Assessment of White Matter Integrity after Pediatric Traumatic Brain Injury

Abstract

White matter (WM) abnormalities, such as atrophy and hyperintensities (WMH), can be accessed via magnetic resonance imaging (MRI) after pediatric traumatic brain injury (TBI). Several methods are available to classify WM abnormalities (i.e., total WM volumes and WMHs), but automated and manual volumes and clinical ratings have yet to be compared in pediatric TBI. In addition, WM integrity has been associated reliably with processing speed. Consequently, methods of assessing WM integrity should relate to processing speed to have clinical application. This study had two goals: (1) to compare Scheltens rating scale, manual tracing, FreeSurfer, and NeuroQuant^® methods of assessing WM abnormalities, and (2) to relate WM methods to processing speed scores. We report findings from the Social Outcomes of Brain Injury in Kids (SOBIK) study, a multi-center study of 60 children with chronic TBI (65% male) from ages 8–13. Scheltens WMH ratings had good to excellent agreement with WMH volumes for both NeuroQuant (ICC = 0.62; r = 0.29, p = 0.005) and manual tracing (ICC = 0.82; r = 0.50, p = 0.000). NeuroQuant WMH volumes did not correlate with manually traced WMH volumes (r = 0.12, p = 0.21) and had poor agreement (ICC = 0.24). NeuroQuant and FreeSurfer total WM volumes correlated (r = 0.38, p = 0.004) and had fair agreement (ICC = 0.52). The WMH assessment methods, both ratings and volumes, were associated with processing speed scores. In contrast, total WM volume was not related to processing speed. Measures of WMH may hold clinical utility for predicting cognitive functioning after pediatric TBI.

Introduction

Neuroimaging is performed routinely on children who sustain a traumatic brain injury (TBI), typically beginning with computed tomography (CT) for initial medical management and triage.^1–3 Follow-up neuroimaging is usually based on magnetic resonance imaging (MRI). The advantage of MRI over CT imaging is its ability to identify a variety of trauma-related pathologies with greater precision and neuroanatomical specification.^1,3,4 In cases of TBI, MRI, especially the T2-weighted fluid attenuated inversion recovery (FLAIR) sequence, is particularly sensitive in detecting white matter (WM) pathology.^1,3,5–7

The WM is susceptible to injury in trauma because of shear-strain and tensile effects on myelinated axons induced by biomechanical deformations, along with secondary neurodegenerative changes.^8–10 In chronic TBI, reductions in WM volume^11,12 and hyperintense signal on the FLAIR image, referred to as a WM hyperintensity (WMH), are common observations.^13,14

The WM volume loss in children who have sustained a TBI has been considered an atrophic indicator, reflective of diminished WM integrity relative to healthy, typical developing, non-injured children.^11,12 Because of the distinctness of the aberrant hyperintense signal against the otherwise uniform appearing WM on the T2-FLAIR sequence, WMHs are readily identifiable in a qualitative manner and their presence, number, and location are reported in clinical radiological reports.¹³ Clinical reporting of WMHs has even been standardized by the use of semi-quantitative rating scales, such as the Scheltens rating method.¹⁵ The WMHs also can be quantified objectively by visual identification for their presence as a region of interest (ROI), with the boundaries of the WMH manually traced and their volume computed.^6,16–18 While manual methods accurately identify and quantify the volume of WM pathology, they are time consuming and laborious to perform.

With recent advancements in neuroimaging quantification, a variety of more automated methods have emerged that are either open source or commercially available and have the ability to identify WMHs as a lesion burden metric, not only counting the number of WMHs but calculating their volume (see NeuroQuant^® at www.cortechslabs.com/neuroquant/; VolBrain at volbrain.upv.es/).

