Skip to main content
NCBI home page
Brain & Spine logoLink to Brain & Spine
. 2022 Nov 14;2:101666. doi: 10.1016/j.bas.2022.101666

Machine learning for image analysis in the cervical spine: Systematic review of the available models and methods

CMW Goedmakers a,b,∗,1, LM Pereboom c,1, JW Schoones d, ML de Leeuw den Bouter e, RF Remis f, M Staring g,h, CLA Vleggeert-Lankamp a,i,j
PMCID: PMC9729832  PMID: 36506292

Highlights

  • •

    Neural network approaches show the most potential for automated image analysis of thecervical spine.

  • •

    Fully automatic convolutional neural network (CNN) models are promising Deep Learning methods for segmentation.

  • •

    In cervical spine analysis, the biomechanical features are most often studied using finiteelement models.

  • •

    The application of artificial neural networks and support vector machine models looks promising for classification purposes.

  • •

    This article provides an overview of the methods for research on computer aided imaging diagnostics of the cervical spine.

Keywords: Cervical spine, Computer aided diagnostics, Radiological imaging, Machine learning, Image analysis

Abbreviations

AAM

Active Appearance Model

ANN

Artificial Neural Network

ASM

Active Shape Model

ASM-M

Active Shape Model Mahalanobis Distance-Based

ASM-RRF

Active Shape Model random regression forest-based

ASM-RCF-AM

Active Shape Model Random Classification Forest-based with ArgMax

Aver

Average

C

Cervical

COG

Center of Gravity

(C)RBM

(Conditional) Restricted Boltzmann Machines

CSF

Cerebrospinal fluid

DNN

Deep Neural Network

DC

Dice Coefficient

DO

Dice Overlap

DR

Detection Rate

FAST

FMRIB Automated Segmentation Tool, Part of FMRIB Software Library (FSL)

GC

Graphical Cut

GHT

Generalized Hough Transform

GLM

Grey-Level Model

GLV

Grey-Level Values

GM

Graphical Model

GT

Ground Truth

HMM

Hidden Markov Model

HD

Hausdorff Distance

HT

Hough Transform

IR

Identification Rate

IQR

Interquartile Rage

J-CNN

Joint learning model Convolutional Neural Network

KDE

Kernel Density Estimation

kNN

k-Nearest Neighbours

LE

Localization Error

MASD

Mean Absolute Surface Distance

MDCP

Mean Distance to the Closest Point

MRF

Markov Random Field

MSE

Root Mean Square Error

NLM

National Library of Medicine

NHANES II

Second National Health and Nutrition Examination Survey

PCA

Principe Component Analysis

QM

Quantitative Morphometry

RANSC

Random Sample Consensus

RCF

Random Classification Forrest

R–CNN

Region Based Convolutional Neural Networks

SC

Spinal Canal

Sens

Sensitivity

SiFC

Sparse intervertebral fence composition

SP

Shape Prior

Spec

Specificity

SRF

Structured Regression Forest

SSAE

Stacked Sparse Autoencoder

SSM

Statistical Shape Model

SVM

Support Vector Machine

TDCN

Transformed Deep Convolutional Neural Network

VolHOG

Histograms of oriented gradients for volumetric data

W(S)

Whole Spine

1. Introduction

Neck pain is the number four cause of physical disability worldwide, and it can be an important symptom in identifying degenerative pathologies of the cervical spine. In most cases, acute neck pain resolves without invasive treatment, but in nearly 50% of patients, the pain returns or develops a chronic nature. With the current ageing population and the relatively high prevalence of neck pain and spine disease, there is increasing demand on radiological image analysis in healthcare (Cohen, 2015). However, the analysis of those visualizations is time-consuming and is subject to significant interobserver variability (Urrutia et al., 2017). Automating parts of the radiological image analysis process can support clinicians in providing a more accurate and consistent image assessment with increased time efficiency.

Over the last decade, the application of artificial intelligence (AI) in medical research has become increasingly popular. Machine Learning (ML) techniques show promise in computer aided diagnostics (CAD), specifically for clinical tasks related to detection and segmentation, as well as classification and prediction (Esteva et al., 2017; Tabibu et al., 2019; Tiulpin et al., 2018; Xu et al., 2019; Zhang et al., 2020). A ML algorithm is able to “learn,” which means in this context that the algorithm can improve performance by previous experience or provided data to give a valid result for never-before-seen data, without being explicitly programmed to do so (Jakubicek et al., 2020).

The majority of the available literature on image analysis concentrates on the thoracic and lumbar spine, while the cervical spine is studied less often. The difference can be partly attributed to the lower incidence of neck pain in the general population, compared to (lower) back pain (Sinnott et al., 2017). Nevertheless, the neck is an essential part of the body with several vital anatomical structures whose functioning can be visualized using radiological imaging. Additionally, considering the relative novelty of the subject matter no systematic reviews have been published, while this could significantly improve the quality of future research on this topic.

Therefore, we aim to create the first overview of the available Machine Learning methods for image analysis of the cervical spine, while weighing and discussing their risks and benefits and providing recommendations for future research in this field. We will divide the systematic review into two sections, one focusing on ML for segmentation and the other on applying ML to automate the study different properties, such as segment mobility and curvature, of the cervical spine on radiological imaging. The overview provided in this systematic review may function as a reference for all authors conducting research on computer aided diagnostics of cervical spine disease.

2. Methods

2.1. Literature search

The initial literature search was performed in PubMed, EMBASE and Web of Science, on December 18th, 2020. Two of the authors (CG, LP) separately evaluated the articles by title, abstract and full text, when necessary, to select the studies that met the predefined selection criteria. As the topic of this review touches both the medical, and the technical research field, both points of view had to be highlighted. Therefore, an additional literature search was performed in the Google Scholar, Scopus, SPIE Digital Library and IEEE Explore databases, to obtain as many articles as possible from both medical and technical journals. The search strategies used in the different databases were based on the search string as shown in Fig. 1.

Fig. 1.

Fig. 1

Open in a new tab

The search strategy used to perform the systematic search in the medical databases.

