Mammograms are x-ray images of the breast. At present, they are the method of choice for screening asymptomatic women for early detection of breast cancer. Such screening will necessarily generate a large number of mammograms that must be viewed and interpreted by a limited number of expert radiologists. Automatic analysis of mammograms by computer, as the first stage in analyzing mammograms, could serve to reduce the workload on radiologists.
Before mammograms can be analyzed by machine, they must be digitized with adequate gray-scale and spatial resolution. It is also vital to ensure that low-intensity features such as the skin-air interface and the nipple are preserved with fidelity during digitization. Assuming that this has been done, the first stage of preprocessing is segmentation of the image into background and object. This should yield a clear and smooth skin-air interface.
The next logical step would be to locate the nipple on the mammogram. There are several reasons why this should be done. Anatomically, the nipple is the only landmark on the breast (see p. 119 of Ref. 1). The glandular structures comprising the lobules, ductules, lobes, and ducts hierarchically converge onto the nipple. Because cancer arises in the glandular tissue of the breast (see p. 121 of Ref. 1), it would be sensible for any automatic search strategy for detecting cancer to begin at the nipple and fan out into the "cone" or "triangle" of glandular tissue that has the nipple as its apex. Moreover, given its singularity, radiologists pay specific attention to the nipple as part of their examination of a mammogram (see p. 22 of Ref. 2 and p. 123 of Ref. 3). Radiologists also compare corresponding regions of the right and left breasts to detect relative anomalies (see p. 22 of Ref. 2). Computer methods that attempt the same task rely heavily on the nipple as an alignment pivot.
This paper reports on a simple method for locating the nipple on a mammogram. It has been tested on a total of 24 images. Sixteen of these are oblique-view mammograms from a digital database made available to researchers by the Mammographic Image Analysis Society (MIAS) of the United Kingdom. The remaining eight images are craniocaudal views from another database of digitized mammograms distributed by the University of California, San Francisco (UCSF) and the Lawrence Livermore National Laboratory (LLNL) in the USA. These two databases shall henceforth be referred to by the acronyms MIAS and UCSF/LLNL, respectively.
Online access to SPIE eBooks is limited to subscribing institutions.