The amount of data that needs to be examined by a radiologist in DCE-MRI to locate suspicious lesions is huge (e.g. 512x512 voxels x 50 slices x 5 volumes). The automatic segmentation of suspicious lesions in DCE-MRI of the breast is thus attractive, because it may dramatically reduce the amount of data that needs to be examined and draws the radiologist�s attention to volumes that have a high probability of containing suspicious lesions. However, while, segmenting the image into the relevant objects and background parts is a crucial step, �it is also, in many cases, one of the more difficult tasks� (Rodenacker and Bengtsson, 2003).
This chapter introduces the challenge of automatic segmentation. Also, it presents a method for the automatic segmentation of enhancing breast tissue. The method is based on seeded region growing and merging. The criteria for growing and merging are based on both the original image intensity values and the fitted parameters of an empiric parametric model of contrast enhancement. The results for the application of the method are also presented on 24 DCE-MRI breast data sets originating from routine clinical breast MRI examinations. The data includes 10 cases of benign enhancement and 14 cases of malignant enhancement (the latter confirmed by histopathology). The results show that the segmentation method has 100% sensitivity for the detection of suspicious regions independently identified by a radiologist. The results suggest that the method has the potential both as a tool to assist the clinician with the task of locating suspicious tissue and as a means for generating quantitative features for the automatic classification of suspicious regions. The core material of this chapter was presented at �Digital Image Computing: Techniques and Applications�, 2007, Adelaide, Australia (Gal et al., 2007a).
�Image segmentation is the process of separating objects from background� (Snyder and Qi, 2004). The segmentation of an image is the partitioning of an image into a set of connected regions, where each region is homogeneous in some sense (e.g. intensity or texture) and is identified by a unique label (Snyder and Qi, 2004). The basic assumption is that the object in the image differs from the background (i.e. everything that is not part of the object) in some properties (e.g. shape, intensity, texture). The result of a segmentation method is usually a list of equivalence classes where each class represents an object or the background.
Classification of objects (e.g. lesions) in the spatial domain is commonly based on the segmentation and different properties of the image, such as morphometric (i.e. shape, size), radiometric (i.e. gray level, histogram) and textural properties. The first step in object classification is usually the segmentation of the object of interest in the image. Robust segmentation is difficult to achieve; thus, the classification process is often expected to overcome the noise and bias that may be introduced by the segmentation step.
Manual segmentation is subjective, and given the vast quantity of data to be analysed in a DCE-MRI data set, the possibility exists that diagnostically-significant regions of enhancement may be overlooked. However, automatic segmentation is challenging, because the temporal and spatial distributions of the contrast agent in suspicious tissue can be highly varied, both for an individual patient and between patients.
In this section, an overview of segmentation methods for greyscale images is presented. These methods provide the basis for MR images segmentation techniques. In general, segmentation techniques can be roughly divided into two types of algorithms: those that find the edges and contours in the image, assuming that once a contour is found, the inner side of the contour will be the object while the outer region is the background; and those that define a criteria of membership for a region and search for connected sets of pixels that satisfy these criteria. Some of these are based on global and local thresholding, which usually involves the analysis of the image histogram