Studies were included when they reported on a form of automated radiologic image analysis focusing on the human cervical spine or whole spine including the cervical vertebrae.

Studies were excluded if they met any of the following criteria: (1) Publications not written in English; (2) Conference abstracts; (3) Narrative reviews; (4) Cadaver studies without proven clinical application; (5) Phantom studies without proven clinical application; (6) Studies that describe a protocol without any form of analysis; (7) Studies on the thoracic or lumbar spine; (8) Studies on radiation dose, artifact reduction, sequence analysis or robotic surgery; (9) Studies on image processing without segmentation, landmarking or any other measurement on the spine involved.

Any discrepancy in selection between the reviewers was resolved in open discussion (CG, LP), and, if needed, a third reviewer was asked to make a final decision (CVL). Reference screening and citation tracking were performed on the identified articles. This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses: the PRISMA Statement (Moher et al., 2009).

2.2. Quality assessment

The methodological quality of all studies was assessed separately by two reviewers (CG, LP), using a version of the Modified New Castle – Ottawa Quality assessment scale for Cohort Studies (Wells et al., 2020). If there was no consensus about the assessment, a third reviewer (CVL) was consulted. The New Castle – Ottawa scale was manually adjusted to better fit human to model comparison studies with a technical nature.

The items reviewed in the assessment were: 1.1 Representativeness of cohort; 1.2 Model selection, development and implementation; 1.3 Comparison made; 1.4 Ground truth assessment and Data extraction; 2. Applicability and Generalizability (data variability, semi-/fully-automatic, different modalities); 3.1 Outcome Assessment (clear split, ground truth objectified); 3.2 Outcome reporting (different outcome measures, uncertainty metrics reported); 3.3 Sharing (data or code sharing). All items could be awarded a maximum of 1 point, except for ‘Applicability and Generalizability’, for which a maximum of 2 points could be given. Studies could maximally be awarded 9 points. Studies were then divided into low (7–9 points), intermediate (5–6 points) or high (4 or less points) risk of bias.

2.3. Data extraction

Data extraction for all included articles was performed by two reviewers separately (CG, LP) and any controversies were resolved by a third reviewer (CVL). From each article the following information was collected: year of publication, image modality, spine region, model description, degree of automation, number of images included, train to test set distribution and description of how the ground truth was acquired. The determination of the ground truth can be done by either one or more clinical experts and can be provided in different formats; i.e. bounding boxes, vertebra centers, or complete pixel-wise segmentations. Only outcomes that were mentioned in the text or tables of a publication were included into the analysis, as extracting outcomes from graphs was deemed too imprecise and time-intensive.

In order to compare model performance, commonly reported outcome measures were extracted from each publication. Outcomes were divided in either the internal comparison group; when the model's performance was compared to the ground truth, or the external comparison group; when the model's performance was compared to model performance from previous publications.

Outcomes of articles in the segmentation category were reported in five major groups:

Accuracy: Accuracy, Identification Rate (IR), Detection Rate (DR)

Error (mm): Localization error (LE), Mean Distance Closest Point (MDCP),

  • Mean Absolute Surface Distance (MASD), Point-to-Surface error,

  • Hausdorff Distance (HD), Center of Gravity (COG)

Overlap: Dice Overlap (DO), Dice Coefficient (DC), Dice Index (DI)

Time: Runtime, Efficiency.

Other: Precision, Sensitivity, Specificity.

The aims of studies included into the second category, cervical spine analysis, can be divided up into five broad categories: 1. Biomechanical analysis; 2. CVM stage; 3. Clinical prognosis/prediction; 4. Image registration/Planning; 5. Clinical/Radiological Feature Detection. Additional variables collected for the second category articles were aim, included vertebrae and key points.

3. Results

3.1. Article selection

Through searching PubMed, EMBASE and Web of Science, using the predefined search strategy, 956 records could be identified. 654 remained after duplicates were removed. An added search in Google Scholar, Scopus, SPIE Digital Library and IEEE Explore yielded an additional 28 publications. The 682 unique records were screened for title and abstract, after which a total of 506 articles could be excluded. The full-texts of the remaining 176 articles were screened, and 125 did not fit all in- and exclusion criteria and were therefore removed. The remaining 51 articles were included in this systematic review and, based on their primary aim, divided into the two main categories; 1. Segmentation (n ​= ​32) and 2. Cervical Spine Analysis (n ​= ​19). The first category was then divided into two subcategories; 1.1 Conventional Machine Learning Segmentation (n ​= ​20) and 1.2 Deep Learning Segmentation (n ​= ​12) (Fig. 2).

Fig. 2.

Fig. 2

Open in a new tab

Flowchart illustrating the inclusion and exclusion process of articles.

Where articles in the first subcategory focus more on the conventional Machine Learning methods for segmentation, studies in the second category deploy the relatively newer, neural networks. In the second category studies were included that did not necessarily focus on segmentation but in some other way analyzed the cervical spine and its radiologic characteristics.

The increasing popularity of Machine Learning for image analysis of the cervical spine is clearly illustrated when the number of included publications in this study is plotted against the year of publication in total and per subcategory (Fig. 3, Fig. 4). The majority of the included articles (n ​= ​39) is published within the last 5 years.

Fig. 3.

Fig. 3

Open in a new tab

Number of publications plotted per year.

Fig. 4.

Fig. 4

Open in a new tab

Number of publications per subcategory per year.

3.2. Quality assessment

In the Conventional Machine Learning Segmentation group there was one study included with a high risk of bias, eleven studies with an intermediate risk of bias and eight with a low risk of bias (Table 1a). In the Deep Learning Segmentation group three studies showed intermediate risk of bias and nine a low risk of bias, while there were no studies included with a high risk of bias (Table 1b). Lastly, in the Cervical Spine Analysis group there was one study with a high, 14 with an intermediate and 4 with a low risk of bias (Table 1c).

Table 1a.

