Elsevier

NeuroImage

Volume 27, Issue 4, 1 October 2005, Pages 795-804
NeuroImage

Probabilistic segmentation of brain tissue in MR imaging

https://doi.org/10.1016/j.neuroimage.2005.05.046Get rights and content

Abstract

A new method has been developed for probabilistic segmentation of five different types of brain structures: white matter, gray matter, cerebro-spinal fluid without ventricles, ventricles and white matter lesion in cranial MR imaging. The algorithm is based on information from T1-weighted (T1-w), inversion recovery (IR), proton density-weighted (PD), T2-weighted (T2-w) and fluid attenuation inversion recovery (FLAIR) scans. It uses the K-Nearest Neighbor classification technique that builds a feature space from spatial information and voxel intensities. The technique generates for each tissue type an image representing the probability per voxel being part of it. By application of thresholds on these probability maps, binary segmentations can be obtained. A similarity index (SI) and a probabilistic SI (PSI) were calculated for quantitative evaluation of the results. The influence of each image type on the performance was investigated by alternately leaving out one of the five scan types. This procedure showed that the incorporation of the T1-w, PD or T2-w did not significantly improve the segmentation results. Further investigation indicated that the combination of IR and FLAIR was optimal for segmentation of the five brain tissue types. Evaluation with respect to the gold standard showed that the SI-values for all tissues exceeded 0.8 and all PSI-values exceeded 0.7, implying an excellent agreement.

Introduction

Magnetic resonance imaging is nowadays a widely applied technique for visualization of the human brain. MR images offer a large utility in diagnosis and treatment of brain diseases. In addition to the detection of abnormalities, there is an increasing interest in quantitative determination of intracranial lesions (i.e. volume rather than number of white matter lesions), cerebro-spinal fluid (for atrophy research), white matter (WM) and gray matter (GM). Therefore, accurate segmentation of different brain tissues and pathologies is of high importance for many studies.

The last decade, several segmentation methods based on signal intensities in MR images have been developed. Some methods focus on the segmentation of WM and GM only (González Ballester et al., 2000, Schnack et al., 2001a, Shattuck et al., 2001, Suckling et al., 1999, Tang et al., 2000). Other methods also segment cerebro-spinal fluid (Amato et al., 2003, Chard et al., 2002, Lemieux et al., 2003, Grabowski et al., 2000, Kovacevic et al., 2002, Marroquin et al., 2002, Mohamed et al., 1999, Ruan et al., 2000, Vaidyanathan et al., 1997, Van Leemput et al., 1999a, Van Leemput et al., 1999b, Zhu and Jiang, 2003). Furthermore, a few techniques exist for segmentation of specific brain structures, like thalamus, putamen, caudate (Barra and Boire, 2001, Fischl et al., 2002), for multiple sclerosis (MS) lesions (Alfano et al., 2000, Goldberg-Zimring et al., 1998, Guttman et al., 1999, Kamber et al., 1995, Mohamed et al., 2001, Van Leemput et al., 2001, Warfield et al., 1995, Wei et al., 2002, Zijdenbos et al., 2002), as well as cerebral white matter lesions (WML) (Anbeek et al., 2004, Jack et al., 2001, Wei et al., 2002). A method for whole brain segmentation including WMLs validated in a quantitative way on a voxel basis would be highly desirable.

The techniques of Chard et al. (2002), González Ballester et al. (2000), Grabowski et al. (2000), Lemieux et al. (2003), Ruan et al. (2000) and Shattuck et al. (2001) are based on one type of MR image only, namely the T1-weighted scan (T1-w). The majority of the abovementioned methods are multispectral. Tang et al. (2000) apply the T1-w together with the T2-weighted scan (T2-w), and Goldberg-Zimring et al. (1998), Suckling et al. (1999) and Wei et al. (2002) use the proton density-weighted (PD) with T2-w scan. A combination of three image types (T1-w, PD, T2-w) was used by Alfano et al. (2000), Amato et al. (2003), Guttman et al. (1999), Kovacevic et al. (2002), Marroquin et al. (2002), Mohamed et al., 1999, Mohamed et al., 2001, Vaidyanathan et al. (1997), Van Leemput et al., 1999a, Van Leemput et al., 1999b, Van Leemput et al., 2001 and Zijdenbos et al. (2002). Another development is the application of the T1-weighted 3D-FFE by Kamber et al. (1995) and Schnack et al. (2001a), which is also used in a technique for the segmentation of ventricles only (Schnack et al., 2001b). Interestingly, other image types, like inversion recovery (IR) or fluid attenuation inversion recovery (FLAIR), used by Jack et al. (2001), are seldom considered for segmentation.

