Validation of a brain MRI relaxometry protocol to measure effects of preterm birth at a flexible postnatal age

This methodological development study made use of a subset of 52 infants with MRI performed within 2 months of term equivalent (defined as 40 weeks postmenstrual age). These infants were part of a prospective, observational study of preterm and term infants at the Monroe Carell Jr. Children’s Hospital at Vanderbilt in Nashville, TN between November 2004 and July 2010. Infants in the current subset were recruited from the Vanderbilt NICU. The goal of the larger study was an observation of common nutritional exposures and included older children born at term who were not part of the current cohort, and were recruited from the Children’s Hospital pediatric radiology outpatient schedule. Ethical approval for this research was obtained from the Vanderbilt Institutional Review Board (protocol #07077 and 04881). All parents of participating infants signed Vanderbilt IRB-approved informed consent documents prior to MRI. Excluded infants were those not expected to survive to the age of 3 months, intubated infants whose clinical instability increased the risks associated with transport from the NICU for the MRI procedure or those with a confirmed diagnosis of hypoxic- ischemic encephalopathy. A pediatric neuroradiologist reviewed clinically indicated MRIs to exclude those with white matter injury or lesions in the deep nuclei.

Gestational age at birth was ascertained from the best obstetric estimate, based on the mother’s last menstrual period, obstetric measurements, and ultrasound measurements taken in the first trimester of pregnancy [19]. Clinical data were obtained from the medical record.

Subjects were placed in a MRI compatible papoose (Universal Medical, Norwood, MA) to minimize movement artifact following published recommendations [20] and scanned on a Philips Achieva 1.5 T MRI scanner (Philips, Amsterdam, Netherlands). Infants were swaddled and fed prior to MRI when possible. Those who fell asleep were scanned without sedation. Those who were fussy received oral chloral hydrate. None of the infants were intubated or given narcotics for the purposes of the study scan. T2-weighted images (dual echo, turbo spin echo, TR = 3 s, TE = 16.8 and 210 ms, in-plane resolution 0.7 × 0.7 mm, number of slices = 18, slice thickness = 4 mm, slice gap = 1 mm, number of acquisitions = 3, echo train length = 16, scan time = 6:06) and T1-weighted images (inversion-prepped turbo spin echo, TR = 8.3 s, TE = 15 ms, in-plane resolution. 0.7 × 0.7 mm, number of slices = 18, slice thickness = 4 mm, slice gap = 1 mm, number of acquisitions = 2, echo train length = 16, scan time = 3:03, inversion delays T1 = 300, 500, 700, 1000, 1500, 2000, and 2500 ms) were acquired. Images were oriented in coronal planes for all subjects.

Tissue T2 relaxation time was calculated for each image voxel, assuming monoexponential signal decay. Tissue T1 was calculated using a two-parameter fit for T1 relaxation time and equilibrium magnetization. First, complex image data were phase-corrected using one of the scans (usually the scan with the shortest inversion delay) as a phase reference. The signal amplitude was then fit to a standard inversion recovery function [21]. Ideally, each inversion delay time (T1) contributed one data point to the fit. However, in some cases, subject motion corrupted the images, producing strong ghost artifacts. The entire image volume was discarded when motion-induced ghost artifacts corrupted the images. If the number of time points with uncorrupted data fell below four, no T1 fit was attempted. All calculations were performed using the MATLAB (The Mathworks, Natick, MA) programming language.

Maps of T1 and T2 relaxation times were aligned in a common space to improve the accuracy of comparisons across subjects. One subject (Figure 1A) was chosen to define a reference set of brain images (the remaining T1 and T2 data for this subject were discarded to avoid bias in the final result). The T1 and T2 maps for other subjects were spatially registered to this reference using linear transformations (i.e., translation, rotation, scaling, and shearing in three dimensions) and maximizing mutual information [22]. We employed a standard affine registration algorithm optimizing normalized mutual information and using three spatial resolution levels [22]. We chose not to smooth the registered image data (beyond the smoothing inherent in interpolation) in order to avoid partial volume averaging/contamination from neighboring structures. T2 maps were aligned to the reference brain’s T1 data by finding the linear spatial transformation that maximized the normalized mutual information between the two data sets.