Complete overview of the Risk of Bias assessment for conventional Machine Learning segmentation articles. Color coded with red (high risk of bias), orange (intermediate risk of bias) and green (low risk of bias) (Al Arif et al., 2016; Banik et al., 2010; Burnett et al., 2004; Chen et al., 2012; Giulietti et al., 2011; Hanaoka et al., 2017; Glocker et al., 2013; Huang et al., 2009; Klinder et al., 2009; Mehmood et al., 2017; Mirzaalian et al., 2013; Schmidt et al., 2007; Zamora et al., 2003).

3.2.

Open in a new tab

Table 1b.

Complete overview of the Risk of Bias assessment for Deep Learning segmentation articles. Color coded with red (high risk of bias), orange (intermediate risk of bias) and green (low risk of bias) (Chen et al., 2015, Chen et al., 2020, Liu et al., 2018, Rak et al., 2019, Suzani et al., 2015, Wang et al., 2019).

3.2.

Open in a new tab

Table 1c.

Complete overview of the Risk of Bias assessment for cervical spine analysis articles. Color coded with red (high risk of bias), orange (intermediate risk of bias) and green (low risk of bias) (Balkovec et al., 2018, Benjelloun and Mahmoudi, 2009, Hopkins et al., 2019, Jin et al., 2019, Lecron et al., 2012, Makaremi et al., 2019, Nikkhoo et al., 2019, Nikkhoo and Mohammad, 2020, Pekar et al., 2007, Rashad et al., 2019, Schmitz et al., 2004, Shin et al., 2020, Srinivasan et al., 2020).

3.2.

Open in a new tab

In general it can be observed that the more recently studies were published, the more likely they were to have a decreased risk of bias. Therefore, the percentage of low risk of bias studies in the Deep Learning Segmentation group is higher than in the Conventional Machine Learning Segmentation group, as the latter includes more recent studies. The same pattern - a decreased risk of bias over time - can be observed in the Cervical Spine Analysis group.

3.3. Qualitative synthesis

3.3.1. Conventional Machine Learning Segmentation techniques

The total number of included studies involving Conventional Machine Learning segmentation techniques is 20, of which 6 studies focused on X-ray images, 6 on MR imaging and 8 studies on CT imaging. The major part, consisting of 14 studies, involved two-dimensional models. The remaining 6 studies used three-dimensional models, of which 4 studies used CT imaging and 2 studies used MRI. The number of images used per study varied widely by image modality. The range of the number of included X-ray images was between 66 (Larhmam et al., 2014) and 10024 ( Xi et al., 2012). The range of included MR images and CTs was diffusely reported, as publications did not only use different numbers of scans but also different numbers of slices, sometimes differentiating per spine region. The number of studies with semi- or fully-automated methods was the same (n ​= ​10) (Table 2a).

Table 2a.

Conventional Machine Learning Segmentation articles overview.

3.3.1.

3.3.1.

3.3.1.

3.3.1.

Open in a new tab

The highest accuracy for MR imaging were reported by Weiss et al. (2006); 96% for the initial model and 100% for the modified model, for the whole spine and the cervical spine, respectively. The study included the entire spine; the vertebrae and intervertebral disks and the ground truth consisted of ‘independent assignments’ of neurologists. In total, 50 ​MR images were included, 27 were used for the initial model and 23 ​MR images were used for the modified model. Image volumes are enhanced with a tophat filter, the program assigns the threshold values and applies a median spatial filter to the search regions. Voxels exceeding threshold values are then subjected to additional constraints and the centroids of these voxel clusters are then connected. 3D linear interpolation and Gaussian filters were applied, the longest disc chains were then analyzed in clusters, which obtained the above mentioned accuracies (De Leener et al., 2015).

The best performing methods are VolHOG for MR images Daenzer et al. (2014), Modified GHT and K-means clustering with the use of X-ray imaging Larhmam et al. (2014) and a statistical and Gaussian shape model in combination with a principal component in combination with CT imaging Clogenson et al. (2015). The research of Daenzer et al. (2014) approached the cervical vertebra detection with a proposed novel machine learning method based on new radiological features, combined with a linear SVM. An accuracy of 98.1% was achieved with the baseline model and improved to 99.1% with the VolHOG. In addition, various levels of artificial noise are used during the performance analysis of the algorithm.

The ground truth in these studies is based on manually determined datapoints by (clinical) experts. All studies reported an internal comparison, comparing the performance of their model to the ground truth, and 9 studies additionally reported some form of external comparison, with earlier publications, published in the years 2003–2009. Of all Conventional Machine Learning segmentation studies, 7 studies reported segmentation results for only the whole spine, while 8 reported results for specifically the cervical spine. In 4 studies, the segmentation results were reported for both the cervical spine specifically and the whole spine. Clogenson et al. (2015) is an exception, just focusing on vertebra C2, which decreases external validity as compared to the other included studies (Table 3a).

Table 3a.

Conventional Machine Learning Segmentation articles extracted outcomes.

3.3.1.

3.3.1.

3.3.1.

3.3.1.

3.3.1.

Open in a new tab

3.3.2. Deep Learning Segmentation techniques

There was a total number of 12 studies included that proposed Deep Learning segmentation techniques, of which two studies focused on MR imaging, 8 studies focused on CT imaging, and just one study used X-ray imaging. The study of Cai et al. (2016) involved both MRI and CT imaging. The majority, consisting of 7 studies, involved three-dimensional models, of which one study Jakubicek et al. (2019) combined 2D and 3D. The remaining 4 studies used two-dimensional models. The number of images used per study varied again widely per image modality, comparable to the Conventional Machine Learning segmentation studies. The range of the number of included CT images was between 41 Bae et al., 2019 and 392 (Jakubicek et al., 2019). The range of used MR images was slightly smaller but comparable; from 60 (Cai et al., 2016) to 245 MRI images (Forsberg et al., 2017). However, the interpretation of this range is difficult as publications, like in the conventional Machine Learning group, did not only use different numbers of scans but also different numbers of slices, sometimes differentiating per spine region.

