Abstract
Development of myelin, a fatty sheath that insulates nerve fibers, is critical for brain function. Myelination during infancy has been studied in postmortem histology, but such data cannot evaluate the developmental trajectory of the white matter bundles of the brain. To address this gap in knowledge, we (i) obtained longitudinal diffusion MRI measures and quantitative MRI measures of T1, which is sensitive to myelin, from newborns to 6-months-old infants, and (ii) developed an automated fiber quantification method that identifies bundles from dMRI and quantifies their T1 development in infants. Here we show that both along the length of each bundle and across bundles, T1 decreases from newborns to 6 months-old’s and the rate of T1 decrease is inversely correlated with T1 at birth. As lower T1 indicates more myelin, these data suggest that in early infancy white matter bundles myelinate at different rates such that less mature bundles at birth develop faster to catch-up with the other bundles. We hypothesize that this development reflects experience-dependent myelination, which may promote efficient and coordinated neural communication. These findings open new avenues to measure typical and atypical white matter development in early infancy, which has important implications for early identification of neurodevelopmental disorders.
Myelin, the fatty sheath that insulates axons that connect different brain regions is essential for brain function, as it enables rapid and synchronized neural communication across the brain. The formation of myelin, or myelination, is a key hallmark of brain development during infancy, and abnormalities in myelination are linked to a plethora of developmental and cognitive disorders1. Classic post-mortem histology reported heterogeneous myelination during infancy2–5. However, histological studies compare postmortem brain samples across individuals, often include pathologies, and use observer-dependent methods6. Thus, classic histology provides a cross-sectional and qualitative glimpse of myelination. While the heterogenous pattern of development has been replicated7,8 with modern quantitative MRI (qMRI) 8,16,17, how and at what rate myelin develops in white matter bundles during infancy is unknown.
Prior data suggest two hypotheses of myelination in infancy. The starts-first/finishes-first hypothesis proposes that postnatal myelination follows prenatal patterns2,3,5, predicting that bundles that are more myelinated at birth will develop faster postnatally and finish myelinating earlier (Supplementary Data 1). This may allow for most important brain functions to mature faster. Alternatively, the catch-up hypothesis7,12 suggests that white matter tracts that are less myelinated at birth will develop faster postnatally (Supplementary Data 1). This development may be experience-dependent13–16 and allow for more efficient and coordinated signal transmission across the entire brain.
Distinguishing between these hypotheses requires in-vivo measurements of the typical, longitudinal developmental of myelin in individual infants and across bundles. While we cannot measure myelin directly in-vivo, qMRI enables the measurement of proton relaxation time (T1 [s]). Notably, 90% of the variance of T1 in the white matter is driven by myelin17, whereby higher myelin content results in lower T1. Thus, we predict that (i) bundles that are more myelinated at birth, will have lower T1 in newborns than less myelinated bundles, (ii) if myelin increases from 0 to 6 months, then T1 will decrease from 0 to 6 months, and (iii) if T1 development follows the starts-first/finishes-first hypothesis T1 will decrease faster in bundles with lower T1 at birth, but if T1 development follows the catch-up hypothesis T1 will decrease faster in bundles with higher T1 at birth.
To test these predictions, we acquired longitudinal measurements of anatomical MRI, diffusion MRI (dMRI), and qMRI in infants during natural sleep at 3 timepoints: newborn (N=9; age: 8-37 days), 3 months (N=10; age: 79-106 days), and 6 months (N=10; age: 167-195 days) of age.
Results
New method for automated fiber quantification in infants
Evaluating the relationship between myelination at birth and its development across bundles necessitates identifying each individual infant’s bundles in their native brain space in a systematic and automated way. A major challenge is that tools developed for adults may not be suitable for infants due to substantial differences in brain size18 and organization19. Thus, we developed a new pipeline for analyzing infant dMRI data and a novel method, baby automated fiber quantification (babyAFQ), for automatically identifying 24 bundles (11 in each hemisphere and 2 between-hemispheres) in each individual infant’s brain and timepoint (Supplementary Data 2-5). We optimized babyAFQ for infants by: (i) generating waypoints (anatomical ROIs for defining bundles) on a newborn brain template (University of North Carolina (UNC) neonatal template20), (ii) decreasing the spatial extent of waypoints compared to adult standard21 to fit the more compact infant brain, and (iii) adding additional waypoints to better define curved bundles.
