Baby brain atlases

Abstract

The baby brain is constantly changing due to its active neurodevelopment, and research into the baby brain is one of the frontiers in neuroscience. To help guide neuroscientists and clinicians in their investigation of this frontier, maps of the baby brain, which contain a priori knowledge about neurodevelopment and anatomy, are essential. “Brain atlas” in this review refers to a 3D-brain image with a set of reference labels, such as a parcellation map, as the anatomical reference that guides the mapping of the brain. Recent advancements in scanners, sequences, and motion control methodologies enable the creation of various types of high-resolution baby brain atlases. What is becoming clear is that one atlas is not sufficient to characterize the existing knowledge about the anatomical variations, disease-related anatomical alterations, and the variations in time-dependent changes. In this review, the types and roles of the human baby brain MRI atlases that are currently available are described and discussed, and future directions in the field of developmental neuroscience and its clinical applications are proposed. The potential use of disease-based atlases to characterize clinically relevant information, such as clinical labels, in addition to conventional anatomical labels, is also discussed.

Graphical Abstract

1. Introduction

An atlas is a set of reference files based on knowledge accumulated through past experiences. A brain atlas, in this review, consists of a brain template image or a set of brain template images with various types of boundaries drawn on them. Traditionally, such boundaries were based on existing knowledge about features of brain anatomy, pathology, and functions. During the past decade, various types of baby brain atlases have been published based on accumulating knowledge about the image representation of the neuroanatomy and functions of developing brains. In this review, the types and roles of the human baby brain MRI atlases that are currently available are described and discussed. We focused on the atlases made from brain MRIs scanned before 12 months of chronological age, including fetus, neonate, and infant, with the goal of elucidating future directions in the field of developmental neuroscience and its potential clinical applications.

2. Technological advancements have enabled the creation of baby brain atlases

A brain atlas in three-dimensional electronic form is an essential tool for modern neuroimaging research. Since the early-90s, anatomical MRI has become widely available and efforts have been made to create brain atlases that represent common features of the human brain (Collins et al., 1994; Lehmann et al., 1991; Mazziotta et al., 1995; Tzourio-Mazoyer et al., 2002). Such atlases have been used as a template to which brain images are spatially normalized. This normalization allows researchers to introduce a map of the human brain, based on which our a priori knowledge about anatomical areas and their functions can be applied on each individual brain. The Talairach atlas (Talairach and Tournoux, 1988), which was created based on post mortem brain sections from a 60-year-old woman, is one of the most referenced atlases, and has been used as the ‘gold standard’ anatomic reference. Although the Talairach atlas is in paper form, it introduced a unique three-dimensional coordinate system using the anterior and posterior commissures at the midsagittal plane, as well as the boundary box of the brain, as the main landmarks. Brain images can be normalized simply by aligning the anterior and posterior commissures, scaling the brain images to the boundary box, and applying the Talairach atlas as a reference to estimate the locations of specific anatomical brain regions. Methodological developments in image transformation in the 1990s allowed the creation of group-averaged brain atlases of the in vivo human brain scanned by high-resolution, structural MRI (Mazziotta et al., 1995). The normalization of the images to an atlas space allows researchers to perform a group analysis of brain images by introducing the template’s ‘common coordinate system,’ which enables statistical analyses at voxel-level granularity. The introduction of a common coordinate system is also useful for multi-center studies, with data collected on different scanners, or for meta-analyses of studies that are performed at multiple institutions.

During the past two decades, image transformation methods and anatomic feature extraction algorithms have become more and more sophisticated, which has resulted in group-representative MRI atlases with detailed anatomical delineation (Akazawa et al., 2016; Kazemi et al., 2007; Lorenzen et al., 2006; Mazziotta et al., 2001a; Mazziotta et al., 2001b; Mori et al., 2008; Oishi et al., 2011c; Oishi et al., 2008; Rohlfing et al., 2010; Shattuck et al., 2008; Van Leemput, 2009; Yeo et al., 2008; Zhang et al., 2014b). The accuracy of image normalization depends on how well the anatomical landmarks are delineated (e.g., sharp boundaries between gray and white matter structures), on both the template and the target brains, as well as the image transformation algorithm applied. Thus, greater anatomical details on the template image facilitate accurate normalization and yet reduce noise derived from misregistration of the images at the voxel level. This advancement in normalization accuracy is favorable for neuroimaging studies because it increases the statistical power to detect biologically meaningful changes in MR signals or morphology.

Despite this advancement in adult and pediatric brain research, the baby brain has remained a challenge because of two major reasons: difficulty in image acquisition and in image processing. Recent advancements in scanner hardware, sequences, and motion control or correction methodologies (Cordero-Grande et al., 2017; Gumus et al., 2014; Herbst et al., 2017; Herbst et al., 2015; Hughes et al., 2017; Kuklisova-Murgasova et al., 2012; Li et al., 2015b; Singh et al., 2015; Zahneisen et al., 2016; Zahneisen et al., 2014) have partly overcome the issues related to image acquisition, yielding higher image resolution, shorter scan times, and easier motion control and correction than was possible previously. The resultant improvement in image quality in the past decade has allowed researchers to optimize image-processing methods developed through adult brain research and translate the approaches to baby brain research. For example, a tissue segmentation tool (MANTiS) (Beare et al., 2016) that works on the Statistical Parametric Mapping SPM software (http://www.fil.ion.ucl.ac.uk/spm/), the tract-based spatial statistics (TBSS) (Ball et al., 2010; Gao et al., 2012), and tract-specific analysis (TSA) (Anjari et al., 2007; Ball et al., 2010; Li et al., 2016; Porter et al., 2010; Seo et al., 2013) have all been customized for neonatal and infantile brains. However, several issues related to the processing of baby brain images remain and will require further explorations (see Sections 9 and 10).

3. The increasing amount of information in the atlas space during early brain development

MRI atlases in the 1990s were mostly based on T1-weighted images. The information to be extracted from the T1-weighted contrast was limited to macroscopic morphological features of the gray matter structures. An important aspect of newer atlases is the inclusion of information beyond gray matter morphology. Atlases based on multiple modalities can convey various types of contrasts, which are linked to the histological- or even molecular-level properties of the brain. For example, currently available MRI modalities from which atlases were generated include diffusion-weighted images, diffusion tensor images (DTI), T2 maps, fluid attenuation inversion recovery sequences, and quantitative susceptibility maps (Lim et al., 2013; Miller et al., 2013; Mori et al., 2013), in addition to conventional T1- and T2-weighted images and echo-planar imaging. The introduction of a common coordinate system is useful for multi-center studies, with data collected on different scanners, or for meta-analyses of studies that are performed at multiple institutions. These atlases in three-dimensional form also serve as templates to which secondary information could be mapped. Landmark accomplishments in this field include maps within the MRI atlas that show brain areas based on the transcriptome (Hawrylycz et al., 2012; Rizzo et al., 2016), genetic influences (Chen et al., 2013; Chen et al., 2012; Chen et al., 2015; Fan et al., 2015; Thompson et al., 2016), cytoarchitecture of the brain (Amunts et al., 2000; Amunts et al., 2004; Bailey et al., 2007; Eickhoff et al., 2006; Eickhoff et al., 2007; Eickhoff et al., 2005), and probabilistic or group-averaged features of functional or anatomical connectivity (Akazawa et al., 2016; Figley et al., 2015; Figley et al., 2017; Shirer et al., 2012; Zhang et al., 2010).