Almost all studies deployed fully automated methods. Only one study used a semi-automated approach Forsberg et al. (2017), which then also achieves highest detection accuracies (98.8–99.8%). Forsberg et al. (2017) focused on both the cervical and lumbar spine, creating two separated training and configuration pipelines, both having the same CNN setup. The CNN uses fully connected layers, drop-out rate of 0.5, a categorical cross-entropy cost function and Nesterov momentum accelerated Stochastic gradient descent (SGD). The included MR images, together with the annotated spine labels, are focused on either the lumbar or cervical part of the spine. The dataset was originated from an image archive. The missed detections were mainly concerning partly visible vertebrae on the available images. This research showed promising results for labeling and detection by a CNN, focusing on both the cervical and lumbar spine (Glocker et al., 2012) (Table 2b).

Table 2b.

Deep Learning Segmentation articles overview.

3.3.2.

3.3.2.

3.3.2.

Open in a new tab

The highest segmentation accuracy was achieved by the SpineCNN from Jakubicek et al. (2019) (93.3%). Thereby, the best performing methods are CNN based methods for both CT and MR images. The study presents a fully automated approach based on 130 CT scans, which includes two CNNs and a spine tracing algorithm, among which a fine-tuned AlexNet and a VGG-16 ​R–CNN. A population approach was used to increase robustness. The novel combination of the CNN and the tracing, results in almost 90% of correctly identified spinal centerlines within 20 ​s of computing time (Forsberg et al., 2017).

The only study focusing on X-ray imaging Al Arif et al. (2018) used a 6-layered FCN, with an accuracy of center localization of 93.7%.

Similar to the Conventional Machine Learning segmentation studies, the ground truth in the Deep Learning segmentation publications is based on manually determined ground truth by (clinical) experts. The majority of the included studies regarding Deep Learning segmentation methods used both internal and external comparison of their results (n ​= ​9). Results were reported for the whole spine and cervical spine, in 2 and 4 studies, respectively. In 6 studies, half of the total number of studies related to Deep Learning segmentation methods included, the results were reported for both the cervical part and the whole spine (Table 3b).

Table 3b.

Deep Learning Segmentation articles extracted outcomes..

3.3.2.

3.3.2.

3.3.2.

Open in a new tab

3.3.3. Cervical Spine Analysis

There was a total number of 19 studies included involving cervical spine analysis, of which four studies focused on MR imaging, two studies focused on CT imaging, and the majority of the studies used X-ray imaging (n ​= ​11). The aims of the included studies could be further divided into five subgroups; 1. Biomechanical analysis (n ​= ​7), 2. CVM stage (n ​= ​3), 3. Clinical prognosis/prediction (n ​= ​2), 4. Image registration/Planning (n ​= ​4), 5. Clinical/Radiological Feature detection (n ​= ​3).

Two studies included both CT and MR imaging, of which du Bois d'Aische et al. (2007) included PET imaging as well. The use of two-dimensional and three-dimensional visualizations were equal (n ​= ​9), and the remaining study of Kage et al. (2020) (Kage et al., 2020) used a combination of 2D and 3D imaging. Most studies included vertebrae C2–C6, of which several expanded with inclusion of the vertebrae C1, C7 or T1. Other studies used a smaller area of the spine, vertebrae C2–C4, which had the aim to determinate the CVM stage Kök et al. (2019) and Amasya et al. (2020)). The study by Dzyubachyk et al. (2013) was the only one to include the entire spine in the analysis model, with the aim to create an automated reconstruction of the complete spine, based on multistation 7T MR images. The authors applied intensity inhomogeneity correction and used coherent local intensity clustering (CLIC) and fuzzy-c-means-clustering. The performance of the model by Dzyubachyk et al. (2013) was validated based on 18 different datasets, which showed a mean registration error of 0.53 ​mm, which was lower than the MR image pixel size and showed thereby sufficient accuracy.

A wide range of methods was deployed. The best method for radiological feature detection is a CNN model, while the SVM model gave the best result in terms of clinical classification. The ANN approach was reported to work best for CVM stage determination and the FE model, in combination with X-ray imaging, is the most-used method for biomechanical analysis of the spine. In the included spine analysis studies, the amount of fully and semi-automated methods was 7 and 12, respectively (Table 4).

Table 4.

Cervical Spine Analysis articles overview and extracted outcomes.

3.3.3.

3.3.3.

3.3.3.

3.3.3.

3.3.3.

Open in a new tab

3.4. Quantitative synthesis

It was considered to pool accuracy rates in the Conventional Machine Learning and Deep Learning segmentation groups, however it was found that outcomes in the included studies were too heterogeneously reported for doing so. Authors chose to report different outcome metrics and the majority did not report on uncertainty metrics (confidence intervals, standard errors, standard deviations or p-values) with their primary outcome. Pooling the data would therefore require statistical imputation for the majority of the uncertainty metrics. Subsequently, this means that heterogeneity tests, such as the I2, were not performed, as data could not be pooled.

4. Discussion

In this systematic review an overview was provided of the literature on the available Machine Learning techniques for automated image analysis of the cervical spine on radiological imaging. The results of the included studies show a wide variety of possibilities in Machine Learning methods, depending on the aim of the application and the available modalities. In segmentation models, Deep Learning methods show promising results with the application of (fully automatic) CNN models using X-ray, CT or MR imaging. Regarding cervical spine analysis, the biomechanical properties are most often studied using finite element models. The application of artificial neural networks and support vector machine models looks promising for other classification purposes.

Most of the published work on image analysis of the spine focusses on the (thoraco-) lumbar spine. This can be explained by the higher prevalence of lumbar spine pathology, as compared to cervical spine pathology. However, this study, focusing on the cervical spine, is the first of its kind and we therefore believe it can be used as a reference study for all researchers aiming to use radiological image analysis for the cervical spine, as well as other diseases in the neck area.

