A new method for structural volume analysis of longitudinal brain MRI data and its application in studying the growth trajectories of anatomical brain structures in childhood
Introduction
Structural age-related change measurements from brain magnetic resonance imaging (MRI) are crucial to the study of normal brain growth and to understanding the biological process underlying brain development. They are also important in assessing the impact of neurological disorders and neurodegenerative diseases on brain growth.
Whereas a cross-sectional study involves only one structural measurement per subject, in a longitudinal study, multiple measurements are taken per subject. Indeed, subjects in cross-sectional MRI studies are scanned only once, while those in longitudinal studies are scanned repeatedly over time. Thus, longitudinal studies allow the observation of individual patterns of change.
To study age-related changes in cerebral structures during childhood, previous studies used either cross-sectional (Lange, 2011) or longitudinal brain MRI scans (Giedd et al., 1999, Lenroot and Giedd, 2006). In studies that used longitudinal pediatric brain MRI scans, each scan of a specific subject was analyzed independently. The structural brain measurements were computed without considering the longitudinal consistency between the scans of a single subject.
Recent methods reduce within-subject variability by taking into account longitudinal consistency whether for registration or tissue segmentation (Reuter et al., 2012, Wu et al., 2012).
Longitudinal registration was first proposed by Shen and Davatzikos (2004), who computed deformations between longitudinally acquired subject scans and a four-dimensional (4D) template. Thereafter, methods using consistent diffeomorphic registration of longitudinal images were presented (Durrleman et al., 2009, Lorenzi et al., 2010), and more recently, the creation of a subject-specific linear template was introduced (Reuter et al., 2012).
Regarding longitudinal classification, one proposed method incorporated longitudinal consistency constraints in a 3D fuzzy clustering segmentation (Xue et al., 2006), while another presented 4D image segmentation with a graph cut algorithm (Wolz et al., 2010).
Inspired by this previous work, we introduce a new method to measure structural volume changes in longitudinal MRI scans in which longitudinal information is used for both registration and segmentation. First, we propose the creation of linear and nonlinear subject-specific templates. Each time point is registered to the subject-specific template, which is registered to the population template, thus making the registration of time points to the population template more consistent. Second, we combine this registration with a 4D expectation-maximization (EM) algorithm for tissue classification, using a priori classes computed by averaging the tissue classes of all time points obtained cross-sectionally. This step allows us to take advantage of the longitudinal consistency of the classification.
The goal of the present paper is twofold. First, we study how longitudinal registration and longitudinal classification improve the longitudinal measurements. The longitudinal pipeline is compared with both a cross-sectional pipeline and a hybrid pipeline that combines longitudinal registration with cross-sectional tissue classification. These three pipelines are applied to a scan–rescan database to study the variability in the measurements and to the National Institutes of Health (NIH) MRI study of normal brain development (Evans, 2006) to study the impact of the variability in measurements on the gray matter (GM) growth curve models.
Second, we apply the longitudinal pipeline to report the growth trajectories of different anatomical brain structures in childhood using the NIH pediatric MRI database. We compare these growth trajectories with previously reported trajectories obtained from cross-sectional and longitudinal datasets.
Experimental results using the scan–rescan data as well as the longitudinal data from a large ensemble of subjects show that the variability of segmented volumes decreases by half as more consistent priors are used for tissue classification across all time points when using the proposed longitudinal image processing procedures.
Section snippets
Longitudinal method (LL)
The longitudinal pipeline (“LL method”: Longitudinal registration and Longitudinal classification) comprises the registration framework and classification framework defined below. Fig. 1 shows a flowchart detailing the steps involved in the LL method and the volume and registration notations used below.
Cross-sectional and hybrid methods
The LL method (longitudinal registration and classification) consists in applying the pipeline detailed above in its entirety. To study the impact of the longitudinal registration and classification steps, we compared the LL method with a standard cross-sectional framework (CC method: cross-sectional registration and cross-sectional classification) and a hybrid framework (LC method: longitudinal registration and cross-sectional classification).
