JOURNAL OF MAGNETIC RESONANCE IMAGING 41:454–459 (2015)
Original Research
Validation of a Semiautomated Spinal Cord Segmentation Method €l Chen, MSc,1,2 Brice Tiret, MSc,1,2 Mohamed-Mounir El Mendili, MSc,1,2* Raphae lanie Pe le grini-Issac, PhD,1,2 Julien Cohen-Adad, PhD,3 Ste phane Me 2,4,5 1,2,6 Pierre-Franc¸ois Pradat, PhD, MD, and Habib Benali, PhD1,2 Lehericy, PhD, MD, Purpose: To validate semiautomated spinal cord segmentation in healthy subjects and patients with neurodegenerative diseases and trauma.
region. DSC was high (0.96) in both cervical and thoracic regions. DSC remained higher than 0.8 even when modifying initial contours by 50%.
Materials and Methods: Forty-nine healthy subjects, as well as 29 patients with amyotrophic lateral sclerosis, 19 with spinal muscular atrophy, and 14 with spinal cord injuries were studied. Cord area was measured from T2weighted 3D turbo spin echo images (cord levels from C2 to T9) using the semiautomated segmentation method of Losseff et al (Brain [1996] 119(Pt 3):701–708), compared with manual segmentation. Reproducibility was evaluated using the inter- and intraobserver coefficient of variation (CoV). Accuracy was assessed using the Dice similarity coefficient (DSC). Robustness to initialization was assessed by simulating modifications to the contours drawn manually prior to segmentation.
Conclusion: The semiautomated segmentation method showed high reproducibility and accuracy in measuring spinal cord area. Key Words: spinal cord; MRI; segmentation; cross-sectional area; atrophy measurement J. Magn. Reson. Imaging 2015;41:454–459. C 2014 Wiley Periodicals, Inc. V
Results: Mean interobserver CoV was 4.00% for manual segmentation (1.90% for Losseff’s method) in the cervical region and 5.62% (respectively 2.19%) in the thoracic region. Mean intraobserver CoV was 2.34% for manual segmentation (1.08% for Losseff’s method) in the cervical region and 2.35% (respectively 1.34%) in the thoracic 1 Inserm U678, UPMC Univ Paris 6 UMR-S 678, Laboratoire d’imagerie fonctionnelle, Paris, France. 2 Univ Paris 11, IFR 49, Institut f ed eratif de recherche en imagerie neurofonctionnelle, DSV/I2BM Neurospin, Gif-sur-Yvette, France. 3 Ecole Polytechnique de Montr eal, Department of Electrical Engineering, Montr eal, Canada. 4 Inserm U975, UPMC Univ Paris 6, UMR-S975, CNRS UMR7225, pinie`re – Centre de recherche de l’Institut du Cerveau et de la Moelle e CRICM, Centre de Neuroimagerie de Recherche – CENIR, Paris, France. 5 APHP, Groupe Hospitalier Piti e-Salp^ etrie`re, Service de neuroradiologie, Paris, France. 6 APHP, Groupe Hospitalier Piti e-Salp^ etrie`re, D epartement des Maladies du syste`me Nerveux, Paris, France. Additional Supporting Information may be found in the online version of this article. Contract grant sponsor: Association franc¸aise contre les myopathies (AFM); Contract grant sponsor: Institut pour la recherche sur la pinie`re et l’enc moelle e ephale (IRME). The research leading to these results has also received funding from the program “Investissements d’avenir” ANR-10-IAIHU-06. *Address reprint requests to: M.-M.E.M., Laboratoire d’Imagerie Fonctionnelle, Facult e de M edecine Pierre et Marie Curie, Site Piti eSalp^ etrie`re, 91 Boulevard de l’H^ opital, F-75634 Paris cedex 13, France. E-mail: [email protected] Received September 26, 2013; Accepted January 3, 2014. DOI 10.1002/jmri.24571 View this article online at wileyonlinelibrary.com. C 2014 Wiley Periodicals, Inc. V
SEVERAL MAGNETIC RESONANCE IMAGING (MRI) studies have shown that spinal cord atrophy may be used as an indicator of tissue destruction in neurological disorders such as multiple sclerosis (MS) (1–3), spinal cord injury (SCI) (4–6), and amyotrophic lateral sclerosis (ALS) (7,8). Atrophy measurement has been facilitated by advances in MRI technology (9–13). In parallel, several segmentation methods with varying degrees of manual intervention have been proposed to evaluate cord atrophy, including measurements of cross-sectional area (CSA) (14–16) and volume of the spinal cord (17–21). However, most studies have used low-resolution 1.5T images and investigated only the upper cervical spinal cord in MS patients. Validation of cord atrophy as a biomarker in spinal cord diseases requires evaluating these segmentation methods in other segments of the spinal cord as well as in diseases other than MS. The threshold-based segmentation method proposed by Losseff et al (16) is one of the most-used approaches to quantify atrophy due to its high reproducibility and accuracy. This method has to be initialized manually by drawing two contours, one to delineate the spinal cord and the other surrounding the cerebrospinal fluid (CSF). However, this method has only been validated at the C2 vertebral level using 1.5T images in MS patients (16,18,22). The present study had two main objectives: 1) To validate Losseff’s segmentation method at both the cervical and thoracic levels and in neurological disorders other than MS. 2) To evaluate its robustness to manual initialization.
