Elsevier

NeuroImage

Volume 98, September 2014, Pages 528-536
NeuroImage

Technical Note
Robust, accurate and fast automatic segmentation of the spinal cord

https://doi.org/10.1016/j.neuroimage.2014.04.051Get rights and content

Highlights

  • Automatic spinal cord segmentation method using propagated deformable models

  • Works on T1- and T2-weighted MR images and any field of view

  • High accuracy, similar to the inter-rater variability

  • Robust towards local lack of contrast between spinal cord and CSF

Abstract

Spinal cord segmentation provides measures of atrophy and facilitates group analysis via inter-subject correspondence. Automatizing this procedure enables studies with large throughput and minimizes user bias. Although several automatic segmentation methods exist, they are often restricted in terms of image contrast and field-of-view. This paper presents a new automatic segmentation method (PropSeg) optimized for robustness, accuracy and speed. The algorithm is based on the propagation of a deformable model and is divided into three parts: firstly, an initialization step detects the spinal cord position and orientation using a circular Hough transform on multiple axial slices rostral and caudal to the starting plane and builds an initial elliptical tubular mesh. Secondly, a low-resolution deformable model is propagated along the spinal cord. To deal with highly variable contrast levels between the spinal cord and the cerebrospinal fluid, the deformation is coupled with a local contrast-to-noise adaptation at each iteration. Thirdly, a refinement process and a global deformation are applied on the propagated mesh to provide an accurate segmentation of the spinal cord. Validation was performed in 15 healthy subjects and two patients with spinal cord injury, using T1- and T2-weighted images of the entire spinal cord and on multiecho T2*-weighted images. Our method was compared against manual segmentation and against an active surface method. Results show high precision for all the MR sequences. Dice coefficients were 0.9 for the T1- and T2-weighted cohorts and 0.86 for the T2*-weighted images. The proposed method runs in less than 1 min on a normal computer and can be used to quantify morphological features such as cross-sectional area along the whole spinal cord.

Introduction

Spinal cord segmentation is commonly done to quantify spinal cord atrophy (Coulon et al., 2002, Lin et al., 2004) or to extract MRI metrics within segments of the spinal cord (Cohen-Adad et al., 2011). Segmentation can also be useful for surgical planning (Mukherjee et al., 2010). A review of possible applications of spinal cord segmentation can be found in Wheeler-Kingshott et al. (2013). However, the spinal cord is a challenging structure to segment. The longitudinal and thin morphology limits the application of basic registration methods (Chen et al., 2013). Moreover, the large subject-dependent curvature makes classical deformable model-based methods (Terzopoulos and Fleischer, 1988) challenging, because these methods assume small structural deformations and a good initialization (McInerney and Terzopoulos, 1996).

Several attempts were made to automatically or semi-automatically segment the spinal cord or the spinal canal in computed tomography imaging (CT). Archip et al. (2002) proposed a method for the extraction of the spinal cord and nearby structures using a knowledge-based segmentation. Specifically, the spinal cord detection was made using the information of the body contour, the lamina to compute seed points of a region growing segmentation of the spinal canal. Nyúl et al. (2005) developed a technique based on active contour on 2D slices requiring one seed point. Rangayyan et al. (2006) proposed an automatic method for the segmentation of the spinal canal using mathematical morphology and region growing. The automatic detection of the spinal canal was provided using the circular Hough transform (Ballard, 1981). Although these methods are suitable for CT images, they do not apply on MRI data because of the large differences between CT and MRI contrast. Challenges of MRI-based spinal cord segmentation include the intensity bias field due to distance from the coil and inhomogeneous transmit field (Lin et al., 2004). Furthermore, missing CSF around the spinal cord as well as the common presence of intervertebral disk calcification can introduce errors in segmentations.

Several MRI-based spinal cord segmentation methods have been proposed. Van Uitert et al. (2005) developed a semi-automatic level-set segmentation method for measuring the spinal cord area in sagittal T2-weighted MR images. However, the method is quite sensitive to contrast-to-noise ratio and requires a morphological closing to remove possible holes in the segmentation. Coulon et al. (2002) used a parameterized active surface to segment the spinal cord in cervical T1-weighted images and Horsfield et al. (2010) proposed a method based on an active surface model with intrinsic smoothness constraints. Both techniques require providing the spinal cord centerline to initialize the surface model of the cord. The user has to provide points distributed with regular spacing along the spinal cord; these points are used to approximate a cubic spline which is computed along the spinal cord centerline. McIntosh et al. (2011) demonstrated the usability of deformable models for the spinal cord segmentation. Their Spinal Crawlers method, based on an artificial life framework with high-level control mechanisms, propagates a mesh along the spinal cord using a minimal path guidance approach and only requires three seed points. More recently, Kawahara et al. (2013) developed a semi-automated spinal cord segmentation method based on a globally minimal path optimization in high dimensions using probabilistic principal component analysis to represent shape variability. Although the previously stated methods are efficient in segmenting the spinal cord, they still require manual intervention.

