# Spinal Cord Morphology in Degenerative Cervical Myelopathy Patients; Assessing Key Morphological Characteristics Using Machine Vision Tools

^{1}

^{2}

^{3}

^{4}

^{5}

^{6}

^{7}

^{8}

^{*}

^{†}

## Abstract

**:**

## 1. Introduction

## 2. Methods

#### 2.1. Computational Tools Used

`Spinal Cord Toolbox, v.5.0.1`[13],

`3D Slicer v.4.10.2`[18],

`SciKit-Learn v.0.23.2`[19],

`SciPy v.1.5.2`[20],

`matplotlib v.3.3.2`[21],

`seaborn v.0.11.1`[22],

`numpy v.1.19.2`[23], and

`pandas v.1.2.0`[24]. As CovBat was still in development at time of this paper’s publication [25], its state at the time of this analysis can be replicated by using the GitHub commit

`23a0429`, available at https://github.com/andy1764/CovBat_Harmonization/commit/23a0429c2a81e7682da94ff2d0f5e634ab91b429 (accessed on 9 June 2020).

#### 2.2. Data Preparation

#### 2.3. Spinal Cord Segmentation

#### 2.4. Metric Extraction and Standardization

`‘sct_process_segmentation’`script to extract metrics from each spinal cord image’s segmentations (both automated and manual). All metrics were taken from the entire spinal cord volume, and included the means and standard deviations of the cross-sectional area of the spinal cord segmentation slices (mm squared), anterior/posterior angle (degrees), right/left angle (degrees), anterior/posterior diameter (mm), right/left diameter (mm), eccentricity (ratio of two prior diameter measurements), orientation (relative angle, image to spine), and solidity (ratio of true and convex-fit cross-sectional area). The total length of the spinal cord (mm) was also obtained, being produced by the same analysis pipeline; given its tenuous-at-best relation to the morphology associated with DCM, this was kept to evaluate SCT’s options in full. That is to say, we did not expect length (sum) to be useful to any model, but included for the sake of being thorough.

#### 2.5. Model Metric Selection

`SelectFdr`function. The scoring function was set to the F-test score of the metric to the mJOA score (evaluated with SciKit-Learn’s

`‘f_regression’`function) or DCM severity category (evaluated with SciKit-Learn’s

`‘f_classif’`function). The F-test was selected for its ability to evaluate whether data would conform well in a regression model; as we kept to simple regression-based models for this study (see below), this fit our use case perfectly. The allowable probability of false discovery was set to $p=0.05$. This feature selection process served both to reduce the list of spinal cord morphological metrics to only those anticipated to be correlated with our target metric (our mJOA score or the mJOA severity categories), but also to filter out assessment methodologies which are likely to be ineffective (by selecting 0 features for them). This resulted in a drastic reduction in valid assessment methodologies, with at most 3 passing this stage per severity category and model type (linear or categorical) and proceeding to the final model assessment.

#### 2.6. mJOA Correlation and Categorization Model Assessment

`‘LinearRegression’`model (for linear metric to mJOA score models) or

`‘LogisticRegression’`model (for DCM severity classification models). These simple models fit linearly to each parameter, allowing for metrics to be evaluated sans-interaction effects, and does so very quickly. This made them ideal for rapid, diverse, and simple assessments, perfect for evaluating the SCT derived metrics on their own. All groups were split into train-test groups using 5-fold shuffle split grouping, and cross-validated by fitting the modeling method to each group in turn. Each resulting model’s effectiveness was then evaluated using ${r}^{2}$ for the linear regression models, and using receiver operating characteristic area under curve (ROC AUC) for categorical models. The effectiveness of the model type was then assessed via the mean score of all resulting models. To confirm that the somewhat experimental CovBat method worked correctly, all processes prior were run on both the standardized-only metric sets and the CovBat-harmonized metric sets as well. Categorical imbalance was also evaluated for each model type via assessing the accuracy of a “dummy” model, which simply guessed the most common category at all times.

