Objective To investigate the predictive value of 1 year subtraction MRI (sMRI) on activity and progression over the next 4 years in early phase multiple sclerosis (MS). To compare sensitivity of sMRI and contrast enhanced MRI towards disease activity.
Methods The study was performed on 127 MS patients with brain MRI within 5 years of symptom onset (y0), after 1 year (y1) and after 5 years (y5). Measures of clinical (Expanded Disability Status Scale, relapse rate) and conventional MRI outcomes (brain parenchyma fraction (BPF); T2 lesion volume (T2LV); contrast enhancing lesions (CEL)) were available at all time points. sMRI was obtained from y1–y0, y5–y1 and y5–y0 image pairs and the number of new, enlarged, resolved and regressed lesions was counted.
Results One year lesion change measured by sMRI predicted sMRI lesion change (p<0.0001), BPF and T2LV (p<0.05) changes, as well as clinical relapse rate (p<0.02) in the subsequent 4 years. sMRI measures were retained in stepwise predictive models that included other candidate MRI predictors. Active lesions on sMRI over a 1, 4 or 5 year interval provided a more sensitive assessment of disease activity than number of CEL at y0, y1 and/or y5: 83%, 93% and 90% of patients without CEL showed sMRI activity during the y1–y0, y5–y1, and y5–y0 intervals.
Conclusions sMRI is a feasible and sensitive tool for detecting MS activity and may provide an alternative to contrast enhanced MRI in clinical practice, particularly in cases where CEL are not available or inconclusive. Furthermore, sMRI metrics combined with conventional MRI outcomes (CEL, T2LV, BPF) can increase the prediction of longer term MRI activity and progression.
Statistics from Altmetric.com
Multiple sclerosis (MS) is a chronic heterogeneous disease, as reflected by clinical and neuroimaging findings.1 2 Clinical and laboratory predictors of activity and long term disability are of essential importance in planning patient management. However, currently available measures are only weakly predictive of individual disease course.3
The evaluation of lesion volume on T2 weighed MRI (T2LV) remains an objective surrogate outcome in the diagnosis and monitoring of MS4–6 although it demonstrates only modest correlation with clinical disability7–10 and lacks sensitivity in evaluating longitudinal changes.11 12 Contrast enhancing lesions (CEL) are commonly assessed at monthly or longer intervals to detect treatment effects in clinical trials. However, since new MS lesions only show enhancement for a few weeks,4 CEL counts from yearly routine MRI are poorly predictive of future MS activity.4 13
Quantification of lesion change based on pairwise subtraction of annual T2 weighted MRI images detected more MS activity than traditional single time point analysis.12 While subtraction MRI (sMRI) metrics have demonstrated sensitivity to subtle changes in lesion burden and increased power in clinical trials,12 14 their relationship with other clinical and neuroimaging measures of MS disease activity and progression has yet to be explored.
The main aim of this study was to investigate whether sMRI would provide a valuable predictive biomarker of clinical and MRI activity and progression when assessed over a 1 year interval early in the course of MS. In addition, we compared the sensitivity of sMRI to gadolinium (Gd) enhancement for the detection of active MRI lesions.
Subjects and methods
MS patients were retrospectively identified among those clinically evaluated at the Partners MS Center (Brigham and Women's Hospital, Boston, USA) between April 1998 and November 2005. Subjects with an initial (y0) brain MRI within 5 years of the first recorded MS symptom and a minimum of two follow-up examinations after 1 (y1) and 5 (y5) years were selected.
A total of 327 MS patients (250 women and 70 men) who fulfilled the international criteria for the disease15 16 and with conventional MRI metrics were found fitting these inclusion criteria. We tested our hypothesis on a smaller group of randomly identified (97 women, 30 men) MS subjects. No significant difference in baseline clinical (disability score, disease duration) and MRI (T2LV, BPF) variables between analysed and not analysed patients was detected (data not shown).
The Expanded Disability Status Scale (EDSS)17 was scored within 3 months of MRI. The occurrence of clinical relapses18 was collected and relapse rate was calculated by taking the average of the annualised relapse rate per patient (for each patient we calculated the number of relapses divided by interval length).
The use of disease modifying therapy (DMT) or immunosuppressant agents during the observational interval was also recorded.
This retrospective medical records review was approved by our institutional review board.
MR images were obtained on a 1.5 T scanner (Signa, General Electric Medical Systems, Milwaukee, Wisconsin, USA). The whole brain was imaged using a dual echo spin echo sequence yielding proton density (PD) and T2 weighted axial images (3 mm thick contiguous sections, interleaved acquisition; TR 3000 ms; TE 30/80 ms; matrix 256×192; nominal in plane resolution 0.94×0.94 mm).