Among various aspects of Information included in atlases, time-axis, cortical convolution, age-specific structures, and diffusion orientation are particularly important for the analysis of baby brains. Baby brains are under rapid growth with the cytoarchitecture, shape, and volume continuously changing. Indeed, the Jacobian determinants that quantify the local volume changes of the brain among 33, 36, and 39 gestational weeks demonstrated dramatic volumetric changes in most gray and white matter structures (Feng et al. 2017). To account for this rapid change, the referenced baby atlases need to be specific to different but narrow age ranges. The inclusion of a time axis can be achieved by creating multiple age-representative atlases during development (Habas et al., 2010a; Huang et al., 2009; Huang et al., 2006; Kuklisova-Murgasova et al., 2011; Shi et al., 2011; Huang and Vasung, 2014; Yu et al., 2014; Ouyang et al., 2015; Zhang et al., 2007), to elucidate the structure-specific features of neurodevelopment.

A notable change that occurs on the surface of the brain is gyrification. The cortical surface starts to fold at the second trimester and accelerates during the third trimester. The major gyri seen in adult brains are mostly observed at term, but further gyrification continues until two years of age (Li et al., 2015a). A two-dimensional representation of the brain surface, which is another feature of the MRI atlas, can display this developmental change. Complex features of the brain, such as surface areas, folding, thickness, or myelination, could be measured and mapped on the surface (Van Essen, 2002; Van Essen and Drury, 1997). The increasing gyrification associated with brain growth also leads to increasing total brain volumes. At birth, the whole brain volume is approximately one-third of healthy adult brain volume, with a growth rate of 1%/day, slowing to 0.4%/day by the end of 3 months when the infant brain is half of the adult brain volume (64% increase), with male infants growing faster than the female infants, the cerebellum grows at the highest rate (more than 200%) while the hippocampus grows the slowest (47%) during the first 3 months of life (Holland et al., 2014).

Major neurodevelopmental changes that occur from 30 gestational weeks to full-term birth (40 gestational weeks) include the disappearance of the ganglionic eminence in the periventricular areas, the organization of the radial glial scaffold in the cortical plate, and the pre-myelination of the white matter fibers. Shrinkage of the ganglionic eminence results in the disappearance of the periventricular hypo-intense areas seen on the T2-weighted images. True myelination of the axonal fibers follows the pre-myelination, rapidly until two years of age and then continues slowly. Diffusion MRI is suitable for the evaluation of well-aligned cytoarchitectures and, therefore, sensitive for the detection of the rapid decrease in fractional anisotropy (FA) of the cortex, which is caused by the reduced radial orientation of the cortical architecture, and decreased radial and axial diffusivities with stable FA values in the white matter, which reflect ongoing pre-myelination.

The perinatal and early postnatal periods are ideal target periods to investigate the genetic influences on brain development or to characterize cytoarchitectonic development. Baby brain atlases incorporated with knowledge derived from genetic or microscopic observations, such as transcriptome (Hawrylycz et al., 2012; Rizzo et al., 2016) or cytoarchitectonic maps (Amunts et al., 2000; Amunts et al., 2004; Bailey et al., 2007; Eickhoff et al., 2006; Eickhoff et al., 2007; Eickhoff et al., 2005), are novel and emerging fields of research. Multimodal atlases, including a diffusion MRI atlas in a common geometrical framework, are essential as references by which to integrate and interpret findings from various types of neuroimaging studies (Faria et al., 2012; Oishi et al., 2011a; Oishi et al., 2009).

4. The atlas as a teaching file

The core value of an atlas is the a priori knowledge it can provide as a teaching file of brain anatomy and function. The advancements in atlas-creation technology and image transformation, described in Section 2, also revealed limitations in atlas-based image analysis. The increasing accuracy of image normalization illuminated the fact that there is no neurobiologically ‘perfect normalization’ that could account for the wide anatomic variations seen in the brain, especially those seen in the cortical areas. The normalization here refers to a process of matching each pixel of the brain image to the neurobiologically identical pixel(s) of the atlas brain. This process cannot be perfect simply because each brain is, like a fingerprint, unique, and structures or spaces (e.g., minor convolution, fissure, bridge between the caudate and the putamen, or the cavum septum pellucidum) seen in one brain may not exist in the atlas brain (Ono et al., 1990). Indeed, this uniqueness of each brain is becoming an active research field for the investigation of factors that determine the behavioral or psychological traits of each individual (Finn et al., 2015; Kaufmann et al., 2017; Miranda-Dominguez et al., 2014; Miskovich et al., 2016). Assessing diseased brains with anatomic features that deviate from or that do not exist in the atlas brain, are more problematic than measuring normal brains, because atlases with ‘common anatomic features’ would not contain ‘disease-specific features.’

The baby brain is the new area in the landscape of anatomic variation, which requires baby-specific teaching files (Chapman et al., 2010) to interpret the neuroradiological findings. The baby brain during development is not a smaller version of the adult brain. Some structures exist in the developing brain with no corresponding structures in the adult brain, or vice versa. Structure-specific variation in brain development (Shi et al., 2010; Shi et al., 2011), which has been articulated by longitudinal studies, is another type of variation that makes the use of a single-atlas space not straightforward. What is becoming clear is that one atlas is not enough to characterize our existing knowledge about anatomic variations, which may be attributable to time-dependent changes or disease-related anatomic alterations. As we move toward this new frontier, more teaching files than ever before are needed. Brain atlases are playing an essential role as the “knowledge database”, beyond their traditional use as simply a template for image analysis.

5. Age definitions used in baby brain atlases

The stage of development from which the age is calculated plays an important role in defining the age of the baby. The words “neonate” and “infant” are, by definition, based on chronological age that refers to the age from the time of birth. However, the time of fertilization should be the origin when evaluating neurodevelopment from the blastocyte. Since accurate identification of the time of fertilization is often impossible except when assisted reproductive technology is performed, the first day of the mother’s last menstrual period (FM) is usually used as the surrogate. Gestational age is the time elapsed after the FM and used to express the age of fetuses. Postmenstrual age, which is calculated from the gestational age at birth plus the chronological age, is the most common term to express the age of the neonatal and infantile brains. There are two issues to be noted when the terms gestational age, postmenstrual age, and chronological age are used. First, the FM contains some inaccuracy, with regard to the estimated time of fertilization, because the duration from the last menstrual period to ovulation or fertilization varies. This variation probably has an influence on developmental trajectories (Wu et al., 2018). Second, the conversion from the chronological age to postmenstrual age varies, depending on the gestational age when the baby was born. The normal duration of gestation is from 37 to 41 weeks. This five week variation causes inconsistency in the conversion if the gestational age at birth is unknown or inaccurate. The inconsistency becomes even more marked for preterm-born babies who were born before 37 postmenstrual weeks. For example, a child scanned at 32 postmenstrual weeks is a neonate, if the child was born at 29 postmenstrual weeks (three weeks old in chronological age), but is a fetus if the child was scanned in utero at 32 gestational weeks, although in both cases the time from fertilization to scan is almost identical.

6. Parcellation maps on baby brain MRI atlases

6.1: Role of the parcellation map in image quantification

The fundamental information to be referenced from a brain atlas is a precise geographical location. Delineation of each anatomic structure in the atlas space is useful for interpreting the coordinate representation in an anatomic ontology. Such a set of boundaries drawn on neuroimaging modalities is called a parcellation map (Fig. 1). The parcellation map with the atlas space can be transformed and registered to individual images to obtain a parcellation map on each individual, based on either a single-atlas (Sections 8.1. and 8.2.) or a multi-atlas approach (Section 8.3.), or variants that include patch-based methods (Liu et al., 2016; Mechrez et al., 2016; Park et al., 2016; Rekik et al., 2015; Rousseau et al., 2011; Wang et al., 2014; Wu et al., 2015; Wu and Shen, 2014; Wu et al., 2013). This transformation is a process that transfers the knowledge incorporated in the atlas (information about the anatomy or function) onto each individual’s image. The parcel introduced onto each image enables measurements of the volume, shape, and intensity of the anatomical structure or unit, which is known as atlas-based or parcel-based image quantification. Using parcellation maps for image quantification has several advantages. The parcellation map functions as an anatomically reasonable spatial filter to increase the signal-to-noise ratio (SNR) of the measures (volume or intensity), which are low at each voxel level (Faria et al., 2012; Faria et al., 2015; Maldjian et al., 2003). Anatomical boundaries obtained from the parcellation map can also be used to define the surface area of each structure, which is used in the surface-based image analysis (Miller et al., 2015; Qiu et al., 2010; Qiu and Miller, 2008; Steinert-Threlkeld et al., 2012; Tang et al., 2015a, b; Wang et al., 2003; Younes et al., 2014) to detect group differences in the shape of the anatomical structure. The increased SNR is beneficial in several ways. For example, in group comparisons, the differences between the groups are more likely to be detected by increasing the SNR. Increased SNR is also beneficial for detecting anatomical areas in which a quantified measure is correlated with the regressor (e.g., neurological function).

