Automatic target recognition application in mammography and synthetic aperture radar images

Automatic Target Recognition (ATR) is an important issue in many military and non-military applications. ATR generally refers to the use of computer processing to detect and recognize target signatures in sensor data. One of the applications of ATR that has drawn a lot of attention recently is the computer-aided diagnosis (CADx) system of breast cancer. This dissertation describes a new CADx system used to analyze regions of interest (ROIs), specifically masses, in digitized mammograms. The algorithm and architecture is based on a framework of dividing the ROI into different ellipsoidal regions after extracting the mass automatically. This approach has never been applied to breast cancer diagnosis. Ellipsoidal ring features and size, shape, contrast, and texture features were extracted. Statistical-based feature saliency techniques were used to determine the best features for discrimination. The regions were then classified using one of four methods. These classifiers are a multilayer perception neural network, a Mahalanobis distance classifier, a quadratic classifier, and a Fisher discriminant linear classifier. The mass detection system performed well in identifying spiculated masses in digitized mammograms. Using the resubstitution method, the CADx system correctly classified 98.28 percent of the spiculated masses with true negative (TN) accuracy of 97.67 percent with a quadratic classifier. On the other hand, using the leave-half-out (LHO) method, the best performance was a true positive (TP) accuracy of 83.79 percent and a TN accuracy of 80.62 percent with a Fisher linear classifier. However, since the LHO method is very pessimistic, the actual error rate is lower. The performance on classifying malignant tumors was also very good. The quadratic classifier reached 100 percent TP accuracy, and 98.52 percent TN accuracy using the resubstitution method. Using the LHO method, the TP accuracy was 93.41 percent and the TN accuracy was 59.82 percent. These results are from the several biopsy-proven database of 245 masses obtained from two hospitals (12 bit, 43.5 micron).; This dissertation also uses Gabor filters and the weighted Mahalanobis distance clustering technique for autonomous segmentation of (1 foot by 1 foot) high-resolution polarimetric synthetic aperture radar (SAR) images. Processing involved correlation between the SAR imagery and Gabor functions. This research used even-symmetric cosine Gabor functions and operated on single polarization horizontal-horizontal magnitude data. Provided are results demonstrating combination of Gabor processing and WMD clustering provide scene segmentation. The WMD algorithm achieved consistent improvements over the competing method in both the visual quality and the “misclassification” rate.