Introduction

Image analysis techniques have been broadly used in computer-aided medical analysis and diagnosis in recent years. Computer-aided image analysis is an increasingly popular tool in medical research and practice, especially with the increase of medical images in modality, amount, size, and dimension. Image segmentation, a process that aims at identifying and separating regions of interests from an image, is crucial in many medical applications such as localizing pathological regions, providing objective quantative assessment and monitoring of the onset and progression of the diseases, as well as analysis of anatomical structures.

Generally speaking, segmentation techniques are application specific and nonuniversal. There exists no approach that works best for all types of images. In fact, some approaches work better on one type of image than others, depending on the modality of the image. For example, images acquired by magnetic resonance imaging (MRI) and radiographic X-ray imaging are quite different from retinal images. In the former, the images are represented by intensity variations proportional to radiation absorption or RF signal amplitude mapped into gray-level values, while for the latter, the images are chromatic and generated optically. Numerous segmentation techniques have been developed for gray scale

1 Department of Electrical and Computer Engineering, Texas Tech University Lubbock, TX 79409-3102

images [1-5], while color image segmentation techniques have been created much later than its gray-level counterpart because of the computational complexity involved with the latter. However, the availability of fast digital processors in recent times allows easy implementations of such complex algorithms. Most of the segmentation techniques applied to gray-level images can also be extended to color images [6, 7].

Clustering is a pattern recognition technique that has been frequently used in image segmentation [8,9]. Similar to the variety of approaches in image segmentation, there are numerous clustering techniques based on statistics, fuzzy logic [10, 11], neural network, or an integration of these [12]. This chapter applies two recently developed advanced clustering algorithms, namely, deterministic annealing (DA) [13] and adaptive fuzzy leader clustering (AFLC) [14] and compares their performances with other standard well-known algorithms in efficient segmentation of medical images. DA is designed on a statistical frame work, while AFLC has a neural network structure embedded with fuzzy optimization. The performances of these two algorithms have been compared with classical clustering techniques such as k-means [15], and fuzzy C-means (FCM) [16]. These clustering algorithms have been applied to segment a few diverse types of medical images. All operations are performed on images in the spatial domain, i.e., pixel intensity will be used as the only feature. For gray-scale images, such as MRI, the feature will be 1D, while for color images, such as the retinal image, the classification is 3D (red, green, and blue components for each pixel).

The major advantage of using clustering for medical image segmentation is that these unsupervised techniques for data partitioning do not require a training set, which is not easy to find in most clinical datasets. The two clustering techniques, namely AFLC and DA, used in our study to investigate the effectiveness and accuracy of these techniques in medical image segmentation can be considered as optimization processes. Both AFLC and DA do not require an initial guess of the actual number of clusters present in a dataset and thus do not suffer from the instability inherent to traditional and well-known clustering algorithms such as k-means.

Several types of medical images are selected and used as examples of clustering application. The first modality we used is MRI. We compared the segmentation of anatomical structures such as gray matter, white matter, and cerebrospinal fluid from simulated MRI. Pathological segmentation of multiple sclerosis with both simulated and real MRI is also performed. The second imaging modality involves stereo retinal imaging for evaluating structural damage in the retina. We demonstrated delineation of blood vessels in these images with DA and applied the result to 3-D disparity mapping and segmentation of optic disk/cup. Traditional 2-D segmentation of the optic disk/cup is also presented and compared with manually segmented optic disk/cup by ophthalmologists. The third imaging modality is direct optical (color) imaging used to investigate precancerous lesions in the cervix. Traditionally, an abnormal Pap smear finding is followed by Cervicographic and Colposcopic examination of the cervix by a gynecologist/oncologist to determine the severity of the lesions and identify the location for a biopsy probe. Segmentation of lesions using selected clustering algorithms is carried out for comparison.

The chapter is organized as follows: A brief description of image segmentation is given in section 6.2. Selected clustering techniques, traditional and advanced, are introduced in section 6.3. In section 6.4, results of applying these techniques for precise segmentation of various modalities of medical images are presented. Section 6.5 contains discussions and conclusions.

0 0

Post a comment