the fetal brain from stacks of MRI slice acquisitions. Continuous progress in fast image acquisition techniques ⁹ and super-resolution methods ^10–12 has improved the quality and success rate of 3D fetal brain reconstruction. The adoption of these techniques has allowed imaging scientists to perform complex computational analyses and make substantive contributions to fetal neuroscience.

Currently, in clinical practice, healthcare professionals visualize individual slices of a stack of images. Therefore, sequences are often repeated multiple times in axial, coronal, and sagittal planes to account for intermittent fetal movements and ensure sufficient diagnostic-quality images ^13,14. Typically, most of the slices obtained are useful for evaluation; however, in cases where there is significant fetal movement or oblique acquisitions, some slices might need to be skipped. Acquiring stacks of slices in standard orthogonal planes is clinically relevant in assessing various conditions, including dysgenesis of the corpus callosum, malformations of the cortical plate, focal cortical dysplasia, ventriculomegaly, parenchymal injury, periventricular nodular heterotopia, cerebellar dysplasia and posterior fossa abnormalities ¹⁵.

Advanced MRI techniques such as diffusion tensor imaging (DTI), MR spectroscopy (MRS), and functional MRI (fMRI) also have a growing footprint in fetal imaging, providing quantitative information about water motion, tissue microstructure, metabolites involved in brain maturation, and functional connectivity. Recent advancements in all these modalities have opened the possibility of conducting large population studies to increase our understanding of brain development and disorders during this critical stage.

STRUCTURAL IMAGING: SINGLE-SHOT T2-WEIGHTED IMAGING

T2-weighted (T2w) imaging is the most commonly used and widely accepted MRI contrast mechanism for visualizing the anatomy of the developing fetal brain ^13,14,16,17. T2-weighting provides excellent differentiation between various developing structures of the fetal brain, such as the germinal matrix and the transient compartments (Figure 1), as well as the cortical plate, developing white matter, and basal ganglia ^18–20. Among the various MRI sampling and spatial encoding techniques that may be used to achieve T2w imaging, single-shot fast spin echo sequence has become widely accepted as the standard for fetal structural imaging ²¹. We will refer to this imaging technique as SST2w in this article. This sequence is popular in fetal MRI due to its excellent T2W contrast and signal-to-noise ratio (SNR), high spatial resolution without geometric distortions, and, most importantly, its relative immunity to fetal motion compared to other techniques ^9,13,14. This sequence is employed to measure structures, asses the anatomy, and identify significant, gross, or local abnormalities or pathologies ^22–25

Figure 1.

Structural image of a 20-week fetus shows different zones, including ganglionic eminence (red), ventricular + subventricular zones (green), intermediate zone (magenta), subplate (blue), and cortical plate (yellow) on T2-weighted (T2w) imaging.

Motion Correction and Image Reconstruction

In fetal imaging, 3D acquisitions allow quantitative evaluation of the cortical plate, assessment of the transient brain zones, and cortical folding, and also allow for automated segmentation. However, fetal motion can cause disruptions in the spatial encoding necessary for 3D image acquisition. Even with multichannel body and spine coils on 3T scanners (achieving a slice thickness of 1.5–2 mm), the acquired SST2w images are often severely disrupted by inter-slice motion, lacking coherent 3D boundaries. Thick slice acquisitions and inter-slice motion artifacts make it difficult to assess the small structures of the fetal brain in 3D.

Over the last 15 years, the imaging research community has demonstrated the feasibility of retrospectively processing clinically acquired, repeated SST2w scans to correct for inter-slice fetal motion and reconstruct 3D images of the fetal brain. The pioneers of these efforts ^26,27(p206) utilized iterations of volume-to-slice registration for inter-slice motion correction and scattered data interpolation to reconstruct one 3D image from a collection of SST2w slices. In 2010, the problem was reformulated to incorporate the physics of MRI slice acquisition as a forward model, leading to the introduction of robust super-resolution volume reconstruction using its inverse solution ¹². Subsequent works have shown the impact of slice outlier removal and point-spread function modeling on super-resolution slice-to-volume reconstruction (SVR) ²⁸, with Deprez and colleagues’ research culminating in the development of SVRTK ^29,30, a robust toolbox that includes options for various sequences and applications as well as additions such as deformable slice-to-volume reconstruction.

In recent years, there have been significant advancements in the tools and algorithms used for SVR. Noteworthy developments in this field include fast SVR through optimized C++ and CUDA implementations ³¹, patch-to-volume reconstruction (PVR) ³², and NiftyMIC ¹⁰. The latter is a fully automated pipeline and toolkit for fetal MRI SVR, which detects and segments the fetal brain on fetal MRI scans and reconstructs a volume in the standard fetal atlas space ¹⁸. NiftyMIC offers a range of options to ensure optimal results, even when working with clinical SST2w data that may be of limited quality.

