Abstract

Reduction of False Positive signals (FPR) is a fundamental, yet awkward, step in computer aided mass detection schemes. This paper describes preliminary results of a filtering approach to FPR based on Support Vector Regression (SVR), a machine learning algorithm arising from a well-founded theoretical framework, the Statistical Learning Theory, which has recently proved to be superior to the conventional Neural Network framework for both classification and regression tasks: indeed, the proposed filtering method belongs to the family of neural filters. The SVR filter is forced to associate subregions extracted from input images, masses and non-masses, to continuous output values ranging from 0 to 1 representing a measure of the presence in the subregion of a mass. A weighted sum of outputs over each image is used to accomplish the FPR task. In the test phase, this approach reached promising results, retaining 87% of masses while reducing False Positives to 62%.