454
Semiautomated Spinal Cord Segmentation
MATERIALS AND METHODS Subjects In this study, 111 subjects were recruited: healthy volunteers (n ¼ 49, mean age 6 SD: 39 6 17 years, 29 females), patients with ALS (n ¼ 29, 53 6 10 years, 7 females, mean disease duration: 27 6 27 months), SMN1-linked spinal muscular atrophy (SMA) patients (n ¼ 19, 37 6 11 years, 9 females, mean disease duration: 25 6 15 months), and SCI (n ¼ 14, 45 6 14 years, 3 females, delay after injury: 25 6 35 months). The local Ethical Committee of our institution approved all experimental procedures of the study and written informed consent was obtained from each participant. MRI Acquisition All scans were performed using a 3T MRI system (TIM Trio, Siemens Healthcare, Erlangen, Germany) using a body coil for signal excitation and a neck/spine coil for signal reception. Subjects were positioned as comfortably as possible and were systematically asked not to move during the acquisition in order to minimize motion artifacts. Furthermore, visual quality control was performed after the acquisition to make sure that there were no visible motion artifacts in the data. The spinal cord was imaged using a T2-weighted 3D turbo spin echo sequence with a slab-selective excitation pulse. Imaging parameters were: voxel size: 0.9 0.9 0.9 mm3; field of view (FOV): 280 280 mm2; 52 sagittal slices; TR/TE ¼ 1500/120 msec; flip angle: 140 ; generalized autocalibrating partially parallel acquisition (GRAPPA) with acceleration factor R ¼ 3; turbo factor: 69 and acquisition time 6 minutes. This sequence provided a high signal-to-noise ratio (SNR) due to 3D acquisition, high resolution owing to isotropic acquisition, short acquisition times by combining parallel acquisition with high turbo factors, and low specific absorption rate due to low flip angle refocusing pulses. Segmentation Method Table 1 shows the number of segmented vertebral levels for each group of subjects. Due to FOV positioning and variable vertebral size, it was not possible to segment the whole cervical and thoracic spinal cord. C1 and C2 vertebral levels were grouped into a single level because the boundary between C1 and C2 was often indistinct. Segmentation was stopped at the T10 level due to the increasing number and size of nerve roots, which made segmentation inaccurate, and to the reduction of SNR in the lower parts of the spinal cord. In the SCI group, only slices rostral and caudal to the lesions were selected and segmented (5). Segmentation and statistical analyses were conducted with MatLab (MathWorks, Natick, MA). The following procedure proposed in (6) was adopted for spinal cord segmentation. Preprocessing First, the sagittal slice where the spinal cord was the most median at the vertebral level to be segmented
455 Table 1 Number of Segmented Vertebral Levels for Each Group of Subjects Groups Level
Controls
SLA
SMA
SCI
C1-C2 C3 C4 C5 C6 C7 T1 T2 T3 T4 T5 T6 T7 T8 T9
31 44 46 46 48 48 48 47 47 45 35 25 10 7 5
28 29 29 29 29 29 29 29 29 28 26 18 3 0 0
19 19 19 18 18 18 18 18 18 18 16 8 1 0 0
12 9 4 7 4 4 11 14 13 11 3 0 0 0 0
was selected manually. A segment was drawn manually parallel to the spinal cord to delimit the vertebral level of interest. The resulting axial slices, perpendicular to the spinal cord, were resampled to a voxel size of 0.3 0.3 0.3 mm3 using 3D cubic interpolation. The slices were automatically cropped to squares of 120 120 voxels centered on the segment drawn manually, then averaged. This whole procedure minimizes the partial volume effect between spinal cord and CSF. Segmentation Method The contours of two regions of interest (ROIs) were drawn manually in the average axial slice to delineate spinal cord and CSF areas, respectively. This outlining step was achieved by two operators (M.-M.E.M., R.C.) blinded to the type of data. One operator was experienced with spinal cord segmentation (ie, 3 years experience) and the other had no previous experience (ie, had been given brief instructions beforehand on five different datasets randomly selected from those used in the present study). The mid intensity of the mean intensities in the two ROIs was used as a threshold for further separating ROI voxels into two classes, the spinal cord and the CSF (16). In parallel, manual segmentation was also carried out by the same two operators in the images resulting from the preprocessing step. The cross-sectional area (CSA) in mm2 was finally defined as the number of pixels in the resulting spinal cord class, multiplied by the pixel size. Reproducibility Reproducibility for both manual segmentation and Losseff’s method was evaluated using inter- and intraobserver coefficients of variation (CoV, defined as the ratio of the standard deviation to the average of CSA values across measurements). To evaluate
456
El Mendili et al.
Figure 1. a: T2-weighted MRI mid-sagittal slice for an SMA patient, showing the anatomical landmarks of the cervical spinal cord. A segment (orange arrow) was drawn manually parallel to the spinal cord to delimit the vertebral level of interest, here C6. b: Contours drawn manually by the experienced operator delineating the CSF and the spinal cord on the resampled and averaged axial slices (C6), corresponding to AR ¼ 0%. d,f,h: Simulated contours corresponding to AR ¼ 15%, 30%, and 50%, respectively. c,e,g,h: Contours resulting from the semiautomated method for AR ¼ 0%, 15%, 30%, and 50%, respectively.
interobserver CoV, the two operators segmented all images. Intraobserver COVs were evaluated on 20 datasets: Five datasets were randomly selected from each group of subjects and segmented twice by both operators at least 2 weeks apart. The operators were blinded to the type of data.
between 0.6 and 0.8 can be interpreted as a substantial agreement, and a DSC value between 0.4 and 0.6 represents a moderate agreement (26).
Accuracy
The robustness of the semiautomated segmentation method to its initialization was assessed by conducting numerical simulations: we simulated modifications to the initial ROIs drawn manually by the experienced operator, which were used as a reference. To do so, we randomly suppressed border voxels of the initial ROIs, thus defining an "attack rate" (AR) as the sum of suppressed border voxels divided by the total number of initial ROI voxels. This operation was repeated 200 times for each AR (from 0% to 50% by steps of 1%), thus simulating a large variety of manually drawn ROIs. The simulated ROIs were then introduced as inputs to the semiautomated method. The procedure is illustrated in Fig. 1. Finally, the DSC was computed as previously to compare final results
Accuracy of the semiautomated segmentation method was assessed using the Dice similarity coefficient (DSC), which evaluates the performance of segmentation methods by measuring their spatial overlaps (25). The DSC is defined as follows: DSC ¼
2 jM \ Lj ; jMj þ jLj
where jMj (respectively jLj) is the CSA in mm2 obtained by manual (respectively Losseff’s) segmentation, and jM \ Lj is the area in mm2 common to both segmentation results. A DSC value higher than 0.8 indicates an almost perfect agreement, a DSC value
Robustness to Initialization
Semiautomated Spinal Cord Segmentation
457
Table 2 Reproducibility of Mean Cross-Sectional Area Measurements in Cervical and Thoracic Regions Mean cord area (mm2) Cord region Cervical Thoracic
Method Manual Semiautomated Manual Semiautomated
Mean coefficients of variation (%)
Operator 1 Mean (SD)
Operator 2 Mean (SD)
73.48 68.85 46.76 43.66
70.10 69.83 43.35 44.71
(8.43) (8.28) (5.65) (5.63)
of the semiautomated method for each AR and manual segmentation. RESULTS Computation time (mean 6 SD) for the preprocessing step was: 5.0 6 1.1 min/subject, 5.5 6 1.5 min/subject for manual segmentation and 3.0 6 0.9 min/subject for the semiautomated segmentation method. This was achieved using a 64-bit Quad-core (Intel Xenon, processor speed: 2.67 GHz) workstation.
(8.52) (8.46) (5.69) (5.64)
Interobserver Mean (SD) 4.00 1.90 5.62 2.19
(2.76) (1.74) (3.40) (2.10)
Intraobserver Mean (SD) 2.34 1.08 2.53 1.34
(0.96) (0.50) (1.17) (0.92)
regions (0.96). The semiautomated method underestimated the mean CSA measured by manual segmentation (considered as the ground truth) by 3.29% in the cervical region and 1.70% in the thoracic region. When considering each group of subjects separately, the semiautomated method underestimated the mean CSA in the cervical region (respectively in the thoracic region) by 3.07% (respectively 1.38%) in the control group, 3.68% (respectively 1.26%) in the ALS group, 3.23% (respectively 2.19%) in the SMA group, and 3.31% (respectively 3.06%) in the SCI group (Supporting Table S2).
