Breast segmentation in screening mammograms using multiscale analysis and self-organizing maps

Previously we presented an unsupervised self-organizing map (SOM) for segmentation of the breast region in screening mammograms. This study improves upon our earlier technique by (1) enhancing the detection of the breast region near the skin line, as well as (2) reducing the computational complexity. Contrary to the initial technique, the improved one exploits global image properties extracted at different scales. These properties were used to both generate the SOM training samples and obtain a preliminary segmentation. Subsequently, a multi-step strategy was implemented to automatically outline a wide band around the skin line for further analysis. This additional step reduces the computational complexity by focusing the analysis on the set of pixels that creates clinically the highest ambiguity. Specifically, the same (already trained) SOM was applied to classify the ambiguous pixels around the skin line. The study was performed on 400 screening mammograms from the digital database for screening mammography (DDSM). Visual examination of the segmentation results confirmed an improvement in the detection of the low-contrast region near the skin line. The performance was consistent regardless of mammographic view and/or breast density. Furthermore, the computational cost of processing can be reduced by up to 80% of the original value.