Recently, innovative deep learning strategies, such as AFFIRM³³, SVORT³⁴, and NeSVOR³⁵, have been developed by several groups to achieve inter-slice fetal MRI motion correction and volume reconstruction. With the advancements in SVR, successful high-quality 3D image reconstruction is highly likely even in the presence of significant fetal motion, given that enough good-quality SST2w images of the fetal brain are acquired (Figure 2).

Figure 2.

Depiction of the T2-weighted (T2w) structural image processing pipeline, which utilizes several T2w images captured from various orthogonal views. A motion-correction superresolution reconstruction technique is then utilized to generate an isotropic T2w image. The resulting image undergoes brain extraction and registration with a preexisting image from a fetal brain atlas that corresponds to the same age, to bring it into a standard space. Finally, the image is automatically segmented for further analysis.

Developing Fetal Brain Structural Atlases

Thanks to SVR techniques, 3D images of the fetal brain have been reconstructed, allowing for the creation of population atlases from in-vivo scans of normally developing fetuses. To account for the rapid growth and changes in the fetal brain during gestation, four-dimensional (spatiotemporal) atlases have been proposed to represent the normal structure and development of the fetal brain^18,36,37 (Figure 3). One such atlas includes detailed tissue segmentation and structural parcellations of the brain into 125 regions at every gestational week between 21 and 37 weeks¹⁸. This atlas, known as the CRL (Computational Radiology Lab) fetal brain MRI atlas, has been widely used in studies on fetal brain development, including automatic fetal brain localization, segmentation, and volumetric analysis on MRI ^10,38,39, quantitative fetal brain analysis ^19,40,41, and connectivity analysis ^42–44.

Figure 3.

Depiction of the available Fetal Brain T2 atlases at seven different gestational ages (GA). Namely the Computational Radiology Lab (CRL) Fetal Atlas (available at http://crl.med.harvard.edu/research/fetal_brain_atlas/), Chinese Fetal Atlas (available at https://github.com/DeepBMI/FBA-Chinese) and the King’s College London Fetal Atlas (available at https://gin.g-node.org/kcl_cdb/fetal_brain_mri_atlas). The first two rows of each panel represent the coronal and axial views, while the third and fourth rows depict tissue segmentations and cortical parcellations (only available for the CRL atlas), respectively.

Newer works on fetal brain T2w MRI atlas construction and segmentation have also yielded promising results. New spatiotemporal atlases ^45,46 have demonstrated compatibility with the CRL atlas despite being built through different image reconstructions and atlas construction frameworks from different populations. These atlases and the CRL atlas have been supplemented with aligned diffusion tensor MRI atlases ^47,48. Most recently, a multi-contrast spatiotemporal MRI atlas of fetal brain development was released based on fetal MRI data from the developing Human Connectome Project (dHCP) (http://www.developingconnectome.org).

Automatic Segmentation of Fetal Brain Tissue

As more 3D fetal brain MRIs are being reconstructed, there is a growing need for computerized methods that can analyze them quickly, accurately, and reliably. Automatic processing can significantly improve the speed and reproducibility of the results, and automatic segmentation of fetal brain tissue types is particularly valuable for assessing normal and abnormal developmental trajectories. Prior to the development of deep learning-based segmentation methods, classical machine learning methods such as support vector machines and random forests were applied to hand-crafted image features to segment the brain tissue ^49,50. However, recent works have shown that deep learning methods can produce consistently more accurate results ⁵¹. A survey of automatic fetal and neonatal brain segmentation techniques in MRI can be found in the manuscript authored by Makropoulos et al. ⁵²

Compared with the adult brain, tissue segmentation for the fetal brain is more challenging because of inherently lower image quality and rapid brain development. Despite these obstacles, several studies have focused on the segmentation of specific structures. Dou et al. used manual segmentations to train and validate a deep, attentive, fully convolutional network (FCN) ⁵³ for the cortical gray matter. Due to the complex topology of the cortex, another study proposed integrating topological constraints into the training loss function ⁵⁴. Meanwhile, Fidon et al. utilized the nnU-Net framework to segment the white matter, ventricles, and cerebellum ⁵⁵. Payette et al. developed a dataset of fetal brain MRI images with labels for seven tissue types ⁵⁶ using 2D UNet. A succession of two FCNs was proposed by Khalili et al. ⁵⁷, with the first FCN extracting the intracranial volume and the second segmenting the brain tissue into seven compartments.