Reproducibility The results are shown in Table 2 for all subjects. The mean interobserver CoV in the cervical region was 4.00% for manual segmentation and 1.90% for the semiautomated method in the cervical region, and 5.62% and 2.19%, respectively, in the thoracic region. The mean intraobserver CoV was 2.34% for manual segmentation and 1.08% for the semiautomated method in the cervical region, and 2.53% and 1.34%, respectively, in the thoracic region. The smallest volume change that the semiautomated method was able to detect regardless of the errors made by operators in the repetitive use of the technique was calculated using the following formula: 1.96 intraobserver CoV mean CSA (16,18). This showed that changes as small as 1.50 mm2 (1.96 SDs) in the cervical cord area and 1.16 mm2 in the thoracic cord area may be detected with 95% confidence. Similar results were obtained when considering separately healthy volunteers and patients (Supporting Table S1). There was no significant correlation between the mean CSA and mean interobserver CoVs in both cervical and thoracic regions (pCervical ¼ 0.35, pThoracic ¼ 0.19), or between the mean CSA and mean intraobserver CoVs (pCervical ¼ 0.29, pThoracic ¼ 0.86). Accuracy The results are shown in Table 3. The mean DSC value was high in both cervical (0.96) and thoracic
Robustness to Initialization The results are shown in Fig. 2. The mean DSC value remained stable for attack rates ranging from 0% to 30% in both cervical and thoracic regions. The variance was larger for AR values less than 10%, then decreased with increasing AR up to 30%. After this plateau, the mean DSC value decreased rapidly followed by an increase in variance. The maximum mean DSC was obtained for AR ¼ 15% in the cervical region (DSCCervical ¼ 0.97) and AR ¼ 17% in the thoracic region (DSCThoracic ¼ 0.97). The minimum mean DSC values were obtained for AR ¼ 50% in both regions (DSCCervical ¼ 0.89, DSCThoracic ¼ 0.90). DISCUSSION The semiautomated segmentation method was validated at 3T in the cervical and thoracic spinal cord in a large set of healthy volunteers and three groups of patients with ALS, SMA, and SCI. The method showed high inter- and intraobserver reproducibility in the cervical and thoracic regions. The mean inter- and intraobserver CoVs were about twice as high for manual segmentation than for the semiautomated method. CoVs were higher in the thoracic than in the cervical region, either in patient groups or in the control group (Supporting Table S1). This result can be explained by, on the one hand, the
Table 3 Accuracy of Mean Cross-Sectional Area Measurements in Cervical and Thoracic Regions Mean Dice similarity coefficient
Mean error estimation (%)
Cord region
Operator 1 Mean (SD)
Operator 2 Mean (SD)
Operator 1 Mean (SD)
Operator 2 Mean (SD)
Cervical Thoracic
0.97 (0.02) 0.96 (0.02)
0.96 (0.02) 0.95 (0.02)
0.28 (3.55) 3.27 (3.27)
4.30 (3.06) 4.99 (3.09)
458
El Mendili et al.
Figure 2. DSC function of the attack rate in cervical and thoracic regions. The solid line represents the mean DSC and the dashed lines the mean 6 SD.
fact that thoracic images presented more physiological artifacts due to respiratory movements inducing errors in manual and semiautomated segmentation, and lower SNRs inherent to the thoracic coil compared to the cervical receive coil and, on the other hand, the difference in operators’ experience in the use of the technique. C1 and C2 vertebral levels could not be segmented separately because the boundary between C1 and C2 was difficult to distinguish in all subjects, despite the high resolution of the images. CoVs should be lower at C2 than for the combined C1–C2 vertebral level. The mean inter- and intraobserver CoVs were 1.30% and 0.60%, respectively. These values were lower than those obtained in the C2 region in Horsfield et al (18) (20 controls and 40 MS patients), where the authors reported an intraobserver CoV of 2.15% and an interobserver CoV of 7.05% for the semiautomated segmentation method. In Losseff et al (16), the interobserver CoV was 0.83%, which is lower than the value obtained in the present study. This can be explained first by our larger sample size (n ¼ 111) as compared with 15 subjects in Losseff et al (16) and, second, by the difference in operators’ experience in the use of the technique. In contrast, we found a lower intraobserver CoV (0.60%) than in Losseff et al (16) (0.83%), probably due to the high spatial resolution obtained at 3T and possible improvement provided by the preprocessing step. CoVs in the present study were similar to those obtained in Horsfield et al (18) in the C2 region (intraobserver CoV ¼ 0.59% using 15 scans; interobserver CoV ¼ 1.36% using 60 scans), where the authors used a method based on image smoothing followed by an active surface model constrained by intensity gradient image information. The DSC was higher than 0.8 in the cervical and thoracic regions, indicating almost perfect agreement
between the semiautomated segmentation method and manual outlining. Despite these high performances, the semiautomated method had a number of limitations. In the preprocessing step, partial volume was still present even if axial slices were resampled, and this led to underestimation of the cross-sectional area. This issue has been outlined previously. Partial volume correction was proposed in Tench et al (15) by calculating the portion of cord in the voxels immediately surrounding the detected edge. This method increased the accuracy of the estimated CSA by 1.45%. Another limitation was that CSF contrast was not homogeneous in all slices, especially in the thoracic region, which presented more motion artifacts. In addition, small structures such as nerve roots, with signal magnitude similar to that of the spinal cord, were included in the final segmentation, leading to an overestimation of the cross-sectional area. These structures may be removed using a region-labeling method combined with priors on the size and shape of the cord (18,27). An additional limitation of the present study was the lack of scan–rescan reproducibility evaluation. This was due to imaging protocol constraints, especially concerning inclusion of patients with large motor deficits (ALS and SCI patients). A comparison between CSA and volume measurements would be of interest, as many recently published studies have used volume measurement instead of CSA to quantify age-related atrophy as well as atrophy in MS patients (28,29). In conclusion, the semiautomated method of Losseff et al (16) showed high reproducibility and accuracy for measuring spinal cord cross-sectional area in both cervical and thoracic regions at 3T in normal subjects and a large set of patients with various spinal cord