A key breakthrough in more automated image analysis came with the T1-weighted processing technique referred to as FreeSurfer (surfer.nmr.mgh.harvard.edu/; see¹⁹), which became the underlying platform for NeuroQuant.²⁰ FreeSurfer is open source, but does not have a built-in normative comparison group for its volumetric findings, as does NeuroQuant. Both NeuroQuant and FreeSurfer generate a measure of overall WM volume via a segmentation process that distinguishes WM from gray matter (GM) and cerebral spinal fluid (CSF). Only NeuroQuant, however, generates an automated image analysis of the FLAIR sequence to calculate a WMH lesion count and volume.

In the routine clinical setting, visual inspection of the MRI scan permits identification of the presence or absence of WMHs, but other than counting them, provides no additional quantification. Scheltens ratings, an ordinal scale, provides a method for rating WMH severity, with the benefit of being rapidly performed and necessitating no special hardware or software, other than a digital imaging and communications in medicine (DICOM) reader, to visualize the digital image.¹⁵ While the Scheltens rating can quickly yield a semiquantitative rating, we do not know how well a clinical rating of WMH captures relevant information regarding TBI in comparison with volumetric quantification methods.

In quantitative neuroimaging analyses, operator controlled, ROI-hand tracing is considered the “gold standard,”^21,22 but this method is particularly time consuming and requires expertise in neuroanatomy combined with the appropriate computer hardware and software. More automated image quantification methods like FreeSurfer and NeuroQuant also yield data about ROI volumetrics, such as total WM volume.^19,23,24

Although the comparability of these approaches to operator-controlled methods has been a topic of several investigations, no comparison has focused on WMHs.^21,25–30 FreeSurfer can be performed relatively quickly and allows for visual examination and manual boundary corrections of ROIs,^19,23,30,31 but all normative comparisons have to be generated separately. In contrast, NeuroQuant image quantification, also performed relatively quickly, generates a volumetric comparison of the individual TBI patient to group normative data, adjusted for head size, age, and sex for various ROIs, all performed with automated algorithms.^25,31–33 NeuroQuant includes a specific protocol designed to assess the T2-FLAIR sequence to perform a WMH count and volume calculation^34,35

Because various methods are available for obtaining quantitative and qualitative WM data involving WMHs and total WM volumes, their clinical similarity (or difference) as well as application in detecting WM pathology in TBI warrants further study. For example, do the results from various WM assessment methods have similar associations with cognitive outcomes? Processing speed, or the rate at which new information can be learned and used to complete cognitive tasks,³⁶ has a well-established association with WM integrity,^3,37–42 with reduced processing speed a common consequence of TBI.^43,44 Reduced processing speed has been found irrespective of the location of WM pathology^37,41 and is most disrupted in moderate to severe TBI.^38,43,44 As such, a comparison of measures of WM integrity should assess their relation to processing speed.

Using existing approaches involving both clinical ratings and volumetric methods derived from FreeSurfer and NeuroQuant, we aimed to assess chronic WM integrity in children with TBI drawn from the Social Outcomes in Brain Injury in Kids (SOBIK) study by identifying WM abnormalities from both a qualitative as well as quantitative perspective. Accordingly, this study explored the following questions: (1) how do quantitative WMH methods using operator-controlled, hand-tracing of WMH volume compare with the automated NeuroQuant algorithm generated WMH volume; (2) how does a clinical rating of WMH based on a modified Scheltens scale compare with quantitative WMH methods; and (3) how do these different methods for assessing WMHs relate to processing speed as measured using the Processing Speed Index (PSI) derived from the Wechsler Intelligence Scale for Children-IV Edition.³⁷

Methods

Participants

This was an archival study that used data extracted from the SOBIK investigation, completed in 2012. Study details have been published elsewhere.^37,45–49 Briefly, children with TBI were recruited from three metropolitan children's hospitals: The Hospital for Sick Children (Toronto, Canada), Rainbow Babies & Children's Hospital (Cleveland, Ohio, USA), and Nationwide Children's Hospital (Columbus, Ohio, USA). Participants were children between eight and 13 years of age who underwent hospitalization for at least 24 h after a TBI.

