MRI in the diagnosis of pediatric multiple sclerosis

Pediatric MS cohorts.

We performed a retrospective analysis of MR images acquired between September 1994 and December 2003 from all children with clinically definite MS (CDMS) followed prospectively in the Demyelinating Disease program at The Hospital for Sick Children in Toronto (study cohort). Patients diagnosed with CDMS between January 2004 and July 2006 were combined with patients with CDMS followed at Texas Children’s Hospital to create the MS group used for criteria validation (validation cohort).

All pediatric MS participants were under age 18 years at MS diagnosis, defined by two or more demyelinating attacks separated by more than 30 days,¹² and all were required to have axial FLAIR or T2-weighted MR images of adequate quality acquired within 6 months of either the initial demyelinating attack or the second MS-defining clinical event. Of the 43 children identified for the study cohort, one was excluded based on clinical features more consistent with a diagnosis of neuromyelitis optica,¹³ and four had inadequate MR images for analyses. Of the remaining 38 patients, 28 had scans available from their initial demyelinating attack, 34 had scans available from their MS-defining clinical event, and 24 had both. All 21 children in the validation cohort met inclusion criteria, and all had images available from their second attack.

OND controls.

MRI scans from children diagnosed with migraine (International Headache Society Criteria for Migraine with or without aura)¹⁴ or systemic lupus erythematosus (SLE)^15,16 with clear documentation of clinical CNS involvement comprised the OND control group. These subjects were age- and sex-matched as a group with the MS study cohort.

Quantitative analysis of MR images.

MR images were acquired on 1.5 Tesla magnets with slice thicknesses between 3 and 5 mm and an interslice gap of up to 2.5 mm. All MR images were scored blinded to clinical diagnosis.

Detailed delineation of the lesion characterization technique employed is provided online (see e-Methods on the Neurology® Web site at www.neurology.org). Briefly, individual lesions were identified on axial T2 and FLAIR images (viewed simultaneously when both available). All lesions were identified and coded for location and measured in their maximum axial diameter. Lesion location was categorized as shown in the figure. Lesions spanning multiple contiguous slices were mapped to their full longitudinal extent on sequential images, and the longitudinal dimension calculated using known slice and interslice dimensions. For each patient, the total number of lesions in each category, as well as the overall total lesion count, was determined. Although each lesion counted only once toward the lesion total, many individual lesions—and essentially all large lesions—contributed to more than one lesion location category.

• Download figure
• Open in new tab
• Download powerpoint

Figure Axial T2-weighted image at the level of the decussation of the genu of the corpus callosum showing representative examples for most of the location categories assessed

Infratentorial lesions and size categories are not displayed.

The number of enhancing lesions was recorded, but since gadolinium enhanced scans were not acquired for all patients, assessment of enhancing lesions was not included in the quantitative analysis. The presence or absence of “black holes” was scored as a dichotomous variable. Lesions were only considered to be black holes if they appeared as hypointense to gray matter on T1-weighted images with a signal intensity similar to CSF, and as bright on T2-weighted images (defined as brighter than the non-T1 hypointense T2 lesions seen on the same image and of a signal intensity similar to CSF). Contrast enhancement of black holes was documented. When available, spinal MRI studies were analyzed to determine total lesion count and number of lesions spanning three or more vertebral segments.

Interrater and intrarater reliability.

A nonselected sample of 20 scans was scored for quantitative variables twice in random sequence by one evaluator (D.C.), and 15 scans were scored separately by two investigators (D.C. and M.S.; see table e-1). Qualitative variables were analyzed by two investigators (D.C. and B.B.) and scored by consensus.

Qualitative analysis of MR images.

Following the quantitative scoring, axial and sagittal T2 or FLAIR images for all patients were reanalyzed for the presence of 1) KIDMUS criteria,¹⁷ 2) revised McDonald criteria,² and 3) for evidence of diffuse, bilateral lesions (hazy, ill-defined, bilaterally asymmetric, and large (>20 mm in either axial or longitudinal dimension).

Statistical analyses.

All statistical analyses were performed using SPSS version 12.0.

Sample size calculations.

Based on previously published MRI characteristics (i.e., expected number of T2 hyperintensities) of children with MS, migraine, and systemic lupus erythematosus,^18–20 a large effect size (0.8) was expected when total lesion counts were compared between groups. Based on this, it was determined that a minimum of 25 patients with MS and OND controls would be required if the α error was set at 0.05, and the β error was set at 0.2.

Comparison of patients with MS and OND controls.