Fetal brain MRI data poses a persistent challenge in obtaining reliable ground truth segmentations for large training datasets. Various methods have been proposed to tackle this issue. One approach involved using an automatic method initially designed for neonatal brains to generate noisy cortical plate segmentations on fetal brain images ⁵⁸. The researchers then refined these segmentations using a human-in-the-loop technique on selected image slices and trained a fully convolutional network with the refined annotations. Another study proposed a unified deep learning framework to estimate a conditional atlas and predict a segmentation jointly ⁵⁹, which aimed to reduce the impacts of motion artifacts and partial volume effects. The rationale for this approach is that the prior knowledge provided by the atlas can guide the segmentation where image quality is low. Karimi et al. proposed a spatially variable label smoothing method to account for uncertain segmentations produced by a multi-atlas segmentation method ⁶⁰.

Significant progress has been made in the field of automatic fetal brain tissue segmentation, with recent developments achieving very high levels of accuracy compared to manual segmentation ^60,61. These advances have been driven by deep learning. However, challenges related to label noise and data scarcity persist, and extensive research has suggested that FCN-type methods “have reached a plateau” ⁶². Encouragingly, new transformers-based models show promise in enhancing performance, as demonstrated in recent preliminary studies ⁶³.

Analysis of Brain Folding and Normal Development Using Structural Imaging

Initial in-utero imaging studies ^64–67 characterized the growth patterns of fetal zones and regional volumes. However, these reports had limitations that prevented simultaneous and automated 3D measurement of multiple zones in a single brain. The first studies to conduct 3D volumetric analysis of the fetal brain using MRI scans were by Scott et al. ⁶⁸ and Andescavage et al. ⁶⁹. Scott et al. found that the supratentorial volume, subplate + intermediate zone, and deep subcortical structures increased linearly with gestational age at a similar rate, while cortical gray matter increased at a greater relative rate and ventricles changed more slowly. Andescavage and colleagues found that the most pronounced changes were seen in the cerebellum, followed by the white matter, cortical gray matter, and deep subcortical structures. In early gestation, all tissue types except the white matter were found to have larger volumes in the left hemisphere, which were then equalized by term. A recent report by Machado-Rivas et al. evaluated age-related volumetric changes in various brain structures and revealed hemispheric asymmetries and sex-related differences, providing valuable insights into fetal brain development ¹⁹.

Furthermore, one of the most remarkable developmental changes in the fetal brain during gestation is the folding of the cortex, also known as gyrification. From approximately 25 weeks, the neocortex rapidly becomes more convoluted, resulting in a significant increase in surface area without a corresponding increase in thickness. Measurements of folding that indicate the degree of gyrification have shown a strong positive correlation with GA, indicating that these measurements are a valuable tool for quantifying cortical folding and neurological development. Furthermore, the link between these folding measurements and GA is stronger than the link between GA and volume ⁷⁰, indicating that neural development is more closely tied to cortical complexity than brain size in this age range.

Similar to what has been seen on direct inspection ⁷¹, studies utilizing 3D reconstructed fetal MR imaging have revealed that at 22 weeks, the brain surface is initially smooth, with the Sylvian fissure being present as a shallow depression. The primary sulci, such as the central sulcus, emerge first around 24–25 weeks, followed by increasing lobar gyrification around 30 weeks. By 34 weeks gestational age, all of the primary and a majority of the secondary sulci are present ⁷².

DIFFUSION IMAGING: ECHO PLANAR IMAGING

Fetal diffusion magnetic resonance imaging (dMRI) provides unparalleled insights into the developing brain. This imaging technique is pivotal in assessing neural connectivity ⁷³ and microstructural changes during gestation ^74,75, a crucial aspect for understanding and potentially intervening in neurodevelopmental disorders. By non-invasively probing the movement of water molecules in the fetal brain, dMRI provides critical data on a variety of microscopic processes that occur during the most dynamic period of human brain development. These data offer a unique opportunity to link these processes to long-term neurodevelopmental outcomes.

Despite its potential, fetal dMRI faces unique challenges, primarily stemming from the dynamic prenatal environment. Factors like maternal breathing, fetal movement, small fetal head, and limitations of MRI technology, such as geometric distortions in echo planar imaging (EPI), signal-to-noise ratio issues, and spin history artifacts from prolonged relaxation times of water-rich fetal brain tissues, pose significant obstacles ^76,77. These challenges cause misalignment in dMRI sequences and can disrupt the assumption of a consistent relationship between image space and anatomy ⁷⁸.

Motion correction of Diffusion Imaging

To overcome the challenges mentioned above in fetal dMRI and enable further data analysis, a variety of fetal-specific motion correction and reconstruction techniques have been developed. Due to the unpredictable nature of fetal and maternal motion, prospective motion correction can be challenging ^79,80. Thus, retrospective motion correction methods emerged as a practical solution for producing reliable dMRI data for further analysis.