## 3. Results

#### 3.1. Spinal Cord Metrics of DCM Patients by mJOA Severity

#### 3.2. Manual vs. Automated Segmentation Metrics

#### 3.3. mJOA Score Regression by Assessment Methodology

`‘f_regression’`function). However, only the T2w contrast, sagittal orientation, and the svm deepseg segmentation algorithm methodology produce a model which had more than 3 parameters significantly related to mJOA score, with 5 total; mean of spinal cross-sectional area ($p=0.007$), mean of anterior/posterior cross-sectional diameter ($p=0.001$), mean right/left spinal angle ($p=0.024$), mean eccentricity ($p=0.031$), and mean solidity ($p=0.013$). For all other groups, a combination of these metrics, with the occasional standard deviation of solidity, angle, or diameter was observed to have significant predictive power with the mJOA score. Notably, however, the T2w contrast, sagittal orientation, propseg segmentation algorithm methodology was the only one to find total summed length of the spinal cord as significantly related, despite our assumption that it would not be found as such. A more detailed overview of the distributions of these p-values has been visualized by metric (Figure 6) and methodology element (Figure 7).

#### 3.4. Linear mJOA Prediction Models

`‘r2_score’`function, which can produce negative ${r}^{2}$ scores which imply that the associated model is worse-than-random. For non-batch compensated data, the ${r}^{2}$ scores hovered around −30, while batch compensated metric derived models resulted in ${r}^{2}$ scores ranging from −25 to −10. False Discovery Rate Feature Selection also tended to choose more features for the harmonized data set (with harmonized models having an average of 2 features selected, versus the 1.33 feature average form models trained on standardized metrics alone). This implies that the harmonization processed removed noise which otherwise masked useful trends, though clearly this was still not enough to lead to a valuable model. Tables summarizing these attributes, for both standardized (Table 4) and harmonized (Table 5), are available for further inspection.

