Using evolutionary programming to configure support vector machines for the diagnosis of breast cancer

Support Vector Machines(s) (SVMs) are new machine intelligence paradigms that use the Structural Risk Minimization (SRM) concept to develop learning machines. SVMs can always be trained to provide global minima, given that the leaning machine parameters are optimally computed. The current most prevalent methods to select these parameters are numerical iterative techniques. While useful, these methods frequently have no basis in theory, and cannot guarantee that the resultant parameters will yield optimum learning machine performance. The purpose of this paper is to discuss and demonstrate the application of Evolutionary Programming (EP) concepts to develop learning machine parameters and demonstrate the effectiveness of this process. This paper will also demonstrate that the applied EP process will reduce the amount of time required to configure a learning machine for those data sets studied, while developing optimal learning parameters with minimal user intervention. Specifically, this research has demonstrated, using the Duke mammogram data set, that SVMs derived using this modified EP process improved the specificity by a significant 45.3% at 100% sensitivity (missing no cancers) as well as improving the specificity by 17.5% at 95% sensitivity (missing 5% of the cancers) when compared to the performance of SVMs whose parameters were computed using the standard iterative method. The practical consequence of these results is that many women, who currently have false positive diagnoses resulting from the application of existing methods, will no longer be required to undergo biopsy with the resultant cost, morbidity and physical disfigurement that frequently results from these procedures. This approach, in addition, may also be used in linear separable; linear, non-separable; and nonlinear, non-separable environments.

Duke Scholars

Author Joseph Yuan-Chieh Lo Radiology