The first motion correction methods for fetal dMRI were introduced by Jiang et al., who extended the slice-to-volume registration techniques initially developed for in-utero structural MRI ⁸¹. This technique addressed contrast variation in dMRI slices that resulted from varying sensitization gradients. The method assumed that a rank-2 tensor model could represent local diffusion properties and register each slice to the simulated volume. This resulted in scattered slices in both spatial and angular domains, as each slice’s sensitizing gradient must be rotated according to the motion parameters. The realigned slices were integrated into a regular grid using a diffusion tensor matrix reconstruction to create the final reconstructed volume. The main drawback of this method was the inability of the tensor model to describe voxels containing crossing fibers, a condition found in many voxels.

Following this, Oubel et al. ⁸² proposed a groupwise registration method, which involved collectively aligning diffusion-weighted images followed by using a derived image to register to the T2-weighted reference image. Using affine transformation matrices allowed for the compensation of motion and eddy-current effects in the original sequences. Next, a dual radial basis function-based interpolation was employed to reconstruct a consistent volume on regular grids from the realigned slices that are scattered in the spatial and gradient domains. Fogtmann et al. then introduced a super-resolution reconstruction method of the diffusion tensors that shared some features with Scherrer et al. work ⁸³ but also had some important differences ⁸⁴. One major innovation was adopting a unified alignment approach that resulted in a diffusion-sensitive slice registration model. Furthermore, they incorporated point spread function deconvolution into the 3D image reconstruction process, which enabled the creation of isotropic 3D datasets from scattered DWI acquisitions in various anatomical planes.

Marami et al. proposed motion correction and reconstruction by tracking temporal head movements using filtered image registration. This method assumed a dynamic model for fetal motion and utilized a Kalman filtering approach in combination with slice-to-volume registration to estimate motion parameters. Moreover, they reconstructed diffusion tensors using a weighted least square fit, accounting for the shape of the point-spread function. Outlier slices were identified using a morphological closing filter operator and a support vector machine classifier. This technique notably played a significant role in developing the first spatio-temporal diffusion tensor fetal atlas ⁴⁸ (Figure 4).

Figure 4.

Depiction of the diffusion-weighted image (DWI) processing pipeline on a 33-week fetus. For each DWI scan, a single b=0 image and multiple b≠0 images (depending on the number of directions) are acquired. All non-diffusion-sensitive images (b=0) from multiple scans are merged using a superresolution approach, as are the b≠0 images. Then, the images are brain extracted, and a tensor model is fitted. The resulting image is registered with a preexisting image from a fetal brain atlas that corresponds to the same age to bring it into a standard space. Afterward, the images are brain extracted, and a tensor model is applied. To standardize the results, the image is registered with a preexisting image from a fetal brain atlas that corresponds to the same age. Finally, the image can be automatically segmented for further analysis, and the tensor image can be utilized for tractography.

A significant advancement in the field has been proposed by Derpez et al. They have introduced a higher-order reconstruction method that combines motion correction with super-resolution reconstruction using spherical harmonics ⁷⁶. This approach has proven to be particularly effective in identifying crossing fibers in the fetal brain. Furthermore, building upon their previous work ²⁸, an additional improvement is the integration of intensity correction within the reconstruction process.

Developing Fetal Brain Diffusion Atlases

Similar to T2w imaging of the fetal brain, advances in fetal diffusion imaging have enabled the creation of population atlases from in-vivo scans of normally developing fetuses. To account for the rapid growth and changes in the fetal brain during gestation, four-dimensional (spatiotemporal) atlases have been proposed to represent the normal structure and development of the fetal brain^47,48,85 (Figure 5). The first such atlas, available since 2019, covers every gestational week between 21 and 37 weeks. This atlas, called the CRL fetal brain dMRI atlas, has been widely used in studies on fetal brain development, including microstructural analysis ^74,86 and connectivity analysis ⁷⁵. Recently, Calixto et al. introduced a label set for this atlas, including segmentations for 24 structures ⁸⁶. Another more recent atlas is the Chinese fetal brain dMRI atlas, which covers gestational weeks 24 to 38 and was created using 3D reconstructed images of 89 normally developing fetuses scanned between 24 and 38 weeks of gestation ⁴⁷. This atlas proposed using a fiber orientation distribution (FOD) based pipeline for generating fetal brain dMRI atlases, which showed better registration accuracy than a diffusion tensor pipeline.

Figure 5.

Depiction of the available Fetal Brain diffusion MRI (dMRI) atlases at seven different gestational ages (GA). Namely the Computational Radiology Lab (CRL) Fetal Atlas (available at http://crl.med.harvard.edu/research/fetal_brain_atlas/), Chinese Fetal Atlas (available at https://github.com/RuikeChen/Fetal-Brain-dMRI-Atlas) and the King’s College London Fetal Atlas (available at https://gin.g-node.org/kcl_cdb/fetal_brain_mri_atlas). The top row in each panel shows the fractional anisotropy map, followed by the mean diffusivity map in the second row. The third row displays the tissue segmentations (only available for the CRL and KCL atlases), and the cortical parcellations (only available for the CRL atlas) are shown in the bottom row.