#### 3.5. Logistic DCM Categorical Models

## 4. Discussion and Conclusions

## Author Contributions

## Funding

## Institutional Review Board Statement

## Informed Consent Statement

## Data Availability Statement

## Acknowledgments

## Conflicts of Interest

## Abbreviations

DCM | Degenerative Cervical Myelopathy |

MRI | Magnetic Resonance Imaging |

SCT | Spinal Cord Toolbox |

FMRIB | Functional Magnetic Resonance Imaging of the Brain |

CSORN | Canadian Spine Outcomes and Research |

DICOM | Digital Imaging and Communications in Medicine |

CNN | Convolutional Neural Network |

SVM | Support Vector Machine |

ROC AUC | Receiver Operating Characteristic Area Under Curve |

## References

- Nouri, A.; Tetreault, L.; Singh, A.; Karadimas, S.K.; Fehlings, M.G. Degenerative cervical myelopathy: Epidemiology, genetics, and pathogenesis. Spine
**2015**, 40, E675–E693. [Google Scholar] [CrossRef] [PubMed] - Davies, B.M.; Mowforth, O.D.; Smith, E.K.; Kotter, M.R. Degenerative cervical myelopathy. BMJ
**2018**, 360. [Google Scholar] [CrossRef] [Green Version] - Tracy, J.A.; Bartleson, J. Cervical spondylotic myelopathy. Neurologist
**2010**, 16, 176–187. [Google Scholar] [CrossRef] [PubMed] - Kovalova, I.; Kerkovsky, M.; Kadanka, Z.; Kadanka, Z., Jr.; Nemec, M.; Jurova, B.; Dusek, L.; Jarkovsky, J.; Bednarik, J. Prevalence and imaging characteristics of nonmyelopathic and myelopathic spondylotic cervical cord compression. Spine
**2016**, 41, 1908–1916. [Google Scholar] [CrossRef] - Bednarik, J.; Kadanka, Z.; Dusek, L.; Kerkovsky, M.; Vohanka, S.; Novotny, O.; Urbanek, I.; Kratochvilova, D. Presymptomatic spondylotic cervical myelopathy: An updated predictive model. Eur. Spine J.
**2008**, 17, 421–431. [Google Scholar] [CrossRef] [PubMed] [Green Version] - Kopjar, B.; Bohm, P.E.; Arnold, J.H.; Fehlings, M.G.; Tetreault, L.A.; Arnold, P.M. Outcomes of surgical decompression in patients with very severe degenerative cervical myelopathy. Spine
**2018**, 43, 1102. [Google Scholar] [CrossRef] - Boogaarts, H.D.; Bartels, R.H. Prevalence of cervical spondylotic myelopathy. Eur. Spine J.
**2015**, 24, 139–141. [Google Scholar] [CrossRef] - Evaniew, N.; Cadotte, D.W.; Dea, N.; Bailey, C.S.; Christie, S.D.; Fisher, C.G.; Paquet, J.; Soroceanu, A.; Thomas, K.C.; Rampersaud, Y.R.; et al. Clinical predictors of achieving the minimal clinically important difference after surgery for cervical spondylotic myelopathy: An external validation study from the Canadian Spine Outcomes and Research Network. J. Neurosurg. Spine
**2020**, 33, 129–137. [Google Scholar] [CrossRef] - Jenkinson, M.; Beckmann, C.F.; Behrens, T.E.; Woolrich, M.W.; Smith, S.M. Fsl. Neuroimage
**2012**, 62, 782–790. [Google Scholar] [CrossRef] [Green Version] - Penny, W.D.; Friston, K.J.; Ashburner, J.T.; Kiebel, S.J.; Nichols, T.E. Statistical Parametric Mapping: The Analysis of Functional Brain Images; Elsevier: Amsterdam, The Netherlands, 2011. [Google Scholar]
- Vincent, R.D.; Neelin, P.; Khalili-Mahani, N.; Janke, A.L.; Fonov, V.S.; Robbins, S.M.; Baghdadi, L.; Lerch, J.; Sled, J.G.; Adalat, R.; et al. MINC 2.0: A flexible format for multi-modal images. Front. Neuroinform.
**2016**, 10, 35. [Google Scholar] [CrossRef] - Tetreault, L.A.; Kopjar, B.; Vaccaro, A.; Yoon, S.T.; Arnold, P.M.; Massicotte, E.M.; Fehlings, M.G. A clinical prediction model to determine outcomes in patients with cervical spondylotic myelopathy undergoing surgical treatment: Data from the prospective, multi-center AOSpine North America study. JBJS
**2013**, 95, 1659–1666. [Google Scholar] [CrossRef] [Green Version] - De Leener, B.; Lévy, S.; Dupont, S.M.; Fonov, V.S.; Stikov, N.; Collins, D.L.; Callot, V.; Cohen-Adad, J. SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data. Neuroimage
**2017**, 145, 24–43. [Google Scholar] [CrossRef] - Martin, A.R.; De Leener, B.; Cohen-Adad, J.; Kalsi-Ryan, S.; Cadotte, D.W.; Wilson, J.R.; Tetreault, L.; Nouri, A.; Crawley, A.; Mikulis, D.J.; et al. Monitoring for myelopathic progression with multiparametric quantitative MRI. PLoS ONE
**2018**, 13, e0195733. [Google Scholar] - Perone, C.S.; Cohen-Adad, J. Promises and limitations of deep learning for medical image segmentation. J. Med. Artif. Intell.
**2019**, 2. [Google Scholar] [CrossRef] - Thakar, S.; Arun, A.; Aryan, S.; Mohan, D.; Hegde, A. Deep flexor sarcopenia as a predictor of poor functional outcome after anterior cervical discectomy in patients with myelopathy. Eur. J. Neurosurg.
**2019**, 161, 2201–2209. [Google Scholar] [CrossRef] - Martin, A.; De Leener, B.; Cohen-Adad, J.; Cadotte, D.; Nouri, A.; Wilson, J.; Tetreault, L.; Crawley, A.; Mikulis, D.; Ginsberg, H.; et al. Can microstructural MRI detect subclinical tissue injury in subjects with asymptomatic cervical spinal cord compression? A prospective cohort study. BMJ Open
**2018**, 8, e019809. [Google Scholar] [CrossRef] - Kikinis, R.; Pieper, S.D.; Vosburgh, K.G. 3D Slicer: A platform for subject-specific image analysis, visualization, and clinical support. In Intraoperative Imaging and Image-Guided Therapy; Springer: Cham, Switzerland, 2014; pp. 277–289. [Google Scholar]
- Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. Scikit-learn: Machine Learning in Python. J. Mach. Learn. Res.