Semiautomated Spinal Cord Segmentation
disorders. Improved image spatial resolution obtained at 3T allowed higher segmentation performances than reported previously. ACKNOWLEDGMENTS We thank all the subjects and their relatives. We thank Kevin Nigaud, Romain Valabre`gue, Alexandre de ric Humbert, Christelle Macia, and Vignaud, Fre Eric Bardinet for helping with the acquisition and Dr. Henrik Lundell for making the code for measuring cord area available to us. REFERENCES 1. Cohen AB, Neema M, Arora A, et al. The relationships among MRI-defined spinal cord involvement, brain involvement, and disability in multiple sclerosis. J Neuroimaging 2012;22:122–128. 2. Furby J, Hayton T, Anderson V, et al. Magnetic resonance imaging measures of brain and spinal cord atrophy correlate with clinical impairment in secondary progressive multiple sclerosis. Mult Scler 2008;14:1068–1075. 3. Lin X, Tench CR, Evangelou N, Jaspan T, Constantinescu CS. Measurement of spinal cord atrophy in multiple sclerosis. J Neuroimaging 2004;14(3 Suppl):20S–26S. 4. Freund P, Weiskopf N, Ward NS, et al. Disability, atrophy and cortical reorganization following spinal cord injury. Brain 2011; 134(Pt 6):1610–1622. 5. Cohen-Adad J, El Mendili MM, Leh ericy S, et al. Demyelination and degeneration in the injured human spinal cord detected with diffusion and magnetization transfer MRI. Neuroimage 2011;55: 1024–1033. 6. Lundell H, Barthelemy D, Skimminge A, Dyrby TB, BieringSïrensen F, Nielsen JB. Independent spinal cord atrophy measures correlate to motor and sensory deficits in individuals with spinal cord injury. Spinal Cord 2011;49:70–75. 7. Cohen-Adad J, El Mendili MM, Morizot-Koutlidis R, et al. Involvement of spinal sensory pathway in ALS and specificity of cord atrophy to lower motor neuron degeneration. Amyotroph Lateral Scler Frontotemporal Degener 2013;14:30–38. 8. Agosta F, Rocca MA, Valsasina P, et al. A longitudinal diffusion tensor MRI study of the cervical cord and brain in amyotrophic lateral sclerosis patients. J Neurol Neurosurg Psychiatry 2009;80: 53–55. 9. Yiannakas MC, Kearney H, Samson RS, et al. Feasibility of grey matter and white matter segmentation of the upper cervical cord in vivo: a pilot study with application to magnetisation transfer measurements. Neuroimage 2012;63:1054–1059. 10. Ozturk A, Aygun N, Smith SA, Caffo B, Calabresi PA, Reich DS. Axial 3D gradient-echo imaging for improved multiple sclerosis lesion detection in the cervical spinal cord at 3T. Neuroradiology 2013;55:431–439. 11. Martin N, Malfair D, Zhao Y, et al. Comparison of MERGE and axial T2-weighted fast spin-echo sequences for detection of multiple sclerosis lesions in the cervical spinal cord. AJR Am J Roentgenol 2012;199:157–162. 12. White ML, Zhang Y, Healey K. Cervical spinal cord multiple sclerosis: evaluation with 2D multi-echo recombined gradient echo MR imaging. J Spinal Cord Med 2011;34:93–98.
459 13. Smith SA, Edden RA, Farrell JA, Barker PB, Van Zijl PC. Measurement of T1 and T2 in the cervical spinal cord at 3 Tesla. Magn Reson Med 2008;60:213–219. 14. Carbonell-Caballero J, Manj on JV, Martı-Bonmatı L, et al. Accurate quantification methods to evaluate cervical cord atrophy in multiple sclerosis patients. MAGMA 2006;19:237–246. 15. Tench CR, Morgan PS, Constantinescu CS. Measurement of cervical spinal cord cross-sectional area by MRI using edge detection and partial volume correction. J Magn Reson Imaging 2005;21: 197–203. 16. Losseff NA, Webb SL, O’Riordan JI, et al. Spinal cord atrophy and disability in multiple sclerosis. A new reproducible and sensitive MRI method with potential to monitor disease progression. Brain 1996;119(Pt 3):701–708. 17. McIntosh C, Hamarneh G, Toom M, Tam RC. Spinal cord segmentation for volume estimation in healthy and multiple sclerosis subjects using crawlers and minimal paths. Proceedings of the First IEEE International Conference on Healthcare Informatics, Imaging and Systems Biology (HISB). 2011;25–31. 18. Horsfield MA, Sala S, Neema M, et al. Rapid semi-automatic segmentation of the spinal cord from magnetic resonance images: application in multiple sclerosis. Neuroimage 2010;50:446–455. 19. Zivadinov R, Banas AC, Yella V, Abdelrahman N, WeinstockGuttman B, Dwyer MG. Comparison of three different methods for measurement of cervical cord atrophy in multiple sclerosis. AJNR Am J Neuroradiol 2008;29:319–325. 20. Hickman SJ, Hadjiprocopis A, Coulon O, Miller DH, Barker GJ. Cervical spinal cord MTR histogram analysis in multiple sclerosis using a 3D acquisition and a B-spline active surface segmentation technique. Magn Reson Imaging 2004;22:891–895. 21. Coulon O, Hickman SJ, Parker GJ, Barker GJ, Miller DH, Arridge SR. Quantification of spinal cord atrophy from magnetic resonance images via a B-spline active surface model. Magn Reson Med 2002;47:1176–1185. 22. Leary SM, Parker GJ, Stevenson VL, Barker GJ, Miller DH, Thompson AJ. Reproducibility of magnetic resonance imaging measurements of spinal cord atrophy: the role of quality assurance. Magn Reson Imaging 1999;17:773–776. 23. Klein JP, Arora A, Neema M, et al. A 3T MR imaging investigation of the topography of whole spinal cord atrophy in multiple sclerosis. AJNR Am J Neuroradiol 2011;32:1138–1142. 24. Gorgey AS, Mather KJ, Poarch HJ, Gater DR. Influence of motor complete spinal cord injury on visceral and subcutaneous adipose tissue measured by multi-axial magnetic resonance imaging. J Spinal Cord Med 2011;34:99–109. 25. Dice LR. Measures of the amount of ecologic association between species. Ecology 1945;26:297–302. 26. Kundel HL, Polansky M. Measurement of observer agreement. Radiology 2003;228:303–308. 27. Haralick RM, Shapiro LG. Computer and robot vision: vol. 1. Boston: Addison-Wesley. 1991. 28. Valsasina P, Horsfield MA, Rocca MA, Absinta M, Comi G, Filippi M. Spatial normalization and regional assessment of cord atrophy: voxel-based analysis of cervical cord 3D T1-weighted images. AJNR Am J Neuroradiol 2012;33:2195–2200. 29. Valsasina P, Rocca MA, Horsfield MA, et al. Regional cervical cord atrophy and disability in multiple sclerosis: a voxel-based analysis. Radiology 2013;266:853–861.