Analysis of Brain Microstructure Using Diffusion Imaging

The potential of diffusion imaging to explore the microstructure of the brain stems from the fact that the diffusion of water molecules changes according to the tissue’s architecture. Key metrics parameters such as FA and MD provide quantitative insights into the density and coherence of neural tracts, as well as the rate of water diffusion.

Several studies have provided insights into the changes in the microstructure of the fetal brain during development. Notably, during the second and early third trimesters, the telencephalic wall is organized into concentric zones, including the marginal zone (not visible on MRI), the cortical plate, the subplate, the intermediate zone, and the proliferative compartments (subventricular zone, ventricular zone, ganglionic eminence). Studies have shown variations in anisotropy and diffusivity in the fetal transient zones during this period that follow different trends ^87–89.

A major limitation of these study types is the need for tissue labels to extract tissue-specific microstructural information. While segmentation algorithms have become available for structural data based on T2-weighted super-resolution reconstructions, they cannot be used for DTI sequences due to differences in tissue contrast and distortions ⁷⁵. For example, motion between DWI and T2-weighted structural acquisitions can cause geometric distortions due to susceptibility-induced B1 field inhomogeneities in echo-planar imaging, ultimately reducing the accuracy of T2 to DWI registration ⁹⁰. To address this issue, ⁴⁸Calixto and colleagues propose a pipeline for automatic segmentation of fetal brain 47, using tensor-based registration ⁹¹ with probabilistic label fusion ⁹². With this approach, Calixto et al. investigated microstructural changes in a healthy cohort of fetal brains during the late second and third trimesters using in vivo motion-corrected diffusion MRI. Their findings are consistent with those of preterm and postmortem studies and are closely linked to cytoarchitectonic changes.

Fetal Brain Tractography

Aside from assessing the tissue microstructure, dMRI has two exciting applications: tract-specific analysis and measuring the brain’s structural connectivity, both enabled by streamlined tractography ⁷³. Classically, the orientation of fibers in tractography has been determined using the diffusion tensor model. However, newer approaches implement diffusion orientation distribution functions (ODFs) and fiber orientation distributions (FODs). The tensor model can only resolve one fiber orientation per voxel, whereas the diffusion ODF can recognize fibers crossing at angles between 45 and 90 degrees ⁹³. FOD tractography outperforms DTI and diffusion ODF methods with its superior resolving power, making it ideal for delineating fibers crossing at a sharp angle. For estimating FODs, Constrained Spherical Deconvolution (CSD) is used, which requires high angular resolution diffusion imaging (HARDI) that has at least one non-zero b-value, generally in the range of 2,500 – 3,000 s/mm^{2 94}.

However, obtaining diffusion data with b values higher than 1000 in fetuses is challenging due to several factors. High b values increase the sensitivity to diffusion but also increase the sensitivity to physiological noise, such as cardiac and respiratory motion. This is particularly relevant in fetal imaging, where the fetus is constantly moving, and the mother’s breathing and cardiac motion can also affect the imaging. Additionally, high b values can lead to a decrease in signal-to-noise ratio (SNR). As the b value increases, the signal intensity decreases, making it difficult to obtain clear images, especially in small structures like the fetal brain. Lastly, using high b values can increase scan time, increasing the likelihood of motion artifacts. Recently, Kebiri et al. have leveraged deep learning and trained a neural network to map dMRI data directly, acquired with as low as six diffusion directions, to FODs for fetuses ⁹⁵.

Tractography has proven to be an effective tool for estimating major white matter bundles and assessing developmental changes in microstructure during the latter half of gestation ^48,96–102. Fetal brain tractography has been employed to depict the radial migration of neurons to the insular cortex ⁴⁰, the fetal visual pathway ¹⁰³, and the maturation of thalamocortical pathways ¹⁰⁴. Moreover, tractography can reveal the occurrence of plastic events leading to the formation of aberrant tracts in the brain, such as in cases of corpus callosum agenesis ¹⁰⁵. Recently, Calixto et al. have developed a new technique for anatomically constrained tractography¹⁰⁶. This method is based on accurate deep learning-based segmentations of the fetal brain tissue directly in the dMRI space. Their approach also involves computing the diffusion orientation distribution function (dODF) from the diffusion tensor, followed by a sharpening step to reduce the incidence of spurious streamlines often encountered with traditional tractography techniques due to low diffusion anisotropy in fetal white matter (Figure 6 shows a depiction of major brain tracts generated using the technique proposed by Calixto et al.)

Figure 6.

Depiction of five major brain tracts on two subjects at different gestational ages (GA). The following tracts are shown: Cortico- Spinal Tract (CST), Forcep Minor (FMajor), Forcep Minor (FMinor), Inferior Occipito- Frontal fascicle (IFO), and Inferior Longitudinal Fascicle (ILF).

