0525
Improving longitudinal spinal cord atrophy measurements in multiple sclerosis using the Generalised Boundary Shift Integral1Department of Neuroinflammation, UCL Queen Square Institute of Neurology, University College London, London, United Kingdom, 2Department of Neurosciences, Federico II University, Naples, Italy, 3National Institute for Health Research (NIHR), University College London Hospitals (UCLH) Biomedical Research Centre, London, United Kingdom, 4Centre for Medical Image Computing (CMIC), Department of Medical Physics and Bioengineering, University College London, London, United Kingdom, 5Universitat Oberta de Catalunya, Barcelona, Spain, 6Neuroimaging Research Unit, Institute of Experimental Neurology, Division of Neuroscience, San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Milan, Italy, 7Neurocenter of Southern Switzerland, Ospedale Regionale di Lugano, Lugano, Switzerland, 8Department of Medical, Surgical, Neurologic, Metabolic and Aging Sciences, University of Campania Luigi Vanvitelli, Naples, Italy, 9MRI Research Center SUN-FISM, Institute of Diagnosis and Care (IDC) Hermitage-Capodimonte, Naples, Italy, 10MR Unit and Section of Neuroradiology, Department of Radiology, Hospital Universitari Vall d'Hebron, Universitat Autònoma de Barcelona, Barcelona, Spain, 11Department of Neurology, Universitätsmedizin Mannheim, University of Heidelberg, Mannheim, Germany, 12Department of Clinical Neurosciences, John Radcliffe Hospital, Oxford, United Kingdom, 13St. Josef Hospital, Ruhr University, Bochum, Germany, 14Department of Radiology and Nuclear Medicine, VU University Medical Center, Amsterdam, Netherlands
Synopsis
Spinal cord atrophy is a clinically-relevant feature of multiple sclerosis (MS), but can be difficult to estimate longitudinally using segmentation-based methods. We applied a fully-automated registration-based technique for spinal cord atrophy measurement (Generalised Boundary Shift Integral-GBSI-) on MS patients (n=282) and controls (n=82), from MAGNIMS and Queen Square cohorts. GBSI provided similar spinal cord atrophy rates, compared with cervical cord cross-sectional area (CSA), but with lower variability and favourable sample size estimates. GBSI performed better than CSA in differentiating cases from controls, and in depicting MS clinical features. GBSI could be used to monitor disease progression and in neuroprotective trials.
Introduction
Spinal cord atrophy is a common and clinically relevant aspect of multiple sclerosis (MS).1,2
Spinal cord atrophy is conventionally estimated with segmentation-based methods (e.g., cervical cord cross-sectional area (CSA)),3,4 that measure cord characteristics at each time point, and longitudinal atrophy rates are estimated by numerical subtraction. This approach is limited by relatively-low reproducibility and responsiveness to change.5 We have developed a registration method for the spinal cord, called GBSI (generalised boundary shift integral), where atrophy is directly measured as a change in the intensity profile. This allows to better capture cord changes (atrophy), since they are seen as continuous change between two time-points,6–8 and not as a hard difference between two masks.
In the present study, we aim to (1) compare spinal cord atrophy measurements by using GBSI and CSA; (2) define the sample size needed to detect variations in spinal cord atrophy using GBSI and CSA; and (3) explore associations between GBSI- and CSA-derived spinal cord measurements and MS clinical features.
Methods
This is a retrospective study including patients and healthy controls, from MAGNIMS and Queen Square cohorts (Figure 1).3,9
3DT1-weighted images of the spinal cord (1x1x1mm3) were segmented with Spinal Cord Toolbox for CSA calculation at the C2-C5 level.10 These segmented images were fed into the GBSI pipeline. GBSI detected the longitudinal intensity changes in the vicinity of the cord boundaries (Figure 2). These changes were weighted by a non-binary pXOR region-of-interest that was adaptively estimated from probabilistic baseline and follow-up cord segmentations. The GBSI clipping intensities were set to 0.4 and 0.96 for all our images. A non-binary XOR region-of-interest was used. Percent annual change was calculated for C2-5 CSA and GBSI.
Age, sex and interval between scans were collected. Disease duration was calculated. Cases were divided into clinically isolated syndrome (CIS), relapsing-remitting (RR) MS and progressive MS (PMS) (including both primary and secondary progressive MS). EDSS progression was recorded (1 point if baseline EDSS≤5.5, and 0.5 point if ≥6.0).
