Algorithms for Segmenting Small Low-Contrast Objects in Images

Abstract

One of the signs of early breast cancer is the presence of microcalcification clusters in mammograms. Automated diagnostic aids for detection of breast cancer in mammograms must determine the presence of microcalcification clusters with high enough accuracy to serve as a clinical tool. To do so, the automated system has to process the digitized mammogram with at least three different types of algorithms. The first is segmentation of individual small candidate structures that are separated from the background tissue. The second is extraction of quantitative features from each candidate structure, to represent size, shape, texture, contrast, etc. The third is classification of candidate structures as microcalcification or other tissue, using the feature values, based on automated training against ground truth provided by clinical experts. An optional fourth type of algorithm, enhancement, may also be used as a preprocessing step, depending on the image quality, to reduce the noise level of the image or to increase the contrast of candidate objects before segmentation is applied. In this multistage process, the role of segmentation is essential because the delineation of each candidate object dictates the extracted features, and subsequently their classification. This chapter focuses on segmentation algorithms that are especially suited for segmenting small, low-contrast, arbitrarily shaped objects such as microcalcifications in mammograms. Segmentation algorithms used to separate objects of interest from the image background are primarily based on three different approaches. Some algorithms determine the pixels of the object by comparing their brightness values to either a global or local threshold. The second type of algorithm first determines the edge pixels of the object using a gradient operator, then an edge-linking technique forms a closed contour and labels all pixels within the edge contour. The third approach is to grow a region of contiguous pixels starting from a seed pixel selected inside the object. The segmentation of microcalcifications in mammograms is especially demanding because each microcalcification is a very small, low-contrast object embedded in a nonuniform background that may have significant spatial high-frequency components. An edge detector applied to a mammogram finds edges throughout the image and may also overlook the edges of some microcalcifications due to their low contrast. Segmentation of small, low-contrast objects such as microcalcifications cannot effectively be based on conventional edge detection, especially because linking edge pixels obtained on a mammogram is not practical. Most algorithms reported for segmentation of microcalcifications are based either on thresholding or region growing, with or without enhancement. The segmentation algorithm suggested in Refs. 1 and 2 compares each pixel to a local threshold computed within a square neighborhood around the pixel. The size of the neighborhood is selected by the user, and the threshold is set as the mean value of pixels in the neighborhood plus their rms value multiplied by a selected coefficient. This threshold represents the highest value that the background pixels in that neighborhood are expected to have. Pixels that have values greater than the threshold are connected to form the segmented region.