Computer diagnosis of mammographic masses

The objective of this work is to provide a probability of malignancy of a mammographic mass to the interpreting physician. Using the location of a mass, it is automatically segmented using fuzzy clustering. Features are extracted from the segmentation results using morphological, first-order statistical, and texture measures. Selection of relevant features is done using sequential selection. Fitness functions are based on the scatter matrices, k-nearest neighbors classifier, or neural network classifier using two-fold cross validation. The diagnosis is then provided by a trained three layer neural network. Feature selection provides a dramatic reduction in the number of required measurements to less than 25 as well as improve the accuracy of the results, from about 70% correct to 82% correct. The area under the ROC curve also increased dramatically. Computer vision applied to mammographic masses results in a very complex data space, that requires careful analysis for the design of a classifier. While further improvements are needed, current results are becoming clinically interesting.