For sample size calculation, we included beta-coefficients of variations in GBSI and CSA from linear regression models using controls as reference, and adjusted by age, sex, acquisition site and disease duration), and respective standard deviation. A range of treatment effects was modelled (e.g., 30%, 60% and 90%) (power=80%; alpha=5%). To compare GBSI and CSA in their ability to predict clinical variables, we obtained areas-under-the-curve (AUC), using either GBSI or CSA as the main explanatory variables; we used bootstrap resampling (1000 repetitions) to calculate pointwise confidence intervals for the ROC curve.
Results
Demographic and clinical features of included cases and controls are reported in Figure 3.
Sample size to detect significant changes in spinal cord volume was consistently lower for GBSI, when compared with CSA, across the treatment effects simulated (Figure 4).
CIS patients were better differentiated from controls using GBSI (95%CI=0.572-0.759), than CSA (95%CI=0.430-0.634). RRMS patients were better differentiated from controls using GBSI (95%CI=0.669-0.800), than CSA (95%CI=0.521-0.668). PMS patients were better differentiated from controls using GBSI (95%CI=0.684-0.868), than CSA (95%CI=0.451-0.647). Patients with EDSS progression were better differentiated from those without progression using GBSI (95%CI=0.526-0.667), than CSA (95%CI=0.433-0.582). AUC and p-values are reported in Figure 5.
Discussion
Though CSA and GBSI provided similar mean rates of spinal cord atrophy in RRMS (-1.737% and -1.740% per year, respectively), CSA yielded larger variability (standard deviation), when compared with GBSI (±4.024% and ±2.568%, respectively), implying that GBSI is a more precise method of measurement. The main reasons for the smaller variability (and four-fold improved sample size) obtained with GBSI is that inconsistent segmentation errors, which introduce variability when working with absolute cross-sectional areas, have less influence on the boundary contours using GBSI.6–8,11
Sample size estimates for spinal cord atrophy measurements with GBSI are of the same order of magnitude as for brain atrophy with registration-based methods.12–15 However, studying the spinal cord could be more relevant than brain, due to its strong clinical correlates.16–18 In keeping with this, GBSI performed better that CSA in detecting patients with EDSS progression and different disease subtypes.
Limitations of the present study include short duration of follow-up, not allowing to draw final conclusions on disability outcomes. Also, cohorts were obtained from sites with advanced MRI technology and specific MS expertise, whilst external validation in the real world is needed.
Conclusion
GBSI provided atrophy rates with less variability than CSA, leading to increased statistical power, and better clinical discrimination. GBSI could be used for future MS research on spinal cord imaging, and for re-analysis of already-existing longitudinal datasets.Acknowledgements
Marcello Moccia has received 2017 ECTRIMS-MAGNIMS Fellowship. Ferran Prados has a non-clinical postdoctoral Guarantors of Brain fellowship. Olga Ciccarelli and Frederik Barkhof are supported by the NIHR UCLH Biomedical Research Centre.References
1 - Brex P, Leary S, O’Riordan J, et al. Measurement of spinal cord area in clinically isolated syndromes suggestive of multiple sclerosis. J Neurol Neurosurg Psychiatry 2001; 70: 544–7.
2 - Gass A, Rocca MA, Agosta F, et al. MRI monitoring of pathological changes in the spinal cord in patients with multiple sclerosis. Lancet Neurol 2015; 14: 443–54.
3 - Kearney H, Yiannakas M, Abdel-Aziz K, et al. Improved MRI quantification of spinal cord atrophy in multiple sclerosis. J Magn Reson Imaging 2014; 39: 617–23.
4 - Prados F, Cardoso M, Yiannakas M, et al. Fully automated grey and white matter spinal cord segmentation. Sci Rep 2016; 6: 36151.
5 - Durand-Dubief F, Belaroussi B, Armspach JP, et al. Reliability of longitudinal brain volume loss measurements between 2 sites in patients with multiple sclerosis: Comparison of 7 quantification techniques. Am J Neuroradiol 2012; 33: 1918–24.
6 - Freeborough P, Fox N. The boundary shift integral: an accurate and robust measure of cerebral volume changes from registered repeat MRI. IEEE Trans Med Imaging 1997; 16: 623–9.
7 - Prados F, Yiannakas M, Cardoso MJ, et al. Computing spinal cord atrophy using the Boundary Shift Integral: a more powerful outcome measure for clinical trials? ECTRIMS 2016.
