Elsevier

Magnetic Resonance Imaging

Volume 31, Issue 9, November 2013, Pages 1507-1514
Magnetic Resonance Imaging

Original contribution
Reduced field-of-view DTI segmentation of cervical spine tissue

https://doi.org/10.1016/j.mri.2013.07.003Get rights and content

Abstract

The number of diffusion tensor imaging (DTI) studies regarding the human spine has considerably increased and it is challenging because of the spine’s small size and artifacts associated with the most commonly used clinical imaging method. A novel segmentation method based on the reduced field-of-view (rFOV) DTI dataset is presented in cervical spinal canal cerebrospinal fluid, spinal cord grey matter and white matter classification in both healthy volunteers and patients with neuromyelitis optica (NMO) and multiple sclerosis (MS). Due to each channel based on high resolution rFOV DTI images providing complementary information on spinal tissue segmentation, we want to choose a different contribution map from multiple channel images. Via principal component analysis (PCA) and a hybrid diffusion filter with a continuous switch applied on fourteen channel features, eigen maps can be obtained and used for tissue segmentation based on the Bayesian discrimination method. Relative to segmentation by a pair of expert readers, all of the automated segmentation results in the experiment fall in the good segmentation area and performed well, giving an average segmentation accuracy of about 0.852 for cervical spinal cord grey matter in terms of volume overlap. Furthermore, this has important applications in defining more accurate human spinal cord tissue maps when fusing structural data with diffusion data. rFOV DTI and the proposed automatic segmentation outperform traditional manual segmentation methods in classifying MR cervical spinal images and might be potentially helpful for detecting cervical spine diseases in NMO and MS.

Introduction

Diffusion tensor imaging (DTI) is an advanced medical technology in human anatomy research. It can provide rich information in a noninvasive way. The cervical spinal canal and cervical spinal cord white matter/grey matter are three basic tissues in the human spine. Spinal segmentation has important applications in clinics studying the structure of the spine [1], [2], [3], [4], such as the central nervous system disorder known as neuromyelitis optica (NMO) and manifestations of multiple sclerosis (MS). However, traditional DTI is technically limited to be utilized in the cervical spinal cord due to the spinal fluid flow effect, low resolution and signal-to-noise ratio(SNR) caused by the 1–1.5 cm diameter of the spinal cord, and the severe distortion caused by the complex tissue–bone tissue interface [5]. Currently, these technical limitations have been rendered by the application of line-scan imaging and triggering to diminish the flow effect [6], [7]. A novel scheme for reduced field-of-view (rFOV) excitation proposed by General Electric (GE) Healthcare allows time-efficient interleaved multi-slice data acquisition which can be used to acquire high-resolution diffusion weighted images and substantially reduce susceptibility artifacts in the cervical spinal cord [5], [8], [9]. The rFOV DTI is feasible in a clinical procedure involving the cervical spinal cord. At a high field 3 T strength, the rFOV technique can greatly and simultaneously reduce the susceptibility effects and motion artifacts in diffusion-weighted imaging, and has higher SNR and resolution relative to 1.5 T MRI [5], [10]. Furthermore, the goal is to obtain a partition of the image that is composed of homogeneous regions, with each partition representing a tissue. It is an important preprocessing step in cervical spinal cord disease research and clinical applications because these contrasts define the boundaries of the cervical spinal cord in addition to manually defining the spinal cord grey matter. Different methods were employed for segmentation of the human brain and imaging data in both conventional magnetic resonance imaging (MRI) and DTI space during the past years, especially from classical, statistical, fuzzy, neural network (NN) and multi-channel fusion techniques [11], [12]. Most prior segmentation approaches include threshold, edge-based and pixel-based techniques which extract tissue that is defined by intensity transitions or by gradients with intensity transitions. Unfortunately, intensity inhomogeneities imply intensity variations in the same class of tissues that are not caused by random noise in MRI, making it difficult to obtain accurate segmentation results in clinical research. The intensities in different substructures, even in the same tissue class of the spinal cord, are also more or less different for the inherent regional differences in imaging-related properties across substructures, e.g. the composition, density, and magnetic properties of different tissues indifferent positions, especially in diseases such as NMO and MS. Due to the mentioned above and the MRI intrinsic adverse impact, it makes intensity distribution within a particular tissue class more flat, resulting in overlapping intensity components among different tissues that are neighbors in the intensity histogram, especially in spinal cord grey matter or cerebellar tissue segmentation. This challenges the traditional single channel image model with a Gaussian mixture and impacts the precision and reliability of automatic intensity based segmentation approaches.