Unfortunately, results in this systematic review were too heterogeneously reported and therefore pooling the results was not possible. Reporting outcomes clearly and homogenously is an important requirement to compare performance among publications. The authors of this review want to plead for more consistent reporting of outcomes, i.e. the same set of outcome variables for every segmentation, classification or prediction study in order to increase the external validity and reproducibility of these type of studies. Several guidelines that describe the appropriate reporting process for Machine Learning studies have been published (Heil et al., 2021; Luo et al., 2016). However, after reviewing the vast amount of data from the included studies in this systematic review it can be concluded additional guidelines for reporting specifically on image analysis studies using machine learning, are needed. Apart from the recommendation to report a minimal of accuracy (in percentages from 0 to 100%) and error (in mm), reporting uncertainty metrics (confidence intervals, standard errors or standard deviations) with the primary outcome metrics should be required, as it is essential in order to unify the reporting process and aids pooling of results from future studies. Another essential recommendation is for authors to share code. The majority of publications included in this systematic review did not share their code. Creating an academic environment in which code sharing is promoted is essential to keep improving the work in this field.

The concept of ‘Grand Challenges’ presents a promising alternative to current comparative research on the topic of image analysis, by eliminating a range of biases. The aim of these public challenges is to let participants apply their algorithms to the provided Grand Challenge task, using the public test set of images provided by the challenge organizers. In a Grand Challenge organized for analysis of breast histology images, a total of 64 submitted algorithms improved the state-of-the-art in classification of microscopy images to an accuracy of 84% (Aresta et al., 2019).

This systematic review demonstrates a solid body of evidence describing effective segmentation of the cervical spine, with CNN achieving highest accuracy combined with the lowest computing times. Additionally, publications on the different applications for cervical spine analysis show high potential for Machine Learning for several classification and prediction tasks. However, the possibilities for implementation are far-reaching and several newer applications still deserve more attention in future research, including; automated detection, localization and classification of degenerative changes, specifically in the cervical spine. On thoracolumbar CT machine learning was used for automated detection of sclerotic metastases and detection, localization and classification of traumatic vertebral body fractures (Burns et al., 2013, 2016), something that has not been done for the cervical spine yet. On thoracolumbar lateral X-rays the intervertebral disc height measurements were conducted for 1186 participants using machine learning (Allaire et al., 2017), while the study included in this review on the same topic for the cervical spine showed results for only 1 patient (Tan et al., 2012).

The challenges in future research are not just in focusing on the cervical vertebrae or increasing the numbers of images, but also in the integration of different models into one fully automated pathway. Incorporating both radiological and clinical parameters into a fully-automatic model and implementing those into the clinical workflow is the end goal. As was established in this review, the detection and segmentation of the cervical spine have achieved sufficient attention in research, but it is the clinically important classification and prediction tasks, and combining those with detection and segmentation into a fully automatic structure, what future research should focus on.

Funding

No funding was received for this research.

Ethical approval

This article does not contain any studies with human participants or animals performed by any of the authors.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Footnotes

Appendix A

Supplementary data to this article can be found online at https://doi.org/10.1016/j.bas.2022.101666.

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Multimedia component 1
mmc1.doc (67KB, doc)