BabyAFQ successfully identifies 24 bundles in each infant and timepoint (example infant: Fig. 1, all infants: Supplementary Data 5), including bundles that have not previously been identified in infants: the posterior arcuate fasciculus22, vertical occipital fasciculus22–24, and middle longitudinal fasciculus25. The 24 bundles have the expected shape and location in all infants even as their brains grow from 0 to 6 months. 3D interactive visualizations at 0 months (http://vpnl.stanford.edu/babyAFQ/bb11_mri0_interactive.html), 3 months (http://vpnl.stanford.edu/babyAFQ/bb11_mri3_interactive.html) and 6 months of age (http://vpnl.stanford.edu/babyAFQ/bb11_mri6_interactive.html) show the 3D structure of bundles in an example infant.
For quality assurance, we compared babyAFQ and AFQ26 (developed in adults and used in prior infant studies27–29) to manually identified bundles (‘gold-standard’). In newborns, bundles identified by babyAFQ substantially overlapped the gold-standard (mean dice coefficient± standard error (SE): 0.61±0.02) and this overlap was significantly higher compared to AFQ (Fig 1b; Fig. Supplementary Data 3,5; 2-way repeated measure analysis of variance (rmANOVA) with AFQ-type and bundle as factors: AFQ-type: F(1,08)=528.60, p<0.0001, bundle: F(19,152) = 11.31, p<0.0001, AFQ-types x bundle: F(19,152)=7.13, p<0.0001; additional 3-way rmANOVA on the 11 bilateral bundles, with AFQ-type, bundle, and hemisphere as factors revealed no effects of, or interaction with, hemisphere). Improvements from babyAFQ were also evident at the other timepoints in qualitative evaluations in individual infants. E.g., the Forceps Major was successfully identified by babyAFQ in 29/29 brains, but identified by AFQ only in 13/29 brains.
T1 develops faster during early infancy in bundles that are less mature at birth
Measurements of mean T1 of the 24 bundles identified by babyAFQ at 0, 3, and 6 months reveal a substantial decrease in T1 from 0 to 6 months-olds (Fig. 2a). Mean T1 across bundles±SE [range]: 0 months: 2.2±0.03s [1.86s-2.39s], 3 months: 1.94±0.03s [1.61s-2.18s], 6 months: 1.64s±0.02s [1.40-1.85s]. This is a profound change, as T1 decreases on average by 0.6s within just 6 months. We modeled T1 development in each bundle using linear mixed models (LMMs) with age as predictor and a random intercept (estimated T1 at birth) for each individual. For all bundles, LMMs revealed a negative slope, indicating that T1 decreases linearly from 0-6 months. Overall, LMMs explained ~90% of the T1 variance across development (adjusted Rs2>0.89, ps<0.0001, for details see Supplementary Table 1).
We next examined if there is a relationship between the rate of T1 development and T1 in newborns across bundles. The starts-first/finishes-first hypothesis predicts a positive relationship, whereas the catch-up hypothesis predicts a negative relationship. Results in Fig 2b reveal: (i) both mean T1 in newborns and rate of T1 development during infancy vary between bundles: e.g., the cortico-spinal tract has lowest newborn T1 and the Forceps Major has the steepest slope of T1 development, and (ii) there is a significant negative correlation (adjusted R2=0.35, p=0.001) between the rate of T1 development (T1 slope) and mean T1 measured in newborns. That is, bundles that have higher newborn T1 (associated with less myelin) have a faster rate of development, which is consistent with the predictions of the catch-up hypothesis.
