Improvement of a Mammographic CAD System for Mass Detection

A previously developed computerized scheme to detect masses has been further revised and several improvements were intended. Mammograms were digitized at a higher resolution with a mammographic laser scanner providing 12 bits. Some steps of the scheme, based on bilateral subtraction technique, were modified. Several new features were designed and a BPN neural network was used to reduce the number of false positives. Results obtained with the training set were encouraging, yielding a sensitivity of 85% and 1.54 mean number of false positives per image before applying false positive reduction. After applying false positive reduction, a sensitivity of 78.3% at a mean number of 0.4 false positives per image was obtained. The area under the AFROC curve was A1 = 0.808.