THE HOTELLING TRACE CRITERION USED FOR SYSTEM OPTIMIZATION AND FEATURE ENHANCEMENT IN NUCLEAR MEDICINE (PATTERN RECOGNITION).
AuthorFIETE, ROBERT DEAN.
AdvisorLaursen, Emmette M.
MetadataShow full item record
PublisherThe University of Arizona.
RightsCopyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction or presentation (such as public display or performance) of protected items is prohibited except with permission of the author.
AbstractThe Hotelling trace criterion (HTC) is a measure of class separability used in pattern recognition to find a set of linear features that optimally separate two classes of objects. In this dissertation we use the HTC not as a figure of merit for features, but as a figure of merit for characterizing imaging systems and designing filters for feature enhancement in nuclear medicine. If the HTC is to be used to optimize systems, then it must correlate with human observer performance. In our first study, a set of images, created by overlapping ellipses, was used to simulate images of livers. Two classes were created, livers with and without tumors, with noise and blur added to each image to simulate nine different imaging systems. Using the ROC parameter dₐ as our measure, we found that the HTC has a correlation of 0.988 with the ability of humans to separate these two classes of objects. A second study was performed to demonstrate the use of the HTC for system optimization in a realistic task. For this study we used a mathematical model of normal and diseased livers and of the imaging system to generate a realistic set of liver images from nuclear medicine. A method of adaptive, nonlinear filtering which enhances the features that separate two sets of images has also been developed. The method uses the HTC to find the optimal linear feature operator for the Fourier moduli of the images, and uses this operator as a filter so that the features that separate the two classes of objects are enhanced. We demonstrate the use of this filtering method to enhance texture features in simulated liver images from nuclear medicine, after using a training set of images to obtain the filter. We also demonstrate how this method of filtering can be used to reconstruct an object from a single photon-starved image of it, when the object contains a repetitive feature. When power spectrums for real liver scans from nuclear medicine are calculated, we find that the three classifications that a physician uses, normal, patchy, and focal, can be described by the fractal dimension of the texture in the liver. This fractal dimension can be calculated even for images that suffer from much noise and blur. Given a simulated image of a liver that has been blurred and imaged with only 5000 photons, a texture with the same fractal dimension as the liver can be reconstructed.
Degree ProgramOptical Sciences