Early micro‐ and macrostructure of sensorimotor tracts and development of cerebral palsy in high risk infants

Abstract Infants born very preterm (VPT) are at high risk of motor impairments such as cerebral palsy (CP), and diagnosis can take 2 years. Identifying in vivo determinants of CP could facilitate presymptomatic detection and targeted intervention. Our objectives were to derive micro‐ and macrostructural measures of sensorimotor white matter tract integrity from diffusion MRI at term‐equivalent age, and determine their association with early diagnosis of CP. We enrolled 263 VPT infants (≤32 weeks gestational age) as part of a large prospective cohort study. Diffusion and structural MRI were acquired at term. Following consensus guidelines, we defined early diagnosis of CP based on abnormal structural MRI at term and abnormal neuromotor exam at 3–4 months corrected age. Using Constrained Spherical Deconvolution, we derived a white matter fiber orientation distribution (fOD) for subjects, performed probabilistic whole‐brain tractography, and segmented nine sensorimotor tracts of interest. We used the recently developed fixel‐based (FB) analysis to compute fiber density (FD), fiber‐bundle cross‐section (FC), and combined fiber density and cross‐section (FDC) for each tract. Of 223 VPT infants with high‐quality diffusion MRI data, 14 (6.3%) received an early diagnosis of CP. The cohort's mean (SD) gestational age was 29.4 (2.4) weeks and postmenstrual age at MRI scan was 42.8 (1.3) weeks. FD, FC, and FDC for each sensorimotor tract were significantly associated with early CP diagnosis, with and without adjustment for confounders. Measures of sensorimotor tract integrity enhance our understanding of white matter changes that antecede and potentially contribute to the development of CP in VPT infants.


