DESIGN OF A HIGH-SENSITIVITY CLASSIFIER BASED ON A GENETIC ALGORITHM - APPLICATION TO COMPUTER-AIDED DIAGNOSIS

A genetic algorithm (GA) based feature selection method was developed for the design of high-sensitivity classifiers, which were tailored to yield high sensitivity with high specificity. The fitness function of the GA was based on the receiver operating characteristic (ROC) parti al area index, which is defined as the average specificity above a giv en sensitivity threshold. The designed GA evolved towards the selectio n of feature combinations which yielded high specificity in the high-s ensitivity region of the ROC curve, regardless of the performance at l ow sensitivity. This is a desirable quality of a classifier used for b reast lesion characterization, since the focus in breast lesion charac terization is to diagnose correctly as many benign lesions as possible without missing malignancies. The high-sensitivity classifier, formul ated as the Fisher's linear discriminant using GA-selected feature var iables, was employed to classify 255 biopsy-proven mammographic masses as malignant or benign. The mammograms were digitized at a pixel size of 0.1 mm x 0.1 mm, and regions of interest (ROIs) containing the bio psied masses were extracted by an experienced radiologist. A recently developed image transformation technique, referred to as the rubber-ba nd straightening transform, was applied to the ROIs. Texture features extracted from the spatial grey-level dependence and run-length statis tics matrices of the transformed ROIs were used to distinguish maligna nt and benign masses. The classification accuracy of the high-sensitiv ity classifier was compared with that of linear discriminant analysis with stepwise feature selection (LDA,a). With proper GA training, the ROC partial area of the high-sensitivity classifier above a true-posit ive fraction of 0.95 was significantly larger than that of LDA(sfs), a lthough the latter provided a higher total area (A(z)) under the ROC c urve. By setting an appropriate decision threshold, the high-sensitivi ty classifier and LDA(sfs) correctly identified 61% and 34% of the ben ign masses respectively without missing any malignant masses. Our resu lts show that the choice of the feature selection technique is importa nt in computer-aided diagnosis, and that the GA may be a useful tool f or designing classifiers for lesion characterization.