Original Research
Validation of a Semiautomated Spinal Cord Segmentation Method €l Chen, MSc,1,2 Brice Tiret, MSc,1,2 Mohamed-Mounir El Mendili, MSc,1,2* Raphae lanie Pe le grini-Issac, PhD,1,2 Julien Cohen-Adad, PhD,3 Ste phane Me 2,4,5 1,2,6 Pierre-Franc¸ois Pradat, PhD, MD, and Habib Benali, PhD1,2 Lehericy, PhD, MD, Purpose: To validate semiautomated spinal cord segmentation in healthy subjects and patients with neurodegenerative diseases and trauma.
region. DSC was high (0.96) in both cervical and thoracic regions. DSC remained higher than 0.8 even when modifying initial contours by 50%.
Materials and Methods: Forty-nine healthy subjects, as well as 29 patients with amyotrophic lateral sclerosis, 19 with spinal muscular atrophy, and 14 with spinal cord injuries were studied. Cord area was measured from T2weighted 3D turbo spin echo images (cord levels from C2 to T9) using the semiautomated segmentation method of Losseff et al (Brain [1996] 119(Pt 3):701–708), compared with manual segmentation. Reproducibility was evaluated using the inter- and intraobserver coefficient of variation (CoV). Accuracy was assessed using the Dice similarity coefficient (DSC). Robustness to initialization was assessed by simulating modifications to the contours drawn manually prior to segmentation.
Conclusion: The semiautomated segmentation method showed high reproducibility and accuracy in measuring spinal cord area. Key Words: spinal cord; MRI; segmentation; cross-sectional area; atrophy measurement J. Magn. Reson. Imaging 2015;41:454–459. C 2014 Wiley Periodicals, Inc. V
Results: Mean interobserver CoV was 4.00% for manual segmentation (1.90% for Losseff’s method) in the cervical region and 5.62% (respectively 2.19%) in the thoracic region. Mean intraobserver CoV was 2.34% for manual segmentation (1.08% for Losseff’s method) in the cervical region and 2.35% (respectively 1.34%) in the thoracic 1 Inserm U678, UPMC Univ Paris 6 UMR-S 678, Laboratoire d’imagerie fonctionnelle, Paris, France. 2 Univ Paris 11, IFR 49, Institut f ed eratif de recherche en imagerie neurofonctionnelle, DSV/I2BM Neurospin, Gif-sur-Yvette, France. 3 Ecole Polytechnique de Montr eal, Department of Electrical Engineering, Montr eal, Canada. 4 Inserm U975, UPMC Univ Paris 6, UMR-S975, CNRS UMR7225, pinie`re – Centre de recherche de l’Institut du Cerveau et de la Moelle e CRICM, Centre de Neuroimagerie de Recherche – CENIR, Paris, France. 5 APHP, Groupe Hospitalier Piti e-Salp^ etrie`re, Service de neuroradiologie, Paris, France. 6 APHP, Groupe Hospitalier Piti e-Salp^ etrie`re, D epartement des Maladies du syste`me Nerveux, Paris, France. Additional Supporting Information may be found in the online version of this article. Contract grant sponsor: Association franc¸aise contre les myopathies (AFM); Contract grant sponsor: Institut pour la recherche sur la pinie`re et l’enc moelle e ephale (IRME). The research leading to these results has also received funding from the program “Investissements d’avenir” ANR-10-IAIHU-06. *Address reprint requests to: M.-M.E.M., Laboratoire d’Imagerie Fonctionnelle, Facult e de M edecine Pierre et Marie Curie, Site Piti eSalp^ etrie`re, 91 Boulevard de l’H^ opital, F-75634 Paris cedex 13, France. E-mail: [email protected] Received September 26, 2013; Accepted January 3, 2014. DOI 10.1002/jmri.24571 View this article online at wileyonlinelibrary.com. C 2014 Wiley Periodicals, Inc. V
SEVERAL MAGNETIC RESONANCE IMAGING (MRI) studies have shown that spinal cord atrophy may be used as an indicator of tissue destruction in neurological disorders such as multiple sclerosis (MS) (1–3), spinal cord injury (SCI) (4–6), and amyotrophic lateral sclerosis (ALS) (7,8). Atrophy measurement has been facilitated by advances in MRI technology (9–13). In parallel, several segmentation methods with varying degrees of manual intervention have been proposed to evaluate cord atrophy, including measurements of cross-sectional area (CSA) (14–16) and volume of the spinal cord (17–21). However, most studies have used low-resolution 1.5T images and investigated only the upper cervical spinal cord in MS patients. Validation of cord atrophy as a biomarker in spinal cord diseases requires evaluating these segmentation methods in other segments of the spinal cord as well as in diseases other than MS. The threshold-based segmentation method proposed by Losseff et al (16) is one of the most-used approaches to quantify atrophy due to its high reproducibility and accuracy. This method has to be initialized manually by drawing two contours, one to delineate the spinal cord and the other surrounding the cerebrospinal fluid (CSF). However, this method has only been validated at the C2 vertebral level using 1.5T images in MS patients (16,18,22). The present study had two main objectives: 1) To validate Losseff’s segmentation method at both the cervical and thoracic levels and in neurological disorders other than MS. 2) To evaluate its robustness to manual initialization.