| INTRODUCTION
Each year, 1 in 10 babies globally are born preterm, and those born very preterm (VPT, ≤32 weeks gestational age [GA]) face the highest risk of mortality or neurodevelopmental morbidities (Vogel et al., 2018). Of the VPT survivors, up to 50% develop mild to severe motor abnormalities such as cerebral palsy (CP), the most common physical disability in children (Spittle et al., 2011). CP prevalence is estimated at around 10% in VPT infants and increases as GA decreases (Himpens, Van den Broeck, Oostra, Calders, & Vanhaesebrouck, 2008;Vincer et al., 2006). Clinical diagnosis of CP and minor motor impairments can be delayed until 2 years of age, partially because early neuroimaging is "normal" in up to 30% of children who develop CP (Benini, Dagenais, & Shevell, 2013;Hadders-Algra, 2014;Hubermann, Boychuck, Shevell, & Majnemer, 2016). A late diagnosis leads to critical time lost for interventions at a point during brain development that is optimal for neuroplasticity. Metrics commonly used for the detection of motor impairment, including qualitative injury scores from structural magnetic resonance imaging (sMRI) and outcomes from the general movements assessment (GMA), are insufficient on their own for early, accurate diagnosis of CP (Datta et al., 2017;Hintz et al., 2015;Parikh, 2018;Van't Hooft et al., 2015). Newer prognostic imaging biomarkers available around the time of birth could promote early diagnosis and enable targeted delivery of neuroprotective interventions to preserve motor function (Parikh, 2016;Spittle, Orton, Anderson, Boyd, & Doyle, 2015). Advanced brain MRI modalities, such as diffusion MRI (dMRI), can provide more sensitive and objective measures of motor injury. dMRI exploits the diffusion of water molecules to obtain detailed information about brain microarchitecture (Alexander, Lee, Lazar, & Field, 2007). The diffusion tensor (DT) model identifies the principal orientation of white matter fibers in three-dimensional space, and DT metrics such as fractional anisotropy (FA) are routinely used to investigate white matter structural connectivity (Tournier, Mori, & Leemans, 2011). DT-based tractography has illuminated the role of white matter injury in the pathophysiology of CP and motor impairments. In children with CP, injury is most often observed in the corticospinal tract (CST), posterior thalamic radiations (PTR), superior thalamic radiations (STR), and regions of the corpus callosum (CC) (Ceschin, Lee, Schmithorst, & Panigrahy, 2015;Hoon et al., 2009;Parikh, Hershey, & Altaye, 2019). However, the DT model has proven inadequate in brain regions containing crossing fibers, two or more fiber bundles with distinct orientation that contribute to a single measured signal. It is estimated that nearly 90% of white matter voxels contain crossing fibers, including the superior longitudinal fasciculus, corona radiata, PTR, and CC (Jeurissen, Leemans, Tournier, Jones, & Sijbers, 2013;Schilling et al., 2017). This poses a serious problem for DT-based tractography methods. If spurious fiber orientations are estimated, tracking can veer off-course, leading to false-positive and false-negative connections . Crossing fibers also make it difficult to attribute FA, a traditional measure of tract integrity, to changes occurring at the microstructural level (Alexander, Hasan, Lazar, Tsuruda, & Parker, 2001).
Newer and more mathematically complex models have been developed to overcome limitations of the DT model and more accurately represent white matter microstructure (Jeurissen, Tournier, Dhollander, Connelly, & Sijbers, 2014;Tournier, Calamante, Gadian, & Connelly, 2004;Wang et al., 2011;Zhang, Schneider, Wheeler-Kingshott, & Alexander, 2012). One such technique is constrained spherical deconvolution (CSD). Using high b-shell (≥2,000) diffusion-weighted data, CSD models the signal in each voxel as a function of all fiber population orientations present within that voxel, aka the fOD.
High angular resolution diffusion imaging (HARDI) signals are thus expressed as the convolution over spherical coordinates of the response function, or the expected signal from a single population of white matter fibers, and the fOD. By performing spherical deconvolution of the diffusion signal with an estimated response function, the fOD is obtained and can be used for more accurate tractography (Tournier et al., 2004;Tournier, Calamante, & Connelly, 2007). Studies have shown the improved sensitivity and specificity of CSD-based tractography for detecting differences in white matter diffusion characteristics, when compared to DT-tractography (Auriat, Borich, Snow, Wadden, & Boyd, 2015;Jeurissen, Leemans, Jones, Tournier, & Sijbers, 2011;Reijmer et al., 2012). CSD is valuable in group studies; by using a single response function to compute fODs for all subjects, population-wide differences in diffusivity can be detected (Dhollander et al., 2020;Raffelt et al., 2012).
Quantitative measures of white matter morphology derived from fODs have been proposed (Raffelt et al., 2012(Raffelt et al., , 2017. These metrics are associated with single fiber populations (i.e., fibers of a single orientation) within individual voxels, also known as "fixels." This "fixelbased (FB) analysis," first proposed by Raffelt et al., provides an advantage over traditional voxel-based analysis for interpreting changes in white matter connectivity, especially in regions with crossing fibers. The mathematical framework of FB analysis allows for the calculation of fiber density (FD) and fiber-bundle cross-section (FC), both of which influence axonal integrity. Observed differences in FD, FC, and the combined fiber density and cross-section (FDC), can be used to better detect aberrant axonal integrity (Raffelt et al., 2017).
Histological analysis has validated the accuracy of the fOD in representing the brain's microarchitecture (Leergaard et al., 2010). Studies have also uncovered correlations between FB metrics and white matter pathology. FD and FC have been associated with damaged fiber populations in the periventricular white matter, hippocampus, cerebellum, and optic chiasm, which in some areas are consistent with histological evidence of white matter hypomyelination and disorganization (Malhotra et al., 2019;Rojas-Vite et al., 2019).
In the preterm infant brain, white matter abnormalities are thought to result from the complex interplay between impaired axonal development and axonal degeneration (Volpe, 2009). Diffusion MRI techniques such as CSD are particularly relevant to uncovering the subtle pathology underlying the noncystic, diffuse form of periventricular leukomalacia that adversely affects the neuronal/axonal bundle but is invisible on standard anatomic MRI (Volpe, 2009). Neonatal risk factors such as bronchopulmonary dysplasia (BPD) and sepsis can also influence (e.g., delay) the microstructure of developing white matter . Brain injury in CP is hypothesized to follow the multifactorial etiology of preterm encephalopathy. By quantifying micro-and macrostructural white matter changes, FB metrics should be able to better capture aberrant white matter development following delays in development or initial direct injury to either preoligodendrocytes or immature axons with or without aberrant recovery in the pathogenesis of CP.
Our goal was to assess the micro-and macrostructural integrity of major sensorimotor white matter tracts in the pathophysiology of CP in a large, prospective cohort of VPT infants. To this end, we computed CSD-derived, FB metrics from term-equivalent age (TEA) dMRI and assessed FD, FC, and FDC as measures of tract integrity for nine sensorimotor tracts (the CC and the bilateral CST, STR-sensory, STRmotor, and PTR), which have been implicated in the development of CP (Parikh et al., 2019;Scheck, Boyd, & Rose, 2012). We hypothesized that FD, FC, and FDC of these sensorimotor tracts at TEA would be negatively associated with CP in VPT infants, diagnosed early at 3-4 months corrected age. 2.3 | Global brain abnormality score A single masked pediatric neuroradiologist performed all qualitative and quantitative MR image assessments with high reliability, as previously described (Tamm, Patel, Peugh, Kline-Fath, & Parikh, 2020).