The abovementioned techniques can be divided into methods based on pattern recognition, and based on other methods. The techniques that do not use pattern recognition mostly perform a type of histogram analysis (Grabowski et al., 2000, Jack et al., 2001, Lemieux et al., 2003, Ruan et al., 2000, Schnack et al., 2001a, Shattuck et al., 2001). Exceptions are the methods of Schnack et al. (2001b), which are mainly based on region growing, and Tang et al. (2000), which combine edge detection, region selection and intensity thresholding. Some pattern recognition based methods make use of unsupervised clustering (Chard et al., 2002, Suckling et al., 1999, Warfield et al., 1995, Zhu and Jiang, 2003). Furthermore, the iterative Expectation-Maximization algorithm is frequently applied (Guttman et al., 1999, Kovacevic et al., 2002, Marroquin et al., 2002, Wei et al., 2002), sometimes in combination with other techniques like surface fitting (González Ballester et al., 2000), or a Markov random field (Van Leemput et al., 1999b, Van Leemput et al., 2001). The remaining group applies supervised pattern recognition methods, such as a minimum distance classifier, a Bayesian classifier, and decision tree classifiers (Kamber et al., 1995), discriminant analysis (Alfano et al., 2000, Amato et al., 2003), artificial neural networks (Goldberg-Zimring et al., 1998, Zijdenbos et al., 2002) or K-Nearest Neighbor (KNN) classification (Mohamed et al., 1999, Mohamed et al., 2001, Vaidyanathan et al., 1997). Some of them combine pattern recognition with shape analysis or thresholding. Training data for the supervised methods are obtained in different ways. Mohamed et al., 1999, Mohamed et al., 2001), Vaidyanathan et al. (1997) andGoldberg-Zimring et al. (1998) select samples of the small tissue areas. However, a training set based on sampling might not be representative for the whole amount of voxels of a particular tissue type. Alfano et al. (2000), Amato et al. (2003) and Kamber et al. (1995) derive model parameters, like the mean intensity value, from segmented volumes. Zijdenbos et al. (2002) uses manually identified MS-lesions for the training set. In our study, we preferred a technique that was suitable for multiple input channels, using information from different scans, which made pattern recognition an appropriate choice. Furthermore, we aimed to be independent of prior assumptions about the intensity distributions in different images. Therefore, we have used the non-parametric technique of KNN-classification. To incorporate the entire tissue information from the images, we have composed the training set from all voxels of the segmented tissue types.

In this paper, we propose a method which is suitable for probabilistic segmentation of five types of brain structure simultaneously: WM, GM, cerebro-spinal fluid without ventricles (CSF), ventricles (VENT) and WML. The technique is based on a supervised K-Nearest Neighbor (KNN) classification method using both signal intensity and spatial information from five types of scans: T1-w, IR, PD, T2-w and FLAIR. Segmentation is performed by estimation of the probabilities of a voxel belonging to each of these tissue types. Having five different scan types available, we investigated which images have a significant influence on the segmentation quality. Furthermore, combinations of the significant images were examined, in order to compose an ideal image combination, with respect to quality and speed, for brain tissue segmentation based on KNN-classification.

Section snippets

Patients and MR imaging

Ten patients, diagnosed with arterial vascular disease were included in this study. Mean age of the patients was 65 (mean ± SD: 65.7 ± 7.7, range: 54–75). Eight patients were male.

MR images were processed on a Philips Gyroscan ACS-NT 1.5 T whole body system (Philips Medical Systems, Best, The Netherlands). All patients underwent the same MR protocol of the brain consisting of transaxial T1-w, IR, T2-w, PD and FLAIR scans (Fig. 1). All scans were performed with a 4 mm slice thickness, no slice

Results

Probabilistic segmentations of the five tissue types (WM, GM, CSF, VENT and WML) have been generated for all 10 patients with six different feature sets. This resulted per patient and per feature set in five probability maps (one for each tissue type). By application of a series of thresholds ranging from 0 to 1 on the probabilistic segmentations, binary segmentations have been generated for all tissue types. The SIs of these segmentations have been calculated for all tissue types, and the

Discussion

KNN-classification combining spatial information with voxel intensities offers a strong and flexible technique for segmentation of brain tissue in MR imaging. It provides the opportunity to segment different tissue types simultaneously with highly accurate results. The choice of the image types in the acquisition protocol is apparently of great importance for the quality of the segmentation. This study has shown that the IR and FLAIR images together provide an optimal combination for

References (40)

  • H.G. Schnack et al.

    Automated segmentation of the ventricular system from MR images of the human brain

    NeuroImage

    (2001)
  • D.W. Shattuck et al.

    Magnetic resonance image tissue classification using a partial volume model

    NeuroImage

    (2001)
  • J. Suckling et al.

    A modified fuzzy clustering algorithm for operator independent brain tissue classification of dual echo MR images

    Magn. Reson. Imaging

    (1999)
  • H. Tang et al.

    MRI brain image segmentation by multi-resolution edge detection and region selection

    Comput. Med. Imaging Graph.

    (2000)
  • M. Vaidyanathan et al.

    Normal brain volume measurements using multispectral MRI segmentation

    Magn. Reson. Imaging

    (1997)
  • S.K. Warfield et al.

    Adaptive, template moderated, spatially varying statistical classification

    Med. Image Anal.

    (2000)
  • C. Zhu et al.

    Multicontext fuzzy clustering for separation of brain tissues in magnetic resonance images

    NeuroImage

    (2003)
  • B. Alfano et al.

    Automated segmentation and measurement of global white matter lesion volume in patients with multiple sclerosis

    J. Magn. Reson. Imaging

    (2000)
  • V. Barra et al.

    Automatic segmentation of subcortical brain structures in MR images using information fusion

    IEEE Trans. Med. Imaging

    (2001)
  • C.M. Bishop

    Neural Networks for Pattern Recognition

    (1995)
  • Cited by (195)

    • Whole brain segmentation with full volume neural network

      2021, Computerized Medical Imaging and Graphics
    View all citing articles on Scopus
    View full text