Pre- and post-contrast T1 weighted images (TR 600–800 ms; TE 20 ms) were acquired 5 min after an intravenous injection of 0.1 mmol/kg Gd-DTPA (Magnevist). CEL were counted by consensus of two experts (ML, PH) according to standard criteria.19
T2LV was determined by semi-automated outlining using local thresholding and manual editing (3D Slicer 3.4, http://www.slicer.org/). Automated template driven segmentation20 identified normal appearing white matter, grey matter and CSF. Head size normalised BPF was calculated as follows: (grey matter + white matter + T2LV)/intracranial volume.20 All patients had T2LV and BPF data at y0, y1 and y5.
sMRI was performed for PD image pairs at y1–y0, y5–y1 and y5–y0, according to our previously described procedure.21 Briefly, subtraction images were obtained after three-dimensional coregistration, intensity normalisation and partial volume filtering of each exam pair. Automated, intensity based coregistration was performed with an affine transformation computed using FLIRT software.22 Intensity normalisation consisted of global histogram matching to correct for scanner drift and other technical sources of intensity bias. A three-dimensional Gaussian filter was applied to reduce partial volume effects associated with anisotropic voxel size.21
From the initial MRI sample, four out of 127 (3.1%) image pairs of the y1–y0 interval were excluded due to low image quality, and about one-third of the y1–y5 and y5–y0 pairs had to be discarded, mainly due to changes in acquisition coil type that caused strong regional intensity differences not addressable with global normalisation. Eighty-three and 94 image pairs were evaluated for the y5–y1 and y5–y0 pairs, respectively. No significant differences in clinical or MRI features were found between the analysed and the excluded samples (data not shown).
Using MIPAV software (http://mipav.cit.nih.gov/index.php), subtraction images were viewed side by side with corresponding coregistered PD and T2 weighted images, and screened for artefacts.21 sMRI lesion activity was identified and characterised by two expert readers (ML, PH) and labelled12 21 after reaching consensus as follows (figure 1):
New lesion: non-artefactual hyperintense area clearly visible against the background and measuring 3 pixels (ca 3 mm) or greater in diameter;
Enlarged lesion: lesion increased in diameter by at least 50%;
Resolved lesion: non-artefactual hypointense focus on sMRI corresponding to total resolution of a lesion at least 3 pixels in size on the baseline image;
Regressed lesion: a lesion decreased in size by at least 50%, but not totally resolved.
Positive sMRI activity was defined as the sum of new and enlarged lesion number, while negative sMRI activity was the sum of resolved and regressed lesion number.
The assessment of imaging activity (both positive and negative) was tallied during multiplanar data assessment of the dataset—that is, active lesions were included only if criteria were met on all slices (slice by slice evaluation).
The primary analysis estimated the association between the numbers of new, enlarged, resolved or regressed lesions between y1–y0 and the corresponding number between y5–y1 using a negative binomial regression model. Since the length of both intervals was variable across patients, the annualised lesion count between y1–y0 was used as the predictor and an offset term for the length of the y5–y1 interval was included in the negative binomial regression model. In the main analysis, the annualised lesion count between y1–y0 was categorised so that the skew in the data would not overly affect the results; the cut-offs for the lesion categories were determined based on what were considered reasonable ranges and to ensure sufficient numbers in each group. For the number of resolved lesions, three categories of annualised lesions were used (0, 0.5–2, >2); for the other measures, four categories were used (0, 0.5–2, 2–5, >5). The p value for multiple group comparison is reported, as well as the differences between each category and the reference category of patients with no new lesions. Multiple group comparisons were made using a likelihood ratio test, and tests for individual regression coefficients (and rate ratios) used a Wald test. A model using the raw annualised lesion count as a continuous predictor was also considered. Multivariate models were fit to control for potential confounding factors, including age, gender and proportion of time on treatment during the y5–y1 interval. For the proportion of time on treatment, all first and second line treatments were considered equally, and the time on any single or combination treatment was identified.