Mean lesion counts for all location and size categories were compared using Student t tests with Bonferroni correction for multiple comparisons. Effect size was measured using Cohen d,²¹ whereby the magnitude of the difference (accounting for SD) was defined as small (<0.5), medium (0.5–0.8), and large (>0.8). Fisher exact testing was used to compare qualitative categories between groups (also using a Bonferroni correction for multiple comparisons). An overall Bonferroni correction for all tests was not utilized.

A forward stepwise conditional logistic regression analysis was used to determine which MRI categories could best separate patients with MS from nondemyelinating controls. Quantitative variables with effect sizes (Cohen d) of 1.2 or greater and all qualitative variables were given the opportunity to enter the regression model (p_in < 0.05, p_out > 0.10). The number of quantitative variables allowed entry was limited to those with the largest effect sizes due to sample size constraints relative to the number of quantitative variables assessed. The threshold of d ≥ 1.2 was chosen post hoc as the mean value being ≥88% of that of the control group. The quantitative variables were entered as continuous and linearity assumptions were considered at the onset of analyses. Variables that were identified as significant were then evaluated with respect to the optimal cutpoints. The categories allowed entry to the final regression equation were then used to develop the proposed diagnostic criteria. In order to maximize accuracy, various category combinations and mean lesion count cutoff values (ranging from 0 to 10 lesions for each numeric variable) were evaluated, with the model producing the highest overall accuracy being used.

Demographics.

No differences were found in gender or age between patients with MS and OND controls (table 1). Seven children in the MS study group experienced a first demyelinating event consistent with ADEM (as defined by polyfocal features and encephalopathy).²² All of these children subsequently experienced multiple, non-ADEM like demyelinating attacks leading to their diagnosis of MS.²²

Rater reliability.

Cohen κ values for all intrarater and interrater reliability comparisons exceeded 0.80, with the majority greater than 0.95 (table e-1). The exception was delineation of lesions located in the internal capsule (intrarater reliability 0.74, interrater reliability 0.82). This category did not contribute to the final criteria.

MRI appearance of pediatric MS.

Table 2 summarizes the mean lesion count and lesion size comparing the MS group at the time of their second MS-defining attack and the OND controls. The number of subjects having at least one lesion in a given lesion location category is represented in table e-2. No differences were found in mean lesion counts between first and second attacks (data not shown). Only two patients failed to show new lesions on the MRI obtained at their second attack (one with optic neuritis, one with unilateral sensory loss). Table 3 summarizes the comparison of qualitative lesion categories between patients with MS at the time of their disease defining attack and the OND controls.

Distinguishing patients with MS from OND controls.

Mean lesion counts of the MS and OND groups were different for all location categories (p < 0.0026 after correction for multiple comparisons) with the exception of juxtacortical white matter and cortical gray matter. The largest effect sizes (d = 1.4) were observed in deep white matter, periventricular white matter, and brainstem.

The regression analysis generated a model (χ² = 85.417, df = 3, R² = 0.887, p < 0.0001) containing mean lesion counts from periventricular white matter (OR = 4.36, CI = 1.11–17.21, p = 0.035), brainstem (OR = 23.01, CI = 1.13–468.23, p = 0.041), and total mean lesion count (OR = 1.15, CI = 1.01–1.32, p = 0.041). Deep white matter lesions did not emerge as contributory to the model, likely owing to overlap with the periventricular category. Application of the regression model to the MS group from which it was derived resulted in a sensitivity of 88%, a specificity of 98%, a positive predictive value of 97%, and a negative predictive value of 92%.

A variety of cutoff values and “and/or” combinations of the periventricular white matter, brainstem, and mean total lesion count were assessed for sensitivity, specificity, positive predictive value, negative predictive value, and receiver operating curve characteristics. The cutoff values were chosen separately for each variable in order to enhance either specificity or sensitivity (e.g., the cutoff value for total lesion count was set at 2 standard deviations above the mean for the OND controls to maximize specificity [see figure e-1]). The optimal set of parameters to maximize group differentiation was determined to be presence of at least two of 1) ≥5 T2 hyperintense lesions, 2) ≥2 periventricular lesions, or 3) ≥1 brainstem lesion. Application of these criteria to the MS group from which they were generated yielded a sensitivity of 85%, specificity of 98%, positive predictive value of 97%, and a negative predictive value of 90%.

Validation of the proposed criteria.

The proposed criteria were applied to MR images (from the second demyelinating attack) of an independent group of 21 pediatric patients with MS (validation cohort). Application of the criteria to MRI scans from the MS validation cohort yielded a sensitivity of 90%.

Comparison of the proposed criteria to published MRI criteria.

Table 4 compares the McDonald MRI criteria for dissemination in space designed for adult patients with MS,^2,4 the pediatric MS criteria proposed by the KIDMUS study group,¹⁷ and the criteria generated in the present work for accuracy in both the study and validation pediatric MS cohorts.