To determine the strength of the structural connection between a predefined set of brain regions, each streamline is weighted by some measure of tissue microstructure. In the second and third trimesters of pregnancy, the fetal brain networks become stronger and more efficient, with a significant increase in global integration and local segregation ⁷⁵. Additionally, the existence of small-world network organizations has been observed as early as 20 weeks. Interestingly, this network reconfiguration occurs more rapidly between 20–35 weeks than between 35–40 weeks and is associated with a notable increase in major long-association white matter fibers ¹⁰⁷.

FUNCTIONAL MAGNETIC RESONANCE IMAGING

fMRI is a non-invasive method for investigating brain activity. It measures changes in blood flow and oxygenation in specific brain regions, indicating local neural activity. There are two types of fMRI: resting-state fMRI and task-based fMRI. Unlike task-based fMRI, resting-state fMRI (rs-fMRI) is obtained while the subject is at rest and not performing any activity ¹⁰⁸. This allows for studying functional brain networks during resting states without requiring active participation from the subject, making it an ideal tool for investigating functional fetal brain development in utero.

Development of Functional Connectivity

Studies using resting-state fMRI have uncovered the presence of primitive functional networks in the fetal brain, including motor, visual, and auditory networks, which can be detected as early as mid-gestation ¹⁰⁹. As gestational age increases, functional connections (FC) between and within hemispheres develop, following a medial-to-lateral ^110,111 and posterior-to-anterior ¹⁷ maturational sequence. Notably, more evidence has emerged on the evolving functional networks ¹¹² and their association with subcortical structures like the thalamus ⁴⁴, which is crucial for the early stages of the brain’s functional specialization (Figure 7).

Figure 7.

(A) shows the percentage of significant functional connections (P<0.05, FDR corrected) in three different age groups of healthy fetuses. The histogram in (B) displays the normalized anatomical distance associated with the significant connections for each age group. Finally, (C) is a chord diagram of significant connections (P<0.05, FDR corrected) in each age group. The parcellation scheme and the associated colormap are based on the CRL atlas of the fetal brain.

Using graph-based inferences, researchers have found that the fetal functional connectome has a small-world, modular, and rich club architecture, which becomes more efficient and less modular during gestation ^{110,113–115}, reflecting the formation of new brain connections. As the neural system undergoes pruning and refinement, more specialized and segregated networks emerge in infants and early childhood ^116,117. Moreover, significant functional asymmetry has been observed in the inferior temporal gyrus, which likely plays a role in the development of the language system before birth ^44,118. Additionally, sexual dimorphism has been identified in FC-age associations ¹¹⁹ and connection patterns involved in the somatomotor and frontal areas, as well as the hippocampus, cerebellum, and basal ganglia ⁴³.

Building on the understanding of these developmental patterns, recent studies have shifted focus to exploring the predictive value of FC measures by linking in-utero FC to behavior and birth outcomes in a longitudinal design^120,121. For instance, increased connectivity between the emerging motor network and regions that support motor function, particularly prefrontal regions, the posterior cingulate, and supplementary motor regions, has been shown to predict the subsequent infant motor ability at certain ages^120,122.

Compared to age-matched healthy fetuses, premature infants have been found to display altered FC patterns ^123,124, with stronger connectivity in sensory input and stress-related areas ¹¹¹. Additionally, prenatal exposure to adversity such as maternal stress ^125–127, depression ¹²⁸, elevated body mass index ¹²⁹, and teratogens like lead ¹³⁰ and alcohol ¹³¹ have also been linked to aberrant FC.

Motion Correction and Postprocessing

The reliability of FC estimates in fetal fMRI can be hindered due to a variety of factors. These include irregular fetal motion, limited signal-to-noise ratio (SNR) caused by physiological noises from maternal tissues and placenta, as well as low temporal and spatial resolutions. The developing brain’s rapid conformational changes introduce even more variability, such as smaller head sizes in younger fetuses, which may bias FC towards stronger short-range connections ¹⁷. Additionally, fetal motion often leads to intensity inhomogeneities and spin-history artifacts ¹³². These factors collectively contribute to the complexity of obtaining robust and consistent FC estimates, requiring advanced image processing techniques specifically tailored to fetal fMRI data’s unique characteristics.