8 - Prados F, Yiannakas M, Cardoso M, et al. Atrophy computation in the spinal cord using the Boundary Shift Integral. ISMRM 2016.
9 - Kearney H, Miszkiel KA, Yiannakas MC, Ciccarelli O, Miller DH. A pilot MRI study of white and grey matter involvement by multiple sclerosis spinal cord lesions. Mult Scler Relat Disord 2013; 2: 103–8.
10 - De Leener B, Lévy S, Dupont SM, et al. SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data. Neuroimage 2017; 145: 24–43.
11 - Prados F, Cardoso MJ, Leung KK, et al. Measuring brain atrophy with a generalized formulation of the boundary shift integral. Neurobiol Aging 2015; 36: S81–90.
12 - Healy B, Valsasina P, Filippi M, Bakshi R. Sample size requirements for treatment effects using gray matter, white matter and whole brain volume in relapsing-remitting multiple sclerosis. J Neurol Neurosurg Psychiatry 2009; 80: 1218–23.
13 - Van Den Elskamp IJ, Boden B, Dattola V, et al. Cerebral atrophy as outcome measure in short-term phase 2 clinical trials in multiple sclerosis. Neuroradiology 2010; 52: 875–81.
14 - Altmann DR, Jasperse B, Barkhof F, et al. Sample sizes for brain atrophy outcomes in trials for secondary progressive multiple sclerosis. Neurology 2009; 72: 595–601.
15 - Anderson VM, Fernando KTM, Davies GR, et al. Cerebral atrophy measurement in clinically isolated syndromes and relapsing remitting multiple sclerosis: A comparison of registration-based methods. J Neuroimaging 2007; 17: 61–8.
16 - Brownlee WJ, Altmann DR, Alves Da Mota P, et al. Association of asymptomatic spinal cord lesions and atrophy with disability 5 years after a clinically isolated syndrome. Mult Scler 2017; 23: 665–74.
17 - Kearney H, Schneider T, Yiannakas MC, et al. Spinal cord grey matter abnormalities are associated with secondary progression and physical disability in multiple sclerosis. J Neurol Neurosurg Psychiatry 2015; 86: 608–14.
18 - Kearney H, Rocca M a, Valsasina P, et al. Magnetic resonance imaging correlates of physical disability in relapse onset multiple sclerosis of long disease duration. Mult Scler 2014; 20: 72–80.
Figures
Figure 1. Study flow diagram.
Figure shows absolute numbers of cases and controls from MAGNIMS and Queen Square cohorts. Exclusion rate from the original cohort is shown after semi-automatic pipeline for spinal cord segmentation and GBSI measurement, and after exclusion of scans with ±5% variation on both GBSI and CSA.
Figure 2. Semi-automatic spinal cord segmentation.
Spinal Cord Toolbox was used for spinal cord segmentation. C2-3, C3-4, and C4-5 reference points were set manually (A). Representative images of semi-automatic spinal cord segmentation output are shown (B). Baseline (C/D) and follow-up (E/F) spinal cord images were denoised (with a fast version of the adaptive non-local means filter), bias field corrected (FWHM=0.05, convergence threshold 0.0001 and maximum number of iterations 1000), straightened, and, ultimately, registered to the half-way space (3D symmetric and inverse-consistent rigid-only (9 DOF) registration). Intensity changes in the vicinity of the cord boundaries were estimated for GBSI calculation (G/H).
Figure 3. Demographic and clinical features.
Table shows demographic and clinical features of included MS cases and controls. P-values are shown from t-test, Mann–Whitney, χ2 test or Fisher's exact test, as appropriate (*p<0.05).
Figure 4. Sample size estimates for CSA and GBSI.
Box-and-whisker plots show percent annual variation in CSA (A) and GBSI (B) in controls and MS phenotypes. Coefficients (Coeff) and p-values are reported from linear regression models using controls as reference, and age, sex, disease duration and MRI acquisition site as covariates. Standard deviation is also shown. Coefficient and standard deviation from each disease subtype were used for sample size calculation. Profile plot (C) and table (D) show sample size estimates for CSA and GBSI in different disease phenotypes. Different treatment effects were hypothesized (e.g., 30%, 60% and 90%) (power=80%; alpha=5%).
Figure 5. ROC curves for CSA and GBSI in relation to clinical variables.
ROC curves for CSA (green) and GBSI (orange) in relation to differentiating CIS (A), RRMS (B), and SPMS (C) from controls, and patients with EDSS progression, from those without EDSS progression (D). AUC and p-values are reported.