Image Segmentation

Image segmentation can be defined as separating the image into similar constituent parts. Given an image I, segmentation of I is a partition P of I into a set of N regions Rn, n = 1,..., N, such that (JN1 Rn = I. The separated regions should be homogeneous and meaningful to the application intended. According to Pham et al. [1] image segmentation techniques can be classified into several categories, such as thresholding, region growing, classifiers, clustering, Markov random field, artificial neural network, fuzzy logic, deformable models, and atlas-guided approaches. The performance efficiency of each approach, however, varies and is dependent on specific application and image modality. When a practical application is concerned, sometimes integration of these techniques is needed to achieve better performance. A number of review papers on image segmentation in general and specifically on medical image segmentation are already available [1-7,9]. In this chapter, we have focused on the impact of recent advanced clustering algorithms in precise segmentation of medical images.

Most of the common medical images such as MRI, positron emission tomography (PET), computed tomography (CT), and ultrasound images are monochromatic. Some of these types of images can be pseudocolored. As mentioned earlier in section 6.1, most of the segmentation techniques developed for gray scale images can be extended to color images. There are quite a few color models [17] that are commonly used in image processing, mainly to comply with color video standards and human perception. RGB (red, green, blue), HSI (hue, saturation, intensity), and CIE L*a*b* are color models that have been frequently used in segmentation. RGB is hardware oriented while HSI and L* a*b* representations are compatible with human visual perception. What is more, perceptual uniformity of the L*a*b* color space is advantageous over RGB and HSI in that the human perception of color difference can be represented as the euclidean distance between color points, a useful property can be used in error functions of some segmentation algorithms. Most color images, such as the ones used in our examples, the color retinal stereo images and the color cervix images, are captured in RGB. If processing in other color space is preferred, color transformation is needed [6, 7].

To verify the efficiency of a segmentation method, segmentation result is compared with the "truth model." Truth models for practical images are often obtained by manual segmentation. In medical imaging, manual segmentation is usually performed by trained medical professionals. However, such truth models are not as accurate since they are prone to subjective variability, and often poor repeatability, leading to some ambiguities in diagnosis. Therefore, as far as validation is concerned, computer-simulated phantoms are preferred. The phantoms are simplified mimics of the real images. Quantative statistics such as the number of misclassified pixels or shape differences can be obtained by comparison segmented result with the phantom in a straightforward manner. Unfortunately, phantoms are not available for all image modalities. In our examples, we used both computer-simulated phantoms (MRI) as well as manual segmentations (stereo retinal image and color cervix images) to validate the clustering algorithms.

Was this article helpful?

0 0

Post a comment