Biomedical Image Segmentation Based on Multiple Image Features

Abstract

Image segmentation is a procedure that partitions an image into disjointing segments with each segment sharing similar properties such as intensity, color, boundary and texture. In general, three main types of image features are used to guide image segmentation, which are intensity or color, edge, and texture. In other words, image segmentation methods generally fall into three main categories: intensity-based (or color-based), edge-based, and texture-based segmentations. Intensity-based segmentation assumes that an image is composed of several objects with constant intensity. This kind of methods usually depends on intensity similarity comparisons to separate different objects. Histogram thresholding (Otsu, 1979; Sahoo et al., 1988), clustering (Bezdek, 1981; Pappas, 1992), and split-and-merge (Tyagi & Bayoumi, 1992; Wu, 1993) are examples of intensity-based segmentation methods. Edge-based segmentation has a strong relationship with intensity-based segmentation, since edges usually indicate discontinuities in image intensity. Edge-based segmentation uses different methods to describe the salient edges in images. Then, the boundaries of objects are detected by edge grouping or edge-driven active contour construction. Widely-used methods in edge-based segmentation include Canny (Canny 1986), watershed (Vincent & Soille, 1991) and snake (Kass et al., 1998; Xu & Prince, 1998). Texture is another important characteristic used to segment objects in an image. Most texturebased segmentation algorithms map an image into a texture feature space, then statistical classification methods (Randen & Husoy, 1999) are usually used to differentiate texture features. Co-occurrence matrix (Zucker & Terzopoulos, 1980), directional gray-level energy (Hsiao & Sawchuk, 1989), Gabor filters (Jain & Farrokhnia, 1991), and fractal dimension (Mandelbrot, 1976; Pentland 1984) are frequently used methods to obtain texture features. Biomedical images usually suffer from certain imaging artifacts stemming from different imaging modalities. Because of imperfect illumination, signal attenuation or signal superposition in biomedical images, intensity-based segmentation methods are often ineffective in differentiating neighboring tissues with similar intensity features. Further, because of noises in biomedical images, detected tissue edges are often discontinuous, obstructed, or obscure. It remains a problem in edge-based segmentation to interpret and connect discontinuous edges. As a high-level image characteristic, texture reflects the spatial