Fig. 1.

Brain parcellation map overlaid on the JHU-neonate brain template.

6.2 Detecting pathological changes seen in early development

One of the major motivations to investigate baby brains and their development is an interest in detecting the abnormalities in brain anatomy that may predate the onset of neurological diseases (Chang et al., 2016b) or psychiatric disorders (Amaral et al., 2008; Becker et al., 2001; Bogerts et al., 1985; Gottlich et al., 2014; Modell et al., 1989; Posner et al., 2014; Sheth et al., 2013; Tamminga et al., 1992). Since many neurological diseases and psychiatric disorders are known to have genetic backgrounds that might result in specific phenotypes of the brain anatomy (endophenotypes) before the onset of symptoms (Chakravarty et al., 2015; Dixson et al., 2014; Fornito et al., 2013; Hajek et al., 2009; Menzies et al., 2008; Nery et al., 2013; Ordonez et al., 2015; Scognamiglio and Houenou, 2014; Zannas et al., 2014), and certain perinatal environmental risk factors also might affect the later onset of psychiatric or learning disorders (Abel and Sokol, 1987; Favaro et al., 2006; Johnco et al., 2016; Kut et al., 2010; Liu et al., 2010), such endophenotypes or perinatal brain injuries may be detected by neuroimaging modalities pre-symptomatically.

To date, baby brain atlases have been applied to study developmental abnormalities related to specific conditions, such as preterm birth (Akazawa et al., 2016; Chang et al., 2016a; He and Parikh, 2013; Oishi et al., 2011c; Pannek et al., 2013), low birth weight with related risk factors (Kersbergen et al., 2014; Rose et al., 2014a; Rose et al., 2014b), prenatal exposure to stimulants (Chang et al., 2016c), and different ethnic backgrounds (Bai et al., 2012) [further reading (Deshpande et al., 2015)].

6.3. Evaluating the development of a baby brain

Since clinical decisions are always made on an individual basis, the applicability of the atlas to an individual baby is important for clinical applications and precision medicine. The advantage of using a parcellation map derived from a brain atlas, compared to voxel-based image analysis, is the applicability of the map to an individual brain. Once a parcel-based group comparison between conditions (e.g., typically developing babies vs. babies with prenatal exposures to neurotoxic substances) is performed and features of the condition-specific developmental alteration are identified in a specific parcel, the same parcel could be applied to a single baby to quantify the neurodevelopment and determine whether the condition-related developmental delay is also present in the baby. To identify an abnormality in brain development, the atlas-based quantification results from typically developing babies are required, from which a normal growth curve for each anatomic structure can be generated as a reference against which to evaluate the results from each individual’s brain (Akazawa et al., 2016; Chang et al., 2016a; Chang et al., 2016c; Oishi et al., 2011c; Rose et al., 2015; Rose et al., 2014a; Roze et al., 2015; Wu et al., 2017a, b). Ongoing longitudinal projects, such as the developing Human Connectome Project (dHCP) that targets prenatal and perinatal brains, and the Baby Connectome Project that targets postnatal brain development, are promising for the collection of such normative datasets. Longitudinal observation is especially encouraged in the analysis of baby brains, because the growth trajectories of neurodevelopment are difficult to predict. For instance, measurements of preterm born infants that showed large deviations from the normative group at a term-equivalent age (40 weeks postmenstrual age) may either normalize (thin red line in the Fig. 2) or remain abnormal (thick red line with * in the Fig. 2) at follow-up evaluations (Akazawa et al., 2016).

Fig. 2.

Atlas-based quantification of baby brains. The mean diffusivity of the left inferior longitudinal fasciculus is plotted against postmenstrual age in weeks. Nineteen full-term and 30 preterm-born babies were each scanned longitudinally at three time points, and line-plots from the same subject are connected by solid lines (Akazawa et al., 2016). Cyan: term-born babies. Red: preterm-born babies. * The thick red lines denote the outliers (the trajectories outside 95% of the term-born trajectories). Note that the outliers can be identified based on the normal developmental trajectories (cyan lines). (Figure from (Akazawa et al., 2016) with permission)

6.4. Evaluating developmental changes in brain connectivity

6.4.1. Anatomical connectivity based on parcellation map

Once parcellation that defines the geometric boundaries of brain structures (primary parcellation) is completed, anatomical pathways that connect primary parcels could be defined. Such pathways could also be parcellated (secondary parcellation) (Oishi et al., 2011b; Zhang et al., 2010). For example, fiber pathways that connect the primary and higher-order visual cortices were parcellated in neonatal DTI (Akazawa et al., 2016). The effect of intrauterine growth restriction on the motor-related connections and cortico-striatal-thalamic connections of infant brains were investigated using the secondary parcellation method. The secondary parcellation method was also used to investigate substructure-specific developmental patterns seen in the thalamo-cortical connections of neonatal brains (Poh et al., 2015).

6.4.2. Functional connectivity based on parcellation map

The parcellation map from the neonatal brain atlas could be used to facilitate research on the functional connectivity of baby brains. The seed regions of interests (ROIs) could be identified with the guidance of anatomical information from baby brain atlases. Fig. 3 shows the whole-brain functional connectivity maps with the connectivity seed placed on the left precentral gyrus, which was determined by the atlas parcellation map. The left precentral gyrus is a region in the primary sensorimotor system. Functional connectivity analysis of the preterm-born babies scanned at 31 to 42 postmenstrual weeks indicated increasing connectivity strength and heterogeneity with age, between the seed region and the other brain areas during brain development (Cao et al., 2017).

Fig. 3.

Developmental changes in functional connectivity between the left precentral gyrus and other brain areas in 40 term and preterm born infants. The first row indicates the brain regions (orange areas) with significant age effects. The second, third, and fourth rows demonstrate the evolving connectivity seen in three different age groups based on their ages at the scans: group 1, 31.3 – 35.3 postmenstrual weeks; group 2, 35.6 – 38.4 postmenstrual weeks; and group 3, 38.7 – 41.7 postmenstrual weeks. Group 1, group 2, and group 3 included both preterm and term-born neonates scanned at the ages described above. Functional connectivity of the term group is shown in the fifth row as a reference. The term group included only term-born neonates (>38.0 postmenstrual weeks at birth) scanned at the age of 38.4–41.7 postmenstrual weeks. A corrected-P < 0.01 was applied as the threshold (Adapted with permission from Cao et al., 2017).

