Abstract: This paper presents an exploratory study of the effectiveness of support vector machines used as a tool for computer-aided breast cancer diagnosis. We explore the discriminatory power of heterogeneous mammographic and sonographic descriptors in solving the classification task. Various feature selection techniques were tested to find a set of descriptors that outperforms those from similar studies. We also explored how choice of the SVM kernel function and model parameters affect its predictive abilities. The kernels explored were linear, radial basis function, polynomial, and sigmoid. The model performance was estimated by ROC analysis and metrics, such as true and false positive rates, maximum accuracy, area under the ROC curve, partial area under the ROC curve with sensitivity above 90%, and specificity at 98% sensitivity. Particular attention was paid to the latter two as lack of specificity causes unnecessary surgical biopsies. Experiments registered that an appropriate reduction of variables can greatly improve the predictive power of the model, as long as the choice of the kernel affects the model performance marginally. We also found that the SVM is superior to the common classification technique used in the field - MLP neural networks.

Keywords: data mining, support vector machines, heterogeneous data; breast cancer diagnosis, computer aided diagnosis.

ACM Classification Keywords: I.5.2- Computing Methodologies - Pattern Recognition – Design Methodology - Classifier design and evaluation.

Link:

SUPPORT VECTOR MACHINES FOR CLASSIFICATION OF MALIGNANT AND BENIGN LESIONS

Anatoli Nachev, Mairead Hogan

http://www.foibg.com/ibs_isc/ibs-26/ibs-26-p22.pdf