**2011**, 12, 2825–2830. [Google Scholar] - Virtanen, P.; Gommers, R.; Oliphant, T.E.; Haberland, M.; Reddy, T.; Cournapeau, D.; Burovski, E.; Peterson, P.; Weckesser, W.; Bright, J.; et al. SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python. Nat. Methods
**2020**, 17, 261–272. [Google Scholar] [CrossRef] [Green Version] - Hunter, J.D. Matplotlib: A 2D graphics environment. Comput. Sci. Eng.
**2007**, 9, 90–95. [Google Scholar] [CrossRef] - Waskom, M.; Gelbart, M.; Botvinnik, O.; Ostblom, J.; Hobson, P.; Lukauskas, S.; Gemperline, D.C.; Augspurger, T.; Halchenko, Y.; Warmenhoven, J.; et al. Mwaskom/Seaborn. 2020. Available online: seaborn.pydata.org (accessed on 23 February 2021).
- Harris, C.R.; Millman, K.J.; van der Walt, S.J.; Gommers, R.; Virtanen, P.; Cournapeau, D.; Wieser, E.; Taylor, J.; Berg, S.; Smith, N.J.; et al. Array programming with NumPy. Nature
**2020**, 585, 357–362. [Google Scholar] [CrossRef] [PubMed] - McKinney, W. Data Structures for Statistical Computing in Python. In Proceedings of the 9th Python in Science Conference, Austin, TX, USA, 29 June 2010. [Google Scholar]
- Chen, A.; Beer, J.; Tustison, N.; Cook, P.; Shinohara, R.; Shou, H. Removal of Scanner Effects in Covariance Improves Multivariate Pattern Analysis in Neuroimaging Data. bioRxiv
**2019**, 858415. [Google Scholar] [CrossRef] [Green Version] - Gorgolewski, K.J.; Auer, T.; Calhoun, V.D.; Craddock, R.C.; Das, S.; Duff, E.P.; Flandin, G.; Ghosh, S.S.; Glatard, T.; Halchenko, Y.O.; et al. The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments. Sci. Data
**2016**, 3, 160044. [Google Scholar] [CrossRef] [PubMed] - De Leener, B.; Kadoury, S.; Cohen-Adad, J. Robust, accurate and fast automatic segmentation of the spinal cord. Neuroimage
**2014**, 98, 528–536. [Google Scholar] [CrossRef] [PubMed] - Gros, C.; De Leener, B.; Badji, A.; Maranzano, J.; Eden, D.; Dupont, S.M.; Talbott, J.; Zhuoquiong, R.; Liu, Y.; Granberg, T.; et al. Automatic segmentation of the spinal cord and intramedullary multiple sclerosis lesions with convolutional neural networks. Neuroimage
**2019**, 184, 901–915. [Google Scholar] [CrossRef] [Green Version] - Fehlings, M.G.; Tetreault, L.A.; Riew, K.D.; Middleton, J.W.; Aarabi, B.; Arnold, P.M.; Brodke, D.S.; Burns, A.S.; Carette, S.; Chen, R.; et al. A clinical practice guideline for the management of patients with degenerative cervical myelopathy: Recommendations for patients with mild, moderate, and severe disease and nonmyelopathic patients with evidence of cord compression. Glob. Spine J.
**2017**, 7, 70S–83S. [Google Scholar] [CrossRef] [Green Version] - Tetreault, L.; Kopjar, B.; Nouri, A.; Arnold, P.; Barbagallo, G.; Bartels, R.; Qiang, Z.; Singh, A.; Zileli, M.; Vaccaro, A.; et al. The modified Japanese Orthopaedic Association scale: Establishing criteria for mild, moderate and severe impairment in patients with degenerative cervical myelopathy. Eur. Spine J.
**2017**, 26, 78–84. [Google Scholar] [CrossRef] - Tetreault, L.A.; Côté, P.; Kopjar, B.; Arnold, P.; Fehlings, M.G.; America, A.N.; Network, I.C.T.R. A clinical prediction model to assess surgical outcome in patients with cervical spondylotic myelopathy: Internal and external validations using the prospective multicenter AOSpine North American and international datasets of 743 patients. Spine J.
**2015**, 15, 388–397. [Google Scholar] [CrossRef] - Gibson, J.; Nouri, A.; Krueger, B.; Lakomkin, N.; Nasser, R.; Gimbel, D.; Cheng, J. Focus: Sensory Biology and Pain: Degenerative Cervical Myelopathy: A Clinical Review. Yale J. Biol. Med.
**2018**, 91, 43. [Google Scholar] - Kameyama, T.; Hashizume, Y.; Ando, T.; Takahashi, A. Morphometry of the normal cadaveric cervical spinal cord. Spine
**1994**, 19, 2077–2081. [Google Scholar] [CrossRef] - Nakashima, H.; Yukawa, Y.; Suda, K.; Yamagata, M.; Ueta, T.; Kato, F. Abnormal Findings on Magnetic Resonance Images of the Cervical Spines in 1211 Asymptomatic Subjects. Spine
**2015**, 40, 392–398. [Google Scholar] [CrossRef] [Green Version]