7. Types of parcellation maps

7.1. Knowledge-based parcellation map

7.1.1. Parcellation map created from baby brain MRIs

Anatomical ontology is a consensus of structural units, derived from a history of detailed observations that have been performed by numerous neuroanatomists. Therefore, the current “gold standard” with which to define the boundaries of brain structures also relies on expert neuroanatomists’ opinion, which is a “knowledge-based delineation.” MRI contrast has been used to guide manual brain parcellation (Desikan et al., 2006; Klein and Tourville, 2012; Oishi et al., 2009; Shattuck et al., 2008; Tzourio-Mazoyer et al., 2002). However, the parcellation of baby brains is especially challenging, because of the limited contrasts between brain structures (Section 7.1.1.) and the partial volume effect that causes blurring of the images, as well as the co-existing myelinated and unmyelinated fibers that result in uncertain structural boundaries.

The type of MRI modality and the contrast defines the type of parcellation map. In well-developed, mature brains (>3 years of age), T1- weighted contrast is typically employed to define the boundaries between the gray and white matter structures, and to define anatomic boundaries based on patterns of cortical folding, such as lobes and gyri. However, delineation of the cortical surface is often difficult in baby brains, because the intensity of the dura mater and vascular structures, which are located adjacent to the cortical surface, is similar to that of the cortex, and often with little to no cerebrospinal fluid space to separate them. For the baby brains, T2-weighted images provide better contrast than T1-weighted images in delineating the cortical surface, and in defining the boundaries between the gray and white matter structures.

There are several challenges in the manual parcellation of baby brains based on T1- or T2-weighted MRIs. Insufficient contrasts between the gray and white matter structures hinder accurate structural delineations, especially during 4 – 8 months of age when many gray and white matter structures share similar intensities (Oishi et al., 2012). The intensity inhomogeneity is another issue. To create an adult brain atlas, the low-frequency component of the intensity inhomogeneity is often mathematically corrected (Belaroussi et al., 2006; Hou, 2006; Sled et al., 1998; Tustison et al., 2010) to obtain better structural delineation. The major sources of the intensity inhomogeneity are those related to MRI hardware, such as static field inhomogeneity and radiofrequency coil non-uniformity. However, baby brain MRI includes biological inhomogeneity, which refers to an inhomogeneity caused by local differences in ongoing myelination—which occurs from inferior to superior, from the center to the periphery, and from posterior to anterior directions of the brain. Therefore, in the axial plane, there is a biologically derived gradient of white matter intensity that is darker in the central and posterior regions, and brighter in the peripheral and anterior regions, on the T2-weighted images. The effect of the mathematical inhomogeneity correction on the parcellation of baby brain MRI is yet to be investigated.

For delineation of the boundaries of the white matter structures or tracts, Diffusion Tensor Imaging (DTI) has been applied (Oishi et al., 2012), since it provides various types of scalar images that could highlight well-aligned structures, such as the white matter bundles, even in neonatal brains in which the majority of the fibers are unmyelinated (Fig. 4). To parcellate baby brains at the finest granularity possible, a multimodality approach, in which co-registered T1-, T2-weighted, and DTI are used as the anatomical guides (Oishi et al., 2011c).

Fig. 4.

Comparison between neonatal brain (A) and adult brain (B). The T1- and T2-weighted images and the DTI are co-registered. DTI provides greater contrasts in the white matter areas, compared to T1- or T2-weighted images. This feature of DTI is particularly advantageous in identifying white matter bundles in the neonatal brain. The anterior limb of the internal capsule (alic, yellow contour) and the posterior limb of the internal capsule (plic, cyan contour) are almost invisible on conventional T1- and T2-weighted images of the neonatal brain, but these tracts are well visualized in the adult brain.

Currently, the number of modalities available for the brain parcellation is limited in babies, compared to older children or adults. For example, quantitative susceptibility mapping (QSM) is useful in delineating the boundaries of the metal-containing deep gray structures, such as the substantia nigra or the red nuclei (Lim et al., 2013), but has not yet been employed to parcellate the baby brains.

7.1.2. Parcellation map derived from an adult brain atlas

A common approach to avoid problems with the delineation of structural boundaries in baby brains is to adapt the existing adult brain atlases (Akiyama et al., 2013; Blesa et al., 2016; Kim et al., 2016; Shi et al., 2011). Typically, a selected adult atlas is transformed to pediatric brain MRIs, and then transformed to the baby brain MRIs to adjust the size and shape of the adult brain parcellation map to that of the baby brain. The resultant parcellation map is useful for comparing the anatomic features of each parcel at multiple ages, e.g., baby, adolescent, and adult, because parcellation maps derived from a single parcellation map are applied as a reference to evaluate the brains of all age groups. Using the same parcellation map avoids any bias derived from applying different atlases to each age group (e.g., differences between the groups might originate from using different atlases as the reference (see Section 9.3). However, a notable shortcoming of this approach is that accurate anatomical delineation is limited, especially for the baby brains, compared to the gold standard manual delineation.

7.2. Data-driven parcellation map

Advancement in MRI technology now allows the identification of anatomical structures based on the degree of myelination, or patterns of neuronal activities or connectivity (Fan et al., 2016; Glasser et al., 2016; Gordon et al., 2016; Kreilkamp et al., 2017), which do not necessarily follow conventional ontological boundaries. Brain parcellation based on such advanced MRI-derived features might represent the cytoarchitecture or functional substrates of the cortex better than conventional landmarks, such as gyri and sulci. An ambitious project that aims to create a comprehensive map of the human brain architecture and connectivity, the Human Connectome Project (HCP, https://www.humanconnectome.org/ and http://www.humanconnectomeproject.org/), has succeeded in creating cortical maps that comprise 180 parcels per hemisphere, based on cortical thickness, myelination, activation, or connectivity (Glasser et al., 2016). A machine-learning algorithm played an important role in determining the reproducibility of these brain parcellation maps. An example of a data-driven approach to the parcellation of baby brains is a resting-state, functional, connectivity-based clustering, which can achieve high signal homogeneity within the parcel (Shi et al., 2017). Although the neuroscientific or clinical significance of these maps still needs to be rigorously evaluated, MRI data-driven brain parcellation has great potential to advance our understanding of the human brain and its disorders. The application of such a data-driven approach to the parcellation of baby brains is one of the highly anticipated future research directions.

8. Baby brain MRI atlases that are currently available

In this review, we specifically focus on the role of baby brain atlases as knowledge to be referenced. In this context, the “atlas” refers to a 3D-brain image with a set of reference labels, which is a parcellation map, as the anatomic knowledge. Baby brain atlases that are open for public use, to date, through websites are listed below and summarized in Table 1. PubMed keyword search was used as an initial screen to capture relevant research articles, using the following search terms: “MRI” and “atlas” and “brain” and (“neonate” or “neonates” or “neonatal” or “fetus” or “fetuses” or “fetal” or “infant” or “infants” or “infantile”) and “English”[language] not “review.” Among 100 articles that passed this screen, articles related to a specific disease or syndrome, or non-human research, were omitted. Finally, we included only articles with the web address to download the atlas clearly described in the manuscript. At these websites, we further searched the institutes’ websites to include atlases that matched our definition. The Neuroimaging Informatics Tools and Resources Clearinghouse (NITRC) “find neuroimaging tools here” function (https://www.nitrc.org/) was also used with the following terms: “MRI” and “atlas” and (“fetus” or “neonate” or “infant”), to include atlases available through the NITRC. Note that templates without segmentation or parcellation are not included. To date, the parcellation map with the highest granularity is still achieved by a single-subject atlas, although there is a recent trend toward the creation of a multi-atlas repository with high granularity parcellation maps. In the following sections, “single atlas” refers to an age-representative atlas created based on group-representative neuroanatomical features that are characteristic of a specific age. Therefore, a so-called “4D atlas” that refers to a set of age-representative atlases across various ages was categorized as a single atlas. “Multi-atlas” refers to a set of single-subject templates with corresponding parcellation maps. The major features of each atlas are summarized in Table 1 and discussed below.

Table 1.