References

  1. Al Arif S., Knapp K., Slabaugh G. Fully automatic cervical vertebrae segmentation framework for X-ray images. Comput. Methods Progr. Biomed. 2018;157:95–111. doi: 10.1016/j.cmpb.2018.01.006. [DOI] [PubMed] [Google Scholar]
  2. Al Arif S.M.M.R., Gundry M., Knapp K., Slabaugh G. Springer International Publishing; Cham: 2016. Improving an Active Shape Model with Random Classification Forest for Segmentation of Cervical Vertebrae; pp. 3–15. [Google Scholar]
  3. Allaire B.T., DePaolis Kaluza M.C., Bruno A.G., Samelson E.J., Kiel D.P., Anderson D.E., Bouxsein M.L. Evaluation of a new approach to compute intervertebral disc height measurements from lateral radiographic views of the spine. Eur. Spine J. 2017;26(1):167–172. doi: 10.1007/s00586-016-4817-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Amasya H., Cesur E., Yıldırım D., Orhan K. Validation of cervical vertebral maturation stages: artificial intelligence vs human observer visual analysis. Am. J. Orthod. Dentofacial Orthop. 2020;158(6):e173–e179. doi: 10.1016/j.ajodo.2020.08.014. [DOI] [PubMed] [Google Scholar]
  5. Aresta G., Araújo T., Kwok S., Chennamsetty S.S., Safwan M., Alex V., Marami B., Prastawa M., Chan M., Donovan M., Fernandez G., Zeineh J., Kohl M., Walz C., Ludwig F., Braunewell S., Baust M., Vu Q.D., To M.N.N., Kim E., Kwak J.T., Galal S., Sanchez-Freire V., Brancati N., Frucci M., Riccio D., Wang Y., Sun L., Ma K., Fang J., Kone I., Boulmane L., Campilho A., Eloy C., Polónia A., Aguiar P. BACH: Grand challenge on breast cancer histology images. Med. Image Anal. 2019;56:122–139. doi: 10.1016/j.media.2019.05.010. [DOI] [PubMed] [Google Scholar]
  6. Bae H.J., Hyun H., Byeon Y., Shin K., Cho Y., Song Y.J., Yi S., Kuh S.U., Yeom J.S., Kim N. Fully automated 3D segmentation and separation of multiple cervical vertebrae in CT images using a 2D convolutional neural network. Comput. Methods Progr. Biomed. 2019;184 doi: 10.1016/j.cmpb.2019.105119. no pagination. [DOI] [PubMed] [Google Scholar]
  7. Balkovec C., Veldhuis J., Baird J.W., Wayne Brodland G., McGill S.M. Digital tracking algorithm reveals the influence of structural irregularities on joint movements in the human cervical spine. Clin. BioMech. 2018;56:11–17. doi: 10.1016/j.clinbiomech.2018.04.015. [DOI] [PubMed] [Google Scholar]
  8. Banik S., Rangayyan R.M., Boag G.S. Automatic segmentation of the ribs, the vertebral column, and the spinal canal in pediatric computed tomographic images. J. Digit. Imag. 2010;23(3):301–322. doi: 10.1007/s10278-009-9176-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Benjelloun M., Mahmoudi S. Spine localization in X-ray images using interest point detection. J. Digit. Imag. 2009;22(3):309–318. doi: 10.1007/s10278-007-9099-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Burnett S.S., Starkschalla G., Stevens C.W., Liao Z. A deformable-model approach to semi-automatic segmentation of CT images demonstrated by application to the spinal canal. Med. Phys. 2004;31(2):251–263. doi: 10.1118/1.1634483. [DOI] [PubMed] [Google Scholar]
  11. Burns J.E., Yao J., Wiese T.S., Muñoz H.E., Jones E.C., Summers R.M. Automated detection of sclerotic metastases in the thoracolumbar spine at CT. Radiology. 2013;268(1):69–78. doi: 10.1148/radiol.13121351. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Burns J.E., Yao J., Muñoz H., Summers R.M. Automated detection, localization, and classification of traumatic vertebral body fractures in the thoracic and lumbar spine at CT. Radiology. 2016;278(1):64–73. doi: 10.1148/radiol.2015142346. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Cai Y., Landis M., Laidley D.T., Kornecki A., Lum A., Li S. Multi-modal vertebrae recognition using transformed deep convolution network. Comput. Med. Imag. Graph. 2016;51:11–19. doi: 10.1016/j.compmedimag.2016.02.002. [DOI] [PubMed] [Google Scholar]
  14. Chen H., Shen C., Qin J., Ni D., Shi L., Cheng J.C.Y., Heng P.-A. Springer International Publishing; Cham: 2015. Automatic Localization and Identification of Vertebrae in Spine CT via a Joint Learning Model with Deep Neural Networks; pp. 515–522. [Google Scholar]
  15. Chen L., Lan Z., Xu X., Lin J., Hu H. Accuracy and repeatability of computer aided cervical vertebra landmarking in cephalogram. J Huazhong Univ Sci Technolog Med Sci. 2012;32(1):119–123. doi: 10.1007/s11596-012-0021-y. [DOI] [PubMed] [Google Scholar]
  16. Chen Y., Gao Y., Li K., Zhao L., Zhao J. Vertebrae identification and localization utilizing fully convolutional networks and a hidden Markov model. IEEE Trans. Med. Imag. 2020;39(2):387–399. doi: 10.1109/TMI.2019.2927289. [DOI] [PubMed] [Google Scholar]
  17. Clogenson M., Duff J.M., Luethi M., Levivier M., Meuli R., Baur C., Henein S. A statistical shape model of the human second cervical vertebra. Int. J. Comput. Assist. Radiol. Surg. 2015;10(7):1097–1107. doi: 10.1007/s11548-014-1121-x. [DOI] [PubMed] [Google Scholar]
  18. Cohen S.P. Epidemiology, diagnosis, and treatment of neck pain. Mayo Clin. Proc. 2015;90(2):284–299. doi: 10.1016/j.mayocp.2014.09.008. [DOI] [PubMed] [Google Scholar]
  19. Daenzer S., Freitag S., von Sachsen S., Steinke H., Groll M., Meixensberger J., Leimert M. VolHOG: a volumetric object recognition approach based on bivariate histograms of oriented gradients for vertebra detection in cervical spine MRI. Med. Phys. 2014;41(8) doi: 10.1118/1.4890587. [DOI] [PubMed] [Google Scholar]
  20. De Leener B., Cohen-Adad J., Kadoury S. Automatic segmentation of the spinal cord and spinal canal coupled with vertebral labeling. IEEE Trans. Med. Imag. 2015;34(8):1705–1718. doi: 10.1109/TMI.2015.2437192. [DOI] [PubMed] [Google Scholar]
  21. du Bois d'Aische A., De Craene M., Geets X., Gregoire V., Macq B., Warfield S.K. Estimation of the deformations induced by articulated bodies: registration of the spinal column. Biomed. Signal Process Control. 2007;2(1):16–24. [Google Scholar]
  22. Dzyubachyk O., Lelieveldt B.P., Blaas J., Reijnierse M., Webb A., van der Geest R.J. Automated algorithm for reconstruction of the complete spine from multistation 7T MR data. Magn. Reson. Med. 2013;69(6):1777–1786. doi: 10.1002/mrm.24404. [DOI] [PubMed] [Google Scholar]