**Figure 1.**The distribution of clinics in Alberta, as well as their relative contribution of the dataset. Larger circles indicate larger contributions (in number of patients), with each circle representing one clinic.

**Figure 2.**An example of the segmentations produced by each of the methodologies tested. The image used was that of a sagittal, T2w image from a patient with severe DCM (as evaluated by mJOA score). The manually segmented example is provided in the bottom center, with all others being produced via automated analyses using SCT [13]. The CNN kernel in particular seems to struggle when faced with spinal cord compressions, with the SVM kernel and propseg method having relatively minor issues in comparison (usually leaking or outright ignoring the compressed areas instead). This pattern appeared to hold true for all segmentations manually reviewed during the process to create Table 1.

**Figure 3.**A violin plot of the distribution metrics extracted from manually segmented spinal cord images for 50 patients via the SCT. Each box represents one of the metrics evaluated by SCT, with the results grouped by mJOA severity classes. When the metric for one mJOA severity class was significantly different from another mJOA severity class (as determined by one-way ANOVA using SciPy’s

`f_oneway`function returning a p-value less than 0.05), a line denoting such is present. A single * with a sparse dotted line denotes $p<0.05$, ** with a tightly dotted line denotes $p<0.01$. Metrics were taken from automated SCT analysis [13] of segmentations from 195 spinal cord MRI images.

**Figure 4.**Visualized distributions of various metrics estimated by various segmentation methods for a subset of 50 patient records. Manual segmentation results are shown as the far-right distribution for each metric. Automated segmentation methods (not “Manual Segmentation”) are denoted with asterisks denoting how significantly different their distribution is from that of the “Manual Segmentation” distribution; ** for $p<0.01$, * for $p<0.05$, as evaluated by one-way ANOVA using SciPy’s

`f_oneway`function (selected for its ease of implementation). Metrics taken from automated SCT analysis [13] of segmentations from 195 spinal cord MRI images.

**Figure 5.**A box plot showing the number of individuals in our study with any given mJOA score. Although not quite ideal, this distribution is relatively balanced across the mid-range of mJOA scores. Note as well that extreme values (mJOA = 18 and mJOA = 8, 9) are rather rare, as would be expected given the acquisition method we used (data taken from those diagnosed with DCM who were undergoing initial assessment).

**Figure 6.**A box plot of the distribution p-values of metric to mJOA score correlations, across all combinations of acquisition contrast, orientation, and segmentation algorithm, as evaluated via SciKit-Learn’s

`‘f_regression’`algorithm (lower is better). Age was included as a control, as it has been previously shown to be correlated with mJOA score [32]. The dotted blue line represents the threshold of significance for this study ($p<0.05$), with whiskers representing the maximum/minimum value of the set, or 1.5 times the inter-quartile range, whichever is shorter.

**Figure 7.**A box plot of the distribution p-values of metric to mJOA score correlations, grouped by acquisition contrast, orientation, and segmentation algorithm, as evaluated via SciKit-Learn’s

`‘f_regression’`algorithm (lower is better). The dotted blue line represents the threshold of significance for this study ($p\le 0.05$), with whiskers representing the maximum/minimum value of the set, or 1.5 times the inter-quartile range, whichever is shorter. Data points outside this range are denoted with green diamonds. Of the methods, it appears that segmentation using deepseg with a svm kernel provided the best results, as did those processed with a T2w contrast along the sagittal plane. However, all but coronal alignment appears capable of statistically significant metric extraction in at least some manner, though the PDw contrast is quite likely a fluke as well (due to its low sample size).

**Table 1.**Total number of segmentations resulting from each algorithm which were found to be “best-of-type” for a given patient. Ties were allowed, enabling one patient image to have up to two “best” segmentations.