| Study design
Briefly, we used the Kidokoro scoring system (Kidokoro, Neil, & Inder, 2013) to derive a global brain abnormality score for each subject, which sums abnormalities in cortical gray matter, cerebral white matter, deep gray matter, and the cerebellum, with higher scores indicating greater abnormalities.

| Motor testing and early diagnosis of CP
The Hammersmith Infant Neurological Examination (HINE) and Prechtl's GMA were performed at 3-to 4-months corrected age by a single masked assessor, who was unaware of clinical history or MRI results and was trained to reliability. The HINE is a standardized clinical neurological battery indicated for infants 2-to 24-months of age.
It generates a summed motor score (on scale 0-78) based on cranial nerve function, posture, muscle tone, and reflexes (Haataja et al., 1999;Romeo et al., 2008). The GMA is meant to identify absent or abnormal general movements of the trunk, limbs, and neck.
Absence of fidgety general movements at 12-to 16-weeks corrected age strongly predicts long-term sensorimotor impairments like CP (Einspieler & Prechtl, 2005). Using international guidelines from Novak et al. (2017), we diagnosed our high-risk VPT infants with CP based on abnormal sMRI at TEA and the two above-mentioned motor tests at 3-to 4-months corrected age. Subjects were labeled as having a diagnosis of CP if they had any of the following combinations of abnormal prognostic tests: (a) global brain abnormality score >7 (indicating moderate or more severe brain abnormality) AND abnormal HINE exam (score <57), (b) global brain abnormality score >7 AND abnormal GMA outcome (i.e., absent fidgety movements), or (c) both abnormal HINE exam AND abnormal GMA outcome, independent of the associated brain injury score.

| MRI preprocessing
All b2000 diffusion-weighted data were preprocessed using MRtrix3 (www.mrtrix3.org), a CSD-enabled software (Tournier et al., 2019) that includes calls to standard FSL (http://fsl.fmrib.ox.ac.uk/fsl/ fslwiki/) preprocessing routines. Preprocessing consisted of PCA denoising and correction for Gibbs-ringing artifacts, motion artifacts, eddy current distortions (using the reverse, anterior-posterior phaseencoded b0 imaging data), and susceptibility-induced off-resonance field. Bias field correction was performed to remove low-frequency intensity inhomogeneities, by estimating the bias present in the b0 image and using it to correct all other images for that subject. Global intensity normalization was the performed across subjects using the median b0 white matter value; we first created a temporary white matter mask for the entire population, which was then aligned to each subject's native space. A lower FA threshold of 0.15 ) was adopted to account for the higher water content of neonatal brains, and the resulting mask was checked to ensure that it covered the major white matter tracts without extending into the CSF. For more details, see MRtrix3's tutorial on calculating FB metrics using single-tissue CSD (Tournier et al., 2019).

