Purpose: Our previous study showed the feasibility of a rule-based automatic breast density estimation method. However, the rule-based technique could not classify the very fatty and very dense breasts consistently with high accuracy because some of these breasts have very similar gray level histograms. This study develops a new neural network (NN) classifier to improve the performance of rule-based breast density estimation.
Methods and Materials: A mammogram is digitized and the pixel size is reduced to 0.8 mm. The breast region is first segmented by an automatic boundary tracking and a pectoral muscle trimming algorithm. An adaptive dynamic range reduction technique is used to reduce the range of the gray levels in the low frequency background and to enhance the separation of the gray levels of the dense and fatty regions. The breast images are first classified by a rule-based method into a class of median dense and a class of combined very dense/fatty breasts based on the characteristics of their gray level histograms. An Expectation-Maximization (EM) algorithm is then applied to the latter class to extract the gray level features. One morphological feature and 12 EM extracted gray level features are input to a feedforward neural network to further classify the mammograms in the combined class into a class of very dense breasts and a class of very fatty breasts. For each class, a gray level threshold is automatically estimated to segment the dense tissue. For comparison, an experienced radiologist provided a manually segmented percent dense area by interactive thresholding.
Results: In this preliminary study, 498 mammograms from 141 patients were used and 243 were classified into the very dense/fatty combined class by the rule-based classifier. With a jackknife method, this class was randomly partitioned into four non-overlapping groups. In each jackknife cycle, three groups were used for training and one group for testing. The overall accuracy for classification of the four test groups into very dense and very fatty breasts reached 99.6% by the neural network, and 84.8% could be reached by our previous rule-based classifier.
Conclusion: The results demonstrate the feasibility of training a NN classifier for the classification of very dense and fatty breasts. The neural network can be trained very well using the morphological feature and the features extracted by the EM algorithm. Combining the rule-based method with the NN classifier, the two-stage classification improved the performance of our previous breast density estimation technique.