Orientation | Contrast | Deepseg (cnn) | Deepseg (3d svm) | Deepseg (svm) | Propseg |
---|---|---|---|---|---|

sagittal | T2w | 2 | 9 | 51 | 7 |

sagittal | T1w | 0 | 0 | 6 | 29 |

sagittal | PDw | 0 | 0 | 0 | 1 |

axial | T2w | 13 | 0 | 63 | 0 |

axial | T1w | 0 | 0 | 1 | 0 |

axial | PDw | 0 | 0 | 0 | 0 |

**Table 2.**Variation of metric measures across mJOA severity classes in the manually segmented subset, summarized. Please note that the ’Mean/STD’ column denotes whether the metric used was the mean of the ’Metric’ column or the ’Standard Deviation’ of said ’Metric’ column. A visualized version of this data, alongside statistical assessments, can be found in Figure 3.

Metric | Mean/STD | Severe | Moderate | Mild |
---|---|---|---|---|

MEAN(area) | mean | 62.223 | 64.066 | 68.393 |

MEAN(area) | std | 10.999 | 14.206 | 12.574 |

STD(area) | mean | 14.710 | 16.319 | 14.879 |

STD(area) | std | 4.599 | 5.341 | 4.587 |

MEAN(angle_AP) | mean | 0.585 | 0.193 | 0.320 |

MEAN(angle_AP) | std | 1.595 | 1.338 | 0.964 |

STD(angle_AP) | mean | 8.331 | 8.274 | 7.018 |

STD(angle_AP) | std | 4.982 | 4.895 | 4.252 |

MEAN(angle_RL) | mean | 8.029 | 6.554 | 5.188 |

MEAN(angle_RL) | std | 8.354 | 9.779 | 8.244 |

STD(angle_RL) | mean | 12.848 | 11.755 | 11.153 |

STD(angle_RL) | std | 5.536 | 5.816 | 4.706 |

MEAN(diameter_AP) | mean | 6.679 | 6.957 | 7.038 |

MEAN(diameter_AP) | std | 0.666 | 0.778 | 0.787 |

STD(diameter_AP) | mean | 1.109 | 1.231 | 1.073 |

STD(diameter_AP) | std | 0.430 | 0.451 | 0.309 |

MEAN(diameter_RL) | mean | 12.332 | 12.113 | 13.030 |

MEAN(diameter_RL) | std | 1.308 | 1.520 | 1.339 |

STD(diameter_RL) | mean | 2.049 | 2.175 | 2.300 |

STD(diameter_RL) | std | 0.644 | 0.609 | 0.818 |

MEAN(eccentricity) | mean | 0.820 | 0.795 | 0.811 |

MEAN(eccentricity) | std | 0.045 | 0.040 | 0.051 |

STD(eccentricity) | mean | 0.085 | 0.108 | 0.099 |

STD(eccentricity) | std | 0.034 | 0.036 | 0.041 |

MEAN(orientation) | mean | 8.222 | 8.692 | 7.331 |

MEAN(orientation) | std | 4.680 | 5.956 | 4.338 |

STD(orientation) | mean | 9.313 | 12.100 | 9.530 |

STD(orientation) | std | 6.215 | 8.879 | 6.452 |

MEAN(solidity) | mean | 0.920 | 0.925 | 0.917 |

MEAN(solidity) | std | 0.031 | 0.028 | 0.034 |

STD(solidity) | mean | 0.046 | 0.043 | 0.049 |

STD(solidity) | std | 0.027 | 0.025 | 0.023 |

SUM(length) | mean | 165.963 | 175.729 | 162.561 |

SUM(length) | std | 59.023 | 63.725 | 46.554 |

**Table 3.**Variation of metric measures across automated segmentation methods. A visualized version of this data, alongside statistical assessments, can be found in Figure 4.