454
Semiautomated Spinal Cord Segmentation
MATERIALS AND METHODS Subjects In this study, 111 subjects were recruited: healthy volunteers (n ¼ 49, mean age 6 SD: 39 6 17 years, 29 females), patients with ALS (n ¼ 29, 53 6 10 years, 7 females, mean disease duration: 27 6 27 months), SMN1-linked spinal muscular atrophy (SMA) patients (n ¼ 19, 37 6 11 years, 9 females, mean disease duration: 25 6 15 months), and SCI (n ¼ 14, 45 6 14 years, 3 females, delay after injury: 25 6 35 months). The local Ethical Committee of our institution approved all experimental procedures of the study and written informed consent was obtained from each participant. MRI Acquisition All scans were performed using a 3T MRI system (TIM Trio, Siemens Healthcare, Erlangen, Germany) using a body coil for signal excitation and a neck/spine coil for signal reception. Subjects were positioned as comfortably as possible and were systematically asked not to move during the acquisition in order to minimize motion artifacts. Furthermore, visual quality control was performed after the acquisition to make sure that there were no visible motion artifacts in the data. The spinal cord was imaged using a T2-weighted 3D turbo spin echo sequence with a slab-selective excitation pulse. Imaging parameters were: voxel size: 0.9 0.9 0.9 mm3; field of view (FOV): 280 280 mm2; 52 sagittal slices; TR/TE ¼ 1500/120 msec; flip angle: 140 ; generalized autocalibrating partially parallel acquisition (GRAPPA) with acceleration factor R ¼ 3; turbo factor: 69 and acquisition time 6 minutes. This sequence provided a high signal-to-noise ratio (SNR) due to 3D acquisition, high resolution owing to isotropic acquisition, short acquisition times by combining parallel acquisition with high turbo factors, and low specific absorption rate due to low flip angle refocusing pulses. Segmentation Method Table 1 shows the number of segmented vertebral levels for each group of subjects. Due to FOV positioning and variable vertebral size, it was not possible to segment the whole cervical and thoracic spinal cord. C1 and C2 vertebral levels were grouped into a single level because the boundary between C1 and C2 was often indistinct. Segmentation was stopped at the T10 level due to the increasing number and size of nerve roots, which made segmentation inaccurate, and to the reduction of SNR in the lower parts of the spinal cord. In the SCI group, only slices rostral and caudal to the lesions were selected and segmented (5). Segmentation and statistical analyses were conducted with MatLab (MathWorks, Natick, MA). The following procedure proposed in (6) was adopted for spinal cord segmentation. Preprocessing First, the sagittal slice where the spinal cord was the most median at the vertebral level to be segmented
455 Table 1 Number of Segmented Vertebral Levels for Each Group of Subjects Groups Level
Controls
SLA
SMA
SCI
C1-C2 C3 C4 C5 C6 C7 T1 T2 T3 T4 T5 T6 T7 T8 T9
31 44 46 46 48 48 48 47 47 45 35 25 10 7 5
28 29 29 29 29 29 29 29 29 28 26 18 3 0 0
19 19 19 18 18 18 18 18 18 18 16 8 1 0 0
12 9 4 7 4 4 11 14 13 11 3 0 0 0 0
was selected manually. A segment was drawn manually parallel to the spinal cord to delimit the vertebral level of interest. The resulting axial slices, perpendicular to the spinal cord, were resampled to a voxel size of 0.3 0.3 0.3 mm3 using 3D cubic interpolation. The slices were automatically cropped to squares of 120 120 voxels centered on the segment drawn manually, then averaged. This whole procedure minimizes the partial volume effect between spinal cord and CSF. Segmentation Method The contours of two regions of interest (ROIs) were drawn manually in the average axial slice to delineate spinal cord and CSF areas, respectively. This outlining step was achieved by two operators (M.-M.E.M., R.C.) blinded to the type of data. One operator was experienced with spinal cord segmentation (ie, 3 years experience) and the other had no previous experience (ie, had been given brief instructions beforehand on five different datasets randomly selected from those used in the present study). The mid intensity of the mean intensities in the two ROIs was used as a threshold for further separating ROI voxels into two classes, the spinal cord and the CSF (16). In parallel, manual segmentation was also carried out by the same two operators in the images resulting from the preprocessing step. The cross-sectional area (CSA) in mm2 was finally defined as the number of pixels in the resulting spinal cord class, multiplied by the pixel size. Reproducibility Reproducibility for both manual segmentation and Losseff’s method was evaluated using inter- and intraobserver coefficients of variation (CoV, defined as the ratio of the standard deviation to the average of CSA values across measurements). To evaluate
456
El Mendili et al.
Figure 1. a: T2-weighted MRI mid-sagittal slice for an SMA patient, showing the anatomical landmarks of the cervical spinal cord. A segment (orange arrow) was drawn manually parallel to the spinal cord to delimit the vertebral level of interest, here C6. b: Contours drawn manually by the experienced operator delineating the CSF and the spinal cord on the resampled and averaged axial slices (C6), corresponding to AR ¼ 0%. d,f,h: Simulated contours corresponding to AR ¼ 15%, 30%, and 50%, respectively. c,e,g,h: Contours resulting from the semiautomated method for AR ¼ 0%, 15%, 30%, and 50%, respectively.
interobserver CoV, the two operators segmented all images. Intraobserver COVs were evaluated on 20 datasets: Five datasets were randomly selected from each group of subjects and segmented twice by both operators at least 2 weeks apart. The operators were blinded to the type of data.
between 0.6 and 0.8 can be interpreted as a substantial agreement, and a DSC value between 0.4 and 0.6 represents a moderate agreement (26).
Accuracy
The robustness of the semiautomated segmentation method to its initialization was assessed by conducting numerical simulations: we simulated modifications to the initial ROIs drawn manually by the experienced operator, which were used as a reference. To do so, we randomly suppressed border voxels of the initial ROIs, thus defining an "attack rate" (AR) as the sum of suppressed border voxels divided by the total number of initial ROI voxels. This operation was repeated 200 times for each AR (from 0% to 50% by steps of 1%), thus simulating a large variety of manually drawn ROIs. The simulated ROIs were then introduced as inputs to the semiautomated method. The procedure is illustrated in Fig. 1. Finally, the DSC was computed as previously to compare final results
Accuracy of the semiautomated segmentation method was assessed using the Dice similarity coefficient (DSC), which evaluates the performance of segmentation methods by measuring their spatial overlaps (25). The DSC is defined as follows: DSC ¼
2 jM \ Lj ; jMj þ jLj
where jMj (respectively jLj) is the CSA in mm2 obtained by manual (respectively Losseff’s) segmentation, and jM \ Lj is the area in mm2 common to both segmentation results. A DSC value higher than 0.8 indicates an almost perfect agreement, a DSC value
Robustness to Initialization
Semiautomated Spinal Cord Segmentation
457
Table 2 Reproducibility of Mean Cross-Sectional Area Measurements in Cervical and Thoracic Regions Mean cord area (mm2) Cord region Cervical Thoracic
Method Manual Semiautomated Manual Semiautomated
Mean coefficients of variation (%)
Operator 1 Mean (SD)
Operator 2 Mean (SD)
73.48 68.85 46.76 43.66
70.10 69.83 43.35 44.71
(8.43) (8.28) (5.65) (5.63)
of the semiautomated method for each AR and manual segmentation. RESULTS Computation time (mean 6 SD) for the preprocessing step was: 5.0 6 1.1 min/subject, 5.5 6 1.5 min/subject for manual segmentation and 3.0 6 0.9 min/subject for the semiautomated segmentation method. This was achieved using a 64-bit Quad-core (Intel Xenon, processor speed: 2.67 GHz) workstation.
(8.52) (8.46) (5.69) (5.64)
Interobserver Mean (SD) 4.00 1.90 5.62 2.19
(2.76) (1.74) (3.40) (2.10)
Intraobserver Mean (SD) 2.34 1.08 2.53 1.34
(0.96) (0.50) (1.17) (0.92)
regions (0.96). The semiautomated method underestimated the mean CSA measured by manual segmentation (considered as the ground truth) by 3.29% in the cervical region and 1.70% in the thoracic region. When considering each group of subjects separately, the semiautomated method underestimated the mean CSA in the cervical region (respectively in the thoracic region) by 3.07% (respectively 1.38%) in the control group, 3.68% (respectively 1.26%) in the ALS group, 3.23% (respectively 2.19%) in the SMA group, and 3.31% (respectively 3.06%) in the SCI group (Supporting Table S2).
