A Fast Breast Segmentation Algorithm with Pectoral Muscle Suppression

The algorithm is explained in six points:

• Construction of the intensities histogram. The histogram of a complete mammographic image has the behaviour shown in Figure :
  • In the left (lower intensities values) there is a large peak corresponding to the background pixels.
  • In the middle (grey values) there are the pixels corresponding to the breast itself.
  • In the right (brightness pixels) there is another peak corresponding to the pectoral muscle and annotations.

  Figure A.1: Typical histogram of a mammogram. Clearly, there are three different zones: background in lowest intensities, breast tissue in medium intensities, and annotations and pectoral muscle in the highest intensities.

  

• A threshold is used to extract the image from the background. The value of this threshold is determined using the minimum value between the first two most important peaks, which are the peaks of the background and the breast tissue.
• A Connected Component Labeling algorithm [38] is used in order to recover the largest region, which will be both the breast and pectoral muscle. In general, this algorithm is useful to find not connected objects in images.
• For segmenting the breast from the pectoral muscle a new histogram of this biggest region is used. This histogram contains two zones: the pectoral muscle and the breast tissue.
• A region growing algorithm is used to extract the pectoral muscle region from the breast. The seed of this region growing is placed inside the pectoral with value between the brightness maximum and the minimum between the two zones of the histogram. An automatic control is used in order to adjust the intensity condition of the region growing that permits to identify a pixel belonging to the region or not.
• The last step is the use of morphological operations in order to smooth the boundary of the breast.

Figure A.2: Sequence of the breast profile segmentation. (a) is the original image, while (b) is the result of thresholding the image. In (c) the CCL algorithm has been applied in order to detect the biggest region, and finally (d) is the segmented image without background and pectoral muscle.

(a)	(b)
(c)	(d)

Figure  shows a typical mammogram segmented using the above described approach. Its histogram is shown in Figure , and a threshold between the two first major peaks is automatically selected in order to binarize the image, obtaining the Figure (b). The result of applying this threshold is a collection of different regions, being the biggest the union of the breast and the pectoral muscle. This biggest region can be extracted using a CCL algorithm (Figure (c)). In the last image, the breast has been extracted from the pectoral muscle using the region growing algorithm above described.

As is shown in Figure , this segmentation results in a minor loss of skin-line pixels in the breast area, but those pixels are deemed not to be relevant for mass segmentation or breast density estimation, as the lost grey-levels are darker than the rest of the pixels of the breast.

Figure A.3: Three different examples of the breast profile segmentation using the fast segmentatio algorithm.