Finding Regions with Similar Tissue

From observing mammographic images one can conclude that pixels from a similar tissue have similar grey-level values, as can be seen in Figure . Hence, as our aim is to cluster those pixels into meaningful regions, the Fuzzy C-Means algorithm (see Section ) is used to group them into two separate categories: fatty tissue and dense tissue. Beforehand, and with the aim to avoid effects from microtexture that could appear in some regions, the breast region is smoothed by using a median filter of size $5 \times 5$ . From our experiments, this filter size is a good compromise between noise reduction and texture preservation of mammographic tissue.

When using partitional clustering algorithms, like Fuzzy C-Means, the placement of the initial seed points is one of the central issues in the variation of segmentation results [78]. Despite their importance, usually seeds for these algorithms are randomly initialized. As we only consider two classes in our approach, the Fuzzy C-Means is initialized using histogram information, with the aim to obtain representative instances of both classes. Hence, we initialized the two seeds with the grey-level values that represent $15\%$ and $85\%$ of the accumulative histogram of the breast pixels of each mammogram (representing fatty and dense tissue, respectively). Although these values were empirically determined, the obtained segmentations do not critically depend on them. Moreover, some mammograms do not have clearly determined dense and fatty components. In these cases, the segmentation result is one cluster grouping the breast tissue and the other cluster grouping regions with less compressed tissue (an elongated region, like a ribbon, following the skin-line). In these cases, the breast texture information is in the breast tissue cluster, while the ribbon does not provide significant information to the system.