In order to develop a method for classifying masses in digitized screening mammograms as benign or malignant, 260 image texture features were measured on 43 images of known malignant masses and 28 images of known benign masses. A genetic algorithm was used to select the optimal subset of 4 features based on Az scores. The Az score for the optimal 4 features was 0.8879. Since feature space reduction can result in optimistic estimates of classifier performance, the significance of this score was estimated by computing the empirical distribution of Az scores in the context of the experimental parameters. Results indicate that probability of obtaining the observed performance by chance alone is 0.0080.
展开▼
