Systematic Segmentation of Mammograms

Mammograms are at present the method of choice for screening the asymptomatic female population above fifty years for breast cancer. One way of assisting radiologists to cope with the increased workload from screening is to ``pre-read'''' or process mammograms ``intelligently'''' by computer. This thesis addresses the fundamental question of segmenting mammograms as the first stage of this processing. It presents systematic methods for : 1. obtaining the skin-air interface, separating breast from background; 2. locating the nipple; 3. delineating the pectoral muscle; and 4. identifying the adipose and fibroglandular tissue on mammograms. These steps represent a logical progression in the segmentation process. In each case, the methods developed recommend themselves on the basis of their simplicity, accuracy, generality or novelty. The separation of breast from background was accomplished by modelling the entire background, and portions of the breast contiguous with it, by a polynomial. The modelled image was subtracted from the original and further processed to yield a binary image of the breast and the background. An original, heuristic technique was developed to locate the nipple automatically on mammograms for which the skin-air interface had already been obtained. It was observed that at the nipple, there is a distinct change in the average gradient of the intensity in a direction normal to the skin-air interface, directed towards the breast. The value of the gradient and its orientation, along with the derivative of each with the vertical co-ordinate, yielded four descriptors that were sufficiently sensitive to locate the nipple, both when it was in profile and when it was not. Edge detection was considered most suitable for delineating the pectoral muscle. A careful analysis of mask-based edge detection revealed it to be the composition of four mathematical operations: conditioning, feature extraction, blending and scaling. Interesting insights were obtained about the feature extraction and blending steps. It was found that statistical measures such as the range and standard deviation of pixel values within a mask could serve as components of the edge feature vector just as well as conventional directed digital gradients. The function used to blend the components of the feature vectors could be defined analytically or computationally. This was particularly useful for mammograms where a single anatomical edge, like that of the pectoral muscle, could vary in strength across its length. By establishing a framework within which a variety of edge images could be systematically generated, this work has relevance beyond the application domain for which it was developed. Texture analysis was deemed appropriate to distinguish the fatty and fibroglandular components of the region within the breast. To this end, a new texture measure, the two-dimensional Hurst operator, was defined. It is a logical mathematical extension of a statistic