Knowledge-based segmentation and labeling of brain structures from MRI images

https://doi.org/10.1016/S0167-8655(00)00135-5Get rights and content

Abstract

In this paper, we propose a new knowledge-based method illustrated in the context of segmentation, which labels internal brain structures viewed by magnetic resonance imaging (MRI). In order to improve the accuracy of the labeling, we introduce a fuzzy model of regions of interest (ROI) by analogy with the electrostatic potential distribution, to represent more appropriately the knowledge of distance, shape and relationship of structures. The knowledge is mainly derived from the Talairach stereotaxic atlas. The labeling is achieved by the regionwise labeling using genetic algorithms (GAs), followed by a voxelwise amendment using parallel region growing. The fuzzy model is used both to design the fitness function of GAs, and to guide the region growing. The performance of our proposed method has been quantitatively validated by six indices with respect to manually labeled images.

Introduction

Magnetic resonance imaging (MRI) plays a crucial role in noninvasive in vivo study of the human brain, for its high three-dimensional spatial resolution, and its remarkable discrimination of soft tissues. Segmentation and labeling of brain neuroanatomical structures are a prerequisite for quantitative morphometric analysis, three-dimensional volume visualization, and measurement of structure–function correlationship, particularly in clinical investigations, such as pathology, diagnosis, therapy, surgery planning and guidance (Zijdenbos and Dawant, 1994, Clarke et al., 1995). As far as we are concerned, our paper is devoted to the developing of an automatic method to segment and label precisely internal brain structures, such as ventricle, caudate, thalamus and putamen, from MRI images.

Automated labeling of structures is however complicated, facing difficulties due to overlapping intensities, anatomical variability in shape, size, and orientation, partial volume effects, as well as noise perturbations, intensity inhomogeneities, and low contrast in images. Therefore, it is inevitable to supplement anatomical knowledge, to achieve labeling like what radiologists do. In recent years, many reports have been published in this direction in terms of atlas-based (model-based, or knowledge-based) segmentation of neuroanatomical structures.

One intuitive strategy to use knowledge for labeling, named as the registration–segmentation paradigm by Collins and Evans (1999), is to register and transfer labels of a pre-labeled atlas onto the MRI images to be segmented. Various registration schemes have been presented for this purpose (Collins and Evans, 1999, Collins et al., 1995, Dawant et al., 1998, Ferrant et al., 1999, Meier et al., 1998, Chen et al., 1998). The performance of this strategy overrelies on the accuracy of the registration employed, which suffers from the limited degrees of freedom of the transformations, and from anatomical variabilities (for instance, in orientation, shape, size, and position). Furthermore, a one-to-one mapping does not always exist (Collins and Evans, 1999).

Another important strategy is to integrate the statistical knowledge of intensity and position into a shape model, and to locate the structures which match the model. Staib et al. (1997) used a gradient-based parametric deformable shape model, integrating region information and prior probability knowledge of mean shape and variation of the structures. Gonzalez Ballester et al. (1998) guided the segmentation by statistical shape knowledge built from datasets of pre-labeled structures. Several researchers used active shape models (ASM), introduced by (Cootes et al., 1994) to label brain structures (Kelemen et al., 1997, Duta and Sonka, 1998). ASM are parametric deformable models of shape and appearance of flexible objects, which restrict the possible deformation using shape template and intensity model, both generated through statistics of training sets. Kelemen et al. (1997) extended ASM to three dimensions and employed parametric representations of object shapes with elliptical harmonics. Duta and Sonka (1998) reported another improvement of ASM by incorporating a priori knowledge of structures. The performance of the shape model-based strategy is spoiled by the mismatching between the geometric model and the MRI gray level data (Collins et al., 1995).

Besides the two strategies aforementioned, other methods are also reported, for instance, GAs-based (genetic algorithms) interpretation (Sonka et al., 1996), atlas-based sequential recognition using information fusion (Géraud et al., 1999) and ANN-based (artificial neural network) identification (Magnotta et al., 1999).

The contribution of this paper is twofold. First, we propose a coarse-to-fine strategy to achieve precise segmentation and labeling of brain structures, based on the structural knowledge from the Talairach stereotaxic atlas (Talairach and Tournoux, 1998), and the statistical information from the MRI images under study. The atlas-based registration is used only to indicate the coarse location of the structures in the images. GAs are applied to search for the optimal labeling of oversegmented regions, the result of which is refined through voxelwise amendment by parallel region growing. Second, we introduce a fuzzy model of regions of interest (ROI) by analogy with the electrostatic potential distribution in the vicinity of hollow structures with uniform surface charge density. We use this model to describe the spatial and geometric knowledge of structures, to estimate the statistical moments, to design the objective function of GAs, and to guide the voxelwise amendment.

