Detection, localization, diagnosis, staging, and monitoring treatment responses are the most important aspects and crucial procedures in diagnostic medicine and clinical oncology. Early detection and localization of the diseases and accurate disease staging can improve the survival and change management in patients prior to planned surgery or therapy. Therefore, current medical practice has been directed toward early but efficient localization and staging of diseases, while ensuring that patients would receive the most effective treatment.

Current disease management is based on the international standard of cancer staging using TNM classification, viz. size, location, and degree of invasion of the primary tumor (T), status of regional lymph node (N), and whether there is any distant metastasis (M). Over the century, histopathology retains its main role as the primary means to characterization of suspicious lesions and confirmation of malignancy. However, the pathologic interpretation is heavily dependent on the experience of the pathologist, and sampling errors may mean that there are insufficient amounts of tissue in the specimens, or the excised tissue does not accurately reflect tumor aggressivity. In addition, some lesions may return nondiagnostic information from the specimens, or they are difficult or too

1 Department of Electronic and Information Engineering, Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong lll dangerous to biopsy. As a result, more invasive and unpleasant diagnostic procedures are sought.

The last few decades of the twentieth century have witnessed significant advances in medical imaging and computer-aided medical image analysis. The revolutionary capabilities of new multidimensional medical imaging modalities and computing power have opened a new window for medical research and clinical diagnosis. Medical imaging modalities are used to acquire the data from which qualitative and quantitative information of the underlying pathophysi-ological basis of diseases are extracted for visualization and characterization, thus helping the clinicians to accurately formulate the most effective therapy for the patients by integrating the information with those obtained from some other possibly morbid and invasive diagnostic procedures. It is important to realize that medical imaging is a tool that is complementary to but not compete with the conventional diagnostic methods. Indeed, medical imaging provides additional information about the disease that is not available with the conventional diagnostic methods, and paves a way to improve our understanding of disease processes from different angles.

Modern medical imaging modalities can be broadly classified into two major categories: structural and functional. Examples of major structural imaging modalities include X-ray computed tomography (CT), magnetic resonance imaging (MRI), echocardiography, mammography, and ultrasonography. These imaging modalities allow us to visualize anatomic structure and pathology of internal organs. In contrast, functional imaging refers to a set of imaging techniques that are able to derive images reflecting biochemical, electrical, mechanical, or physiological properties of the living systems. Major functional imaging modalities include positron emission tomography (PET), single-photon emission computed tomography (SPECT), fluorescence imaging, and dynamic magnetic resonance imaging such as functional magnetic resonance imaging (fMRI) and magnetic resonance spectroscopy (MRS). Fundamentally, all these imaging techniques deal with reconstructing a three-dimensional image from a series of two-dimensional images (projections) taken at various angles around the body. In CT, the X-ray attenuation coefficient within the body is reconstructed, while in PET and SPECT our interest is in reconstructing the time-varying distribution of a labeled compound in the body in absolute units of radioactivity concentration.

Despite the differences between the actual physical measurements among different imaging modalities, the goal of acquiring the images in clinical environment is virtually the sameā€”to extract the patient-specific clinical information

Figure 3.1: The steps and the ultimate goal of medical image analysis in a clinical environment.

and their diagnostic features embedded within the multidimensional image data that can guide and monitor interventions after the disease has been detected and localized, and ultimately leading to knowledge for clinical diagnosis, staging, and treatment of disease. These processes can be represented diagrammatically as a pyramid, as illustrated in Fig. 3.1. Starting from the bottom level of the pyramid is the medical image data obtained from a specific imaging modality, the ultimate goal (the top level of the pyramid) is to make use of the extracted information to form a set of clinical knowledge that can lead to clinical diagnosis and treatment of a specific disease. Now the question is how to reach the goal. It is obvious that the goal of the imaging study is very clear, but the solution is not. At each level of the pyramid, specific techniques are required to process the data, extract the information, label, and represent the information in a high level of abstraction for knowledge mining or to form clinical knowledge from which medical diagnosis and decision can be made. Huge amounts of multidimensional datasets, ranging from a few megabytes to several gigabytes, remain a formidable barrier to our capability in manipulating, visualizing, understanding, and analyzing the data. Effective management, processing, visualization, and analysis of these datasets cannot be accomplished without high-performance computing infrastructure composed of high-speed processors, storage, network, image display unit, as well as software programs. Recent advances in computing technology such as development of application-specific parallel processing architecture and dedicated image processing hardware have partially resolved most of the limiting factors. Yet, extraction of useful information and features from the multidimensional data is still a formidable task that requires specialized and sophisticated techniques. Development and implementation of these techniques requires thorough understanding of the underlying problems and knowledge about the acquired data, for instance, the nature of the data, the goal of the study, and the scientific or medical interest, etc. Different assumptions about the data and different goals ofthe study will lead to the use of different methodologies. Therefore, continuing advances in exploitation and development of new conceptual approaches for effective extraction of all information and features contained in different types of multidimensional images are of increasingly importance in this regard.

Image segmentation plays a crucial role in extraction of useful information and attributes from images for all medical imaging applications. It is one of the important steps leading to image understanding, analysis, and interpretation. The principal goal of image segmentation is to partition an image into regions (or classes) that are homogeneous with respect to one or more characteristics or features under certain criteria [1]. Each ofthe regions can be separately processed for information extraction. The most obvious application of segmentation in medical imaging is anatomical localization, or in a generic term, region of interest delineation whose main aim is to outline anatomic structures and (pathologic) regions that are "of interest." Segmentation can be accomplished by identifying all pixels or voxels that belong to the same structure/region or based on some other attributes associated with each pixel or voxel. Image segmentation is not only important for feature extraction and visualization but also for image measurements and compression. It has found widespread applications in medical science, for example, localization of tumors and microcalcifications, delineation of blood cells, surgical planning, atlas matching, image registration, tissue classification, and tumor volume estimation [2-7], to name just a few.

Owing to issues such as poor spatial resolution, ill-defined boundaries, measurement noise, variability of object shapes, and some other acquisition artifacts in the acquired data, image segmentation still remains a difficult task. Segmentation of data obtained with functional imaging modalities is far more difficult than that of anatomical/structural imaging modalities, mainly because of the increased data dimensionality and the physical limitations of the imaging techniques. Notwithstanding these issues, there have been some significant progresses in this area, partly due to continuing advances in instrumentation and computer technology. It is in this context that an overview of the technical aspects and methodologies of image segmentation will be presented. As image segmentation is a broad field and because the goal of segmentation varies according to the aim of the study and the type of the image data, it is impossible to develop only one standard method that suits all imaging applications. This chapter focuses on the segmentation of data obtained from functional imaging modalities such as PET, SPECT, and fMRI. In particular, segmentation based on cluster analysis, which has great potential in classification of functional imaging data, will be discussed in great detail. Techniques for segmentation of data obtained with structural imaging modalities have been covered in depth by other chapters of this book, and therefore, they will only be described briefly in this chapter for the purpose of completeness.

0 0

Post a comment