| CSD and FB analysis
After data preprocessing, we used a custom CSD pipeline in MRtrix3 to fit a white matter fOD in each voxel for each dMRI scan. We first applied the Tournier algorithm to estimate the response function for each subject (Tournier et al., 2019). As recommended, we upsampled the preprocessed, intensity-normalized b2000 data to an isotropic resolution of 1.3 mm 3 (from native resolution of 2.0 mm 3 ) before using the population-average response function as the deconvolution kernel to derive a white matter fOD for each subject. A group average fOD template was generated from a subset of 40 participant (20 male and 20 female) fODs selected so that (a) the fOD did not contain any visible artifact or overt severe brain injury and (b) the PMA at MRI and GA at birth of the participant closely matched the group average values. Notably, we included a few subjects with mild ventriculomegaly (a common preterm malformation) to create the template, so that it would generalize well to the entire study cohort. The resulting fOD template was segmented to produce a group fixel template Each subject's fOD was then warped and registered to the template and segmented to generate fixels. We reoriented each subjects' fixels using information stored in the warp and assigned subject fixels to template fixels, to establish a 1-to-1 correspondence. We directly computed measures of FD, FC, and FDC (discussed in Section 2.8), and performed probabilistic whole-brain tractography from the group fixel template.

| Tract segmentation
We extracted nine sensorimotor tracts of interest (the CC and the bilateral CST, STR-sensory, STR-motor, and PTR) from the wholebrain tractograph (Figure 1d-f) using a multiple region-of-interest (ROI) seed selection approach. In MRtrix3 ROI Editor, we created seed point, waypoint, and exclusion masks distinct to each tract, which were used to initiate tracking and to retain or exclude fibers passing through specific ROIs, respectively. We defined the initial position of the ROIs using neuroanatomical landmarks, according to our previously published methods (Kaur, Powell, He, Pierson, & Parikh, 2014;Parikh et al., 2019) and information from the group fixel template (1 month after the initial segmentation), to evaluate intra-rater reliability. The first segmentation attempt for each tract was used in FB analyses.
F I G U R E 1 Group fixel template and whole brain tractograph used to segment sensorimotor white matter tracts. (a-c) Group average fixel map showing the fiber orientation distribution (fOD) for all voxels in axial, coronal, and sagittal views, respectively; (d-f) Corresponding whole-brain tractograph produced from the group average fOD template. Color indicates fiber trajectory. Each tractography figure shows green (anterior to posterior), red (left to right), and blue/purple (superior to inferior) fibers 2.8 | FD, FC, FDC The fODs derived from CSD consist of multiple lobes representing individual fiber bundles. The amplitude of the fOD along a given fiber orientation is proportional to the radial diffusion-weighted signal, and is therefore proportional to the intra-axonal volume of fibers. FD is calculated by integrating the fOD of each lobe (Raffelt et al., 2012(Raffelt et al., , 2017. However, a change in intra-axonal volume may not always reflect a change in FD. Volume differences can also be accounted for by changes in morphology occurring perpendicular to the fiber orientation (i.e., a reduced fiber-bundle cross-section, FC). FC is the determinant of the Jacobian matrix required to spatially warp from subject to template space, with respect to fixel orientation. Last, FDC, the product of FD and FC, provides a more robust measure of axonal integrity by combining information from both metrics (Raffelt et al., 2017). For all tracts and all subjects, we generated a tract mask in template space and extracted the mean FD, FC, and FDC for fixels lying within the tract mask. These metrics served as our biomarkers of microscopic and macroscopic fiber integrity. Additionally, because white matter morphology changes rapidly during the first few months of life, we corrected our FB metrics for PMA at MRI scan according to the following formula: *The slope was derived from linear regression of each FB metric with PMA.