Automatic segmentation would improve the workflow of radiological exams and enable multicenter studies with large throughput and no user-bias. Mukherjee et al. (2010) proposed a method to automatically segment spinal cords of cats using symmetric boundary tracing and adaptive active contour. Their method performs the segmentation on 2D axial slices and reconstructs a 3D segmentation by minimizing an energy equation integrating the symmetry of the detected contours and the smoothness between slices. Koh et al. (2010) developed a 2D active contour based method to extract the spinal cord on sagittal T2-weighted MR images using gradient vector flow field. Later, Koh et al. (2011) proposed an automatic 2D spinal canal segmentation method based on active contours and an attention model. Based on the Gabor filter, this attention model extracts contextual information in an inspired way of the primary visual cortex.

To summarize, most existing fully automated methods are adequate with a particular field of view (FOV) or image contrast, but fail when imaging protocols slightly differ or when in presence of challenging artifacts (e.g. lack of CSF around the spinal cord or calcification). To our knowledge, the automatic framework presented by Chen et al. (2013) is the only method adaptable for different MR sequences and various FOV. The method uses a combination of atlas-based deformable registration and topology preserving intensity classification. One limitation however is the need to create a protocol-specific atlas, which depends on the acquisition sequence and the FOV. Moreover, the technique assumes that CSF surrounds the spinal cord, which is not always the case.

In this paper, we present a fully automatic segmentation method for the spinal cord, which supports T1-, T2- and T2*-weighted types of contrast and which works for cervical, thoracic and lumbar levels. The automation is provided by an independent spinal cord detection module, using the circular/elliptical shape of the spinal cord and the symmetry of the body to perform an elliptical Hough transform. The segmentation is based on an iterative propagation of a deformable model with adaptive contrast mechanisms. This approach aims to be flexible towards highly curved structures by computing the spinal cord orientation in an iterative fashion. Our method performs well for an unsupervised segmentation and is robust towards lack of CSF around the spinal cord.

Section snippets

Materials and methods

The main inspiration of our method is the deformable model approach (Kaus et al., 2003) and consists in the iterated propagation of a tubular deformable model. Adaptive contrast properties are included within the deformable model framework so to appropriately deal with potential lack of signal-to-noise ratio (McInerney and Terzopoulos, 1996). The initialization of the propagation generates a small elliptical tubular mesh near the center of the spinal cord and is properly oriented in the

Spinal cord detection

The sensitivity (proportion of successful detection) of the spinal cord detection method, without the validation procedure, was 95.28% for T1-weighted images and 88.18% for T2-weighted images (N  450). The detection was performed on slices spanning the entire spinal cord, each of them being separated by 2 cm. Failures usually appeared where the contrast between the CSF and the spinal cord was low. Moreover, failures appeared when other contrasted elliptical structures were present in the image,

Discussion

This study presented a spinal cord segmentation method based on MR images. To our knowledge, this is the first fully-automated method for segmenting the whole spinal cord, without particular constraint on the size or location of the field of view. The method proved to be robust towards three different contrasts: T1-, T2- and T2*-weighting and worked equally well on isotropic and anisotropic voxels. The software notably outputs a binary mask of the segmentation as well as values of

Conclusion

This paper presented a method for automatic segmentation of the spinal cord on MR volumes. The proposed method yielded successful results on T1-, T2- and T2*-weighted contrasts as well as on different image resolutions and fields of view, including the cervical, thoracic and lumbar spinal cord. The initialization of the algorithm is provided by an independent module that automatically detects the spinal cord center on axial slices. The module itself can also be adapted for automatizing other

Acknowledgments

The authors would like to thank Alexandru Foias for the manual segmentations, as well as the volunteers. This work was funded by the SensoriMotor Rehabilitation Research Team (SMRRT) of the Canadian Institute of Health Research [229269], the National MS Society [FG1892A1/1], the Fonds de Recherche du Québec — Santé (FRQS) [27130, 28826], the Quebec BioImaging Network (QBIN) [5886] and the Natural Sciences and Engineering Research Council of Canada (NSERC) [435897-2013].

References (29)

  • N. Aspert et al.

    Mesh: measuring errors between surfaces using the Hausdorff distance

  • O. Coulon et al.

    Quantification of spinal cord atrophy from magnetic resonance images via a B-spline active surface model

    Magn. Reson. Med.

    (2002)
  • L.R. Dice

    Measures of the amount of ecologic association between species

    Ecology

    (1945)
  • P.E. Gill et al.

    Practical Optimization

    (1981)
  • Cited by (135)

    • U-NetCTS: U-Net deep neural network for fully automatic segmentation of 3D CT DICOM volume

      2022, Smart Health
      Citation Excerpt :

      Nevertheless, automatic segmentation can play a key role in enhancing the procedures and the workflow of the medical process. Furthermore, it can speed up the clustering process and region detection with a high level of accuracy and without the need for user intervention (De Leener, Kadoury, & Cohen-Adad, 2014). Artificial neural network (ANN) and deep learning (DL) methods have been widely used in image segmentation, classification, and other data analytics tasks.

    View all citing articles on Scopus
    View full text