A performance study of different soft-computing procedures for automatic detection of breast cancer malignancy

Breast cancer is one of the most common cancers in women, and its early detection increase the patient's rate of survival. Several studies have shown that computer assisted diagnosis can improve detection of breast cancer, or at least detect those suspicious cases that are worth studying in detail by an expert.

In this paper we analyze the performance of different neural classifiers that are designed for identifying malignant microcalcifications on mammograms. Specifically, we have used different techniques, such as Multi Layer Perceptrons, Radial Basis Functions and Support Vector Machines, comparing the capabilities of the resulting systems with other approaches that can be found in the literature.

We have used information collected from mammograms of 210 patients in which 229 clustered microcalcifications were detected. This information has been automatically extracted, and is related to characteristics of the cluster and to the individual microcalcifications. Additionally, biopsy results from each patient determined that 46% of the cases were malignant and 54% were benign tumors.

Analysis of performance is based on receiver operating characteristic (ROC) curves, showing that these general purpose classifiers achieve results that are similar to those of previous automatic designs. Hopefully, ad hoc neural designs could improve these results even further. Finally, a comparison with the opinion of three human experts shows that these technologies can be of great help to assist doctors in the clinical decision process.