| Statistical analyses
The intra-rater reliability and reproducibility of tract segmentation was determined using intraclass correlation coefficients (ICC) and Dice similarity index. FB metrics were calculated for all tracts in each of two segmentation attempts (1 month apart) and then used to determine ICC. To calculate Dice similarity index, we used MRtrix3's mrcalc command to create an intersection mask of the two segmentations, based on the voxel-wise overlap of the corresponding binary tract masks. Then, we quantified the number of voxels present in each tract mask, using FSL's fslstats command. We calculated Dice similarity index as: To identify confounders of sensorimotor development in our cohort, we examined group differences in variables known to be associated with CP, between study infants with and without CP diagnosis. Sum Test or a Chi Square Test, as appropriate, to identify significant group differences (p <.05 indicated statistical significance). These tests were also used to assess statistical differences in baseline variables between the excluded subjects and the final cohort used in the analysis. The same baseline variables were compared across the representative subset of 40 subjects used to create the fOD template and the total population.
We used logistic regression analysis to determine the relationship of PMA-corrected FB metrics for each tract of interest and early CP diagnosis, with and without controlling for the significant confounders.

| Demographics
Of the initial cohort of 263 infants with b2000 dMRI acquired data, six subjects were excluded at the global intensity normalization step, due to bright or dark artifacts that interfered with normalization across all subjects. Twenty-three additional subjects were excluded because they were missing small brain regions at the periphery of their diffusion-weighted scans (necessary because a 1-to-1 correspondence is required between subject and template fixels, and FB analysis in MRtrix3 can only be performed for brain regions shared by all subjects

| Covariate selection
Between-group analyses for infants with CP diagnosis (n = 14) and without CP diagnosis revealed significant differences in severe BPD, postnatal corticosteroids for BPD, and postnatal sepsis (Table 1). Several variables, including severe BPD and postnatal sepsis, are well-known neonatal risk factors that have been shown to influence white matter microstructure . These potential confounders, along with GA, were included as covariates in all logistic regression analyses. As expected, severe BPD, postnatal corticosteroid use, and postnatal sepsis were associated with increased CP risk. There were no statistically significant differences in sex, antenatal steroids, maternal magnesium therapy, caffeine therapy, and severe ROP between groups.

