Overview and Fundamentals of Medical Image Segmentation

Jadwiga Rogowska 1 Introduction 69

Harvard Medical School 2 Thresholding 70

2.1 Global Thresholding • 2.2 Local (Adaptive) Thresholding • 2.3 Image Preprocessing and Thresholding

3 Region Growing 73

4 Watershed Algorithm 74

5 Edge-Based Segmentation Techniques 75

6 Multispectral Techniques 77

6.1 Segmentation Using Multiple Images Acquired by Different Imaging Techniques •

6.2 Segmentation Using Multiple Images Acquired over Time

7 Other Techniques 79

8 Concluding Remarks 80

References 80

1 Introduction

The principal goal of the segmentation process is to partition an image into regions (also called classes, or subsets) that are homogeneous with respect to one or more characteristics or features [11,16,20, 30,36,66,77, 96,107,109]. Segmentation is an important tool in medical image processing and it has been useful in many applications. The applications include detection of the coronary border in angiograms, multiple sclerosis lesion quantification, surgery simulations, surgical planning, measuring tumor volume and its response to therapy, functional mapping, automated classification of blood cells, studying brain development, detection of microcalcifications on mam-mograms, image registration, atlas-matching, heart image extraction from cardiac cineangiograms, detection of tumors, etc. [8,14,15,35,38,41a, 61,71,88,109,115,132].

In medical imaging, segmentation is important for feature extraction, image measurements, and image display. In some applications it may be useful to classify image pixels into anatomical regions, such as bones, muscles, and blood vessels, while in others into pathological regions, such as cancer, tissue deformities, and multiple sclerosis lesions. In some studies the goal is to divide the entire image into subregions such as the white matter, gray matter, and cerebrospinal fluid spaces of the brain [67], while in others one specific structure has to be extracted, for example breast tumors from magnetic resonance images [71].

A wide variety of segmentation techniques has been proposed (see surveys in [11,20,30,41,77,83,127]). However, there is no one standard segmentation technique that can produce satisfactory results for all imaging applications. The definition of the goal of segmentation varies according to the goal of the study and the type of the image data. Different assumptions about the nature of the analyzed images lead to the use of different algorithms.

Segmentation techniques can be divided into classes in many ways, depending on classification scheme:

• Manual, semiautomatic, and automatic [101].

• Pixel-based (local methods) and region-based (global methods) [4].

• Manual delineation, low-level segmentation (thresholding, region growing, etc), and model-based segmentation (multispectral or feature map techniques, dynamic programming, contour following, etc.) [109].

• Classical (thresholding, edge-based, and region-based techniques), statistical, fuzzy, and neural network techniques [87].

The most commonly used segmentation techniques can be

Copyright © 2000 by Academic Press.

All rights of reproduction in any form reserved.

classified into two broad categories: (1) region segmentation techniques that look for the regions satisfying a given homogeneity criterion, and (2) edge-based segmentation techniques that look for edges between regions with different characteristics [16,36,77,96,107].

Thresholding is a common region segmentation method [25,83,98,107,127]. In this technique a threshold is selected and an image is divided into groups of pixels having values less than the threshold and groups of pixels with values greater or equal to the threshold. There are several thresholding methods: global methods based on gray-level histograms, global methods based on local properties, local threshold selection, and dynamic thresholding. Clustering algorithms achieve region segmentation [13,27,37,54] by partitioning the image into sets or clusters of pixels that have strong similarity in the feature space. The basic operation is to examine each pixel and assign it to the cluster that best represents the value of its characteristic vector of features of interest. Region growing is another class of region segmentation algorithms that assign adjacent pixels or regions to the same segment if their image values are close enough, according to some preselected criterion of closeness [77,85].

The strategy of edge-based segmentation algorithms is to find object boundaries and segment regions enclosed by the boundaries [16,36,41,72,96]. These algorithms usually operate on edge magnitude and/or phase images produced by an edge operator suited to the expected characteristics of the image. For example, most gradient operators such as Prewitt, Kirsch, or Roberts operators are based on the existence of an ideal step edge. Other edge-based segmentation techniques are graph searching and contour following [6,14,106].

Traditionally, most image segmentation techniques use one type of images (MR, CT, PET, SPECT, ultrasound, etc.). However, the performance of these techniques can be improved by combining images from several sources (multi-spectral segmentation [29,89,117]) or integrating images over time (dynamic or temporal segmentation [71,93,108]).

The following sections will present some of the segmentation techniques that are commonly used in medical imaging. In Section 2 we will discuss several thresholding techniques. Section 3 will describe region growing techniques. The watershed algorithm will be reviewed in Section 4. Section 5 will present edge-based segmentation techniques. A discussion of multispectral segmentation methods will be given in Section 6.