Currently, clinical research in the DTI field is undergoing rapid expansion to depict human brain and spinal microstructural features that weave these sites together into a system with spatially interacting neural elements, and it is increasingly applied to related disease studies. In particular, diffusion tensor maps are typically computed by fitting the signal intensities from diffusion weighted images as a function of their corresponding b-matrices according to the multivariate least squares regression modal proposed by Basser et al. [13]. Many usage approaches have been proposed to compensate the intensity inhomogeneities. Among them, a multi-channel fusion technique was employed to define accurate tissue maps when dealing with fused structural and diffusion information from SPGR and DTI data. Usually, the measurements of grey matter diffusivity based on the anatomical information in the SPGR image may fail to reveal the real diffusion for the problem of heterogeneous voxels [14], [15]. Compared with traditional segmentation methods, the multi-channel fusion technique performs estimation and segmentation in the fused way from multiple DTI data channels, such as the apparent diffusion coefficients (ADC), fractional anisotropy (FA) values, axonal damage related axial diffusivity (AD) values and radial diffusivity (RD) values associated with demyelination, so that intermediate information gained from current segmentation can be used to improve the estimation results, which can in turn lead to more accurate segmentation [16], [17]. During cervical spine segmentation, the white matter boundaries of the DTI image are crossing the cerebrospinal fluid (CSF) in the cervical spinal canal of the ADC image and the grey voxels in the DTI image correspond to both white and grey matter voxels in the ADC image. Such a problem can occur for a variety of reasons, including geometric distortion in DTI and partial volume effect. Fortunately, the Volume Ratio (VR) values in the CSF are more than twice those of spinal grey matter and white matter, leading to visible differences between CSF and non-CSF tissues; FA, VR and Linearity Anisotropy (CL) images are used to separate white matter, since highly directional white matter structures have much larger fractional anisotropy values; many other channels from rFOV DTI measurements also have the ability to separate spinal tissues, such as compositional Kullback–Leibler anisotropy (KLA), AD and spherical anisotropy (CS). However, the MRI signal not only is influenced by thermal noise, but also is sensitive to small spatially and temporally varying artifacts in rFOV DTI datasets. All these factors might result in multiple NEX low SNR images and lead to substantial influence on accuracy of the artificial tensor estimation in every channel of DTI data. If all of the features from multiple channels of DTI data were selected as the characteristic for each voxel, the information on spinal tissues in every channel can be compensated for other advantages in distinguishing specific spinal tissues from other channels leading to the consensus from all of the channels. Subsequently, we need to construct an effective classifier for spinal tissue segmentation, which presents a challenge to current researchers in the field of cervical spinal cord diseases.

In recent years, researchers proposed two categories of approaches to obtain this goal. The first category is supervised classification, such as the support vector machine (SVM), k-NN and Bayesian discriminant classifier [18], [19]. The other category is unsupervised classification, such as the self-organization feature map (SOFM) and fuzzy c-means [20], [21]. While both of these methods achieved satisfactory results, supervised classification performed better than unsupervised classification in terms of a successful classification rate. In this work, a novel pattern classification approach–Bayesian discriminant based classifier which utilized the distribution characteristics of the manually-defined samples in each spinal cord tissue. The Bayesian discriminant in the subspace spanned by the eigenvectors, which are associated with the smaller eigenvalues in each class, was adopted as the classification criterion for the human spine grey and white matter tissues. The method depends on knowledgeable information and the discrimination power of the features. Classification accuracy depends mainly on the quality of features, which should be robust with maximum discrimination power and must encompass most of the information available in the rFOV DTI data. As mentioned above, many other channels with a useful meaning in physics were adopted to be used as the texture features. The principal component analysis was applied to fourteen channel features to reduce the dimension of the generated feature vectors, and the bigger eigenvalues corresponding to eigenvectors could be obtained and used to extract the optimal feature vectors for spinal cord tissue segmentation. The effects of cervical spinal cord segmentation results were quantitatively evaluated and applications of the proposed classifier on spinal tissue segmentation demonstrated that they were promising in clinical applications such as for: NMO and MS, respectively.

Section snippets

Overview

The framework of human spinal tissue segmentation based on DTI data consists of five steps, as summarized in Fig. 1. The rFOV DTI pulse sequence used in this study is a ‘work-in-progress (WIP)’ software of GE Healthcare, and we gained the data in Step 1. Step 2 is the pre-processing used to perform eddy current correction by using the FSL FDT tools (http://www.fmrib.ox.ac.uk/fsl/), tensor calculation and channel image generation from the WIP in FuncTool software of GE Healthcare. Fourteen

Results

The algorithm was implemented in a C environment on a Linux system on an AW workstation from GE Healthcare. It has been tested on human cervical spinal rFOV DTI images from 10 subjects including 8 healthy subjects, and 2 more NMO and MS patients. In this study, the standard cervical spinal canal and spinal cord grey matter and white matter were manually segmented from the DTI measures, which were a combination of multi-channel images. The 2D merge image was taken as a rough guide and reference

Discussion and conclusion

The DTI technique is an important method towards unraveling the microstructural features of the human central nervous system, including the brain and cervical spine, but problems are common in traditional DTI acquisitions for poor tissue contrast and low image resolution, especially in human spinal cord studies, which can result in erroneous diffusion tensor values. Efforts have been undertaken to address various aspects of this application, but this field is still poorly explored by targeting

Acknowledgments

The authors would like to thank the anonymous reviewers for their significant and constructive comments and suggestions which greatly improved the paper. We are also gratefully thankful for the support and assistance from Drs. Ajit Shankaranarayanan and Suchandrima Banerjee of the ASL Team of GE Healthcare.

