Elsevier

Magnetic Resonance Imaging

Volume 23, Issue 7, September 2005, Pages 817-828
Magnetic Resonance Imaging

Original contribution
A soft-segmentation visualization scheme for magnetic resonance images

https://doi.org/10.1016/j.mri.2005.05.003Get rights and content

Abstract

Prevalent visualization tools exploit gray value distribution in images through modified histogram equalization and matching technique, referred to as the window width/window level-based method, to improve visibility and enhance diagnostic value. The window width/window level tool is extensively used in magnetic resonance (MR) images to highlight tissue boundaries during image interpretation. However, the identification of different regions and distinct boundaries between them based on gray-level distribution and displayed intensity levels is extremely difficult because of the large dynamic range of tissue intensities inherent in MR images. We propose a soft-segmentation visualization scheme to generate pixel partitions from the histogram of MR image data using a connectionist approach and then generate selective visual depictions of pixel partitions using pseudo color based on an appropriate fuzzy membership function. By applying the display scheme in clinical examples in this study, we could demonstrate additional overlapping regions between distinct tissue types in healthy and diseased areas (in the brain) that could help improve the tissue characterization ability of MR images.

Introduction

Interpreting 2D magnetic resonance (MR) images is an important area in the diagnosis of a disease condition. An image visualization tool that could shorten the time required for making a diagnosis, improve diagnostic accuracy and improve throughput of doctors is much needed. Identifying different types of tissues in MR images and boundaries is the most challenging problem for interpretation and diagnosis. Difficulties in identifying tissue types based on gray scale intensity in 2D MR images are due to partial volume averaging, tissue inhomogeneity and the nature of tissue types besides factors related to imaging parameters [1], [2]. We propose an interactive segmentation-based visualization tool that aims to circumvent these difficulties and to define a boundary region between distinct tissue types, which can be clinically useful with the help of anatomical knowledge and experience. The tool exploits a network-based clustering scheme and fuzzy logic to generate alternate and selective segmentation of 2D MR images for detailed analysis.

The main aim of MR image enhancement techniques is to improve visual information for human interpretation. These techniques are applied to spatial domains or histograms [3]. All the MR diagnostic consoles and offline workstations are equipped with 2D window width/window level and other image enhancement schemes [4], [5], [6], [7]. This interactive scheme can generate multiple interpretations of 2D MR images by changing window widths and window levels or by automatically finding the optimum value [8], [9], [10], [11], [12], [13].

In addition, several stand-alone commercially developed tools are also used for offline analysis and visualization. Some of these are 3DVIEWNIX [14], Analyze AVW [15], [16], [17], ApX, AVS/Express, IDL, IRIS Explorer, Khoros, MEDx, PV-WAVE, ROSS, Slicer Dicer, VoxelView, VTK [18] and BRAINS2 toolbox [19]. These software tools on offline workstations can be used in visualizing and processing biomedical images. The images can be viewed in different modes such as single slice, interactive, multimodal, multidimensional and multiplanar reformation, among others. These tools can also perform magnification, multicolor maps and reslicing in multidimensional visualization. In addition, these tools are used to perform surface rendering and volume rendering, among others, for MR volume-based image analysis. Other tools have been developed by various investigators to visualize biologic processes: AFNI [20] and VolVis [21]. AFNI (analysis of functional neuroimages) is used for functional magnetic resonance imaging (MRI) whereas VolVis is used for volume visualization. Different cut views of the brain in any plane can be generated for surgical planning. Kang et al. [22] proposed the use of interactive 3D editing tools for image segmentation after selecting the volume of interest based on visualization.

The BRAINPLOT tool [23] is used for visualization of structures related to a variety of neurological disorders. Clarke et al. [24] compared visual metric and automatic segmentation techniques for segmentation of MR images in brain tumor cases. Jaaski et al. [25] and Suri et al. [26] proposed to view 3D reconstructed blood vessels in brain, which are used for treatment planning and tumor surgery. Pommert and Hohne [27] reviewed the state of the art for the validation of medical volume visualization.

Ohhashi [28] proposed an approach based on neural network–based learning to set the window width/window level automatically. The learning data are formed based on actual image data and window width/window levels set by skilled operators. The patent further claims that the neural network can perform a learning system for individual display devices for all hospital equipment.

Most of the existing approaches in the literature use image enhancement techniques to generate various image interpretations of 2D MR images using histogram modification and threshold. Intensity variation in MR images, however, allows better differentiation between white matter (WM), gray matter (GM) and cerebral spinal fluid (CSF) in normal brain images of a diseased brain (e.g., brain tumor); intensity variation between tissue types is not crisp because pixels can belong to a tumor, a complex edema around it with or without tumor infiltration and healthy brain tissues. Image analysis of tissue boundaries can be of high clinical significance and play an important role in detecting early tissue changes, planning surgery and monitoring therapy response, among others. Existing visualization techniques do not use soft partitioning and rely on visual examination of image data. To overcome this shortcoming, we propose a neurofuzzy algorithm that works in two phases: first, using the connectionist approach [29], finding prototypes in the gray-level histogram and, second, using the fuzzy approach, generating segments interactively in 2D MR images for visual analysis of tissue anatomy. Our prototype-finding algorithm takes into consideration not only the peaks but also the data distribution and slope based on accumulative distribution along the histogram. This feature is not available in any other technique. The present technique, on the other hand, allows analysis of all the cluster prototypes that are present in the histogram. Furthermore, this adopts soft computation for segmentation in the identification of tissue types in gray-tone MR images associated with ambiguous and fuzzy intensity values with ill-defined boundaries.

