Deformable Model-Based Medical Image Segmentation

Abstract

Image data is of immense practical importance in medical informatics. Automated image segmentation, which aims at automated extraction of region boundary features, plays a fundamental role in understanding image content for searching and mining in medical image archives. A challenging problem is to segment regions with boundary insufficiencies, i.e., missing edges and/or lack of texture contrast between regions of interest (ROIs) and background. To address this problem, several segmentation approaches have been proposed in the literature, with many of them providing rather promising results. In this chapter, we focus on a specific category of image segmentation methods widely used in medical vision, namely the deformable models. We first review two general classes of deformable models, i.e., (1) the parametric deformable models, or active contours, and (2) the geometric or implicit models. Then we describe the feature extraction, i.e., the estimation of image features based on which the segmentation is performed. We show the most common approaches of how the image data is transformed into compact (higher level) numerical representations, which are integrated into the deformable models to play the image-based driving factor for the segmentation. Since these features can be used in a deterministic or probabilistic manner, we describe the basic principles of transforming these numerical representations into probabilities, using two known probabilistic graphical models, namely the Markov Random Fields (MRFs) and the Conditional Random Fields (CRFs). Then we show how these machine learning methods are integrated with deformable models to yield robust segmentation results. To illustrate the latter case, we describe a geometric model, which is integrated with a CRF: the deformable model is driven by probability fields estimated from the images (features), rather than being driven by the image features directly, with the main advantage being the increased robustness in cases of feature ambiguities, i.e., noise. We show different examples of medical data deformable model-based segmentation, we draw general conclusions from the methods described in this chapter, and we give future directions for solving challenging and open problems in medical image segmentation.