Several other analyses were completed using similar models. The association between positive active annualised sMRI lesion count during y1–y0 (ie, rates of new and enlarged lesions over the first year interval) and the number of CEL at y5 and the number of new clinical relapses between y1 and y5 were analysed using multivariate negative binomial regression analysis controlling for the confounders listed above. In addition, multivariate linear regression modelling was used to estimate the association between annualised lesion count at y1–y0 and annualised changes in BPF and T2LV over the interval y5–y1, again adjusting for the confounders listed above. A proportional odds model assessed the association between number of lesions during y1–y0 and EDSS at y5, controlling for EDSS at y1 and the other potential confounders listed above. A logistic regression model was also applied to test for the presence of EDSS progression.23 Finally, for each of these outcomes and the number of new lesions over y5–y1, a stepwise selection based on Akaike's Information Criteria was used to develop models with multiple potential predictors.24 For all models, age, gender and proportion of time on treatment were forced in as covariates. The candidate predictors included all first year categorised subtraction measures, CEL count at y0 and y1 categorised into three groups (0, 1, ≥2), T2LV and BPF at y1, and first year changes in T2LV and in BPF. T2LV and BPF at y0 were not explicitly included as candidate predictors as they are linear combinations of the already included y1 and first year change measures of T2LV and BPF. The number of patients with y1–y0 sMRI activity was compared with the number of patients with activity based on the presence of CEL using McNemar's test. All statistical analysis was completed in R (http://www.r-project.org) and using the MASS library.24
Clinical and MRI demographics
At the first MRI, disease categorisation of the MS study population was: n=2 clinically isolated syndrome, n=112 relapsing–remitting, n=10 secondary progressive and n=3 primary progressive MS. Other features of the MS patients were as follows: mean age 37 years (SD 9.4), mean disease duration 3.1 years (SD 2.9), mean EDSS 1.8 (SD 1.5; range 0–6), mean annualised relapse rate 0.46 (SD 0.41), mean T2LV 6.8 ml (SD 8.7), mean BPF 0.88 (SD 0.04) and mean CEL 0.41 (SD 1.05). Mean EDSS was 1.7 (SD 1.5) at y1 and 1.9 (SD 1.6) at y5, with the same median (1.5) at all time points. In line with other reports,25 the mean per cent change in BPF from y0 was −0.46% (SD 1.2) at y1 and −2.7% (SD 2.7) at y5. Finally, mean T2LV was 7.5 ml (SD 9.2) at y1 and 8.2 ml (SD 9.1) 4 years later.
Ninety-one per cent of patients were treated with DMT for a median proportion of time equivalent to 91.6% of the studied interval; at the y1 visit, 70% were treated (23.6% with glatiramer acetate, 40.9% with interferon β and the remaining 5.5% with other DMT).
Detection of MRI activity in MS
All MS patients with CEL also showed activity on sMRI at all intervals (100% positive predictive value). No CEL were observed at either y0 or y1 in 80 subjects, yet 66 of these 80 (83%) CEL negative patients demonstrated activity on sMRI during the 1 year interval (figure 2). Thus sMRI was significantly more sensitive to disease activity than CEL (p<0.0001, McNemar's test). In the interval y5–y1, 55 MS patients with Gd data showed no CEL at y1 and y5, yet 51 of these 55 (93%) demonstrated sMRI activity (figure 2). In the interval y5–y0, 50 MS patients had no CEL at y0, y1 and y5 but 45 of these (90%) were active on sMRI.
Prediction of sMRI defined lesion activity
sMRI activity in the 1 year interval significantly predicted the lesion count in the subsequent 4 years (adjusted p<0.0001 for comparison among four groups) (table 1).
Looking at the details, positive activity during y1–y0 significantly predicted the positive activity during the interval y5–y1 (adjusted p<0.0001) (table 1). Similarly, negative MS activity identified on y1–y0 sMRI significantly correlated with negative active lesions in the following 4 years (adjusted p<0.0001 for comparison among four groups) (table 1). Spearman's correlation test also showed that the standardised positive and standardised negative sMRI activity evaluated in the y1–y0 interval were significantly correlated (rs=0.45, p<0.0001).
All of the mentioned analyses were confirmed by treating lesion count as a continuous variable, and performed for each sMRI lesion types (see supplementary tables 1 and 2 available online only).
Prediction of BPF and T2LV changes
Linear regression analysis showed that first year (y1–y0) sMRI active lesions were significantly associated with annualised y5–y1 BPF changes (adjusted p=0.0018 for four group comparison) (table 2). When lesion counts were treated as a continuous variable, a significant impact on both BPF and T2LV change was observed (p<0.031). Patients with more sMRI lesions had both higher atrophy as well as greater T2LV increase (see supplementary table 2 available online).