This article is outlined as follows: Clustering MRI Data, the section that describes the clustering algorithm to determine clusters using the neural network and its analysis; Fuzzy Membership of Partitions, the section that describes a soft visualization approach to soft segment MR image data using the fuzzy approach; Results of Soft Visualization of Healthy and Diseased Cases, which shows results; and Conclusion.

Section snippets

Clustering MRI data

Clustering, in the context of the present problem, is the process of grouping pixels based on perceived similarities. These clusters can provide natural partitions of pixels corresponding to different regions in an image. Traditional clustering algorithms require a priori knowledge about the number of clusters, nature of data, clustering criteria, and so on. In algorithms such as K-means clustering [30], Fuzzy c-means clustering [31], Kohonen's self-organizing map [32], [33], and so on, the

Fuzzy membership of partitions

The prototypes found by the connectionist scheme define partitions of pixels in the image. To enable construction of soft partitions so that alternative possibilities and inherent variability in the pixel values can be taken care of, a fuzzy membership function is defined for the partitions.

Each intensity value is associated with multiple fuzzy memberships (GLS) using its distance from each prototype along the histogram as given by Eq. (6). If an intensity value is a prototype, its membership

Results of soft visualization of healthy and diseased cases

In this section, we present with clinical examples the effectiveness of the visualization scheme proposed. Prototype distribution in the axial T2-weighted MR image of the brain of a healthy volunteer in Fig. 2(A) broadly shows three tissue types: WM, GM and CSF. The cluster prototypes in the 1D histogram of the same image in Fig. 3(A) are located at 74, 91, 102, 113, 125 and 142 intensity values. The intensity values of WM lie in the dark gray region; of GM, in the gray region; and of CSF, in

Conclusion

Intensity variation in MR images at tissue boundaries in healthy cases is gradual and not sharp. It is more marked in diseased regions, with pixels belonging to either healthy or diseased tissues, causing poor differentiation of tissue interfaces and interfering image interpretation. Clinically, MR image analysis of tissue boundaries plays an important role and is significant in diagnosis and treatment planning. Existing visualization techniques do not support soft partitioning but depend

References (34)

  • Watanabe S. Image display method and image display device. Patent No. JP08278873, Published...
  • Ohhashi A. Digital display apparatus. Patent No. EP0409206, Published...
  • Shimazaki T, Yamaguchi K, Watanabe Y, Yamada N. Image diagnosis apparatus. Patent No. US5058176, Published...
  • Shinsuke I, Hiroshi S, Koji K. Medical image display device. Patent No. JP08096125, published...
  • Noriyuki N. Digital image display device. Patent No. JP 05012432 Published...
  • Sidiropoulos ND, Baras JS. Computer aided determination of window and level settings for film-less radiology. Patent...
  • Nishikawa M, Kabushiki KT, Oikawa D. Image display system. Patent No. EP0390164, Published...
  • Cited by (12)

    • A clustering fusion technique for MR brain tissue segmentation

      2018, Neurocomputing
      Citation Excerpt :

      The resulting images provide significant anatomical information about various body tissues, particularly the brain tissues, which enable researchers to study the pathology of the brain in an accurate manner [6]. However, the most difficult issue when analyzing inhomogeneous tissues, such as those of the brain, is dividing the image region based on gray scale intensity values [7]. This characterizes the concept of image segmentation, which is considered an essential step in many medical image processing applications [8].

    • Tissue classification in magnetic resonance images through the hybrid approach of Michigan and Pittsburg genetic algorithm

      2011, Applied Soft Computing Journal
      Citation Excerpt :

      Such constructions contain an uncertainty model of the type vagueness rather than randomness, and consequently they correspond to an explicit deterministic model even if it is not known. The segmentation of human brain images by fuzzy logic, neural network, evolutionary method [10–15] can produce results for specific problems. Reddick et al. [12] present an automated segmentation and classification scheme for multi-spectral MR images using artificial neural network.

    • Soft-computing based diagnostic tool for analyzing demyelination in magnetic resonance images

      2010, Applied Soft Computing Journal
      Citation Excerpt :

      This is a serious limitation since reliable estimate of the number of clusters in the MR image, in particular, for diseased cases, may not be available a priori. In order to overcome these limitations, we have adopted the connectionist-clustering scheme proposed in [17] for partitioning the MR images. This architecture makes use of contextual information available in the form of density estimates over a neighborhood in the histogram.

    • Bayesian mixture models of variable dimension for image segmentation

      2009, Computer Methods and Programs in Biomedicine
      Citation Excerpt :

      Due to limitations of the image acquisition process, image intensity values are typically corrupted by non-uniformities and noise [16]. Reliable estimates of the number of clusters in MR images, especially MR images of abnormal brains, may not be available a priori [17]. As a consequence of the finite resolution of the imaging process and the complexity of tissue boundaries, there is a partial volume effect which causes many voxels in MRI images to contain a mixture of more than one tissue type.

    • Handcrafted fuzzy rules for tissue classification

      2008, Magnetic Resonance Imaging
      Citation Excerpt :

      Here each slope value will be assigned a linguistic fuzzy value as L, M or H. A connectionist-based scheme for identifying prototypes in the histogram is proposed for soft segmentation of MR images [26]. These prototypes can induce partitions of pixels based upon gray-level distribution.

    View all citing articles on Scopus
    View full text