The catch-up hypothesis also predicts that the variability of myelination across bundles will decrease with age, as less mature bundles develop faster. To test this, we compared the standard deviation (SD) of T1 across bundles for newborns and 6-month-olds. Results indicate that SD of mean T1 across bundles significantly decreases (two-sample t-test: t(17)=7.49, p<0.0001) from newborns (0.14s±0.0009s, SD±SE) to 6-months-olds (0.11s±0.0007s), consistent with this prediction.
T1 varies across the length of a given bundle in early infancy
Our data show that bundles that are less mature in newborns develop faster than those that are more mature in newborns. As white matter bundles are large structures that connect cortical regions across brain lobes, an important question is whether T1 development varies across the length of bundles.
Analysis of T1 along bundles (Fig 3) using babyAFQ reveals three main findings: (i) Some bundles illustrate substantial variations in T1 (e.g., cortico-spinal tract), while others exhibit only modest variations (e.g., vertical occipital fasciculus). (ii) Consistent with the prior analyses, across the lengths of bundles, T1 systematically decreases from newborns (Fig 3-dotted line) to 3-month-olds (Fig 3-dashed line) to 6-months-olds (Fig 3-solid line). (iii) The fluctuation in T1 among nearby points along bundles decreases from newborns to 6-month-olds. That is, the variability in T1 between nonoverlapping, nearby positions along the length of each bundle (sum of squared difference (SSD) of T1 values between positions that are 10 nodes apart) significantly decreased (two-sample t-test: t(17)=3.29 p=0.004) from 0.08s±0.001s (mean SSD across bundles±SE) in newborns to 0.07s±0.0007s in 6-months-olds.
Segments of infants’ bundles with less mature T1 at birth develop at a faster rate
We next determined the rate of T1 development across the length of each bundle, by using LMMs to relate T1 to age at 100 equidistant locations (nodes) (one LMM per node and bundle; random intercepts for individuals). Examination of the rate of T1 development (Fig 4-dashed lines) relative to the measured T1 in newborns (Fig 4-solid lines, left y-axis), reveals that (i) even as the slopes are negative throughout, the rate of T1 decrease varies across the length of the bundles and (ii) segments of bundles that are less mature in newborns (higher T1) have a steeper rate of T1 decrease (more negative slopes) than segments than are more mature in newborns. E.g., the superior aspect of the cortico-spinal tract has higher T1 in newborns than its inferior aspect, and correspondingly, a more negative slope.
We quantified the relationship between the slope of T1 development and the measured T1 in newborns at nonoverlapping positions (every 10th node) along all bundles (LMM relating T1 slope to measured T1 in newborns; random intercepts for each bundle). This analysis reveals a significant negative relationship (Fig 4b, adjusted R2=0.64, p<0.0001) between T1 development rate and measured T1 in newborns along the length of these bundles. Results suggest that segments of bundles that are more mature at birth develop slower than segments that are less mature at birth as predicted by the catch-up hypothesis.
Discussion
By combining a novel approach for white matter bundle delineation in individual infant brains (babyAFQ) with new longitudinal measures of quantitative T1, we find a substantial decrease in T1 across all investigated bundles during early infancy. Notably, both within and across bundles, the rate of T1 development shows a negative relationship with the initial T1 in newborns. As T1 is inversely correlated with myelination, this suggests that bundles and their segments that are less myelinated in newborns develop faster, consistent with the predictions of the catch-up hypothesis of infant myelin development.
The finding that less mature white matter at birth myelinates faster during infancy is important for several reasons. First, our data not only provides empirical evidence against the classic view that white matter develops in a strictly hierarchically manner from early sensory to higher-level cognitive regions2,3, but it also offers a new parsimonious explanation for the heterogenous nature of white matter development in infancy. As myelination is experience-dependent13–16, our data suggests that the new postnatal environment and experiences may produce a flurry of myelination during the first 6 months of life, overtaking the earlier prenatal gradients. For example, projection bundles associated with movement receive input already in utero and develop slowly after birth, while bundles that connect sensory or higher order regions may only begin to receive input after birth and develop quickly thereafter. Due to this, myelination may also be fine-tuned based on each individual infant’s experience. Second, we further hypothesize that the resulting negative relationship between myelination at birth and the rate of myelin development is functionally relevant. Due to this, some level of myelin will arise in all bundles during early infancy, which may enable more coordinated and effective communication across the brain. Third, our data help interpret developmental trajectories of diffusion metrics in infants11,12,30,31. Specifically, diffusion metrics that develop similarly to T1 may be more closely related to myelination than metrics with a different developmental trajectory. Thus, future studies combining multiple quantitative and diffusion MRI metrics32–34 may disentangle multiple aspects of white matter microstructural development including not only myelination but also fiber organization, packing, and diameter.