The analysis was also performed for positive and negative sMRI activities separately. Positive active lesions assessed with y1–y0 sMRI significantly predicted annualised y5–y1 BPF changes (adjusted p=0.0038 for four group comparison) (table 2) and showed a trend towards an effect on T2LV change (adjusted p=0.055) when the categories for lesion count were used (table 1). Patients with the highest number of lesions had more brain atrophy and larger T2LV increase, and these results were driven by new sMRI lesions (see supplementary tables 1 and 2 available online).
First year negative activity on sMRI predicted subsequent brain atrophy (adjusted p=0.030 for four group comparison) (table 2), showing that patients with the highest sMRI negative activity had the greatest brain atrophy. Looking at details, this finding was driven by regressed lesions (see supplementary tables 1 and 2 available online).
Prediction of MS clinical measures
First year new, as well as enlarged, sMRI lesions were significant predictors of relapse rate during the following 4 years (adjusted p=0.023 for new lesions and p=0.015 for enlarged lesions, after four group comparison) (see supplementary tables 1 and 2 available online only). Although significant group differences were observed between lesion number categories, this analysis did not show a consistent dose–response trend. When lesion count was considered as a continuous variable, both new and positive active lesions were significant predictors of relapse rate (see supplementary tables 1 and 2 available online), and patients with larger numbers of lesions had higher relapse rates.
Proportional odds logistic regression for sMRI active lesions at y1–y0 failed to demonstrate significant correlation with EDSS at y5 when the analysis was controlled for EDSS at y1 (adjusted p>0.1 for all comparisons) (table 2). Similar results were observed when progression on the EDSS was the outcome variable (data not shown).
Relative predictive value of sMRI and other candidate predictors
Firstly, we investigated the potential MRI predictors of the number of new lesions between y1 and y5 and found that their best association was with first year positive active sMRI lesions (p<0.0001) and new sMRI lesions (p<0.0001). When stepwise selection was employed, the final model included the number of positive active sMRI lesions, BPF and T2LV at y1 and CEL at y1.
Similarly, BPF change in the y1–y5 interval was best predicted by CEL at y1 (p<0.0001) and early new sMRI lesions (p=0.002). A combined stepwise selection retained the number of CEL at y1, number of regressed lesions on y0–y1 sMRI and first year change in T2LV (R2 value for the final model was 0.27).
In multivariate modelling, the best predictors of 4 year (y1–y5) T2LV change were first year T2LV change (p=0.0028) and new lesions on first year sMRI (p=0.0092). First year T2LV change, new and enlarged lesions on first year sMRI, T2LV and BPF at y1 were all retained in full stepwise selection modelling (overall R2=0.30).
Finally, in multivariate modelling to predict relapse rate between y1 and y5, the predictors with the highest association were CEL at y0 (p=0.004) and enlarged sMRI lesions (p=0.035). Retained in the full model were number of CEL at y0, first year T2LV change, resolved and enlarged sMRI lesions.
All analyses are detailed in supplementary tables 3 and 4 available online.
At the early stages of MS, 1 year of lesion change measured by sMRI predicted lesion change, brain atrophy and clinical relapse rate in the subsequent 4 years. Furthermore, active lesions on sMRI provided a strikingly more sensitive assessment of MRI activity than contrast enhancement over a 1, 4 or 5 year interval. These results underscore the sensitivity of sMRI in assessing MS radiological activity, and suggest its potential utility in predicting and monitoring disease activity in individual patients.
T2 weighted MRI derived lesion metrics are particularly sensitive in capturing subclinical MS activity, particularly in the early years after symptom onset.7 26–33 However, longitudinal evaluation of total T2LV reflects only net changes and underestimates effective lesion activity, especially when activity remains low compared with an already large MS lesion burden at baseline.27
sMRI is capable of detecting even small white matter changes in longitudinal evaluations with high intra- and inter-observer agreement12 14 and a sensitivity significantly higher than that provided by serial total T2LV change.21 In a recent re-analysis of a placebo controlled clinical trial, monthly CEL counts were compared with 9 month interval change measures using sMRI between the first and last month MRI examinations: sMRI showed markedly increased power to detect a treatment effect, with the additional advantage of being safer as it did not require gadolinium administration.14
The present work lends further support to the feasibility and usefulness of sMRI in everyday clinical practice. It proves valuable in assessing individual disease activity on MRI as well as inferring the relative risk of future disease activity and progression. We report in fact that a substantial proportion of MS patients with positive and/or negative sMRI activity were considered inactive when assessing CEL at the beginning and end of each interval, in spite of concomitant DMT. Not only does sMRI appear to surpass contrast enhancement in terms of sensitivity to disease activity, it also becomes particularly useful as a surrogate tool in the analysis of longitudinal MS activity in the absence of post-Gd data.
sMRI analysis of 1 year change in early MS patients decidedly also provided the best prediction of future new sMRI lesion rate. The ability of sMRI to predict and monitor MS activity on MRI in the very early years of the disease carries great potential as an aid to patient specific selection of early DMT. It is also remarkable that sMRI was already a significant predictor for very low levels of annualised activity (category ‘0.5–2’ in table 1).