Individuals with a Glasgow Coma Scale (GCS) of 13 to15 were included only if day-of-injury (DOI) CT was positive for some type of trauma-related abnormality (i.e., fracture, hemorrhage, edema, etc.). Thus, these cases would meet criteria for what is referred to as “complicated” mTBI.⁵⁰ All other children had a moderate or severe TBI, defined based on a GCS of 3–12. Injuries occurred 1 to 4 years prior to study enrollment, and participants were not eligible if the TBI occurred before age four. Detailed inclusion/exclusion criteria have been published previously.^37,45–48

Written informed consent was obtained from the participant's parent or legal guardian, and Institutional Review Board (IRB) approval was obtained from each participating location, including Brigham Young University, where all post-processing of imaging data was performed using de-identified data. Processing speed was assessed using the Processing Speed Index (PSI) from the Wechsler Intelligence Scale for Children-IV Edition.⁵¹

MRI

Participants completed magnetic resonance (MR) scanning during the chronic phase of injury (average = 2.6 years post-injury) using a uniform image acquisition protocol that has been described previously.^37,46 Briefly, MRI studies conducted in Cleveland or Toronto used a 1.5 T GE Signa Excite scanner, whereas a Siemens Symphony MRI scanner was used in Columbus. The T1-weighted (T1), FLAIR, and dual-echo proton density T2-weighted (T2) images were collected at each site (Fig. 1). The current study conducted image analyses using T1 images (repetition time [TR] = 3000 ms; echo time [TE] = 4.0 ms; flip angle = 8; 170 slices at 1.2 mm thick) to collect total WM volumes, FLAIR images (TR = 3000 msec; TE = 100 msec; flip angle = 90; 24 slices at 5 mm thick) to collect WMH volumes, and T2 images (TR = 3000; TE1 = 12 msec; TE2 = 100 msec; flip angle = 90; 3 mm interleaved) to verify WMHs on FLAIR images in manual tracings. All pre-processing procedures are detailed in Supplementary Material A.

FIG. 1.

Summary of white matter neuroimaging methods. Scheltens use unedited fluid attenuated inversion recovery (FLAIR) images for visual ratings of white matter hyperintensities (WMH). Manual tracings use pre-processed FLAIR, T2, and T1 images to obtain WMH volumes. NeuroQuant^® uses FLAIR and T1 images in the dynamic atlas with a thresholding algorithm to obtain WMH volumes. NeuroQuant and FreeSurfer use T1 images to obtain total WM volumes. FreeSurfer obtains volumes using the recon-all function; NeuroQuant uses the dynamic atlas. *Hyperintensity in right lateral frontal lobe is an old contusion classified previously as gray matter hyperintensity; †Image does not align with remaining images because we are unable to manipulate the provided image from the report.

To adjust for variations in head size, the ratio of WMH and total WM volumes to overall total intracranial volume (TICV) were computed and multiplied by 100, so whole numbers could be presented. NeuroQuant and FreeSurfer automatically collect TICV volumes; manual tracings of TICV were performed following methods outlined by Eritaia and associates,⁵² for a within-method head size correction.

Figure 1 provides an overall schematic of what was examined in this study. On the left side, three axial views are presented from a child who sustained a severe TBI (GCS = 3). All images, which are in radiological perspective, depict a prominent region of focal left frontal encephalomalacia, with the FLAIR sequence at the top showing a distinct region of surround hyperintense signal on the boundary of the encephalomalacia. The different titled methods to the right depict which sequence was used, and how rated or processed. Additional details are provided in the figure legend.

Scheltens rating scale

The Scheltens rating scale¹⁵ was modified slightly to exclude GM hyperintensities, changing the total scale to 0–30 (Table 1). Two independent raters scored all participant images with high interrater reliability (Dice similarity coefficient > .9). An expert rater and author (EDB) reviewed and approved all scan images and ratings before analysis.

Table 1.