The CC method is the standard pipeline. Each time
Data
A set of four scan–rescan datasets of T1w data from 20 young normal subjects was used (the 20NC-4 V database). The MRI data were acquired with a 1.5 T Siemens Sonata Vision clinical scanner (Siemens Medical Systems, Erlangen, Germany), using the standard head coil. Subjects were immobilized with a head restrainer. The protocol comprised four conventional whole-head high-resolution T1w scans. Scans were acquired using a 3D spoiled gradient echo sequence [TR = 22 ms, TE = 9.2 ms, α = 30 deg], providing
Scan–rescan reliability — 20NC-4 V database
Fig. 3 shows the GM classes obtained with the 3 methods (left:CC, middle:LC, right:LL) on two visits (V1 and V4) on the same subject (randomly chosen) of the 20NC-4 V database. In purple are the voxels of the GM class identical for the two scans, in blue are the voxels of the GM class in scan 1 but not in scan 4 and in red are the voxels of the GM class in scan 4 but not in scan 1. Ideally, given this scan–rescan data, no changes are expected, so no red and blue voxels should be found. We can
Discussion
The scan–rescan experiments with the 20NC-4 V data demonstrate that the LL pipeline, and the LC method to a lesser extent, give smaller variance and reduced bias in comparison with the CC pipeline. These results indicate that the processing pipeline is stable and reduces noise associated with cross-sectional analysis by taking advantage of the longitudinal consistency of the data. The within-subject alignments improve the registration consistency between each time point and the ICBM152 template.
Conclusion
We proposed a longitudinal method to analyze longitudinal brain MRI datasets and to measure structural age-related changes. The framework comprises two innovative parts: a longitudinal registration step and a longitudinal classification step. The results showed that both steps of the longitudinal framework reduce the variability in and improve the accuracy of the measurements. We also applied the method to the NIH pediatric dataset and reported growth trajectories of some brain structures that
Acknowledgments
Canadian Institutes of Health Research (MOP-111169 & 84360) and les Fonds de Recherche Santé Québec.
Data used in the preparation of this article were obtained from the NIH Pediatric MRI Data Repository created by the NIH MRI Study of Normal Brain Development. This is a multisite, longitudinal study of typically developing children from ages newborn through young adulthood conducted by the Brain Development Cooperative Group and supported by the National Institute of Child Health and Human
References (28)
- et al.
A fully automatic and robust brain MRI tissue classification method
Med. Image Anal.
(2003) - et al.
Robust Rician noise estimation for MR images
Med. Image Anal.
(2010) - et al.
BEaST: brain extraction based on nonlocal segmentation technique
NeuroImage
(2012) The NIH MRI study of normal brain development
NeuroImage
(2006)- et al.
Unbiased average age-appropriate atlases for pediatric studies
NeuroImage
(2011) - et al.
Brain development in children and adolescents: insights from anatomical magnetic resonance imaging
Neurosci. Biobehav. Rev.
(2006) - et al.
Within-subject template estimation for unbiased longitudinal image analysis
NeuroImage
(2012) - et al.
Measuring temporal morphological changes robustly in brain MR images via 4-D template warping
NeuroImage
(2004) - et al.
Measurement of hippocampal atrophy using 4D graph-cut segmentation: application to ADNI
NeuroImage
(2010) - et al.
Registration of longitudinal brain image sequences with implicit template and spatial–temporal heuristics
NeuroImage
(2012)
CLASSIC: consistent longitudinal alignment and segmentation for serial image computing
NeuroImage
Geometric means in a novel vector space structure on symmetric positive-definite matrices
SIAM J. Matrix Anal. Appl.
Consistent groupwise non-rigid registration for atlas construction
Real longitudinal data analysis for real people: building a good enough mixed model
Stat. Med.
Cited by (129)
From mother to child: How intergenerational transfer is reflected in similarity of corticolimbic brain structure and mental health
2023, Developmental Cognitive NeuroscienceBrain regional homogeneity and its association with age and intelligence in typically developing youth
2023, Asian Journal of PsychiatryAn open-source tool for longitudinal whole-brain and white matter lesion segmentation
2023, NeuroImage: ClinicalVentricular features as reliable differentiators between bvFTD and other dementias
2022, NeuroImage: Clinical