List of publically available baby brain atlases.

name of atlas	status	age-range	modalities*	number of participants*	number/types of parcels*
JHU-neonate-SS atlas	term-born	2days after birth	T1-w, T2-w, DTI	1	130
In utero 3D statistical atlas	fetus	20.57–22.86 GW	T2-w	14	4 (GM, WM, Germ, Vent)
Spatiotemporal atlas of the fetal brain	fetus	20.57–27.86 GW	T2-w	40	4 (GM, WM, Germ, Vent)
Dynamic 4D probabilistic atlas of the developing brain	preterm-born	29 to 44 PMW	T2-w	153	6 (Cx, WM, sGM, CSF, BS, Cbl)
Multi-channel 4D probabilistic atlas of the developing fetal brain	fetus	21.7 – 38.7 PMW	T2-w	80	4 (Cx, hemspheres, CSF, Vent)
Typical 6-Month-Old Infants Atlas	infants	177 –230 days	T1-w	60	AAL
Consistent high-definition spatio-temporal neonatal brain atlas	preterm-born	28 and 44	T1-w, T2-w	204	ALBERT
UNC Infant 0-1-2 Atlases	term-born	0 month - 2 years	T1-w, T2-w	95	AAL
UNC detail-preserved longitudinal 0-3-6-9-12 months-old atlas	term-born	0 month - 1 year	T1-w, T2-w	35	3 (GM, WM, CSF)
UNC 4D Infant Cortical Surface Atlas	term-born	1 – 72 months	surface	50	Desikan-Killiany and HCP multimodal
JHU-neonate-linear and JHU-neonate-nonlinear atlases	term-born	0–4days after birth	T1-w, T2-w, DTI	20 (DTI &T2-w), 14 (T1-w)	130
UNC/UCI neonate hippocampus amygdala multi-atlas	term-born	within 5 weeks after birth	T1-w, T2-w	6	7 (GM, WM, CSF, bil-Hippo, bil-amyg)
A label-based encephalic ROI template (ALBERT)	term and preterm	36 – 45 PMW	T1-w, T2-w	20	50
Edinburgh neonatal atlas (ENA33)	term-born	37 – 41 PMW	T1-w, T2-w, DTI	33	107 (SRI24/TZO)
M-CRIB atlas	term-born	40 – 43 PMW	T2-w	10	100 (Desikan-Killiany)
MRICloud neonate multiatlas repository	term and preterm	38–42 PMW	T1-w, T2-w, DTI	7	30

^*modalities and number of participants are based on the downloadable atlases and do not necessarily identical to the cited papers.

GW=gestational week, PMW=postmenstrual week, GM=gray matter, WM=white matter, Germ=germinal matrix, Vent=ventricle, Cx=cortex, sGM=subcortical gray matter, BS=brainstem, CSF=cerebrospinal fluid, Cbl=cerebellum, bil=bilateral, hippo=hippocampus, amyg=amygdala.

8.1. Single-subject atlas

JHU-neonate-SS atlas (Oishi et al., 2011c); http://cmrm.med.jhmi.edu/cmrm/Data_neonate_atlas/atlas_neonate.htm

This atlas is derived from a typically developing term-born baby scanned two days after birth. The size and shape is matched to that of a population-averaged, JHU-neonate linear template. The atlas consists of co-registered T1-, T2- weighted, and DTI with a parcellation map that includes 130 anatomic labels, comprising 54 cortical areas, 56 white matter structures, 14 deep gray structures, and six brain stem structures (total 130 structures) that cover the entire brain. The manually-drawn parcellation map follows the adult JHU-MNI parcellation map (Oishi et al., 2009), which enables the structural comparison between adult and neonate brains. The advantage is the multi-modal capability and the finest granularity among the current parcellation maps of baby brains. Group-averaged T1- and T2-weighted and DTI templates (JHU-neonate-linear and JHU-neonate-nonlinear) are also provided.

8.2. Single atlas with group-representative features

8.2.1. In utero 3D statistical atlas (Habas et al., 2010b); http://depts.washington.edu/bicg/research/fba.php

This is a single probabilistic atlas of tissue distribution, which was constructed from 14 clinical MRIs of fetal brains at 20.57–22.86 weeks gestational age. These fetuses were borne and developed normally after birth. The atlas consists of the average shape and intensity from a T2-weighted template, and four tissue probability maps, for the gray matter, white matter, germinal matrix, and the ventricle. The tissue probability maps were generated from a set of manual delineations. This atlas has now been extended to include a time domain to form a spatiotemporal atlas (8.2.2).

8.2.2. Spatiotemporal atlas of the fetal brain (Habas et al., 2010a); http://depts.washington.edu/bicg/research/fba.php

This is a temporal extension of the atlas described in 8.2.1, which was constructed from 40 normally developing fetal brains at 20.57–27.86 weeks gestational age. The atlas consists of the age-specific average shape and intensity from T2-weighted templates, and the age-specific tissue probability maps, for the gray matter, white matter, germinal matrix, and the ventricle. The tissue probability maps were generated from a set of manual delineations. The important knowledge and information provided by the atlas is the age-specific T2-weighted intensity and the tissue probability.

8.2.3. Dynamic 4D probabilistic atlas of the developing brain (Kuklisova-Murgasova et al., 2011); http://brain-development.org/brain-atlases/neonatal-brain-atlas/

This is a set of probabilistic atlases for neonates of 29 to 44 weeks postmenstrual age, created from the segmentations of 153 neonatal subjects at different ages. The atlas consists of the average T2-weighted intensity template and the corresponding tissue probability maps, with the correct sizes and shapes of the structures dynamically generated. The tissue probability maps are provided for six structures—cortex, white matter, subcortical gray matter, cerebrospinal fluid, brainstem, and cerebellum. The tissue segmentation was based on the intensity and parcellation of the subcortical gray matter, brainstem, and cerebellum, which was derived from manual parcellation of three representative images that were transformed to other images.

8.2.4. Multi-channel 4D probabilistic atlas of the developing fetal brain (Serag et al., 2012); http://brain-development.org/brain-atlases/a-multi-channel-4d-probabilistic-atlas-of-the-developing-fetal-brain/

This is a spatiotemporal atlas constructed from 80 fetuses with normal-appearing brains, between 21.7 and 38.7 weeks gestational age at the time of the scan. The downloadable atlas consists of group-averaged, T2-weighted image templates and tissue probability maps of the brain mask, cortex, hemispheres, cerebrospinal fluid space, and ventricles, for ages between 23–37 weeks of gestation. An automated tissue segmentation algorithm based on probabilistic atlas priors [8.2.3., (Kuklisova-Murgasova et al., 2011)] was used to create the tissue probability maps.

8.2.5. Typical 6-Month-Old Infants Atlas (Akiyama et al., 2013); http://ilabs.washington.edu/6-m-templates-atlas

This is a single probabilistic atlas constructed from 60 typically developing, term-born infants at 177 –230 days old in chronological age. The atlas consists of a group-averaged T1-weighted template, and the parcellation map propagated from the Automated Anatomical Labeling (AAL) map (Tzourio-Mazoyer et al., 2002), created on the Colin27 adult single-subject atlas (Collins et al., 1998).

8.2.6. Consistent high-definition spatio-temporal neonatal brain atlas (Makropoulos et al., 2016; Serag et al., 2012); http://brain-development.org/brain-atlases/consistent-high-definition-spatio-temporal-neonatal-brain-atlas/ and http://brain-development.org/brain-atlases/multi-structural-neonatal-brain-atlas/

This atlas is an extension of the fetal atlas (8.2.4) to neonatal brains between 28 and 44 weeks post-menstrual age at the time of the scan. The atlas consists of group-averaged T1- and T2-weighted MRI templates derived from 204 preterm-born neonates. The resulting 4D atlas can serve as a good representative of the population of interest as it reflects both global and local changes. The template also includes 87 labeled structures of the developing brain. The ALBERTs multi-atlas (Gousias et al., 2012) was used as prior knowledge for the segmentation.