  23. Esteva A., Kuprel B., Novoa R.A., Ko J., Swetter S.M., Blau H.M., Thrun S. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115–118. doi: 10.1038/nature21056. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Forsberg D., Sjöblom E., Sunshine J.L. Detection and labeling of vertebrae in MR images using deep learning with clinical annotations as training data. J. Digit. Imag. 2017;30(4):406–412. doi: 10.1007/s10278-017-9945-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Giulietti G., Summers P.E., Ferraro D., Porro C.A., Maraviglia B., Giove F. Semiautomated segmentation of the human spine based on echoplanar images. Magn. Reson. Imaging. 2011;29(10):1429–1436. doi: 10.1016/j.mri.2011.08.006. [DOI] [PubMed] [Google Scholar]
  26. Glocker B., Feulner J., Criminisi A., Haynor D.R., Konukoglu E. Automatic localization and identification of vertebrae in arbitrary field-of-view CT scans. Med. Image Comput. Comput. Assist Interv. 2012;15(Pt 3):590–598. doi: 10.1007/978-3-642-33454-2_73. [DOI] [PubMed] [Google Scholar]
  27. Glocker B., Zikic D., Konukoglu E., Haynor D.R., Criminisi A. Vertebrae localization in pathological spine CT via dense classification from sparse annotations. Med. Image Comput. Comput. Assist Interv. 2013;16(Pt 2):262–270. doi: 10.1007/978-3-642-40763-5_33. [DOI] [PubMed] [Google Scholar]
  28. Hanaoka S., Masutani Y., Nemoto M., Nomura Y., Miki S., Yoshikawa T., Hayashi N., Ohtomo K., Shimizu A. Landmark-guided diffeomorphic demons algorithm and its application to automatic segmentation of the whole spine and pelvis in CT images. Int. J. Comput. Assist. Radiol. Surg. 2017;12(3):413–430. doi: 10.1007/s11548-016-1507-z. [DOI] [PubMed] [Google Scholar]
  29. Heil B.J., Hoffman M.M., Markowetz F., Lee S.-I., Greene C.S., Hicks S.C. Reproducibility standards for machine learning in the life sciences. Nat. Methods. 2021;18(10):1132–1135. doi: 10.1038/s41592-021-01256-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Hopkins B.S., Weber K.A., 2nd, Kesavabhotla K., Paliwal M., Cantrell D.R., Smith Z.A. Machine learning for the prediction of cervical spondylotic myelopathy: a post hoc pilot study of 28 participants. World Neurosurg. 2019;127:e436–e442. doi: 10.1016/j.wneu.2019.03.165. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Huang S.H., Chu Y.H., Lai S.H., Novak C.L. Learning-based vertebra detection and iterative normalized-cut segmentation for spinal MRI. IEEE Trans. Med. Imag. 2009;28(10):1595–1605. doi: 10.1109/TMI.2009.2023362. [DOI] [PubMed] [Google Scholar]
  32. Jakubicek R., Chmelik J., Jan J., Ourednicek P., Lambert L., Gavelli G. Learning–based vertebra localization and labeling in 3D CT data of possibly incomplete and pathological spines. Comput. Methods Progr. Biomed. 2020;183 doi: 10.1016/j.cmpb.2019.105081. [DOI] [PubMed] [Google Scholar]
  33. Jakubicek R., Chmelik J., Ourednicek P., Jan J. Annu Int Conf IEEE Eng Med Biol Soc. 2019. Deep-learning-based fully automatic spine centerline detection in CT data; pp. 2407–2410. 2019. [DOI] [PubMed] [Google Scholar]
  34. Jin R., Luk K.D., Cheung J.P.Y., Hu Y. Prognosis of cervical myelopathy based on diffusion tensor imaging with artificial intelligence methods. NMR Biomed. 2019;32(8):e4114. doi: 10.1002/nbm.4114. [DOI] [PubMed] [Google Scholar]
  35. Kage C.C., Akbari-Shandiz M., Foltz M.H., Lawrence R.L., Brandon T.L., Helwig N.E., Ellingson A.M. Validation of an automated shape-matching algorithm for biplane radiographic spine osteokinematics and radiostereometric analysis error quantification. PLoS One. 2020;15(2) doi: 10.1371/journal.pone.0228594. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Klinder T., Ostermann J., Ehm M., Franz A., Kneser R., Lorenz C. Automated model-based vertebra detection, identification, and segmentation in CT images. Med. Image Anal. 2009;13(3):471–482. doi: 10.1016/j.media.2009.02.004. [DOI] [PubMed] [Google Scholar]
  37. Kök H., Acilar A.M., İzgi M.S. Usage and comparison of artificial intelligence algorithms for determination of growth and development by cervical vertebrae stages in orthodontics. Prog. Orthod. 2019;20(1):41. doi: 10.1186/s40510-019-0295-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Larhmam M.A., Benjelloun M., Mahmoudi S. Vertebra identification using template matching modelmp and K-means clustering. Int. J. Comput. Assist. Radiol. Surg. 2014;9(2):177–187. doi: 10.1007/s11548-013-0927-2. [DOI] [PubMed] [Google Scholar]
  39. Lecron F., Benjelloun M., Mahmoudi S. Cervical spine mobility analysis on radiographs: a fully automatic approach. Comput. Med. Imag. Graph. 2012;36(8):634–642. doi: 10.1016/j.compmedimag.2012.08.004. [DOI] [PubMed] [Google Scholar]
  40. Liu X., Yang J., Song S., Cong W., Jiao P., Song H., Ai D., Jiang Y., Wang Y. Sparse intervertebral fence composition for 3D cervical vertebra segmentation. Phys. Med. Biol. 2018;63(11) doi: 10.1088/1361-6560/aac226. [DOI] [PubMed] [Google Scholar]
  41. Luo W., Phung D., Tran T., Gupta S., Rana S., Karmakar C., Shilton A., Yearwood J., Dimitrova N., Ho T.B., Venkatesh S., Berk M. Guidelines for developing and reporting machine learning predictive models in biomedical research: a multidisciplinary view. J. Med. Internet Res. 2016;18(12):e323. doi: 10.2196/jmir.5870. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Makaremi M., Lacaule C., Mohammad-Djafari A. Deep learning and artificial intelligence for the determination of the cervical vertebra maturation degree from lateral radiography. Entropy. 2019;21(12):24. [Google Scholar]
  43. Mehmood A., Akram M.U., Tariq A. International Conference on Communication, Computing and Digital Systems. C-CODE); 2017. Vertebra localization and centroid detection from cervical radiographs; pp. 287–292. 2017. [Google Scholar]
  44. Mirzaalian H., Wels M., Heimann T., Kelm B.M., Suehling M. Annu Int Conf IEEE Eng Med Biol Soc. 2013. Fast and robust 3D vertebra segmentation using statistical shape models; pp. 3379–3382. 2013. [DOI] [PubMed] [Google Scholar]