The paper is organized as follows. In Section 2, we detail the proposed method, involving preprocessing, fuzzy Markov random field (MRF) based oversegmentation, coarse-to-fine labeling. The validation is implemented quantitatively in Section 3, using manually labeled MRI images as reference. In Section 4, we make the conclusion.

Section snippets

Outline

The Talairach atlas is well accepted in medical image processing, owing to its contribution to the delineation and labeling of numerous brain neuroanatomical structures. As the Talairach atlas, denoted by T0, is sketch-based rather than intensity-based (one sample slice of this atlas in axial direction is shown in Fig. 1(a)), no algorithm exists to automatically superimpose it onto an MRI volume. Therefore, an MRI volume VT, which has been registered interactively into the stereotaxic

Results and quantitative validation

In our study, the subjects were scanned with a GE Signa 1.5 Tesla scanner, employing a T1-weighted spoiled gradient recalled (SPGR) pulse sequence. The parameters of the SPGR sequence were TR=30,TE=7ms,flipangle=40°; each dataset (volume) consists of 256×256×124 voxels, i.e., 124 axial slices with 256×256 voxels in each slice; the size of each voxel is 0.94×0.94×1.2mm3.

We employ six different indices of quantitative measure to validate the accuracy and reliability of this method, compared with

Conclusion

An automatic, knowledge-based method to segment and label brain neuroanatomical structures (for instance, ventricle, caudate, thalamus and putamen) from MRI images has been developed, using structural knowledge derived from the Talairach stereotaxic atlas, statistical and singularity information from the images to be labeled. To achieve a precise labeling of the desired structures, a coarse-to-fine strategy has been proposed, involving regionwise labeling using GAs followed by a voxelwise

Acknowledgements

We wish to express our gratitude to Prof. B. Mazoyer (CYCERON) for having provided the MRI data utilized in the present study. We are also thankful to Dr. J.M. Constans (CHRU de Caen) for his valuable suggestions, and to Dr. J.M. Fadili, for the three-dimensional visualization.

References (29)

  • L Clarke et al.

    MRI segmentation: methods and application

    Magnetic Resonance Imaging

    (1995)
  • B Moretti et al.

    Phantom-based performance evaluation: application to the brain segmentation from magnetic resonance images

    Medical Image Analysis

    (2000)
  • F Salzenstein et al.

    Parameter estimation in hidden fuzzy Markov random fields and image segmentation

    Graphical Models Image Process.

    (1997)
  • Chen, M., Kanade, T., Pomerleau, D., Schneider, J., 1998. 3D deformable registration of medical images using a...
  • D Collins et al.

    Animal: automatic nonlinear image matching and anatomical labeling

  • D Collins et al.

    Automatic 3D model-based neuroanatomical segmentation

    Human Brain Mapping

    (1995)
  • T Cootes et al.

    The use of active shape models for locating structures in medical images

    Image and Vision Comput.

    (1994)
  • B Dawant et al.

    Automatic 3D segmentation of internal structures of the head in MR images using a combination of similarity and free-form transformations

    Proc. SPIE

    (1998)
  • N Duta et al.

    Segmentation and interpretation of MR brain images: an improved active shape model

    IEEE Trans. Medical Imaging

    (1998)
  • M Ferrant et al.

    Multiobject segmentation of brain structures in 3D MRI using a computerized atlas

    Proc. SPIE

    (1999)
  • Géraud, T., Bloch, I., Maı̂tre, H., 1999. Atlas-guided recognition of cerebral structures in MRI using fusion of fuzzy...
  • Géraud, T., Mangin, J., Bloch, I., Maı̂tre, H., 1995. Segmenting internal structures in 3D MR images of the brain by...
  • G Gerig et al.

    Nonlinear anisotropic filtering of MRI data

    IEEE Trans. Medical Imaging

    (1992)
  • Cited by (27)

    • Whole brain segmentation with full volume neural network

      2021, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Brain tissue segmentation provides a quantitative tool for brain morphological analysis, surgical planning, disease detection and a wide range of clinical applications. Region-based approaches (Xue et al., 2001; Gui et al., 2012) take advantage of the homogeneity property of similar voxels in the same region for boundary detection. Contour-based deformable models (Huang et al., 2009; Kapur et al., 1996), region growing methods (Tang et al., 2000), level set methods (Wang et al., 2010, 2013; Chen et al., 2008) and graph-based approaches (Song et al., 2006) all fall under the region-based category.

    • A review on brain structures segmentation in magnetic resonance imaging

      2016, Artificial Intelligence in Medicine
      Citation Excerpt :

      Probably, one of the most well established region based techniques is region growing, which is the most frequently used in brain structures segmentation. Based on this technique, Xue et al. [159] proposed a method that performed regionwise labeling by means of GAs followed by voxelwise refinement using parallel region growing. They first over-segmented the target image into three brain tissues (WM, GM and CSF) and also got a coarse location of the structures by registering an atlas to the image.

    View all citing articles on Scopus
    View full text