Segmentation of mammograms using multiple linked self-organizing neural networks.

A possible first stage in the analysis of the mammographic scene is its segmentation into four major components: background (the nonbreast area), pectoral muscle, fibroglandular region (parenchyma), and adipose region. An algorithm has been developed for this task. It is based on the classification of a feature vector constructed from statistical measures of texture calculated at two window sizes. Separate self-organizing neural networks are trained on sample data taken from each of the four regions. The feature vectors from the entire mammogram are then classified with the trained networks linked via a decision logic. To overcome the variability of texture between mammograms the algorithm uses data from a mammogram to classify itself in a staged approach consisting of several binary decisions. The training regions for each successive stage are determined from geometric information produced by the previous stages. The dataset in the study consisted of thirty (fifteen pairs) digitized normal mammograms of variable radiographic appearance. As a measure of performance, the outlines of the parenchyma were compared to those drawn by a radiologist experienced in reading mammograms. Comparison of the areas and perimeters generated by the human and computer observers gives a relationship with correlation coefficients of 0.74 and 0.59 for each measure, respectively. The overlapping areas of the parenchymas segmented by the observers normalized by the combined area was also calculated for each case. The mean and standard deviation of this measure was 0.69 +/- 0.12.