Deepseg (cnn) | Deepseg (svm) | ||||||
---|---|---|---|---|---|---|---|

Metric | Mean/Deviation | 2d | 3d | 2d | 3d | Manual | Propseg |

MEAN(area) | mean | 47.140 | 56.110 | 46.721 | 31.736 | 65.567 | 54.437 |

MEAN(area) | std | 11.938 | 71.798 | 16.471 | 18.332 | 12.525 | 13.785 |

STD(area) | mean | 13.528 | 24.376 | 14.993 | 16.562 | 15.366 | 13.336 |

STD(area) | std | 5.627 | 37.167 | 5.033 | 7.937 | 4.841 | 4.717 |

MEAN(angle_AP) | mean | −0.099 | −0.045 | −0.173 | 0.273 | 0.374 | 0.039 |

MEAN(angle_AP) | std | 4.842 | 8.273 | 3.535 | 3.917 | 1.283 | 1.381 |

STD(angle_AP) | mean | 16.065 | 16.594 | 20.933 | 20.005 | 7.820 | 5.138 |

STD(angle_AP) | std | 12.614 | 12.492 | 15.748 | 10.252 | 4.704 | 2.664 |

MEAN(angle_RL) | mean | 5.600 | 4.448 | 5.036 | 5.475 | 6.639 | 5.166 |

MEAN(angle_RL) | std | 10.255 | 12.174 | 7.908 | 8.534 | 8.556 | 8.035 |

STD(angle_RL) | mean | 15.907 | 13.974 | 18.742 | 18.717 | 12.053 | 12.502 |

STD(angle_RL) | std | 11.184 | 13.479 | 10.722 | 9.312 | 5.349 | 4.553 |

MEAN(diameter_AP) | mean | 5.673 | 5.677 | 5.738 | 4.477 | 6.920 | 7.618 |

MEAN(diameter_AP) | std | 0.835 | 4.638 | 1.102 | 1.752 | 0.736 | 1.498 |

STD(diameter_AP) | mean | 1.107 | 1.863 | 1.362 | 1.690 | 1.127 | 1.617 |

STD(diameter_AP) | std | 0.572 | 2.302 | 0.535 | 0.652 | 0.383 | 0.629 |

MEAN(diameter_RL) | mean | 10.387 | 9.934 | 9.955 | 7.685 | 12.578 | 9.410 |

MEAN(diameter_RL) | std | 2.019 | 6.107 | 2.701 | 2.921 | 1.423 | 1.537 |

STD(diameter_RL) | mean | 2.346 | 2.828 | 2.353 | 2.834 | 2.189 | 1.243 |

STD(diameter_RL) | std | 0.948 | 2.133 | 0.798 | 0.972 | 0.713 | 0.495 |

MEAN(eccentricity) | mean | 0.815 | 0.829 | 0.792 | 0.784 | 0.810 | 0.683 |

MEAN(eccentricity) | std | 0.057 | 0.086 | 0.055 | 0.054 | 0.046 | 0.084 |

STD(eccentricity) | mean | 0.090 | 0.092 | 0.116 | 0.141 | 0.096 | 0.121 |

STD(eccentricity) | std | 0.042 | 0.058 | 0.054 | 0.054 | 0.038 | 0.037 |

MEAN(orientation) | mean | 9.424 | 17.098 | 12.619 | 15.474 | 7.805 | 27.025 |

MEAN(orientation) | std | 8.231 | 16.024 | 9.849 | 9.238 | 4.850 | 18.971 |

STD(orientation) | mean | 12.068 | 15.077 | 15.893 | 20.170 | 10.081 | 20.863 |

STD(orientation) | std | 8.249 | 11.239 | 11.318 | 8.440 | 7.103 | 9.367 |

MEAN(solidity) | mean | 0.938 | 0.883 | 0.934 | 0.908 | 0.920 | 0.933 |

MEAN(solidity) | std | 0.017 | 0.070 | 0.016 | 0.031 | 0.031 | 0.041 |

STD(solidity) | mean | 0.030 | 0.063 | 0.040 | 0.076 | 0.046 | 0.032 |

STD(solidity) | std | 0.012 | 0.032 | 0.020 | 0.027 | 0.024 | 0.021 |

SUM(length) | mean | 126.828 | 63.919 | 188.960 | 167.814 | 167.805 | 171.913 |

SUM(length) | std | 80.717 | 57.264 | 92.204 | 112.005 | 55.064 | 72.697 |

**Table 4.**The attributes of our linear models fit on metric data, which was standardized to a common scale, but did not become harmonized by scanner used via CovBat. Orientation, contrast, and segmentation represent the acquisition methodology associated with the model. Features contains the list of features used to train the model, as selected by SciKit-Learn’s

`SelectFdr`function.

