Comparison of rule-based and artificial neural network approaches for improving the automated detection of clustered microcalcifications in mammograms

Forty-six thousnad women die each year in the US from breast cancer. Mammography is the best method of detecting breast cancer and has been shown to reduce breast cancer mortality in randomized controlled studies. Clustered microcalcifications are often the first sign of breast cancer in a mammogram. The use of a second reader may improve the sensitivity of detecting clustered microcalcifications. Our laboratory has developed a computerized scheme for the detection of clustered microcalcifications that is undergoing clinical evalution. This paper concerns the feature analysis stage of the computerized scheme, which is designed to remove false-computer detections. We have examined three methods of feature analysis: rule-based (the method currently used in the clinical system), an artificial neural network (ANN), and a combined method. To compare the three methods, the false-positive (FP) rate at a sensitivity of 85% was measured on two separate databases. The average number of FPs per image were: 0.54 for rule-based, 0.44 for ANN, and 0.31 for the combined method. The combined method had the highest performance and will be incorporated into the clinical system.