8.2.7. UNC Infant 0-1-2 Atlases (Shi et al., 2011); http://www.med.unc.edu/bric/ideagroup/free-softwares/unc-infant-0-1-2-atlases and https://www.nitrc.org/projects/pediatricatlas

This is a set of spatio-temporal atlases for neonates, one- and two-year-old infants, and toddlers, created from 95 neonates who were scanned within five weeks after birth, and longitudinally scanned at one and two years of age. This longitudinal design is a significant feature that avoids the selection bias, which is inevitable in a cross-sectional design. The atlas consists of group-averaged T1- and T2-weighted images that include tissue probability maps of the gray matter, white matter, and the cerebrospinal fluid space, created using an automated image segmentation method applied to each image. This atlas also contains the parcellation map propagated from AAL (Tzourio-Mazoyer et al., 2002).

8.2.8. UNC detail-preserved longitudinal 0-3-6-9-12 months-old atlas (Zhang et al., 2016); https://www.nitrc.org/projects/infant_atlas_4d/

This is a set of spatio-temporal atlases age representative for 0-, 3-, 6-, 9-, and 12-month-old infants, created from 35 healthy infants who were longitudinally scanned every three months until one year of chronological age. The atlas consists of group-averaged T1- and T2-weighted images that include tissue probability maps of the gray matter, white matter, and the cerebrospinal fluid space. The major significance is an introduction of simultaneous temporal and spatial constraints to generate atlases with rich structural details and longitudinal consistency. An inclusion of 3-, 6-, and 9-month atlases is beneficial for analyzing infantile MRIs since the T1- and T2-weighted contrast between gray and white matter is still poor for most brain regions during the first year of life.

8.2.9. UNC 4D Infant Cortical Surface Atlas (Li et al., 2015a); https://www.nitrc.org/projects/infantsurfatlas/

This is a set of age-representative, spatio-temporal cortical surface atlases of 1, 3, 6, 9, 12, 18, 24, 36, 48, 60, and 72 months of chronological age, created from 339 longitudinally scanned MRIs from 50 healthy infants. The parcellation maps include those based on (Desikan et al., 2006) and on the HCP multi-modal parcellation (Glasser et al., 2016). The prominent feature is that the surface atlases have vertex-to-vertex cortical correspondence across all ages, which allows cross-sectional and longitudinal analysis.

8.3. Multi-atlas repository

One of the major problems in applying a single atlas (Sections 8.1. and 8.2.) as the reference is the existence of anatomic variations normally seen in developing brains. For example, the cavum septum pellucidum is a cavity between the left and right anterior horns of the lateral ventricles, surrounded by a pair of thin membranes, the bilateral septum pellucidum. This cavity is often seen in the neonatal brain, but typically disappears by four to six weeks of chronological age because the two septi pellucidum fuse during development (Fig. 5), although there are substantial variations in the timing of fusion and the degree of the residual cavity. If a single atlas with a prominent cavum septum pellucidum is used as the reference to label a target brain without a cavum, the topological difference could cause confusion in defining the labels for the anterior ventricles on the target brain. A set of multiple atlases that covers the entire landscape of anatomic variations has the potential to solve this issue. For example, the inclusion of atlases both with and without cavum septum pellucidum in the atlas repository enables the selection of the appropriate set of atlases within the repository as references with which to label the target image. Currently, several multi-atlas repositories are available, all derived from typically developing babies or babies with radiologically normal MRIs.

Fig. 5.

Example of neonates with and without the cavum septum pellucidum (indicated by red arrows)

8.3.1. UNC/UCI neonate hippocampus amygdala multi-atlas; https://www.nitrc.org/projects/unc_brain_atlas/ and https://www.med.unc.edu/psych/research/niral/download/download-data

This is a set of T1- and T2-weighted images from six typically developed neonates scanned within five weeks after birth, which include segmentation of the gray matter, white matter, and the cerebrospinal fluid space, as well as the hippocampi and amygdalae.

8.3.2. A label-based encephalic ROI template (ALBERT) (Gousias et al., 2012); http://brain-development.org/brain-atlases/neonatal-brain-atlas-albert/

This is a set of baby brain atlases at different ages, which includes T1- and T2-weighted images from 20 term- and preterm-born infants scanned at postmenstrual weeks 36 – 45, with a parcellation map that consists of 50 structures that are manually delineated.

8.3.3. Edinburgh neonatal atlas (ENA33) (Blesa et al., 2016); http://brainsquare.org/ and http://www.brainsimagebank.ac.uk/

This is a set of co-registered T1- and T2-weighted images, DTI, and the parcellation map of 33 typically developing neonates born at term, who were scanned at 37 – 41 postmenstrual weeks. The parcellation map consists of 107 anatomic regions, which were propagated from an adult brain atlas (SRI24/TZO) (Rohlfing et al., 2010); https://www.nitrc.org/projects/sri24). This has the finest granularity among the current multi-atlas repositories of baby brains.

8.3.4. M-CRIB atlas (Alexander et al., 2017); https://github.com/DevelopmentalImagingMCRI/M-CRIB_atlas

This is a set of T2-weighted images from 10 healthy term-born neonates, accompanied by parcellation maps that replicate the Desikan-Killiany adult cortical atlas, which parcellates the brain into 100 regions. The parcellation was performed manually. Group-averaged T2-weighted structural templates are also provided.

8.3.5. MRICloud neonate multiatlas repository (Otsuka et al., 2017); http://lbam.med.jhmi.edu/ and www.MRICloud.org

This is a set of T1-weighted images of seven healthy term-born neonates, accompanied by parcellation maps that replicate the MRICloud adult atlas repository, which parcellates the whole head into 30 regions. The parcellation was performed manually. The whole-head coverage allows the quantification of intracranial volume, which is not available from other baby atlases.

9. Common issues in the application of baby brain MRI atlases

Due to lack of clear anatomical landmarks, the missing information of parcellation of early developing brain before or around middle fetal stage (e.g. Yu et al., 2016; Song et al., 2017) warrants further research. Below are common issues in the application of baby brain in late fetal stage and around birth.

9.1. Methods by which atlases are applied to segment or parcellate baby brains

Numerous methods can incorporate parcellation maps on each atlas to segment or parcellate baby brain MRIs from each individual. The choice of the method depends on the purpose of the study, the structures of interest, and the anatomical features of the target population (e.g., radiologically normal brains vs. diseased brains). Technology and algorithm development to achieve optimal segmentation and parcellation with combined atlases is an active research field in biomedical engineering. Details of this area of research are, however, outside the focus of this review, but readers who are interested in this topic are encouraged to read the most recent comprehensive review (Makropoulos et al., 2017).

