Feature and Classifier Selection for Automatic Classification of Lesions in Dynamic Contrast-Enhanced MRI of the Breast

Abstract

The clinical interpretation of breast MRI remains largely subjective, and the reported findings qualitative. Although the sensitivity of the method for detecting breast cancer is high, its specificity is poor. Computerised interpretation offers the possibility of improving specificity through objective quantitative measurement. This paper reviews the plethora of such features that have been proposed and presents a preliminary study of the most discriminatory features for dynamic contrast-enhanced MRI of the breast. In particular the results of a feature/classifier selection experiment are presented based on 20 lesions (10 malignant and 10 benign) from 20 routine clinical breast MRI examinations. Each lesion was segmented manually by a clinical radiographer and its diagnostic status confirmed by cytopathology or histopathology. The results show that textural and kinetic, rather than morphometric, features are the most important for lesion classification. They also show that the SVM classifier with sigmoid kernel performs better than other well-known classifiers: Fisher's linear discriminant function, Bayes linear classifier, logistic regression, and SVM with other kernels (distance, exponential, and radial).