Modified Scheltens Rating Scale for White Matter Hyperintensities¹⁵

Region	Scale	Classification
Periventricular hyperintensities (PVH 0–6)
Caps occipital	0 to 2	0 = absent
frontal	0 to 2	1 = ≤ 5 mm
Bands lateral ventricles	0 to 2	2 = >5 mm and <10 mm
White matter hyperintensities (WMH 0–24)
Frontal	0 to 6	0 = absent
Parietal	0 to 6	1 = <3 mm, n ≤ 5
Occipital	0 to 6	2 = <3 mm, n > 6
Temporal	0 to 6	3 = 4–10 mm, n ≤ 5
Internal capsule	0 to 6	4 = 4 mm, n > 6
		5 = >11 mm, n > 1
		6 = confluent
Total score		0 to 36

n = number of lesions.

Semi-quantitative rating of signal hyperintensities by brain region, with the range of the scale, between brackets.

Manual tracing

Although manual tracings are considered the gold standard for volumetric research,^53,54 no standard procedure exists for performing manual tracings of WMHs.⁵⁵ This is likely because of the vast heterogeneity of WMHs in size, shape, and locale, as well as differences that may relate to etiology. Other publications that include WMH volumes simply note the use of expert raters to manually trace WMHs.^53,54

In this study, two independent expert raters, blinded to injury severity, manually traced WMHs on the pre-processed FLAIR images, using the program Analysis of Functional NeuroImaging (AFNI). Our tracing procedures followed the boundaries of the WMHs (Fig. 1). To be identified as a WMH, the hyperintensity had to be located in WM and appear distinctly hyperintense on the T2 image.

As a rule, periventricular regions were not traced manually except when they were prominently asymmetrical and extended into deep WM, because certain aspects of normal periventricular regions frequently show as hyperintense on FLAIR images and do not necessarily signify pathological damage from TBI.^56–58 Disagreements between tracings were adjudicated by an expert rater (EDB) and were retraced until a consensus was reached. All manual tracings were reviewed and approved by EDB.

NeuroQuant

NeuroQuant has several available reports that provide volumetric data for various structural areas of the brain. For this study, the Multi Structure Atrophy report provided total WM volumes using T1 images and the LesionQuant report provided WMH volumes using the T1 and FLAIR images (see Fig. 1).

FreeSurfer

FreeSurfer's v6.0.0 (surfer.mnr.mgh.harvard.edu) recon-all function was used to collect total WM volumes from the T1 images following the procedures from Bigler (Fig. 1).²¹ Visual inspection of the results ensured the accuracy of the skull strip, image registration, and WM segmentation. Although the focus of the current investigation was on WMHs, total WM volumes were also computed using both FreeSurfer as well as NeuroQuant.

Statistical analyses

For analyses comparing methods, both correlation and agreement were assessed.^30,59,60 A Pearson R or Kendall tau correlation was used, depending on the normality of the data, and considered significant at p ≤ 0.05. Normality was assessed using the Kolmogorov-Smirnov test. Absolute agreement between methods was assessed using a two-way mixed intraclass correlation (ICC), where an ICC below 0.40 indicated poor agreement, between 0.40 and 0.59 indicated fair agreement, between 0.60 and 0.74 indicated good agreement, and between 0.75 and 1.00 indicated excellent agreement.^30,61,62

Outliers were identified as data greater than three standard deviations (SD) from the mean, for each method. Bland-Altman plots with 95% limited agreement (or two SD confidence interval [CI]) were also included to illustrate the agreement between methods.⁶³ Non-normally distributed data were log transformed to meet assumptions for the Bland-Altman plot and ICC, and analyses including Scheltens ratings data were converted to z-scores because of differing metrics (i.e., ratings, volumes).

To assess the clinical application of the WM methods, a standard error of the mean (SEM) model was used to assess the association between all WM methods and processing speed. An SEM model comparison was preferred over multiple individual regression analyses to reduce Type 1 error. The maximum likelihood estimate was used to account for missing data. Data were converted to z-scores, and a Benjamini-Hochberg test was used to correct for Type 1 error, with a false discovery rate of 0.05.⁶⁴

Descriptive statistics were analyzed and reported as means with SDs, or percent as appropriate. A t test was used to assess differences between injury severity groups. Correlations, ICC, and SEM were performed using Statistical Package for the Social Sciences (SPSS) 24.0.0.0 on an Apple Macintosh computer with OS X “Yosemite” 10.10.5 operating system. Bland-Altman plots were created using the batplot command program in STATA 14.2 using a PC computer with Windows 10, version 1703.