Crucially, due to the quantitative nature of T17–9, we can compare our measurements to other populations. E.g., in our newborn bundles, T1 varies between 1.86s-2.39s, which is lower than T1 of 2.75s-3.5s observed in the white matter of preterm infants35. This observation suggests some myelination in all evaluated bundles in full-term newborns, which contrasts with classic histological studies2–5 that reported perinatal myelination in only a few white matter bundles. As classic studies used qualitative visual inspection of myelin stains, our data underscore the utility of quantitative T1 measurements. Our measurements also reveal that T1 in bundles of 6-months-olds ranges between 1.40s-1.85s, which is higher than the 0.8s-1.2s rage reported in adults36,37, suggesting that none of the investigated bundles are fully myelinated by 6 months of age. Future longitudinal investigations over a longer period are necessary to determine when these bundles reach adult-like myelination. Finally, we find that mean T1 across bundles decreases on average by 0.6s within just 6 months, which is 10 times larger than the decrease of ~0.05s observed between 8 and 18 years of age36, which highlights the profound changes occurring in early infancy.
Our study has important societal implications. First, T1 values are quantitative and have units that can be numerically compared across scanners, populations, and individuals9. Thus, our measurements in typically-developing infants provide a key foundation for large-scale studies of infant brain development in typical38,39 and clinical populations such as preterm infants40, infants with cerebral palsy41, or fetal alcohol spectrum disorders42. Second, our methodology is translatable to clinical settings as it is performed during natural sleep. Third, we developed an automated pipeline that simultaneously provides high throughput and high precision in individual infants. This level of precision may enable early identification of developmental impairments in at-risk infants, which in turn may improve the efficacy of interventions43.
In conclusion, we find that during early infancy less mature white matter at birth develops faster than more mature white matter, equalizing myelination across white matter bundles. This finding offers a new parsimonious explanation of white matter development in early infancy. We hypothesize that this pattern of myelination in infancy is driven by experience and ensures that a minimal amount of myelin becomes quickly available throughout the brain, which may serve to promote efficient and coordinated communication across the brain.
Methods
Participants
16 full-term and healthy infants (7 female) were recruited to participate in this study. Three infants provided no usable data because they could not stay asleep once the MRI sequences started and hence, we report data from 13 infants (6 female) across three timepoints: newborn (N=9; age: 8-37 days), 3 months (N=10; age: 79-106 days), and 6 months (N=10; age: 167-195 days). Two participants were re-invited to complete scans for their 6-months session that could not be completed during the first try. Both rescans were performed within 7 days and participants were still within age range for the 6-months timepoint. The participant population was racially and ethnically diverse reflecting the population of the Bay Area, including two Hispanic, nine Caucasian, two Asian, and three multiracial participants. Six out of the 13 infants participated in MRI in all three timepoints (0, 3, 6 months). Due to the Covid-19 pandemic and restricted research guidelines, data acquisition was halted. Consequently, the remaining infants participated in either 1 or 2 sessions.
Expectant mothers and their infants in our study were recruited from the San Francisco Bay Area using social media platforms. We performed a two-step screening process for expectant mothers. First, mothers were screened over the phone for eligibility based on exclusionary criteria designed to recruit a sample of typically developing infants and second, eligible expectant mothers were screened once again after giving birth. Exclusionary criteria for expectant mothers were as follows: recreational drug use during pregnancy, significant alcohol use during pregnancy (more than 3 instances of alcohol consumption per trimester; more than 1 drink per occasion), lifetime diagnosis of autism spectrum disorder or a disorder involving psychosis or mania, taking prescription medications for any of these disorders during pregnancy, written and spoken English ability insufficient to participate in the study, and learning differences that would preclude participation in the study. Exclusionary criteria for infants were: preterm birth (<37 gestational weeks), low birthweight (<5 lbs 8 oz), small height (<18 inches), any congenital, genetic, and neurological disorders, visual problems, complications during birth that involved the infant (e.g., NICU stay), history of head trauma, and contraindications for MRI (e.g., metal implants).