2 Thresholding

Several thresholding techniques have been developed [16,25,36,41,51,96-98,107,127], Some of them are based on the image histogram; others are based on local properties, such as local mean value and standard deviation, or the local gradient, The most intuitive approach is global thresholding, When only one threshold is selected for the entire image, based on the image histogram, thresholding is called global, If the threshold depends on local properties of some image regions, for example local average gray value, thresholding is called local, If the local thresholds are selected independently for each pixel (or groups of pixels), thresholding is called dynamic or adaptive,

2.1 Global Thresholding

Global thresholding is based on the assumption that the image has a bimodal histogram and, therefore, the object can be extracted from the background by a simple operation that compares image values with a threshold value T [25,107], Suppose that we have an image f (x, y) with the histogram shown on Fig, 1,

The object and background pixels have gray levels grouped into two dominant modes, One obvious way to extract the object from the background is to select a threshold T that separates these modes,

The thresholded image g (x, y ) is defined as g(x, y) =

The result of thresholding is a binary image, where pixels with intensity value of 1 correspond to objects, while pixels with value 0 correspond to the background,

Figure 2 shows the result of segmentation by thresholding, The original image (Fig, 2A) contains white cells on a black background, Pixel intensities vary between 0 and 255, The threshold T = 127 was selected as the minimum between two modes on a histogram (Fig, 2B), and the result of segmentation is shown in Fig, 2C, where pixels with intensity values higher than 127 are shown in white, In the last step (Fig, 2D) the edges of the cells were obtained by a 3 x 3 Laplacian (second-order derivative [36]; also see description in Section 5), which was applied to the thresholded image on Fig, 2C,

There are many other ways to select a global threshold, One

FIGURE 1 An example of bimodal histogram with selected threshold T.

FIGURE 2 An example of global thresholding. (A) Original image, (B) histogram of image A, (C) result of thresholding with T = 127, (D) outlines of the white cells after applying a 3x3 Laplacian to the image shown in C.

FIGURE 3 An example of the sensitivity of the threshold level selection. (A) Cross-sectional intensity profile of a light object on a dark background with three thresholding levels T1, T2, and T3, and three other levels generated by adding a small value AT; (B) a hypothetical plot of the area (A) or perimeter (P) versus thresholding level T.

FIGURE 2 An example of global thresholding. (A) Original image, (B) histogram of image A, (C) result of thresholding with T = 127, (D) outlines of the white cells after applying a 3x3 Laplacian to the image shown in C.

of them is based on a classification model that minimizes the probability of error [77]. For example, if we have an image with a bimodal histogram (e.g., object and background), we can calculate the error as the total number of background pixels misclassified as object and object pixels miscalssified as background. A semiautomated version of this technique was applied by Johnson et al. [56]) to measure ventricular volumes from 3D magnetic resonance (MR) images. In their method an operator selects two pixels — one inside an object and one in the background. By comparing the distribution of pixel intensities in the circular regions around selected pixels, the threshold is calculated automatically and it corresponds to the least number of misclassified pixels between two distributions. The result of the thresholding operation is displayed as a contour map and superimposed on the original image. If needed, the operator can manually modify any part of the border. The same technique was also applied to extract lymph nodes from CT images and was found to be very sensitive to user positioning of interior and exterior points [95]. Some of the threshold selection techniques are discussed in Refs. [25,96,127].

In many applications appropriate segmentation is obtained when the area or perimeter of the objects is minimally sensitive to small variations of the selected threshold level. Figure 3A shows the intensity profile of an object that is brighter than

FIGURE 3 An example of the sensitivity of the threshold level selection. (A) Cross-sectional intensity profile of a light object on a dark background with three thresholding levels T1, T2, and T3, and three other levels generated by adding a small value AT; (B) a hypothetical plot of the area (A) or perimeter (P) versus thresholding level T.

background, and three threshold levels for segmentation: T1, T2, and T3. A small variation AT in the lowest threshold level will cause a significant change in the area or perimeter of the segmented object. The same is true for the highest threshold level. However, a change of AT in the middle level will have minimal effect on the area or perimeter of the object. The object area A(T) and perimeter P(T) are functions of the threshold T that often exhibit the trend shown in Fig. 3B. Therefore, the threshold level that minimizes either dA(T)/dT or dP(T)/dT is often a good choice, especially in the absence of operator guidance and when prior information on object locations is not available.

A related technique that evaluates multiple thresholds is based on an estimate of the gradient magnitude around the segmented object edge [16]. The average gradient magnitude is given by

0 0

Post a comment