Computer-Aided Classification of Malignant and Benign Lesions on Mammograms.
Berkman Sahiner · Defense Technical Information Center (DTIC) · 1997
Computerized classification methods were developed for characterization of mammographic lesions. For mass characterization, features related to the degree of spiculation were developed. These features were combined with morphological features related to the computer-segmented mass shape for classification using stepwise feature selection and linear discriminant analysis. For different views of the same mass, the malignancy score provided by the classifier were combined by averaging. The classification accuracy was measured using the area Az under the receiver operating characteristics (ROC) curve. The trained classifier achieved a test Az value of 0.87 on an independent data set of 45 masses. For microcalcification characterization, morphological features were extracted from computer-identified leisons. Morphological and texture features were combined using stepwise feature selection and linear discriminant analysis. The classifier was tested using the leave-one-case- out method. On a data set of 112 pairs of mammograms, the test Az value of the computer was 0.83. In an ROC study, 7 experienced breast radiologists read the same 112 pairs of mammograms. The area Az under the average ROC curve for radiologists was 0.71. The Az value of the computer was higher than that of all radiologists, and the difference was statistically significant for three of the radiologists (p=O.03).