  45. Moher D., Liberati A., Tetzlaff J., Altman D.G., The P.G. Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. PLoS Med. 2009;6(7) doi: 10.1371/journal.pmed.1000097. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Nikkhoo M., Cheng C.-H., Wang J.-L., Khoz Z., El-Rich M., Hebela N., Khalaf K. Development and validation of a geometrically personalized finite element model of the lower ligamentous cervical spine for clinical applications. Comput. Biol. Med. 2019;109:22–32. doi: 10.1016/j.compbiomed.2019.04.010. [DOI] [PubMed] [Google Scholar]
  47. Nikkhoo, Mohammad, et al. The biomechanical response of the lower cervical spine post laminectomy: geometrically-parametric patient-specific finite element analyses. J. Med. Biol. Eng. 2020;41:59–70. doi: 10.1007/s40846-020-00579-8. [DOI] [Google Scholar]
  48. Pekar V., Bystrov D., Heese H.S., Dries S.P., Schmidt S., Grewer R., den Harder C.J., Bergmans R.C., Simonetti A.W., van Muiswinkel A.M. Automated planning of scan geometries in spine MRI scans. Med. Image Comput. Comput. Assist Interv. 2007;10(Pt 1):601–608. doi: 10.1007/978-3-540-75757-3_73. [DOI] [PubMed] [Google Scholar]
  49. Rak M., Steffen J., Meyer A., Hansen C., Tönnies K.D. Combining convolutional neural networks and star convex cuts for fast whole spine vertebra segmentation in MRI. Comput. Methods Progr. Biomed. 2019;177:47–56. doi: 10.1016/j.cmpb.2019.05.003. [DOI] [PubMed] [Google Scholar]
  50. Rashad A., Heiland M., Hiepe P., Nasirpour A., Rendenbach C., Keuchel J., Regier M., Al-Dam A. Evaluation of a novel elastic registration algorithm for spinal imaging data: a pilot clinical study. Int. J. Med. Robot. 2019;15(3) doi: 10.1002/rcs.1991. [DOI] [PubMed] [Google Scholar]
  51. Schmidt S., Kappes J., Bergtholdt M., Pekar V., Dries S., Bystrov D., Schnörr C. Spine detection and labeling using a parts-based graphical model. Inf. Process Med. Imaging. 2007;20:122–133. doi: 10.1007/978-3-540-73273-0_11. [DOI] [PubMed] [Google Scholar]
  52. Schmitz B., Pitzen T., Beuter T., Steudel W.I., Reith W. Regional variations in the thickness of cervical spine endplates as measured by computed tomography. Acta Radiol. 2004;45(1):53–58. doi: 10.1080/02841850410000755. [DOI] [PubMed] [Google Scholar]
  53. Shin Y., Han K., Lee Y.H. Temporal trends in cervical spine curvature of south Korean adults assessed by deep learning system segmentation, 2006-2018. JAMA Netw. Open. 2020;3(10) doi: 10.1001/jamanetworkopen.2020.20961. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Sinnott P.L., Dally S.K., Trafton J., Goulet J.L., Wagner T.H. Trends in diagnosis of painful neck and back conditions. 2002 to 2011, Medicine. 2017;96(20) doi: 10.1097/MD.0000000000006691. [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Srinivasan S, Kumar S D, R S, Jebaseelan D D, Yoganandan N, RajasekaranS. Effect of heterotopic ossification after bryan-cervical disc arthroplasty on adjacent level range of motion: A finite element study. J. Clin. Orthop. Trauma. 15, 99–103. 10.1016/j.jcot.2020.10.027.33717922. [DOI] [PMC free article] [PubMed]
  56. Suzani A., Seitel A., Liu Y., Fels S., Rohling R.N., Abolmaesumi P. In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Navab N., Hornegger J., Wells W.M., Frangi A.F., editors. Springer International Publishing; Cham: 2015. Fast automatic vertebrae detection and localization in pathological CT scans - a deep learning approach; pp. 678–686. [Google Scholar]
  57. Tabibu S., Vinod P.K., Jawahar C.V. Pan-Renal Cell Carcinoma classification and survival prediction from histopathology images using deep learning. Sci. Rep. 2019;9(1) doi: 10.1038/s41598-019-46718-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  58. Tan S., Yao J., Yao L., Ward M.M. Annu Int Conf IEEE Eng Med Biol Soc. 2012. High precision semi-automated vertebral height measurement using computed tomography: a phantom study; pp. 1554–1557. 2012. [DOI] [PubMed] [Google Scholar]
  59. Tiulpin A., Thevenot J., Rahtu E., Lehenkari P., Saarakkala S. Automatic knee osteoarthritis diagnosis from plain radiographs: a deep learning-based approach. Sci. Rep. 2018;8(1):1727. doi: 10.1038/s41598-018-20132-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  60. Urrutia J., Zamora T., Yurac R., Campos M., Palma J., Mobarec S., Prada C. An independent inter- and intraobserver agreement evaluation of the AOSpine subaxial cervical spine injury classification system. Spine. 2017;42(5) doi: 10.1097/BRS.0000000000001302. [DOI] [PubMed] [Google Scholar]
  61. Wang X., Zhai S., Niu Y. Automatic vertebrae localization and identification by combining deep SSAE contextual features and structured regression forest. J. Digit. Imag. 2019;32(2):336–348. doi: 10.1007/s10278-018-0140-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Weiss K.L., Storrs J.M., Banto R.B. Automated spine survey iterative scan technique. Radiology. 2006;239(1):255–262. doi: 10.1148/radiol.2383050456. [DOI] [PubMed] [Google Scholar]
  63. Wells G., Shea B., O'Connel D., Welch P.J.V., Losos M., Tugwell P. The NewcastleOttawa Scale (NOS) for assessing the quality of nonradomised studies in metaanalysis. http://www.ohri.ca/programs/clinical_epidemiology/oxford.asp
  64. Xi X., Hong-Wei H., Xu-Cheng Y., Ning L., Shafin S.H. IJCNN); 2012. Automatic Segmentation of Cervical Vertebrae in X-Ray Images, the 2012 International Joint Conference on Neural Networks; pp. 1–8. [Google Scholar]
  65. Xu Y., Hosny A., Zeleznik R., Parmar C., Coroller T., Franco I., Mak R.H., Aerts H. Deep learning predicts lung cancer treatment response from serial medical imaging. Clin. Cancer Res. 2019;25(11):3266–3275. doi: 10.1158/1078-0432.CCR-18-2495. [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Zamora G., Sari-Sarraf H., Long R.L. Proc.SPIE; 2003. Hierarchical Segmentation of Vertebrae from X-Ray Images. [Google Scholar]
  67. Zhang Y., Lobo-Mueller E.M., Karanicolas P., Gallinger S., Haider M.A., Khalvati F. CNN-based survival model for pancreatic ductal adenocarcinoma in medical imaging. BMC Med. Imag. 2020;20(1):11. doi: 10.1186/s12880-020-0418-1. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Multimedia component 1
mmc1.doc (67KB, doc)

Articles from Brain & Spine are provided here courtesy of Elsevier

RESOURCES