Breast cancer detection from FNA using SVM with different parameter tuning systems and SOM-RBF classifier

In this paper, we consider the benefits of applying support vector machines (SVMs), radial basis function (RBF) networks, and self-organizing maps (SOMs) for breast cancer detection. The Wisconsin diagnosis breast cancer (WDBC) dataset is used in the classification experiments; the dataset was generated from fine needle aspiration (FNA) samples through image processing. The 1-norm C-SVM (L1-SVM), 2-norm C-SVM (L2-SVM), and υ-SVM classifiers are applied, for which the grid search based on span error estimate (GSSEE), gradient descent based on validation error estimate (GDVEE), and gradient descent based on span error estimate (GDSEE) are developed to improve the detection accuracy. The gradient descent (GD) tuning method based on the span error estimate (SEE) is employed for the L2-SVM classifier because of its reachable smooth nonlinearity. Such a GDSEE tuning system also has the advantage of saving available samples for the training procedure. The SOM-RBF classifier is developed to improve the performance of only the SOM learning procedure based on distance comparison, in which the RBF network is employed to process the clustering result obtained by the SOM. Experimental results demonstrate that SVM classifiers with the proposed automatic parameter tuning systems and the SOM-RBF classifier can be efficient tools for breast cancer detection, with the detection accuracy up to 98%.