Reproducibility The results are shown in Table 2 for all subjects. The mean interobserver CoV in the cervical region was 4.00% for manual segmentation and 1.90% for the semiautomated method in the cervical region, and 5.62% and 2.19%, respectively, in the thoracic region. The mean intraobserver CoV was 2.34% for manual segmentation and 1.08% for the semiautomated method in the cervical region, and 2.53% and 1.34%, respectively, in the thoracic region. The smallest volume change that the semiautomated method was able to detect regardless of the errors made by operators in the repetitive use of the technique was calculated using the following formula: 1.96 intraobserver CoV mean CSA (16,18). This showed that changes as small as 1.50 mm2 (1.96 SDs) in the cervical cord area and 1.16 mm2 in the thoracic cord area may be detected with 95% confidence. Similar results were obtained when considering separately healthy volunteers and patients (Supporting Table S1). There was no significant correlation between the mean CSA and mean interobserver CoVs in both cervical and thoracic regions (pCervical ¼ 0.35, pThoracic ¼ 0.19), or between the mean CSA and mean intraobserver CoVs (pCervical ¼ 0.29, pThoracic ¼ 0.86). Accuracy The results are shown in Table 3. The mean DSC value was high in both cervical (0.96) and thoracic
Robustness to Initialization The results are shown in Fig. 2. The mean DSC value remained stable for attack rates ranging from 0% to 30% in both cervical and thoracic regions. The variance was larger for AR values less than 10%, then decreased with increasing AR up to 30%. After this plateau, the mean DSC value decreased rapidly followed by an increase in variance. The maximum mean DSC was obtained for AR ¼ 15% in the cervical region (DSCCervical ¼ 0.97) and AR ¼ 17% in the thoracic region (DSCThoracic ¼ 0.97). The minimum mean DSC values were obtained for AR ¼ 50% in both regions (DSCCervical ¼ 0.89, DSCThoracic ¼ 0.90). DISCUSSION The semiautomated segmentation method was validated at 3T in the cervical and thoracic spinal cord in a large set of healthy volunteers and three groups of patients with ALS, SMA, and SCI. The method showed high inter- and intraobserver reproducibility in the cervical and thoracic regions. The mean inter- and intraobserver CoVs were about twice as high for manual segmentation than for the semiautomated method. CoVs were higher in the thoracic than in the cervical region, either in patient groups or in the control group (Supporting Table S1). This result can be explained by, on the one hand, the
Table 3 Accuracy of Mean Cross-Sectional Area Measurements in Cervical and Thoracic Regions Mean Dice similarity coefficient
Mean error estimation (%)
Cord region
Operator 1 Mean (SD)
Operator 2 Mean (SD)
Operator 1 Mean (SD)
Operator 2 Mean (SD)
Cervical Thoracic
0.97 (0.02) 0.96 (0.02)
0.96 (0.02) 0.95 (0.02)
0.28 (3.55) 3.27 (3.27)
4.30 (3.06) 4.99 (3.09)
458
El Mendili et al.
Figure 2. DSC function of the attack rate in cervical and thoracic regions. The solid line represents the mean DSC and the dashed lines the mean 6 SD.
fact that thoracic images presented more physiological artifacts due to respiratory movements inducing errors in manual and semiautomated segmentation, and lower SNRs inherent to the thoracic coil compared to the cervical receive coil and, on the other hand, the difference in operators’ experience in the use of the technique. C1 and C2 vertebral levels could not be segmented separately because the boundary between C1 and C2 was difficult to distinguish in all subjects, despite the high resolution of the images. CoVs should be lower at C2 than for the combined C1–C2 vertebral level. The mean inter- and intraobserver CoVs were 1.30% and 0.60%, respectively. These values were lower than those obtained in the C2 region in Horsfield et al (18) (20 controls and 40 MS patients), where the authors reported an intraobserver CoV of 2.15% and an interobserver CoV of 7.05% for the semiautomated segmentation method. In Losseff et al (16), the interobserver CoV was 0.83%, which is lower than the value obtained in the present study. This can be explained first by our larger sample size (n ¼ 111) as compared with 15 subjects in Losseff et al (16) and, second, by the difference in operators’ experience in the use of the technique. In contrast, we found a lower intraobserver CoV (0.60%) than in Losseff et al (16) (0.83%), probably due to the high spatial resolution obtained at 3T and possible improvement provided by the preprocessing step. CoVs in the present study were similar to those obtained in Horsfield et al (18) in the C2 region (intraobserver CoV ¼ 0.59% using 15 scans; interobserver CoV ¼ 1.36% using 60 scans), where the authors used a method based on image smoothing followed by an active surface model constrained by intensity gradient image information. The DSC was higher than 0.8 in the cervical and thoracic regions, indicating almost perfect agreement
between the semiautomated segmentation method and manual outlining. Despite these high performances, the semiautomated method had a number of limitations. In the preprocessing step, partial volume was still present even if axial slices were resampled, and this led to underestimation of the cross-sectional area. This issue has been outlined previously. Partial volume correction was proposed in Tench et al (15) by calculating the portion of cord in the voxels immediately surrounding the detected edge. This method increased the accuracy of the estimated CSA by 1.45%. Another limitation was that CSF contrast was not homogeneous in all slices, especially in the thoracic region, which presented more motion artifacts. In addition, small structures such as nerve roots, with signal magnitude similar to that of the spinal cord, were included in the final segmentation, leading to an overestimation of the cross-sectional area. These structures may be removed using a region-labeling method combined with priors on the size and shape of the cord (18,27). An additional limitation of the present study was the lack of scan–rescan reproducibility evaluation. This was due to imaging protocol constraints, especially concerning inclusion of patients with large motor deficits (ALS and SCI patients). A comparison between CSA and volume measurements would be of interest, as many recently published studies have used volume measurement instead of CSA to quantify age-related atrophy as well as atrophy in MS patients (28,29). In conclusion, the semiautomated method of Losseff et al (16) showed high reproducibility and accuracy for measuring spinal cord cross-sectional area in both cervical and thoracic regions at 3T in normal subjects and a large set of patients with various spinal cord