Data Acquisition Procedure
Data collection procedure was developed in a recent study44. All included participants completed the multiple scanning protocols needed to obtain anatomical MRI, qMRI, and dMRI data. Data were acquired at two identical 3T GE Discovery MR750 Scanners (GE Healthcare) and Nova 32-channel head coils (Nova Medical) located at Stanford University: (i) Center for Cognitive and Neurobiological Imaging (CNI) and (ii) Lucas Imaging Center. As infants have low weight, all imaging was done with first level SAR to ensure their safety. Study protocols for these scans were approved by the Stanford University Internal Review Board on Human Subjects Research.
Scanning sessions were scheduled in the evenings close in time to the infants’ typical bedtime. Each session lasted between 2.5 – 5 hours including time to prepare the infant and waiting time for them to fall asleep. Upon arrival, caregivers provided written, informed consent for themselves and their infant to participate in the study. Before entering the MRI suite, both caregiver and infant were checked to ensure that they were metal-free and caregivers changed the infants into MR safe cotton onesies and footed pants provided by the researchers. The infant was swaddled with a blanket with their hands to their sides to avoid their hands creating a loop. During sessions involving newborn infants, an MR safe plastic immobilizer (MedVac, www.supertechx-ray.com) was used to stabilize the infant and their head position. Once the infant was ready for scanning, the caregiver and infant entered the MR suite. The caregiver was instructed to follow their child’s typical sleep routine. As the infant was falling asleep, researchers inserted soft wax earplugs into the infant’s ears. Once the infant was asleep, the caregiver was instructed to gently place the infant on a makeshift cradle on the scanner bed, created by weighted bags placed at the edges of the bed to prevent any side-to-side movement. Finally, to lower sound transmission, MRI compatible neonatal Noise Attenuators (https://newborncare.natus.com/products-services/newborn-care-products/nursery-essentials/minimuffs-neonatal-noise-attenuators) were placed on the infant’s ears and additional pads were also placed around the infant’s head to stabilize head motion.
An experimenter stayed inside the MR suite with the infant during the entire scan. For additional monitoring of the infant’s safety and lack of motion, an infrared camera was affixed to the head coil and positioned for viewing the infant’s face in the scanner. The researcher operating the scanner monitored the infant via the camera feed, which allowed for the scan to be stopped immediately if the infant showed signs of waking or distress. This setup also allowed tracking the infant’s motion; scans were stopped and repeated if there was excessive head motion. To ensure scan data quality, in addition to real-time monitoring of the infant’s motion via an infrared camera, MR brain image quality was also assessed immediately after acquisition of each sequence and repeated if necessary.
Data Acquisition Parameters and Preprocessing
Anatomical MRI
T2-weighted images were acquired and used for tissue segmentations. T2-weighed image acquisition parameters: TE=124 ms; TR = 3650ms; echo train length = 120; voxel size = 0.8mm3; FOV=20.5cm; Scan time: 4 min and 5 sec.
