15-05 Image Segmentation and Multispectral Analysis

Space and military technologies were the forerunners of many image-processing applications which later found their way into medical imaging. One of the most important was the Landsat program of Nasa. Landsat created sets of images of the earth consisting of four or more images of different spectral windows (us­ual­ly, two within the visible spectrum and two within the infrared spectrum). Si­mi­lar ap­proa­ches are used today in medical image-processing. Image seg­men­ta­tion is one of the most important tools in automated image analysis [⇒ Bezdek, ⇒ Lundervold].

Plain and postcontrast T1-weighted, T2-weighted, and diffusion images can be used as multispectral images. Reducing the representation of an image to a small number of components was one of the image-processing projects based on such pictures, a process called feature extraction. It permits the separation of the basic parts of an image by sets of features that can be extracted from the ima­ge (or several images) and, in turn, and then can be used to calculate other features such as edges and textures.

Segmentation is also applied in preprocessing of images for multimodality image registration. Image segmentation can be used in static images and, quite important for the use of contrast agents, in dynamic time-varying images. The detection of gray-level discontinuity allows the highlighting of points, lines, and edges in an image.

Similarity techniques reveal areas of similar signal intensities using thres­hold­ing, region growing, as well as region splitting and merging [⇒ Gonzales]. An overview of the components of an image-analysis system is given in Figure 15- 10. A detailed description of segmentation is beyond the limits of this chapter, but is available in other treatises [⇒ Clarke, ⇒ Gonzales, ⇒ Russ].

Figure 15-10: Steps in image analysis: preprocessing improves the quality of the image by re­duc­ing artifacts; feature extraction and se­lec­tion provide the measurement vectors on which segmentation is based. After seg­men­ta­tion, classification and description allow pattern recognition.

Multispectral models can be divided into supervised and unsupervised mo­dels [⇒ Alaux]. An unsupervised classification (like cluster analysis) into connected regions is generally sufficient to provide good partition of an image into relevant com­po­nent structures. Supervised pattern recognition is mainly successful where a re­li­able classification can be expected on the basis of a priori knowledge of the tis­sue parameters.

The meanwhile classical example is gray- and white-matter separation on the basis of relaxation time data (Figure 15-11).

Figure 15-11: (a) original brain imaging, and (b) segmented image presenting 90 different tissues.

Practical Applications. In medicine, segmentation is applied for the division of images into components reflecting the same or similar tissues. Today, the con­cept of segmentation and its application of volumetry have become fast and cli­ni­cal­ly usable. Segmentation allows identification of anatomical areas of interest for diagnosis and therapy, for instance for planning of surgery. Measurement of tumor volume before and after treatment has become a relatively easy task with image segmentation. Among other applications are quantitative measurements of brain atrophy in patients with Alzheimer's disease or alcoholic brain damage.

In cardiac MR imaging segmentation methods and contour-detection methods have been successfully applied to detect the borders of myocardium in order to calculate parameters like ejection fraction and myocardial mass. Automatic con­tour detection can be used for the three-dimensional depiction of bone or soft tis­sue structures, e.g., to produce prostheses.

15-06 Three-Dimensional Visualization

As with most imaging modalities, MR imaging data are normally presented as two-dimensional gray-scale images. However, MR imaging is essentially a 3D method and can produce three-dimensional data sets of virtually any body organ [⇒ Aichner].

The simplest way of visualizing such data sets is by letting the ra­dio­lo­gist flip through the data set displayed as slices of 2D images, leaving it up to the ra­dio­lo­gist to visualize the structures. Whereas this approach is often suitable for some purposes, like diagnosis, it is less suitable for other purposes, like surgical plan­ning or radiotherapy planning. Thus, there is a need for 3D visualization tech­ni­ques. By performing segmentation, surface- or volume-rendering techniques can be applied [⇒ Maintz]. The advantage of surface-rendering techniques is that they are easy and fast to visualize and manipulate (by rotation, zooming, etc.).

Since a segmentation has been done, it is possible to manipulate the 3D data set by removing tissues, request volumes and sizes, etc. (Figure 15-12). The dis­ad­van­tage is that segmentation of the data is required before visualization can be performed, and that some information is lost in the segmentation step. An alternative technique is volume-rendering. Volume-rendering does not require segmentation. However, the method requires more powerful computers to be fully interactive, and normally requires some interaction to visualize structures of interest.

Figure 15-12: Dissection of an MR imaging-based head model. A wire mesh is used to define cut planes [⇒ Tiede]. Similar reconstructions can be used in surgery and ra­dia­tion the­ra­py plan­ning. One of the main problems in 3D image-processing is that objects within the 3D domain may obscure each other. Therefore any visualization must be pre­ced­ed by a segmentation step in which 3D regions belonging to an organ must be iden­ti­fied.