Semiautomated Spinal Cord Segmentation
disorders. Improved image spatial resolution obtained at 3T allowed higher segmentation performances than reported previously. ACKNOWLEDGMENTS We thank all the subjects and their relatives. We thank Kevin Nigaud, Romain Valabre`gue, Alexandre de ric Humbert, Christelle Macia, and Vignaud, Fre Eric Bardinet for helping with the acquisition and Dr. Henrik Lundell for making the code for measuring cord area available to us. REFERENCES 1. Cohen AB, Neema M, Arora A, et al. The relationships among MRI-defined spinal cord involvement, brain involvement, and disability in multiple sclerosis. J Neuroimaging 2012;22:122–128. 2. Furby J, Hayton T, Anderson V, et al. Magnetic resonance imaging measures of brain and spinal cord atrophy correlate with clinical impairment in secondary progressive multiple sclerosis. Mult Scler 2008;14:1068–1075. 3. Lin X, Tench CR, Evangelou N, Jaspan T, Constantinescu CS. Measurement of spinal cord atrophy in multiple sclerosis. J Neuroimaging 2004;14(3 Suppl):20S–26S. 4. Freund P, Weiskopf N, Ward NS, et al. Disability, atrophy and cortical reorganization following spinal cord injury. Brain 2011; 134(Pt 6):1610–1622. 5. Cohen-Adad J, El Mendili MM, Leh ericy S, et al. Demyelination and degeneration in the injured human spinal cord detected with diffusion and magnetization transfer MRI. Neuroimage 2011;55: 1024–1033. 6. Lundell H, Barthelemy D, Skimminge A, Dyrby TB, BieringSïrensen F, Nielsen JB. Independent spinal cord atrophy measures correlate to motor and sensory deficits in individuals with spinal cord injury. Spinal Cord 2011;49:70–75. 7. Cohen-Adad J, El Mendili MM, Morizot-Koutlidis R, et al. Involvement of spinal sensory pathway in ALS and specificity of cord atrophy to lower motor neuron degeneration. Amyotroph Lateral Scler Frontotemporal Degener 2013;14:30–38. 8. Agosta F, Rocca MA, Valsasina P, et al. A longitudinal diffusion tensor MRI study of the cervical cord and brain in amyotrophic lateral sclerosis patients. J Neurol Neurosurg Psychiatry 2009;80: 53–55. 9. Yiannakas MC, Kearney H, Samson RS, et al. Feasibility of grey matter and white matter segmentation of the upper cervical cord in vivo: a pilot study with application to magnetisation transfer measurements. Neuroimage 2012;63:1054–1059. 10. Ozturk A, Aygun N, Smith SA, Caffo B, Calabresi PA, Reich DS. Axial 3D gradient-echo imaging for improved multiple sclerosis lesion detection in the cervical spinal cord at 3T. Neuroradiology 2013;55:431–439. 11. Martin N, Malfair D, Zhao Y, et al. Comparison of MERGE and axial T2-weighted fast spin-echo sequences for detection of multiple sclerosis lesions in the cervical spinal cord. AJR Am J Roentgenol 2012;199:157–162. 12. White ML, Zhang Y, Healey K. Cervical spinal cord multiple sclerosis: evaluation with 2D multi-echo recombined gradient echo MR imaging. J Spinal Cord Med 2011;34:93–98.
459 13. Smith SA, Edden RA, Farrell JA, Barker PB, Van Zijl PC. Measurement of T1 and T2 in the cervical spinal cord at 3 Tesla. Magn Reson Med 2008;60:213–219. 14. Carbonell-Caballero J, Manj on JV, Martı-Bonmatı L, et al. Accurate quantification methods to evaluate cervical cord atrophy in multiple sclerosis patients. MAGMA 2006;19:237–246. 15. Tench CR, Morgan PS, Constantinescu CS. Measurement of cervical spinal cord cross-sectional area by MRI using edge detection and partial volume correction. J Magn Reson Imaging 2005;21: 197–203. 16. Losseff NA, Webb SL, O’Riordan JI, et al. Spinal cord atrophy and disability in multiple sclerosis. A new reproducible and sensitive MRI method with potential to monitor disease progression. Brain 1996;119(Pt 3):701–708. 17. McIntosh C, Hamarneh G, Toom M, Tam RC. Spinal cord segmentation for volume estimation in healthy and multiple sclerosis subjects using crawlers and minimal paths. Proceedings of the First IEEE International Conference on Healthcare Informatics, Imaging and Systems Biology (HISB). 2011;25–31. 18. Horsfield MA, Sala S, Neema M, et al. Rapid semi-automatic segmentation of the spinal cord from magnetic resonance images: application in multiple sclerosis. Neuroimage 2010;50:446–455. 19. Zivadinov R, Banas AC, Yella V, Abdelrahman N, WeinstockGuttman B, Dwyer MG. Comparison of three different methods for measurement of cervical cord atrophy in multiple sclerosis. AJNR Am J Neuroradiol 2008;29:319–325. 20. Hickman SJ, Hadjiprocopis A, Coulon O, Miller DH, Barker GJ. Cervical spinal cord MTR histogram analysis in multiple sclerosis using a 3D acquisition and a B-spline active surface segmentation technique. Magn Reson Imaging 2004;22:891–895. 21. Coulon O, Hickman SJ, Parker GJ, Barker GJ, Miller DH, Arridge SR. Quantification of spinal cord atrophy from magnetic resonance images via a B-spline active surface model. Magn Reson Med 2002;47:1176–1185. 22. Leary SM, Parker GJ, Stevenson VL, Barker GJ, Miller DH, Thompson AJ. Reproducibility of magnetic resonance imaging measurements of spinal cord atrophy: the role of quality assurance. Magn Reson Imaging 1999;17:773–776. 23. Klein JP, Arora A, Neema M, et al. A 3T MR imaging investigation of the topography of whole spinal cord atrophy in multiple sclerosis. AJNR Am J Neuroradiol 2011;32:1138–1142. 24. Gorgey AS, Mather KJ, Poarch HJ, Gater DR. Influence of motor complete spinal cord injury on visceral and subcutaneous adipose tissue measured by multi-axial magnetic resonance imaging. J Spinal Cord Med 2011;34:99–109. 25. Dice LR. Measures of the amount of ecologic association between species. Ecology 1945;26:297–302. 26. Kundel HL, Polansky M. Measurement of observer agreement. Radiology 2003;228:303–308. 27. Haralick RM, Shapiro LG. Computer and robot vision: vol. 1. Boston: Addison-Wesley. 1991. 28. Valsasina P, Horsfield MA, Rocca MA, Absinta M, Comi G, Filippi M. Spatial normalization and regional assessment of cord atrophy: voxel-based analysis of cervical cord 3D T1-weighted images. AJNR Am J Neuroradiol 2012;33:2195–2200. 29. Valsasina P, Rocca MA, Horsfield MA, et al. Regional cervical cord atrophy and disability in multiple sclerosis: a voxel-based analysis. Radiology 2013;266:853–861.