We generated gray/white matter tissue segmentations of all infants and time-points and used them to optimize tractography (anatomically constrained tractography, ACT45). The T2-weighted anatomy, and a synthetic T1-weighted whole brain image generated from the SPGRs and IR-EPI scans using mrQ software (https://github.com/mezera/mrQ) were aligned and used for segmentations. Multiple steps were applied to generate accurate segmentations of each infant’s brain at each timepoint44. (1) An initial segmentation of gray and white matter was generated from the T1-weighted brain volume using infant FreeSurfer’s automatic segmentation code that expects T1-weighted input (infant-recon-all; https://surfer.nmr.mgh.harvard.edu/fswiki/infantFS46). (2) The T2-weighted anatomical images, which have a better contrast between gray and white matter in infants, were used in an independent brain extraction toolbox (Brain Extraction and Analysis Toolbox, iBEAT, v:2.0 cloud processing, https://ibeat.wildapricot.org/47–49) to generate another, more accurate, white and gray matter segmentation. (3) The iBEAT segmentation was manually corrected to fix segmentation errors (such as holes and handles) using ITK-SNAP (http://www.itksnap.org/). (4) The iBEAT segmentation was then reinstalled to FreeSurfer and the resulting segmentation in typical FreeSurfer format was used to optimize tractography.
Quantitative MRI
Spoiled-gradient echo images (SPGRs) were used together with the Inversion-recovery EPI (IR-EPI) sequence to estimate T1 relaxation time at each voxel and to generate whole-brain synthetic T1-weighted images. We acquired 4 SPGRs whole brain images with different flip angles: α = 4°, 10°, 15°, 20°; TE=3ms; TR =14ms; voxel size=1mm3; number of slices = 120; FOV=22.4cm; Scan time: 4 times ~5 minutes. We also acquired multiple inversion times (TI) in the IR-EPI using a slice-shuffling technique50: 20 TIs with the first TI=50ms and TI interval=150ms as well as a second IR-EPI with reverse phase encoding direction. Other acquisition parameters were: voxel size=2mm3; number of slices=60; FOV=20cm; in-plane/through-plane acceleration=1/3; Scan time=two times 1:45 min.
IR-EPI data were used to estimate T1 relaxation time at each voxel. First, as part of the preprocessing, we performed susceptibility-induced distortion correction on the IR-EPI images using FSL’s top-up and the IR-EPI acquisition with reverse phase encoding direction. We then used the distortion corrected images to fit the T1 relaxation signal model using a multi-dimensional Levenberg-Marquardt algorithm51. The signal equation of T1 relaxation of an inversion-recovery sequence is an exponential decay: where t is the inversion time, a is proportional to the initial magnetization of the voxel, b is the effective inversion coefficient of the voxel (for perfect inversion b=2). To work with magnitude images, we took the absolute value of the above signal equation and used it as the fitting model. The output of the algorithm is the estimated T1 in each voxel.
Diffusion MRI
We obtained dMRI data with the following parameters: multi-shell, #diffusion directions/b-value = 9/0, 30/700, 64/2000; TE = 75.7 ms; TR=2800ms; voxel size = 2mm3; number of slices=60; FOV=20cm; in-plane/through-plane acceleration = 1/3; Scan time: 5:08 min. We also acquired a short dMRI scan with reverse phase encoding direction and only 6 b=0 images (scan time 0:20 min).
dMRI preprocessing was performed in accordance with recent work from the developing human connectome project52,53, using a combination of tools from MRtrix354,55 (github.com/MRtrix3/mrtrix3) and mrDiffusion (http://github.com/vistalab/vistasoft). We (i) denoised the data using a principal component analysis56, (ii) used FSL’s top-up tool (https://fsl.fmrib.ox.ac.uk/) and one image collected in the opposite phase-encoding direction to correct for susceptibility-induced distortions, (iii) used FSL’s eddy to perform eddy current and motion correction, whereby motion correction included outlier slice detection and replacement57 and (iv) performed bias correction using ANTs58. The preprocessed dMRI images were registered to the whole-brain T2-weighted anatomy using whole-brain rigid-body registration and alignment quality was checked for all images. dMRI quality assurance was also performed. Across all acquisitions, less than 5% ± 0.72% of dMRI images were identified as outliers by FSL’s eddy tool. We found no significant effect of age across the outliers (no main effect of age: F(2,26) = 1.97, p=0.16, newborn: 1.07+0.88%; 3 months: 0.4+0.40%; 6 months: 0.67+0.85%), suggesting that the developmental data was well controlled across all time-points.