Positive active and regressed sMRI lesions in the first year interval predicted BPF changes, after adjustment for concomitant DMT. This finding is particularly compelling in view of the known correlation between brain atrophy and clinical progression.34 35 MS patients with a higher number of early sMRI active lesions (positive or negative) showed more decrease in BPF values in the subsequent 4 year interval, underscoring that either positive or negative lesion changes reflected MS pathological activity that is related to a poor outcome. Multivariate analysis revealed that in the presence of y1 CEL count, regressed sMRI lesions, rather than positive lesion activity, added to the model for subsequent BPF change. This is consistent with regressed lesions, probably reflecting activity at the beginning of the observed interval (closer to y0) and therefore providing complementary information to y1 CEL counts.36 37
Of potential clinical relevance is the observation that first year new and enlarged sMRI lesions significantly predicted the clinical relapse rate in the following 4 years. However, the analysis using sMRI lesions as categorical variables showed no clear dose effect, as reported for other outcomes. Further investigations are in progress to evaluate new approaches in image acquisition (ie, using three-dimensional images) and analysis to improve the accuracy and sensitivity of sMRI.38
First year sMRI findings did not predict later clinical disability in this cohort. The studied patients had relatively low baseline disability, as assessed by EDSS, and showed virtually no progression on this metric over the 5 year follow-up, probably reflecting the high use of DMT. The relatively short 1 year observational window35 and the known weakness of correlation between T2 lesion burden and EDSS7 29 31 32 39 40 may have further hampered demonstration of significance. We also note that when assessing the correlation between first year sMRI and EDSS at y5 without correcting for EDSS at y1, only enlarged lesions correlated with later clinical disability (data not shown), in line with other reports.13 Given the heterogeneity of EDSS values at y1 in our sample, we chose not to overemphasise this result.
Clearly, changes in the MRI protocol over time and the need for sophisticated image post-processing reduce the sensitivity and practical feasibility of sMRI at present. In this retrospective study, subtraction images were successfully obtained on a majority of exam pairs several years apart, despite the increasing exclusion rates for longer intervals due to inevitable changes in scanner hardware and protocol. While this does not disqualify the use of sMRI over longer time intervals, it does represent a limitation to the extent that sMRI sensitivity towards activity is reduced whenever significant changes to the acquisition protocol are made.
Another limitation concerns size and type dependence in the sensitivity of selected sMRI change, derived from the inclusion criteria for sMRI lesion change. Negative activity has an inherently lower temporal sensitivity than positive activity, as it is captured only while lesions undergo active change, which lasts approximately 10 weeks and depends on lesion size.36 37 Lesions were evaluated as three-dimensional objects, but for the given anisotropy in voxel size, the same criterion (50% change) was applied to in plane change, for both positive and negative activity.
Multivariate modelling demonstrated that sMRI metrics were better than single time point T2LV or BPF at y1 in predicting subsequent (y1–y5) new sMRI lesions, changes in BPF and clinical relapse rate. Our analysis also showed that sMRI metrics may complement conventional MRI outcomes (CEL, T2LV, BPF) by increasing their predictive power for subsequent MRI activity and progression of MS.
In conclusion, sMRI appears to provide a sensitive tool for detecting MS activity over long intervals and in particular a viable, safe and cost effective alternative to contrast enhanced MRI in clinical practice. sMRI can also provide retrospective markers of activity where post-contrast data are not available or inconclusive. Development of robust, user friendly sMRI techniques, as well as measurement of lesion change size, may become particularly relevant in supporting clinical patient management.
The authors wish to thank Dr Svetlana Egorova on behalf of the many experts at the Centre for Neurological Imaging who participated in image analysis and quality control of the data presented. The authors also thank Ms Mariann Polgar-Turcsanyi for technical assistance in the identification of the MS study population.
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Files in this Data Supplement:
- Data supplement 1 - Online tables
Competing interests None.
Ethics approval This study was conducted with the approval of Partners Healthcare, Boston, Massachusetts, USA.
Provenance and peer review Not commissioned; externally peer reviewed.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.