Recent advancements in fetal fMRI processing leverage deep learning models based on 2D U-Net 29, 3D U-Net ¹³³, and the attention U-Net ¹³⁴. These models automate the brain extraction from motion-corrupted fMRI images, greatly reducing processing time and forming the foundation for the entire processing pipeline (Figure 8). Spin-history artifacts have also been addressed by estimating low-frequency multiplicative fields on a slice-wise basis and imposing smoothness, either parametrically through a linear combination of basis functions ¹³⁵ or non-parametrically through a smoothness penalty ¹³⁶. Additionally, a range of techniques, including a weighted registration algorithm based on 2nd-order edge features ¹³⁷, latent state modeling of motion via Markov assumption ¹³⁸, and super-resolution reconstruction ^139,140, have been proposed to mitigate fetal excessive motion and improve the quality of fMRI images. Despite these advances, challenges remain, highlighting the need for robust and reproducible approaches ¹⁴¹ to ensure optimal acquisition and accurate analysis of the fetal functional connectome.

Figure 8.

Example of the estimated framewise realignment parameters, including translation and rotation along each dimension, for a 34-week fetus.

MAGNETIC RESONANCE SPECTROSCOPY

Magnetic resonance spectroscopy (MRS) is a specialized MRI technique used to measure concentrations of brain metabolites ¹⁴². It has been used to measure changes in metabolites such as N-acetylaspartate (NAA), choline (Cho), creatine (Cr), myo-inositol (Ins), lactate (Lac), and glutamate (Glu) in the developing brain of a fetus ^143–147. However, due to the significantly lower concentration of these metabolites in the brain compared to water and fat molecules (about four orders), MRS requires techniques quite different from most other MRI techniques. In the simplest approach, called single voxel MRS, three orthogonal slice selective pulses define a relatively large voxel region (typically, 1–1.5 cm3), where the water signal is suppressed, and metabolic information is extracted. The peak signals of various metabolites appear at specific frequencies, or ppm values, related to those metabolites on a proton spectrum (Figure 9).

Figure 9.

Example of a Long Echo Time (270 ms) Magnetic Resonance Spectroscopy (MRS) using the single-voxel technique of a 19-week fetus with a mitochondrial disorder. The white arrow shows lactate elevation.

Two commonly used techniques for single-voxel MRS are stimulated echo acquisition mode (STEAM) and point-resolved spectroscopy (PRESS). STEAM uses three slice selective 90° pulses ¹⁴⁸, while PRESS uses a slice selective 90° pulse and two orthogonal slice selective 180° refocusing pulses ¹⁴⁹. Generally, STEAM is preferred for short echo times (20–30ms) due to its ability to provide richer metabolic content information, while PRESS is preferred for longer echo times (~135ms) for its more reproducible spectra ¹⁵⁰. With typical repletion times of 1.5–2.5 seconds and signal averages of 20–30, the total acquisition time for a single-voxel MRS is around one minute. Additional time is required before signal acquisition for adjustments and shimming.

Several studies have utilized single voxel MRS to evaluate brain metabolite levels in live fetuses in utero. However, it is much more challenging to perform than structural fetal MRI. The accuracy and reliability of fetal MRS is hard to determine, but feasibility studies have reported 50–68% of cases as “useable” or “readable” without sedation ^143,151,152. However, these estimations may be optimistic as some studies excluded cases with detected gross motion. In general, performing MRS in younger fetuses and fetuses with the head in non-cephalic positions is challenging.

Performing fetal magnetic resonance spectroscopy (MRS) poses several challenges due to the small size of the fetal brain and its structures compared to the relatively large MRS voxel sizes, as well as fetal motion. Large voxels are necessary to achieve a reasonable signal-to-noise ratio, and these often cover a large area of the developing brain, including cortical plate, deep brain nuclei, white matter, and cerebrospinal fluid in the lateral ventricles. There are inevitable delays between the scans used as a reference for MRS voxel placement and the actual measurement of MRS signal. These can include operator delays in voxel placement, position refinement (~1 minute), lead time for MRS frequency and shim adjustments, and measurements for water and fat suppression (~1–2 minutes). During this time, fetal motion can occur, which can shift the prescribed voxel area partially or entirely outside of the fetal brain. Additionally, fetal motion during the acquisition may also happen, leading to lower signal quality.

Clinical Application

Changes in brain metabolites may precede and predict functional and structural changes in the brain. Detecting and estimating the levels of certain metabolites, such as lactate and glutamate, during the early stages of fetal development could prove beneficial in assessing and understanding brain development in fetuses at risk of neurodevelopmental disorders such as fetuses at risk of hypoxia and ischemia. Deviation from typical brain metabolite levels throughout gestation may offer insight into the extent of neuronal damage in fetuses that have suffered from brain injury.

During gestation, brain metabolites change as the fetal brain matures and axons myelinate. As gestational age advances, NAA, Cr, and Cho have been shown to gradually increase due to myelination in the third trimester ^144,146,153. On the contrary, fetuses with gastroschisis and intra-uterine growth restriction (IUGR) have been reported to have elevated levels of Lac ^154,155, while fetuses with congenital heart disease (CHD) have relatively high levels of Cho and low NAA:Cho ratios compared to healthy fetuses scanned in the third trimester ¹⁵⁶. These findings are consistent with previous studies in neonates with CHD, which reported decreased NAA:Cho ratios and increased Lac and Lac:Cho ratios in newborns with CHD compared to healthy controls ¹⁵⁷. Fetuses with transposition of the great arteries (TGA) and single ventricle (SV) heart defects have also been found to have increased levels of Lac compared to fetuses with normal physiology. These elevated Lac levels have been significantly associated with death before discharge in the TGA and SV subgroups of CHD.

