Support vector machines for diagnosis of breast tumors on US images.

RATIONALE AND OBJECTIVES Breast cancer has become the leading cause of cancer deaths among women in developed countries. To decrease the related mortality, disease must be treated as early as possible, but it is hard to detect and diagnose tumors at an early stage. A well-designed computer-aided diagnostic system can help physicians avoid misdiagnosis and avoid unnecessary biopsy without missing cancers. In this study, the authors tested one such system to determine its effectiveness. MATERIALS AND METHODS Many computer-aided diagnostic systems for ultrasonography are based on the neural network model and classify breast tumors according to texture features. The authors tested a refinement of this model, an advanced support vector machine (SVM), in 250 cases of pathologically proved breast tumors (140 benign and 110 malignant), and compared its performance with that of a multilayer propagation neural network. RESULTS The accuracy of the SVM for classifying malignancies was 85.6% (214 of 250); the sensitivity, 95.45% (105 of 110); the specificity, 77.86% (109 of 140); the positive predictive value, 77.21% (105 of 136); and the negative predictive value, 95.61% (109 of 114). CONCLUSION The SVM proved helpful in the imaging diagnosis of breast cancer. The classification ability of the SVM is nearly equal to that of the neural network model, and the SVM has a much shorter training time (1 vs 189 seconds). Given the increasing size and complexity of data sets, the SVM is therefore preferable for computer-aided diagnosis.