Results

Demographic data

Of the 72 children with TBI scanned in the SOBIK study, nine had significant artifact and three had incomplete or missing MRI data.³⁷ A total of 60 (86%) children with TBI from the SOBIK study met inclusion criteria for this study, with one participant having only T1 images suitable for analysis (see Table 2 for demographic data for study participants). Participants were an average of 7.9 years old at time of injury and an average of 10.4 years old at time of MRI acquisition. There were 32 participants with complicated-mild TBI, nine with moderate TBI, and 19 with severe TBI. The injury severity groups did not differ in terms of sex, age at injury, ethnicity, time from injury to testing, or age at testing. As expected, however, length of hospital stay was significantly different between complicated-mild TBI and moderate (p = 0.03) and severe (p = 0.001) TBI.

Table 2.

Demographics and Medical Data of Traumatic Brain Injury Children in the Social Outcomes of Brain Injury in Kids Dataset

Demographics of TBI children (N = 60)	M (SD) or n; %	Range
Age at testing (years)	10.5 (1.5)	8.1–13.2
Age at injury (years)	7.9 (1.9)	3.2–12.1
Time since injury (years)	2.6 (1.2)	1.0–5.2
Hospital length of stay (days)	5.5 (9.3)	1–58
Sex (% male)	n; 65%
Ethnicity
White	n; 78%
Black	n; 7%
Hispanic	n; 7%
Multi-racial/non-specific	n; 8%
GCS score
Complicated-mild (n = 32)	14.8 (0.5)	13–15
Moderate (n = 9)	10.3 (1.0)	9–12
Severe (n = 19)	3.7 (1.4)	3–7
MRI
Both T1 & FLAIR (n = 59)
Only T1 (n = 1)
Cognitive test scores
Processing speed (WISC-IV PSI)	100.98 (13.05)	62–128

TBI, traumatic brain injury; M, mean; SD, standard deviation; GCS, Glasgow Coma Scale; MRI, magnetic resonance imaging; FLAIR, fluid attenuated inversion recovery; WISC-IV PSI, Weschler Intelligence Scale for Children- 4th edition, Processing Speed Index.

Rating and volumetric descriptive data

Table 3 presents rating and volumetric data by injury severity and total group for the five different WM assessment methods. Visually identifiable WMHs showed an injury severity gradient, with the severe TBI group displaying the highest WMH Scheltens rating, which was significantly larger than the complicated-mild group (p = 0.00) and moderate group (p = 0.00). The moderate group was significantly larger than the complicated-mild group for manual WMH volumes (p = 0.00). Participants with severe TBI had significantly smaller FreeSurfer total WM volumes compared with complicated-mild TBI (p = 0.01).

Table 3.

Ratings or Volumetric Data Adjusted by Total Intracranial Volumes, Grouped by Injury Severity and Total Group

Method	Complicated-mild M (SD)	Moderate M (SD)	Severe M (SD)	Total group M (SD)
Scheltens WMH ratings	1.90 (1.81)	4.22 (2.33)	5.74 (7.29)b	3.49 (4.69)
Manually traced WMH volumes	0.01 (0.01)	0.02 (0.01)a	0.10 (0.30)	0.04 (0.17)
NeuroQuant®–WMH volumes	0.07 (0.07)	0.04 (0.03)	0.10 (0.11)	0.07 (0.08)
NeuroQuant–total WM volumes	27.83 (2.42)	28.68 (2.47)	28.26 (2.56)	28.10 (2.45)
FreeSurfer–Total WM volumes	29.04 (1.43)	28.37 (1.30)	27.65 (2.05)a	28.51 (1.72)

M, mean; SD, standard deviation; WMH, white matter hyperintensities; WM, white matter.