SUMMARY

This review summarizes the progress made in fetal brain MRI imaging, which has significantly enhanced our ability to image and assess the developing fetal brain and has led to an improved understanding of fetal brain development. Over the past two decades, these advancements have enabled a thorough assessment of fetal brain development, from structural imaging that allows for assessing cortical folding and changes in volume to diffusion imaging to assess microstructural changes throughout gestation and even changes in tract configurations. Although there is still much work to be done, including developing better technologies for MRI spectroscopy, we can expect future research to significantly enhance all fetal imaging techniques, incorporate them into clinical workflows, and offer a deeper understanding of the early stages of human brain development.

CLINICS CARE POINTS

• SST2w is widely used for its superior contrast, signal-to-noise ratios, and ability to resist fetal motion. It allows for detailed structural evaluation of the fetal brain, including detecting abnormalities and assessing cortical folding and changes throughout gestation.

• Diffusion MRI is particularly useful for examining fetal brain development and changes in brain microstructure. Progress in motion correction and image reconstruction techniques has made it possible to assess these changes throughout gestation.

• fMRI is a non-invasive technique that measures blood flow and oxygenation changes in specific brain regions, reflecting local neural activity. Resting-state fMRI is especially valuable for studying the fetal brain as it does not require active participation. It is vital for understanding the development and specialization of brain functional networks in utero.

• Magnetic Resonance Spectroscopy is instrumental in measuring concentrations of various brain metabolites in the developing fetal brain. It offers critical insights into the biochemical environment of the fetal brain, aiding in the understanding of its development and early detection of potential neurodevelopmental issues.

Key Points:

1. Advancements in image processing, including motion correction, automatic segmentation, and the development of fetal brain atlases in different MRI techniques (including structural imaging, diffusion imaging, and functional MRI ), have greatly enhanced our ability to image the structure and function of the developing fetal brain in-utero.

2. T2-weighted (T2w) imaging, precisely the single-shot fast spin echo sequence (SST2w), is highly regarded for its excellent contrast and signal-to-noise ratios, relatively high spatial resolution, and ability to withstand fetal motion. SST2w is predominantly employed for structural evaluation and detecting abnormalities in the developing brain. Thanks to advancements in motion correction and 3D image reconstruction, cortical folding and normal volume changes can now be assessed throughout gestation.

3. Fetal diffusion magnetic resonance imaging (dMRI) is essential for examining fetal brain development. Advancements in motion correction and image reconstruction techniques have enabled the assessment of changes in brain microstructure and white matter architecture (tractography) throughout gestation.

4. Functional Magnetic Resonance Imaging (fMRI) is a non-invasive technique that measures changes in blood flow and oxygenation in specific brain regions, indicating local neural activity. Resting-state fMRI is particularly promising for studying the fetal brain as it does not require active participation from the subject, allowing characterization of spontaneous brain activity and functional brain networks during resting states.

5. Magnetic resonance spectroscopy (MRS) measures concentrations of various brain metabolites. MRS is instrumental in detecting changes in metabolites in the developing fetal brain. This technique offers insights into the biochemical environment of the fetal brain, complementing other established modalities, aiding in the understanding of fetal development, and potentially contributing to the early detection of neurodevelopmental disorders in fetuses.

Synopsis.

Over the last 20 years, there have been remarkable developments in fetal MRI analysis methods. This article delves into the specifics of structural imaging, diffusion imaging, functional MRI, and spectroscopy, highlighting the latest advancements in motion correction, fetal brain development atlases, and the challenges and innovations. Furthermore, the article explores the clinical applications of these advanced imaging techniques in comprehending and diagnosing fetal brain development and abnormalities.

Abbreviation List

DTI

Diffusion Tensor Imaging

FA

Fractional Anisotropy

FC

Functional Connections

FCN

Fully Convolutional Network

FOD

Fiber Orientation Distribution

MD

Mean Diffusivity

MRI

Magnetic Resonance Imaging

MRS

Magnetic Resonance Spectroscopy

ODF

Orientation Distribution Function

PRESS

Point Resolved Spectroscopy

RF

Radiofrequency

SNR

Signal-to-Noise Ratio

SST2w

Single-Shot T2-weighted Imaging

SVR

Slice-to-Volume Reconstruction

fMRI

Functional Magnetic Resonance Imaging

rs-fMRI

resting-state Functional Magnetic Resonance Imaging