9.2. Issues in cross-sectional analysis and atlas selection

Rapid brain development during the perinatal period must be considered in the cross-sectional design. Even two weeks difference in the postmenstrual age is enough to make a significant change in brain anatomy that is detectable through MRI (Serag et al., 2012). For group comparison studies, accurate matching of the postmenstrual ages between groups and the selection of an age-appropriate brain atlas to be used are more important than those for adult studies. Hence, chronological age is not recommended for group comparison studies of neonates (Section 5.), although it could be included as a co-variate in the study design to evaluate the effects of the postnatal environment, which has a significant impact on neurodevelopment (Wu et al., 2017a, b). The atlas-of-choice in designing a cross-sectional study to investigate the effects of neurodevelopment is also an important issue, because brain images from a wide age-range are often included in such studies. Using the mean or median age of the study population to choose an age-appropriate atlas is a possible solution, but this may lead to a greater risk for misregistration of the images from babies younger or older than the selected age. This might be a potential problem because, while the effect of age is of scientific interest, the magnitude of misregistration might also relate to the age differences. Although highly elastic transformation or diffeomorphic transformation might ameliorate the effect of misregistration, this atlas selection issue has not been fully investigated thus far. Potential approaches to solve this issue include the use of spatiotemporal atlases. The spatiotemporal atlases introduced in Section 8.2., especially the atlases created with spatial and temporal constraints (8.2.8. and 8.2.9), provides the necessary information for the analyses of baby brain MRIs from a wide age-range, because these atlases offer the ability to converge images that have been normalized to the age-appropriate atlases to a single atlas space. A multi-atlas approach or other automated image parcellation/segmentation approaches are alternatives to solve this issue. These approaches are used to parcellate/segment brain MRIs on each individual space, rather than normalizing all images to a common template space, by transforming multiple atlases or local patches to the target brain. A label fusion method follows this process to obtain hard boundaries that define the anatomical structures-of-interest. Robustness to the age-range is ensured by including multiple-atlases or local patches that cover the age-range of a study population. To accomplish this, multiple atlases with high-quality image parcellation, or tissue priors for various age-ranges are needed.

9.3. Issues in longitudinal analysis

Longitudinal analysis is recommended for the investigation of neurodevelopmental trajectories. However, several issues remain regarding the selection of an atlas derived from repeated scans at different time-points. Similar to the issue raised in cross-sectional analysis (5.1.), the atlas-of-choice is not always straightforward because there is no guarantee that a single atlas for a specific age can accommodate brains at different ages over time. One of the ways to ameliorate this issue in adult studies is to create a customized template derived from all time-points for each individual (Reuter et al., 2012), although the application to this approach to study the developing brain has not been rigorously tested, to date. Another issue is the existence of developmental stage-specific anatomic structures. For example, the ganglionic eminence is visible in most brain images at 28 postmenstrual weeks of age, but mostly disappears at term. The selection of a single atlas as a reference in studies that follow babies throughout this period could be potentially biased by the atlas selection because some of the anatomic structures visible in the atlas-of-choice and those in a target image might be different. The spatiotemporal atlases, multi-atlas approach, or other automated image parcellation procedures are, again, a potential solution, as mentioned in 9.2. However, an atlas selection bias might affect the results of multi-atlas-based image quantification if an automated atlas selection algorithm (Jingbo et al., 2015) is applied. Namely, a different set of atlases could be selected to parcellate brain MRIs scanned at different time-points. In such an instance, the difference in the selected atlas might affect the difference in image parcellation, leading to erroneous results for the growth trajectories being studied. Using multiple atlases would also increase the computational burden. The number of atlases needed to obtain robust image parcellation depends on the variations seen in the structure, which also may depend on the variations seen in a study population. A method designed to estimate the appropriate number of atlases for robust and accurate image parcellation needs to be investigated.

9.4. Issues in atlas modality selection

The choice of atlas modalities depends on the image parameters of interest, which depend on the biological processes of interest. For example, research targeting white matter development often employs DTI, because the diffusion property of water in the brain parenchyma sensitively captures neurodevelopmental changes, including axonal fiber organization or orientations, pre-myelination, and myelination (Dubois et al., 2014). An issue related to the limited availability of DTI atlases (Oishi et al., 2011c) is that non-DTI templates, such as T2-weighted images or b0 images, are often used to normalize DTI. Since the majority of algorithms use image contrast to guide image transformation, the accuracy of the structural co-registration is not guaranteed if there is little or no contrast between different brain structures (Oishi et al., 2012). This inaccuracy could be a problem when T2-weighted contrast is used to transform DTI to the atlas space, since the homogeneous appearance of myelinated or unmyelinated white matter areas on T2-weighted images contain various fiber bundles with different directions, which might not be adequately co-registered after normalization (Fig. 6). Currently, only T1- and T2-weighted and DTI atlases are available for baby brains, which cannot provide accurate analysis of other modalities, such as QSM.

Fig. 6.

Comparison between T2-based image transformation and DTI-based image transformation. The original FA map from a 40-postmenstrual-weeks neonate is shown in the upper row, left. T2-LDDMM: The original b0 image was transformed to the JHU-neonate-b0 template, and the resultant transformation matrix was applied to the original tensor field, from which the transformed FA map was calculated (upper row, middle, with the magnified view in the lower row, left). DTI-LDDMM: The original FA map was transformed to the JHU-neonate FA template, and the resultant transformation matrix was applied to the original tensor field, from which the transformed FA map was calculated (upper row, right, with the magnified view in the lower row, right). Both images are overlaid by the parcellation map that qualitatively demonstrated the registration accuracy. Note that the anterior limb of internal capsule (ALIC, high FA string indicated by red arrowheads) was not well co-registered to the atlas space when the T2-LDDMM was applied, but was well co-registered when the DTI-LDDMM was applied.

9.5. Multimodality integration based on parcellation maps

9.5.1. Co-registration between multi-contrast images

Multimodality images could be analyzed using a common anatomic framework based on the structural parcellation map in an atlas space or in each individual’s space (Faria et al., 2012). Each modality is usually acquired with different image contrasts, resolutions, and matrix sizes. In order to collectively use the multi-modality information, co-registration must be performed in the first step, using mutual information-based image transformations (Nam et al., 2009). Accurate co-registration is also important to generate a myelin map (Glasser and Van Essen, 2011), which could be calculated from co-registered T1- and T2-weighted images. The intensity inhomogeneity seen in T1- and T2-weighted images must be corrected to perform better tissue segmentation. However, the intensity inhomogeneity correction might also affect the results of the myelin map, which is a T1-weighted/T2-weighted ratio map. There is an ongoing effort to optimize the image processing workflow to obtain biologically appropriate myelin maps (Ganzetti et al., 2014).

9.5.2. Multi-contrast image parcellation

The lower contrast-to-noise ratio and signal-to-noise ratio of baby brain MRIs, compared to that of adult brain MRIs, is a major problem in achieving accurate parcellation of brain structures. Since different modalities provide complementary information that defines the boundaries of the structural definitions, structural parcellation based on multiple contrasts might provide a solution for this problem, especially for baby brains. For example, although 3D T2-weighted images often provide higher resolution and sharper structural definitions, white matter fiber bundles with different orientations, which are not visible on T2-weighted images, are clearly delineated on DTI (Fig. 4). A state-of-the-art, multi-contrast, multi-atlas parcellation method (Tang et al., 2014) has the potential to improve the accuracy of image parcellation, but needs to be rigorously tested on baby brains. As a prerequisite for this method, a multiple-contrast, multi-atlas template for the baby brain is required.

9.6. Differences between fetus and preterm-born babies

The applicability of the same brain atlas for both fetuses and preterm-born babies of identical biological age (age counted from time of conception) is unknown. Considering the huge differences in image acquisition, motion during the scan, and different structures surrounding the head of fetuses versus babies, different atlases are needed. Moreover, differences in brain anatomy also exist between fetuses and preterm-born babies. Brain development might be affected by interactions between immaturity at birth, as shown by slower or delayed neurodevelopment, and accelerated development after birth that might be triggered by environmental stimuli (e.g., light, sound, smell, taste, and somatosensory inputs and increased oxygenation from respiration after birth). Importantly, such interaction effects vary depending on the anatomical structure (Wu et al., 2017a, b).