^aSignificantly different from complicated-mild,; ^bsignificantly different from both complicated-mild and moderate severity groups

Volumes are adjusted for total intracranial volumes.

Outliers

Outliers were found for Scheltens ratings, manual WMH volumes, and NeuroQuant WMH volumes), but no outliers were identified in either total WM methods (NeuroQuant and FreeSurfer). Only four participants met the criteria for WM outliers (see Supplementary Fig. SB1 in Supplementary Material B); one participant had outliers for both manual WMH volumes and Scheltens ratings methods, and three participants had outliers for NeuroQuant WMH volumes. Participants with WMH volumes more than three SDs also had the largest brain pathology.

Because this study focused on the comparsion of WM methods, the exclusion of outliers would restrict this comparison to participants with less WM pathology, and hence would not provide a representative picture of pediatric TBI. In other words, excluding participants with particularly large WMH volumes would provide an incomplete comparison of the WM methods and their ability to capture WM pathology in pediatric TBI, as well as restrict the range of WMH volumes in correlations with processing speed.

We understand, however, the statistical issue posed by outliers, and therefore performed analyses both with and without outliers. The analyses performed with outliers excluded can be found in Supplementary Material B. In this article, we focus on the results from the entire sample, and only note when results differed when outliers were excluded.

Comparison of quantitative WMH methods

The Kendall tau correlation between NeuroQuant and manually traced WMH volumes was positive but not significant (r = .12, p = 0.21, n = 56; Fig. 2), and the agreement was poor (ICC [39] = 0.24, CI = -0.19 to 0.56). The Bland-Altman plot, however, had a mean difference of 0.79 with only 2/39 data points outside the 95% limit of agreement of −0.04 and 1.92, suggesting some degree of overall agreement despite different absolute volumes derived from these two methods (Fig. 2).

FIG. 2.

For white matter (WM) method comparisons, correlational scatterplots are to the left and Bland-Altman plots assessing agreement are to the right. The scatterplots include a “line of best fit” (dotted line) for the correlation. The Bland-Altman plots show upper and lower limit (dotted lines) defined as 1.96 standard deviation and the mean difference (black line). The mean difference is close to zero for all comparisons, with the farthest difference for the comparison between NeuroQuant^® and manual tracings. WMH, white matter hyperintensities.

Comparing qualitative versus quantitative WMH methods

Kendall tau correlations were significant between manually traced WMH volumes and Scheltens WMH ratings (r = .50, p = 0.00), and NeuroQuant and Scheltens WMH ratings (r = .29, p = 0.005; Fig. 2). Agreements were excellent between manually traced and Scheltens ratings (ICC [57] = 0.91, CI = 0.85 to 0.95) and NeuroQuant and Scheltens ratings (ICC [56] = 0.79, CI = 0.64 to 0.88). The Bland-Altman plot comparing manual tracing WMH volumes and Scheltens rating had a mean difference of −0.15 with 3/37 data points outside the 95% limit of agreement of −1.59 and 1.30 (Fig. 2). The Bland-Altman plot comparing NeuroQuant WMH volumes and Scheltens rating had a mean difference of 0.06 with 2/49 data points outside the 95% limit of agreement of −1.89 and 2.02 (Fig. 2).

Comparing total WM volume methods

The Pearson r correlation between NeuroQuant and FreeSurfer total WM volume (r = .38, p = 0.004, Fig. 2) was significant, and the agreement was fair (ICC [58] = 0.52, CI = -0.20 to 0.72). The Bland-Altman plot had a mean difference of −0.007 with 0/58 data points outside the 95% limit of agreement of −0.08 to 0.07 (Fig. 2).

Relationship of WM methods to processing speed

Of the multiple associations assessed in the SEM model, Scheltens ratings (r = -.41, p = 0.004), manual tracing (r = -.44, p = 0.002), and NeuroQuant WMH volumes (r = -.38, p = 0.01) were negatively correlated with processing speed (Table 4). In contrast, FreeSurfer and NeuroQuant total WM volumes were not significantly associated with processing speed. When outliers were excluded from the analyses, no significant correlations were found between WM methods and processing speed (see Supplementary Material B).