Regions of interest (ROIs) were defined in the globus pallidus internus (GP), putamen (PN) and centrum semiovale (CS) in both hemispheres of the reference brain image set (Figure 1B). Areas were chosen to have no overlap of grey and white matter in the reference brain and in the neuroanatomical references [23]. Choice of GP and PN ROIs was motivated by the important role these play in motor function and implicit learning, functions that undergo rapid development in the postnatal period. Choice of GP also motivated the use of coronal section for clearest possible definition. Choice of CS was directed by need for white matter representation as related to motor function, as well as inter-hemispheric connectivity, without the drawbacks of using vulnerable areas in the direct periventricular region. The net areas of the ROIs were 13 mm² (GP), 27 mm² (PN) and 12 mm² (CS). When overlaid on the transformed T1 and T2 maps for other subjects, the ROIs in some cases revealed small displacements relative to the intended structure, due to limitations of linear transformations between brain images. These displacements were corrected by shifting the ROIs slightly along the image row and/or column direction so they became centered on the structure of interest. The mean relaxation time (T1 or T2) was calculated for each region and subject.

In addition to the region-of-interest analysis, a univariate linear regression was used in voxel-by-voxel tests for the effects of GA on T1 relaxation time:

This analysis was performed in the common space described above. Voxels were thresholded at the p < 0.001 level and permutation testing [24] was used to control false positive errors at the cluster level (p < 0.01).

Separate linear regression models were used to estimate the association of gestational age and days to MRI with T1 and T2 at each ROI. Univariable and multivariable models were estimated and summarized by the R², Akaike Information Criterion (AIC), and, for univariable models, the statistical significance of the predictor. R² summarized the percentage of the variability in the outcome explained by the predictors in the model. The AIC is a method of model selection that allowed for comparing non-nested models, in this case, comparing a model with GA to a model with days to MRI, while penalizing for model complexity. AIC values were used to compare these models with different predictors but the same outcome, with lower values of AIC indicating a better model fit.

Of the initial sample of 52 patients, 7 were excluded due to evidence of either white matter injury or lesions in the deep nuclei, leaving a normative sample of 45 patients for the study. Three of the 45 study subjects were full term infants recruited from the Children’s Hospital rather than from the NICU. Study infant characteristics are shown in Table 1. The infant subjects were born at gestational ages representing the full spectrum of early preterm to full term birth (24–40 weeks). The median number of days of age at MRI was 49, corresponding to a median corrected GA of 38 weeks at testing (interquartile range 34–41 weeks).To confirm the choice of regions of interest for the quantitative analysis, we compared the statistical parametric map showing the voxel-wise dependence of T1 on GA to the regions of interest. The effect of GA on T1 values in the coronal slice containing the standard ROIs can be visualized using a map of p values for the dependence on GA, where higher significance appears in warmer (redder) colors (Figure 1A). This demonstrates that the strongest effect of GA appears in the white matter, PN, and GP.

To develop a quantitative model for clinical use, we then focused on the defined ROIs in the GP, PN as well as the CS as a comparison for myelinated tissue (Figure 1B). For all ROI and all indices except GP T2, the R² for univariable models including GA were much larger than the R² for univariable models containing days to MRI. In these same regions, the AIC was also lower for the models with GA than the models with days to MRI, indicating that the GA models had better overall fit (Table 2). The multivariable models for PN and CS had both higher R² and lower AIC than the unadjusted models. The exception was GP T1 in which the multivariable model was similar to the univariable model using GA as a predictor (R² 0.57 AIC 532 vs. R² 0.50 AIC 538). As gestational age at birth increased, T1 in the GP, PN and CS decreased, with age at MRI having a much smaller influence in the GP than in the PN or CS. These relationships are illustrated in Figure 2 (A-C), where T1 is plotted as a function of GA, with circle size representing age at MRI.