Orientation | Contrast | Segmentation | Samples No. | Features | ${\mathit{r}}^{2}$ |
---|---|---|---|---|---|

acq-axial | T2w | deepseg_cnn_3d | 395 | STD(angle_RL), MEAN(angle_AP) | −30.492 |

acq-sag | T2w | deepseg_svm | 329 | STD(angle_AP) | −29.873 |

acq-sag | T2w | propseg | 308 | MEAN(diameter_AP) | −30.576 |

**Table 5.**The attributes of our linear models fit on metric data which was standardized to a common scale and harmonized by scanner used via CovBat. Orientation, contrast, and segmentation represent the acquisition methodology associated with the model. Features contains the list of features used to train the model, as selected by SciKit-Learn’s

`SelectFdr`function.

Orientation | Contrast | Segmentation | Samples No. | Features | ${\mathit{r}}^{2}$ |
---|---|---|---|---|---|

acq-sag | T2w | deepseg_svm | 329 | STD(angle_AP), MEAN(angle_AP), STD(angle_RL) | −10.329 |

acq-sag | T2w | deepseg_svm_3d | 329 | MEAN(angle_AP), MEAN(diameter_RL) | −15.927 |

acq-sag | T2w | propseg | 308 | MEAN(orientation) | −25.549 |

**Table 6.**The attributes of logistic models fit on metric data, which was standardized to a common scale, but not and harmonized by scanner used via CovBat. Severity indicates the class attempting to be distinguished from all others (binary classification), while orientation, contrast, and segmentation represent the acquisition methodology associated with the model. Features contains the list of features used to train the model, as selected by SciKit-Learn’s

`SelectFdr`function.

Severity | Orientation | Contrast | Segmentation | Sample No. | Features | AUC |
---|---|---|---|---|---|---|

severe | acq-axial | T2w | propseg | 413 | MEAN(eccentricity), STD(area) | 0.713 |

severe | acq-sag | T2w | deepseg_cnn | 269 | STD(area) | 0.519 |

moderate | acq-axial | T2w | deepseg_cnn | 420 | MEAN(area) | 0.568 |

moderate | acq-axial | T2w | deepseg_svm_3d | 420 | STD(solidity) | 0.549 |

mild | acq-sag | PDw | deepseg_svm_3d | 27 | MEAN(angle_RL) | 0.920 |

**Table 7.**The attributes of logistic models fit on metric data which was standardized to a common scale and harmonized by scanner used via CovBat. Severity indicates the class attempting to be distinguished from all others (binary classification), while orientation, contrast, and segmentation represent the acquisition methodology associated with the model. Features contains the list of features used to train the model, as selected by SciKit-Learn’s

`SelectFdr`function.

Severity | Orientation | Contrast | Segmentation | Samples No. | Features | AUC |
---|---|---|---|---|---|---|

severe | acq-sag | T2w | deepseg_svm_3d | 329 | MEAN(diameter_RL) | 0.630 |

moderate | acq-axial | T2w | deepseg_svm_3d | 420 | STD(solidity) | 0.538 |

mild | acq-sag | PDw | deepseg_svm_3d | 27 | STD(diameter_RL) | 0.75 |

mild | acq-sag | T2w | deepseg_svm | 329 | STD(angle_RL) | 0.558 |

mild | acq-sag | T2w | deepseg_svm_3d | 329 | STD(orientation), MEAN(eccentricity) | 0.592 |

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |

© 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Ost, K.; Jacobs, W.B.; Evaniew, N.; Cohen-Adad, J.; Anderson, D.; Cadotte, D.W.
Spinal Cord Morphology in Degenerative Cervical Myelopathy Patients; Assessing Key Morphological Characteristics Using Machine Vision Tools. *J. Clin. Med.* **2021**, *10*, 892.
https://doi.org/10.3390/jcm10040892

**AMA Style**

Ost K, Jacobs WB, Evaniew N, Cohen-Adad J, Anderson D, Cadotte DW.
Spinal Cord Morphology in Degenerative Cervical Myelopathy Patients; Assessing Key Morphological Characteristics Using Machine Vision Tools. *Journal of Clinical Medicine*. 2021; 10(4):892.
https://doi.org/10.3390/jcm10040892

**Chicago/Turabian Style**

Ost, Kalum, W. Bradley Jacobs, Nathan Evaniew, Julien Cohen-Adad, David Anderson, and David W. Cadotte.
2021. "Spinal Cord Morphology in Degenerative Cervical Myelopathy Patients; Assessing Key Morphological Characteristics Using Machine Vision Tools" *Journal of Clinical Medicine* 10, no. 4: 892.
https://doi.org/10.3390/jcm10040892