Table 4.

Standard Error of the Mean Co-varied Comparison between White Matter Methods and Processing Speed

SEM Comparison	R	p
Scheltens <-> PS	−.41	0.004a
Manual <-> PS	−.44	0.003a
NeuroQuant® WMH <-> PS	−.38	0.007a
FreeSurfer total WM <-> PS	−.02	0.880
NeuroQuant total WM <-> PS	−.15	0.250

SEM, standard error of the mean; PS, processing speed; WMH, white matter hyperintensity; WM, white matter.

^aSignificant after the Benjamini-Hochberg test with a false discovery rate set at 0.05.

Discussion

White matter pathology is a known consequence of TBI. As would be expected, and is apparent in Table 3, visibly identifiable WMHs based on Scheltens ratings as well as WMH lesion volume showed a clear injury severity gradient. Children with severe TBI had the greatest degree of WMHs, regardless of method for identification or quantification; however, agreement was limited between NeuroQuant WMH volumetric findings and manual tracing in terms of absolute volumes.

Nonetheless, in the Bland-Altman analyses, while correlations between the different WM metrics varied and were often non-significant, very few to no outliers were detected relative to the 95% CI. Understandably, given the differences in how lesion boundaries involving WMHs are defined based on operator-controlled versus automated neuroimaging technology, some differences are to be expected. The fact that most of the variability occurred within a 95% CI indicates that the different methods are generally identifying similar pathology.

These observations add to the ongoing interest in examining the comparability of existing volumetric and clinical methods for assessing WM integrity in TBI and determining why differences exist.^21,25–30 Specific to WMHs in pediatric TBI, their identification and quantification, as undertaken in this investigation, implies some degree of broad agreement across approaches. In addition, whether volumes were traced manually, calculated via NeuroQuant automated algorithms, or qualitatively rated using the Scheltens scales, WMH methods were similarly related to processing speed. The lack of relationship found between WMH methods and processing speed when outliers were removed likely reflects the reduced range of WMH volumes.

Further comment on the differences between manual tracing of WMHs and NeuroQuant results is in order. The observation that NeuroQuant and manual tracing of WMH volumes were not strongly related likely is explained by the differences in methods, including what is identified as a WMH. The Bland-Altman plot indicated a consistent bias, such that NeuroQuant WMH volumes were consistently larger than manually traced WMH volumes, yet there were only two values that fell outside of the 95 percent CI for the two methods.

Accordingly, NeuroQuant's larger WMH volumes may not be a major concern for most purposes,³⁰ once it is understood that automated methods may include hyperintense signal findings not necessarily specific to WM pathology. For example, returning to Figure 1, a right frontal hyperintense signal can be seen, likely reflective of residual gliosis from previous surface contusion, that was not included in the manual tracing but likely would be included in automated detection. Also, NeuroQuant will include periventricular WMHs, whereas with manual tracing, that was true only if there was asymmetry or extension into deeper regions of the surrounding WM.

Accordingly, manual tracings were much more conservative in identifying WMHs, and in total, identified 18 participants with no WMHs, whereas NeuroQuant was much more liberal or inclusive of WMHs and identified WMHs in all participants. This likely explains why, in Table 3, the complicated-mild TBI group had a considerably higher WMH volume with NeuroQuant than what was observed with manual tracing.

The size of the WMH likely further contributes to differences between manual and automated approaches to image analysis. Especially at the mild end of the injury spectrum, WMHs tend to be small, punctate abnormalities. Hernández and associates²⁷ recently found greater variance in agreement when hyperintensity volumes were small, and better, more consistent agreement when volumes were larger.

In Table 3, the manually traced WMH volumes and those derived from NeuroQuant are identical in the severe TBI group. In addition, Luo and colleagues³⁴ found strong correlations (r = .98) between NeuroQuant WMH volumes and manually traced WMH volumes in a sample with multiple sclerosis that only included participants with large WMH volumes (>5 cm³). While our largest WMH volume was 1.23 cm³, the majority of the SOBIK participants had WMHs, when present, that were less than 2 mm³.

