Classification of Breast Density in Digital Mammograms

In this paper we investigate a new approach to the classification of mammo graphic images according to breast type based on the underlying texture contained within the breast tissue. Three methods for quantifying the texture are considered and used as input in the evaluation of four different classifiers. In this study we examine two classification tasks, a three-class classification problem between dense, glandular and fatty breast types and a two-class problem, differentiating between dense and fatty breast types. We use Receiver Operating Characteristic (ROC) analysis to evaluate the performance of the two-class problem. The data set used in this study is the Mammographic Image Analysis Society (MIAS) MINIMIAS database containing Medio-Lateral Oblique (MLO) views for each breast for 161 patients. For the three-class problem using a 3-layer feed-forward artificial neural network trained with conjugate gradient descent and 10-fold cross validation, we obtain a recognition rate on test of 70.4%. For the two-class problem test using a k-nearest neighbour classifier and 10-fold cross validation we obtain the area under the ROC curve Az equal to 0.832. This study demonstrates a high sensitivity in the classification of breast types justifying the use of this prior knowledge for the detection of lesions in a proposed CAD system.