Introduction

In many medical imaging applications three-dimensional data sets have to be segmented. In these cases, segmentation involves categorizing voxels into object classes based on their local intensity, spatial location, neighborhood, or shape characteristics of a certain object class. The term "classification" is sometimes loosely used in lieu of "segmentation".

Segmentation plays an important role in biomedical image processing. It is often the starting point for other processes, including registration, shape analysis, motion detection, visualization, and quantitative estimations of linear distances, areas, and volumes. It is clear, however, that a single segmentation technique is not capable of yielding acceptable results for all different types of biomedical images. Quite often, methods are optimized to deal with specific medical imaging modalities such as magnetic resonance (MR) imaging and X-ray computed tomography (CT), or modeled to segment specific anatomic structures such as the brain, the liver, and the vascular system.

In this chapter, we provide an overview of segmentation methods used for quantitative volumetric analysis of human brain images. Measurement of tissue volumes is an increasingly important goal in medical imaging for studying structural changes over time and correlating anatomical information with functional activity or pathology. We describe and numerically compare several methods that seek to segment MR images of the brain into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). Comparisons are made between different methods solely for the purpose of tissue volumetrics.

However, these methods may be employed for different tasks, including diagnosis, localization of pathology, anatomy delineation, computer-aided neurosurgery, treatment planning, or correlation analysis with functional data. In addition, these methods, with appropriate modifications, can be applied to different anatomic regions, to different species, and to different imaging modalities.

Section 2 provides a general overview of the segmentation of MR brain images for quantification of tissue volumes. Next, in Section 3, we describe several image segmentation algorithms, including more recent techniques such as the adaptive Bayesian [12,18] and the adaptive fuzzy c-means [34] algorithms. In Section 4, the volumetric accuracy of these segmentation methods is tested. Next, in Section 5; the concept of transforming the segmentation problem into a registration task is introduced for volumetric quantification of specific brain structures. As an example, we describe a recently proposed method based on stereotaxic normalization that can be used for segmenting specific subcortical nuclei such as the hippocampus, lenticular nucleus, and caudate nucleus [12]. Finally, in the concluding remarks we highlight the trends and research opportunities in the medical image segmentation field.

0 0

Post a comment