| Reliability and reproducibility
Intra-rater reliability, measured via ICC and Dice similarity index, is displayed in  We identified significant negative associations between FB metrics and CP diagnosis in nine sensorimotor tracts, including the CC, and the bilateral CST, STR (motor and sensory), and PTR. Prior pilot T A B L E 3 Linear regression analysis between sensorimotor tract fiber density and cross-section (FDC) and HINE for infants with low-risk CP, with and without adjustment for clinical confounders DT-based studies have implicated injury or immaturity to various tracts in white matter abnormalities found in preterm infants and children with CP (Pannek, Scheck, Colditz, Boyd, & Rose, 2014;Parikh et al., 2019;Scheck et al., 2012). Our results support these findings and CP. The CC is likewise critical to controlling motor function, being responsible for transferring information across the hemispheres. Our findings corroborate prior studies (Barnett et al., 2017;Parikh et al., 2019;Thompson et al., 2011) showing reduced FA or increased RD and MD of the CC in at-risk infants, corresponding to worse motor outcomes. The PTR connects the thalamus to the parietal and occipital lobes and has been implicated in motor function related to proprioception and touch threshold (Hoon et al., 2009). In our analysis, FB metrics of the PTR were significantly associated with CP diagnosis. Studies have reported mixed results regarding the PTR in motor development, with a prior pilot study (Parikh et al., 2019) finding no association between DT metrics of the PTR and CP, and others finding a negative association between FA in older children and CP (Yoshida et al., 2010(Yoshida et al., , 2011. Collectively, our findings are consistent with these smaller studies. However, individually, many of them report associations in only a few ROIs or full sensorimotor tracts, while we report widespread associations across all nine tracts examined. This discrepancy likely results from the higher statistical power of our large study and the higher sensitivity of CSD with HARDI-acquired data as a method for investigating white matter tract integrity. To our knowledge, only one other study (Pannek et al., 2020) has investigated FB metrics of sensorimotor tracts in relation to motor outcomes in VPT infants.
Pannek et al. (2020) (Williams, Lee, & Anderson, 2010). In a separate analysis, we found significant positive associations between FB metrics of sensorimotor tracts and HINE score in our cohort's low-risk group (Table 3). This finding is unique and suggests that micro and macrostructural metrics of the neonatal brain at term may have the potential to detect milder motor impairments than CP, which may nevertheless benefit from early intervention. Our study had a number of strengths, including our large, multicenter, regional cohort of 223 VPT infants. We used comprehensive methods that queried the major sensorimotor tracts that are negatively impacted in older children with CP. Using CSD and FB analysis to assess the integrity of white matter tracts was a major strength of the study. An estimated 90% of voxels contain crossing fibers, which make it difficult to accurately interpret DT-based analyses, as changes at the microstructural level may be undetected or misinterpreted due to a lack of fiber specificity. FB metrics have higher sensitivity than voxel-wise DT metrics in detecting group-wide differences in white matter (Dhollander et al., 2020). We also used meticulous tractextraction methods based on previously-published methods (Kaur et al., 2014;Parikh et al., 2019;Teli et al., 2018) and our own systematic, iterative process. The high correlation between our FB metrics across all sensorimotor tracts and the high intra-rater reliability of the segmentation attest to the quality of our methods. Finally, we adjusted our logistic regression analyses for several confounders that are known to influence motor development, thus demonstrating the independent significance of these FB metrics in the development of CP.
There are several limitations to our study. To facilitate early detection of CP, we used the Novak international early CP diagnosis guideline that combines sMRI at TEA with 3 to 4-month corrected age HINE or GMA scores (Novak et al., 2017). These tests represent the earliest and most predictive tests of motor outcomes available and are used clinically by many centers. However, this combination of tests has yet to be independently validated with later diagnosis of CP (Parikh, 2018). Moreover, our findings suggest that our early CP definition and/or FB metrics may lack the ability to differentiate the type/severity of CP. However, this question can only be answered once we correlate FB metrics (in comparison with DT-based metrics) with more confirmatory diagnosis at 2 years corrected age, as we are currently doing. Prior studies using DT-based metrics have shown an association between FA from the CST and CP severity (Cho et al., 2013;Rose et al., 2007). Nevertheless, infants with a combination of abnormal tests are at higher risk for developing CP. An advantage of using this guideline for early diagnosis is that it is less confounded by early intervention therapies and treatments/exposures following NICU discharge and before age 2. We lacked an analysis of targeted ROIs within tracts, which could reveal more detailed information about the underlying associations. Finally, while the present analyses focused on investigating the association between FB metrics and CP, the results indicate, to an extent, the potential value of these metrics for predicting motor outcomes. Brain morphometric biomarkers that we and others have reported, such as brain volumes and cortical surface measures (Dubois et al., 2008;Kline, Illapani, He, Altaye, et al., 2020;Parikh et al., 2020), may be used in conjunction with FB metrics to enhance prediction of motor outcomes.

| CONCLUSION
In summary, we have shown that CSD derived, FB measures of axonal integrity from key sensorimotor tracts at TEA are independently associated with the early diagnosis of CP in VPT infants. This is an important finding, as it demonstrates the role of multiple sensorimotor tracts in the pathophysiology of CP and related motor impairments.
VPT infants remain at high risk of motor impairments, and there is an urgent need to establish validated biomarkers that can be used for early detection and intervention.

ACKNOWLEDGMENTS
The authors sincerely thank the parents of infants that participated in

CONFLICT OF INTEREST
The authors declare no potential conflict of interest.

DATA AVAILABILITY STATEMENT
Datasets generated from this study and code used in the analysis are available from the corresponding author upon request.