Feature selection for computer-aided diagnosis of breast cancer using dynamic contrast-enhanced magnetic resonance images

The ability to accurately identify abnormal tissue is a fundamental issue in all cancer diagnoses. Dynamic contrast-enhanced magnetic resonance imaging (DCEMRI) has shown promise as an adjunct to mammography for breast cancer detection and can be assisted by computer-aided diagnosis systems.;The final features selected enabled a receiver operating characteristic (ROC) curve with an area under the curve of 0.75238 using the maximum correlation coefficient nearest neighbour classifier. These results outperformed the results of the signal enhancement ratio having an area under the ROC curve of 0.70595, plus a support vector machine having an area under the ROC curve of 0.67341.;Here the fast orthogonal search (FOS) algorithm is used to select features from DCE-MRI breast images without having a priori knowledge of which features are best suited to improve the classification of breast lesions. The initial feature set was extracted from sets of five DCE-MRI images, including one precontrast image and four postcontrast images from 63 benign and 20 cancerous lesions. A leave-one-out trial was conducted using FOS to select features which were then used to create a set of cross product features. The cross product feature set was created by taking the second order cross products of the features selected from the initial feature set. FOS was then reapplied using the combined feature set. The features which best classified the exams as either cancerous or benign were selected.