In addition, methodological differences in handling artifact may help explain the reduced level of agreement and correlation found between NeuroQuant and manual tracings. Scanner and motion artifact are commonly found on FLAIR images, from which WMH volumes were computed. As already mentioned, non-TBI related hyperintense signal may be noted in periventricular regions,^23,57,65 where image thresholding technology may not distinguish artifact or normal variant from WMHs.^6,66

In both clinical and research contexts, automated methods like NeuroQuant may be overinclusive of WMH lesion burden, but this bias would likely be consistent across participants, whereas errors using manual methods have the potential to be more varied because of rater experience, expertise, and bias. In addition, automated methods are frequently improved, addressing consistent errors identified in earlier versions. Our study collected WMH volumes using NeuroQuant 2.0. Additional research using the most recent update of NeuroQuant is needed to reassess the comparability between WMH volumes collected from NeuroQuant versus manual tracings.^67,68

Scheltens ratings correlated and had good agreement with WMH volumes collected from NeuroQuant and manual tracings. While a similar finding has been reported in previous studies in healthy aging,^69–71 this is the first published study involving pediatric TBI. NeuroQuant as well as manual tracing methods require adherence to specific scan sequence parameters necessary for image acquisition, whereas a clinical rating can be performed on a FLAIR sequence not meeting those requirements.

Because NeuroQuant is a commercial product, analyses involved a financial cost in performing, whereas Scheltens rating involved minimal to no cost, other than time for reviewing the scan and being properly trained. Both methods are similarly rapid. With a high-speed connection, DICOM MRI files can be uploaded for image quantification in less than a minute, with an analyzed computer printout and graphical analysis generated in less than 7 min (www.cortechslabs.com/resources/frequently-asked-questions/).

NeuroQuant and FreeSurfer total WM volumes correlated and agreed well. Similar to Ochs and coworkers,³¹ our NeuroQuant total WM volumes were slightly larger than FreeSurfer volumes. Older versions of FreeSurfer and NeuroQuant were found to be comparable previously.³¹ Regardless of multiple updates and the addition of NeuroQuant's dynamic atlas, NeuroQuant and FreeSurfer results remain comparable for standard total WM volume computation.

The T1-weighted image was used for quantification of total WM volume in NeuroQuant or FreeSurfer. The WM pathology is often poorly detected by the T1-weighted sequence, so that even in the presence of a potential lesion within the WM, as identified in the FLAIR sequence, the T1-weighted total WM volume will include the lesion. This means that the total WM volumes derived from FreeSurfer or NeuroQuant will include WM pathology. This may explain why we observed consistent associations between WMH volume and processing speed regardless of the method used, but an absence of any association with overall WM volume. Total WM volume may not be as clinically useful as WMH volume in predicting neuropsychological outcomes, such as processing speed.

Strength and Limitations

This is the first article to compare these different approaches to assess WM integrity in pediatric TBI. Despite the important finding that regardless of the technique, the presence of WMHs predicted reduced processing speed; the minimal correlations between some of the measures were surprising. Reasons for a lack of strong associations across WM measures may reflect several technical factors. For example, neuroimaging was based on MR field strength of 1.5 Tesla. Higher field strength has been shown to improve T2-FLAIR detection of WMHs⁷² and, likewise, segmentation and image classification may be superior at higher field strength. Injury severity appeared to be associated with larger WMHs, but the study was statistically underpowered to fully assess the influence of severity.

Funding Information

Supported by the National Institute of Child Health and Human Development (Grant Nos. 5R01HD048946 and 3R01HD048946-05S1) and college grant funds from Brigham Young University.

Author Disclosure Statement

No competing financial interests exist.

Supplementary Material

Supplementary Material A

Supplementary Material B

 Supplementary Figure SB1

 Supplementary Figure SB2

 Supplementary Table SB1

 Supplementary Table SB2

 Supplementary Table SB3