Next, voxel-wise fiber orientation distributions (FODs) were calculated using constrained spherical deconvolution (CSD) in MRtrix354 (Supplementary Data 2). We used the Dhollander algorithm59 to estimate the three-tissue response function, and we lowered the FA threshold to 0.1 to account for the generally larger FA in infant brains. We computed FODs with multi-shell multi-tissue CSD60 separately for the white matter and the CSF. As in previous work52, the gray matter was not modeled separately, as white and gray matter do not have sufficiently distinct b-value dependencies to allow for a clean separation of the signals. Finally, we performed multi-tissue informed log-domain intensity normalization.
We used MRtrix354 to generate a whole brain white matter connectome for each subject. Tractography was optimized using the tissue segmentation from anatomical MRI (anatomically-constrained tractography, ACT45). We argue that this approach is particularly useful for infant data, as gray and white matter cannot be separated in the FODs. For each connectome, we used probabilistic fiber tracking with the following parameters: algorithm: IFOD1, step size: 0.2 mm, minimum length: 4 mm, maximum length: 200 mm, FOD amplitude stopping criterion: 0.05, maximum angle: 15°. Seeds for tractography were randomly placed within the gray/white matter interface (from anatomical tissue segmentation), which enabled us to ensure that tracts reach the gray matter. Each connectome consisted of 2 million streamlines.
Bundle delineation with baby automated fiber quantification (babyAFQ)
Here we developed a new toolbox (babyAFQ) for the identification of white matter bundles in individual infants, that is openly available as a novel component of AFQ26 (https://github.com/yeatmanlab/AFQ/tree/master/babyAFQ). BabyAFQ identifies the following bundles (Fig. 1): anterior thalamic radiation (ATR), cortico-spinal tract (CS), posterior arcuate fasciculus (pAF), vertical occipital fasciculus (VOF), forceps major (FcMa), forceps minor (FcMi), arcuate fasciculus (AF), uncinate fasciculus (UCI), superior longitudinal fasciculus (SLF), cingulum cingulate (CC), inferior longitudinal fasciculus (ILF), inferior frontal occipital fasciculus (IFOF) and the middle longitudinal fasciculus (MLF).
BabyAFQ uses anatomical ROIs as waypoints for each bundle, that is, a given tract is considered a candidate for belonging to a bundle if it passes through all waypoints. The waypoint ROIs were adjusted from those commonly used in adults21 to better match the head size and white matter organization of infants (Supplementary Data 3). Specifically, we: (i) spatially restricted some of the waypoint ROIs, (ii) introduced a third waypoint for curvy bundles, (iii) changed the waypoint ROIs for the VOF from surface ROIs to volumetric ROIs (Supplementary Data 4), as cortical surface reconstructions in infants are challenging to date and (iv) added way-point ROIs for the identification of the MLF, which was not included in prior AFQ versions. Critically, these waypoints were defined in a neonate infant template brain (UNC Neonatal template20) and are transformed from this template space to individual infant brain space before bundle delineation in each infant’s brain. The use of an infant template brain is critical as commonly used adult templates, such as the MNI brain, are substantially larger and difficult to align to infant data. In cases where a given tract is a candidate for multiple bundles, a probabilistic atlas, which is also transformed from infant template space to individual infant brain space, is used to determine which bundle is the better match for the tract. Bundles are then cleaned by removing tracts that exceed a gaussian distance of 4 from the core of the bundle.
Critically, babyAFQ was designed to seamlessly integrate with AFQ, so that additional tools for plotting, tract profile evaluation and statistical analysis can be applied after bundle delineation.