References (42)

  • C.G. Strickland et al.

    Development of subject-specific geometric spine model through use of automated active contour segmentation and kinematic constraint-limited registration

    J Digit Imaging

    (2011)
  • A. Dalca et al.

    Segmentation of nerve bundles and ganglia in spine MRI using particle filters

    Med Image Comput Comput Assist Interv Int Conf Med Image Comput Comput Assist Interv

    (2011)
  • W. Qian et al.

    Quantitative assessment of the cervical spinal cord damage in neuromyelitis optica using diffusion tensor imaging at 3 tesla

    J Magn Reson Imaging

    (2011)
  • S.E. Maier et al.

    Diffusion tensor imaging of the spinal cord

    Ann N Y Acad Sci

    (2005)
  • P. Summers et al.

    A preliminary study of the effects of trigger timing on diffusion tensor imaging of the human spinal cord

    AJNR Am J Neuroradiol

    (2006)
  • E.K. Jeong et al.

    High-resolution DTI with 2D interleaved multislice reduced FOV single-shot diffusion-weighted EPI (2D ss-rFOV-DWEPI)

    Magn Reson Med

    (2005)
  • B.J. Wilm et al.

    Diffusion-weighted imaging of the entire spinal cord

    NMR Biomed

    (2009)
  • J. Luts et al.

    Nosologic imaging of the brain: segmentation and classification using MRI and MRSI

    NMR Biomed

    (2009)
  • D.P. Mukherjee et al.

    Automatic segmentation of spinal cord MRI using symmetric boundary tracing

    IEEE Trans Inf Technol Biomed

    (2010)
  • M. Cercignani et al.

    Magnetisation transfer ratio and mean diffusivity of normal appearing white and grey matter from patients with multiple sclerosis

    J Neurol Neurosurg Psychiatry

    (2001)
  • V.F. Newcombe et al.

    Concordant biology underlies discordant imaging findings: diffusivity behaves differently in grey and white matter post acute neurotrauma

    Acta Neurochir Suppl

    (2008)
  • Cited by (23)

    • Automatic segmentation of the spinal cord and intramedullary multiple sclerosis lesions with convolutional neural networks

      2019, NeuroImage
      Citation Excerpt :

      Hence, there is a need for robust and automatic segmentation tools for the spinal cord and the intramedullary MS lesions. Various automatic spinal cord segmentation methods have been proposed in the past few years, including active contours and surface-based approaches (De Leener et al., 2015; Koh et al., 2010), and atlas-based methods (Carbonell-Caballero et al., 2006; Chen et al., 2013; Pezold et al., 2015; Tang et al., 2013). While these methods have shown good performance (De Leener et al., 2016), they often require a specific region of interest and/or are limited to a specific contrast and resolution.

    • Identification of ghost artifact using texture analysis in pediatric spinal cord diffusion tensor images

      2018, Magnetic Resonance Imaging
      Citation Excerpt :

      A number of methods to detect ghosting artifacts and address its deleterious influence on tensor estimation have been proposed [8,10,24,33]. The methods, however, share two major limitations: (1) they are computationally intensive, usually in direct proportion to their accuracy, sensitivity and robustness; (2) they distinguish detailed textural features of the images only indirectly, from measures other than those contained in the raw DWI data, potentially missing information that is vitally important for artifact detection [9,10,32,33]. The proposed system addresses both of these limitations by using artificial intelligence in determining the texture features capable of differentiating TC from GC while not sacrificing accuracy at the expense of computation time and also with this technique there is no need to reference or calibration scans to align echoes.

    • Fully-integrated framework for the segmentation and registration of the spinal cord white and gray matter

      2017, NeuroImage
      Citation Excerpt :

      In parallel, there are ongoing developments for processing MRI data, and in particular for segmenting the WM/GM (De Leener et al., 2016b). Some methods are based on fuzzy logic and require manual steps for initialization (Yiannakas et al., 2012; Ellingson et al., 2007b) and some were developed for specific applications (diffusion tensor imaging, DTI) and are therefore not applicable to a variety of contrasts (Ellingson et al., 2007b; Tang et al., 2013). Another approach consists in registering an MR-based template and WM/GM atlas to the data using non-linear transformations (Taso et al., 2015; De Leener et al., 2015).

    • Feature-ranking-based Alzheimer's disease classification from structural MRI

      2016, Magnetic Resonance Imaging
      Citation Excerpt :

      In this context, the proposed method can be employed, as it is the combination of t-test feature ranking and the Fisher Criterion, which not only reduces the dimensionality, but also increases the discriminability. Principal component analysis is a statistical dimensionality reduction method that extracts a set of orthogonal principal components (PCs) from an original dataset [62,63]. In this work, a 10-fold cross validation is used for measuring the performance of the classifiers.

    View all citing articles on Scopus
    View full text