10. Current limitations

10.1. Limited number of shared multimodal multi-atlas libraries in the research community

Manual parcellation of anatomic structures requires in-depth knowledge about both neuroanatomy and neuroradiology. As the number of parcels and image resolution increase, the labor and cost for the manual parcellation increases. Currently, the largest library for the manually parcellated MRI atlases is the Johns Hopkins University MRICloud database (www.MRICloud.org), which contains more than 200 atlases derived from T1-weighted images and 50 DTI atlases that encompass brain measurements of humans from 3 – 85 years of age, and seven neonatal T1-T2-DTI multi-contrast atlases. Although this is a wide range of ages with a relatively large number of subjects, the number of baby brain atlases is still limited, and no atlas is available for subjects between 0.5 and 3 years of age. The lack of atlases for this age-range is mainly due to the difficulty of manual parcellation for these images, as well as the difficulty in obtaining high-quality, motion-free images during the image acquisition of typically developed toddlers without sedation. For the baby brain atlas, only a few repositories are available (Sections 8.3.1to 8.3.5). These atlases are based on babies without any medical or neurological deficits and no radiological abnormalities. The construction of a shared repository that contains more atlases with more modalities is needed to encompass the variations seen in typically developing brains. The necessity of baby brain atlases derived from babies with brain disorders is discussed in Section 11.

10.2. “Ground truth” of MRI atlases

The lack of a ground truth is a major issue for in vivo brain MRI atlases. For example, little is known about whether the boundary between gray and white matter structures visible on T1- and T2-weighted MRIs is identical to that defined by the cytoarchitecture. This issue is especially problematic for brains undergoing development. The gray-white contrast on MRI is mainly determined by the amount of lipid present in each brain region, which is more abundant in the white matter than the gray matter of adult brains, but is the opposite in neonatal brains. The rich lipid content in myelin is the major determinant of this contrast, and active postnatal myelination causes the inverted contrast that is observed during infancy. Cortical thickening during infant development might, therefore, be affected by both histological changes, true cortical thickening, and by ongoing myelination that occurs around the border of the gray and the white matters. An MRI-histological comparative study is required to enrich our knowledge about MRI-based brain anatomy, which is currently ongoing for the study of adult brains (Amunts et al., 2013). Such a comparative study in neonates and infants will expand the field of human neurodevelopment. Currently, efforts are ongoing to create histology-MRI atlases for developing non-human primates (Hikishima et al., 2013). A cross-species comparison would be one of the expected directions in this field.

10.3. Resolution of MRI atlases

High-resolution MR images, with high contrast-to-noise ratio sequences, are needed to create baby brain atlases with clear anatomical delineations. Since the baby brain is smaller than the adult brain, multiple structures may be included within each pixel, which may cause partial volume effects and blurring of the structural boundaries on the baby brain images, more so than those in the adult brain images, even when the image resolution is the same. The resolution of baby brain MRI is mainly constrained by two factors: scan time and body weight. In general, if the same scan sequence is used, a longer scan time is needed to acquire an image with higher resolution without sacrificing the signal-to-noise ratio. This relationship between scan time and resolution is a big dilemma, because a shorter scan time is desirable for acquiring images with less motion artifacts, especially when sedation is not used. Recent advancements in fast scanning protocols and prospective real time motion-correction methods during the scan acquisitions, or for post-processing, are addressing this issue (Cordero-Grande et al., 2017; Gumus et al., 2014; Herbst et al., 2017; Herbst et al., 2015; Hughes et al., 2017; Kuklisova-Murgasova et al., 2012; Li et al., 2015b; Singh et al., 2015; Zahneisen et al., 2016; Zahneisen et al., 2014). Many of these techniques are being applied to ongoing projects, including the developing Human Connectome Project, which targets preterm and term-born neonates 20 to 44 weeks postmenstrual age. The small body size of babies is another issue, especially when scanning babies that are small-for-gestational-age, since the Specific Absorption Rate (SAR) limits the scanning time for babies. The SAR is defined by watts per kilogram; the upper limit of which for babies is less than 3W/kg/10 min exposure averaged (US Food and Drug Administration). Therefore, high-resolution images are difficult to obtain from babies with low body weight because higher resolution would require a longer scan time, which is accompanied by an increase in the SAR.

11. Expansion of the role of atlases for future clinical applications

The biggest challenge in the clinical application of atlases is the heterogeneity of the clinical images. The source of heterogeneity can be classified into two major causes: technical and biological heterogeneity. The technical heterogeneity refers to variations derived from different scanners and scanning protocols, and the biological heterogeneity refers to variations in age, sex, race, diseases the babies might have suffered from, and the severity of the disease. Biological heterogeneity is a particular important factor in assessing the baby brain, due to the rapid and structure-specific nature of developmental changes and the variations in the signal and volumes of brain structures, as well as variations in pathological changes related to pediatric diseases. An atlas created based on clinically and radiologically normal individuals may not cover all of these variations. To account for the heterogeneity, multiple atlases that cover the entire landscape of the heterogeneity would be needed to ideally serve as the anatomical references, which are unavailable currently. The Bayesian template of neonates, which was generated from a single atlas with a parcellation map, and modified based on the anatomic features of the target population, might be applied for the analysis of clinical images (Zhang et al., 2014a), as long as the clinical images share the same set of anatomic structures with the atlas. For diseases with lesions that do not exist in typically developing brains (e.g., tumors) or lack structures that exist in typically developing brains (e.g., malformations), it is critically important to have an atlas repository that contains a huge amount of imaging data with the associated clinical information. The role of atlases in this case is more than the traditional role of the atlas as a geographical reference. Rather, the role of the atlas repository expands as a clinical reference or teaching file (Qin et al., 2013; Wu et al., 2016), from which statistics about clinically relevant information, such as diagnosis, prognosis, or response to treatments, could be obtained (Miller et al., 2013; Mori et al., 2013). Fig. 7 demonstrates the concept of a brain atlas repository with brain disorders. In this example, an atlas repository stores a set of MRI images containing various types and stages of diseases scanned using various parameters. Atlases with features similar to a given patient would be clustered and retrieved from the repository, from which statistics about potential diagnoses could be obtained. Again, what is important here is the clinical labeling (diagnosis, prognosis, or response to treatment) that accompanies the anatomical labeling (parcellation map). To rigorously test this concept, a database of anonymized clinical images with a wide variety of diagnoses, which is shared within the research community, is highly desirable.

Fig. 7.

Concept of the disease atlas library. Each MRI within the library has non-image information, such as demographics (e.g., sex and age), clinical (e.g., diagnosis, prognosis, and responsiveness to treatments), or genetic (e.g., gene mutation or SNP) information attached. Suppose there are N atlases within the library. New images with the non-image information attached are mixed into the library (now N +1) and clustered based on the image and non-image features. The new images would iteratively be directed to the atlases (stored in the library) that has similar image and non-image features. From the atlases within the same cluster, probabilities and statistics about possible diagnosis or prognosis can be calculated (red arrows). This algorithm is similar to the multi-atlas label fusion methods.

12. Summary

Brain atlases play an essential role as references or teaching files to help researchers or clinicians interpret their findings according to a priori knowledge about brain anatomy. The baby brain is in a state of active neurodevelopment and is one of the frontiers in neuroscience; therefore, the construction of various types of baby brain atlases is ongoing. Multi-atlas repositories of baby brains with fine granularity parcellation maps are becoming available, but several issues related to the creation, selection, and application of baby brain atlases still remain. The lack of brain atlases for disease conditions is a major hurdle in clinical applications. The potential to include clinically relevant information into the atlas as clinical labeling, in addition to conventional anatomic labeling, is critically important and may lead to improved clinical diagnosis and treatment monitoring or disease prognosis.

Highlights.

• The types and roles of the human baby brain MRI atlases that are currently available are described and discussed.

• The core value of an atlas is that it provides a priori anatomical knowledge to serve as a teaching file of brain anatomy.

• One atlas is not sufficient to encompass or capture all the spatial-temporal variations in normal brain anatomy and the possible alterations.

• The potential of disease-based atlases to characterize clinically relevant information is discussed.