BabyAFQ quality assurance
In order to evaluate the quality of the bundle delineation in babyAFQ, we compared the identified bundles to manually delineated “gold-standard” bundles. Manual bundle delineation was performed for the newborns in DSI Studio (http://dsi-studio.labsolver.org/) by 2 anatomical experts who were blinded to the results of babyAFQ. As a benchmark, we also delineated bundles with AFQ developed using adult data and compared these bundles to the manual bundles. For both babyAFQ and AFQ we quantified the spatial overlap between the automatically identified bundles and the manual bundles using the dice coefficient61 (DC): where |A| are voxels of automatically-identified bundles, |B| are voxels of the manual bundles, and |A⋂B| is the intersection between these two sets of voxels (Fig. 1b). We compared dice coefficients between babyAFQ and AFQ in two rmANOVAs. First, a 2-way rmANOVA with AFQ-type and bundle as factors allowed us to evaluate the effect of AFQ type across all bundles. Second, a 3-way rmANOVA with AFQ-type, bundle and hemisphere as factors, that only included bilateral bundles, enabled us to test for hemispheric differences. Finally, we also used the dice coefficients to test if tracts identified to be part of the VOF are similar across methods – i.e., using volumetric way-point ROIs vs. surface ROIs (Supplementary Data 4).
In addition to the quantitative evaluation, we examined all bundles delineated using babyAFQ and AFQ qualitatively at all time-points (Supplementary Data 5), by evaluating how well they match the typical spatial extent and trajectory. We also provide an interactive 3D visualization of an example infant’s bundles (created with pyAFQ62).
Modeling T1 developement
After identifying all bundles with babyAFQ, we modeled their T1 development using mixed linear models (LMMs). First, we modeled mean T1 development within each bundle using LMMs with age as predictor and a random intercept (estimated T1 at birth) for each individual (Fig 2a). We used model comparison (likelihood ratio tests) to determine that LMMs allowing different slopes for each individual do not better explain the data compared to LMMs using a single slope across individuals. To distinguish between the starts-first/finishes-first hypothesis and the catch-up hypothesis, we then related the developmental slopes from the LMMs and the T1 in newborns across bundles (Fig 2b). Finally, we compared the standard deviation in T1 across bundles between newborns and 6 months-olds with 2-sample t-tests.
Next, we evaluated the development of T1 across the length of each bundle. For this, we divided each bundle into 100 equidistant locations (nodes) and visually inspected T1 at each time-point across these nodes (Fig 3). We observed that the fluctuation in T1 among nearby nodes decreased with age, and quantified this observation by comparing the sum of squared difference (SSD) between positions that are 10 nodes apart in the newborns and the 6-months-olds with 2-sample t-tests.
We then determined the rate of T1 development across the length of each bundle by fitting LMMs that relate T1 to age at each node (one LMM per bundle; random intercepts for each individual as above, Fig 4a). Finally, we evaluated the relationship between the slope of T1 development and the measured T1 in newborns at nonoverlapping positions (every 10th node) along all bundles (LMM relating T1 slope to measured T1 in newborns, random intercepts for each bundle, Fig 4b). We used model comparison (likelihood ratio test) to determine that a LMM allowing different slopes for each bundle does not better explain the data compared to this LMM.
Data and code availability
The data were analyzed using open source software, including mrDiffusion and MRtrix354. We developed a new toolbox for automatic fiber quantification in individual infants (babyAFQ) and make it openly available (https://github.com/yeatmanlab/AFQ/tree/babyAFQ/babyAFQ). Code for reproducing all figures is made available in GitHub as well (https://github.com/VPNL/CatchUp). The data generated in this study will be made available by the corresponding author upon reasonable request.
Author contribution
MR, HK, and FRQ collected the data. MR, VN, HK and FRQ generated gray/white matter segmentations and T1 maps. HW developed scanning sequences. MG and JDY developed babyAFQ and data analysis pipeline. MG, JDY and KGS analyzed data. MG and KGS wrote initial draft of the manuscript. All authors edited and improved the initial draft.
Competing Interests
The authors declare no competing interests.
Acknowledgements
The research was funded by: Wu Tsai Neurosciences Institute Big Idea Neurodevelopment Grant, R21 EY030588 grant and the Center for Mind, Brain and Behavior (CMBB, Marburg, Germany).
We would like to thank all participating families, as well as, KK Barrows, Amy Kang, Javier Lopez, Laura Villalobos, Nancy Lopez-Alvarez, and Lois Williams for their help with white/gray matter segmentations of infant brains. We would also like to thank Jiyeong Ha for her contributions towards data quality assurance and Caitlyn Estrada for her contribution to data collection.