A normative spatiotemporal MRI atlas of the fetal brain for automatic segmentation and analysis of early brain growth

Abstract

Longitudinal characterization of early brain growth in-utero has been limited by a number of challenges in fetal imaging, the rapid change in size, shape and volume of the developing brain, and the consequent lack of suitable algorithms for fetal brain image analysis. There is a need for an improved digital brain atlas of the spatiotemporal maturation of the fetal brain extending over the key developmental periods. We have developed an algorithm for construction of an unbiased four-dimensional atlas of the developing fetal brain by integrating symmetric diffeomorphic deformable registration in space with kernel regression in age. We applied this new algorithm to construct a spatiotemporal atlas from MRI of 81 normal fetuses scanned between 19 and 39 weeks of gestation and labeled the structures of the developing brain. We evaluated the use of this atlas and additional individual fetal brain MRI atlases for completely automatic multi-atlas segmentation of fetal brain MRI. The atlas is available online as a reference for anatomy and for registration and segmentation, to aid in connectivity analysis, and for groupwise and longitudinal analysis of early brain growth.

Figure 1

The preprocessing steps in fetal brain MRI analysis: (a) shows the out-of-plane view (coronal view) of an original axial T2wSSFSE scan, (b) is the volumetric image obtained from iterations of inter-slice motion correction and robust super-resolution volume reconstruction, (c) shows the brain mask obtained through supervised levelset segmentation and manual refinement, and (d) is the reconstructed image, reoriented and co-registered to the common atlas coordinate space after N4 bias field correction and intensity normalization.

Figure 2

Frequency distribution of subjects contributed to atlas construction at each gestational age point in weeks. The number of subjects used in atlas construction at lower GAs was, on average, smaller than the numbers at higher GAs, which was acceptable as the fetal brain has less features and variability at lower GAs compared to higher GAs.

Figure 3

The procedure to generate atlas labels and segmentations: Step 1: labels and tissue-type segmentation of 20 neonatal ALBERTS atlases^, were refined manually. Step 2: The segmented neonatal atlases were used to generate initial labels on the spatiotemporal fetal brain MRI atlas at higher GAs (35–37 weeks) through multiatlas segmentation using probabilistic label fusion. Step 3: Fetal brain MRI labels were manually defined and propagated in iterations from the higher GAs to the lower GAs. Step 4: Atlases within one week of any query subject were used to generate an initial segmentation of that subject. The initial segmentations of all subjects within one week of the query subject were used to segment that subject.

Figure 4

The spatiotemporal fetal brain MRI atlas (CRL fetal brain atlas) at six representative GAs: 22, 25, 28, 31, 34, and 37 weeks. Axial, coronal, and sagittal views of the atlas have been shown at each age point. Note that the spatiotemporal atlas construction process is a time-continuous process, therefore the atlas can be constructed at any continuous age point within the age range of the subjects used in the atlas construction process.

Figure 5

Visual comparison of our spatiotemporal fetal brain MRI atlas (A) and the atlas from brain-development.org (B). We did not register or resample the atlases to avoid artificial blur; we instead tried to compare the closest planes in each view at three ages (24 weeks, 30 weeks, and 36 weeks GA). The images in (B) are generally smoother than those in (A) but lack anatomic details compared to the images in (A). Red circles and markers point at some of the areas with relatively blurred anatomy on the atlas from brain-development.org but with more details on the CRL fetal brain atlas. Both atlases are available online.

Figure 6

Tissue segmentation and structural labels defined and overlaid on the spatiotemporal fetal brain MRI atlas at six representative GAs: 22, 25, 28, 31, 34, and 37 weeks. The labels visible on these axial sections include the developing white matter, cerebrospinal fluid (CSF), corpus callosum, and left and right gray matter (cortical plate), ventricles, thalami, hippocampi, and lenticular and caudate nuclei.

Figure 7

Visual assessment and comparison of atlas-based segmentation of fetal brain MRI using multiple atlases: (1) segmentation using the spatiotemporal fetal brain MRI atlas (STA), (2) segmentation using the combination of the spatiotemporal atlas and individual-subject atlases (STA + ISA), and (3) reference standard. The circles and squares point at some of the areas in which the methods performed differently. Overall the results were satisfactory. As expected due to the use of a larger number of atlases, slightly better performance was observed for STA + ISA. For quantitative comparison of methods (STA and STA + ISA) we relied on the analysis of average DSC metrics reported in Table 1. Labels that are visible on these images are CSF, corpus callosum, developing white matter, and left and right cortical plates, ventricles, thalami, hippocampi, amygdalae, and caudate nuclei.

Figure 8

DSC metrics computed for brain tissue and structures for leave-one-out automatic multi-atlas segmentation evaluation applied to seven subjects with reference manual segmentations (the test set), averaged in three GA ranges: less than 27 weeks, between 28 and 34 weeks, and more than 35 weeks. Overall, the results indicated that relatively accurate automatic segmentation of the main brain structures and tissue were achieved by using the developed spatiotemporal fetal brain MRI atlas, symmetric diffeomorphic deformable registration for label propagation, and probabilistic STAPLE for label fusion. Labels are: Thalam L: left thalamus, Thalam R: right thalamus, CC: corpus callosum, Vent L: left lateral ventricle, Vent R: right lateral ventricle, Brainstem, CP L: left cortical plate, CP R: right cortical plate, WM L: left developing white matter, WM R: right developing white matter, and CSF: cerebrospinal fluid.