Automated Analysis of Mammography Phantom Images

The present work stems from the hypothesis that humans are inconsistent when making subjective analyses of images and that human decisions for moderately complex images may be performed by a computer with complete objectivity, once a human acceptance level has been established. The following goals were established to test the hypothesis: (1) investigate observer variability within the standard mammographic phantom evaluation process; (2) evaluate options for high-resolution image digitization and utilize the most appropriate technology for standard mammographic phantom film digitization; (3) develop a machine-based vision system for evaluating standard mammographic phantom images to eliminate effects of human variabilities; and (4) demonstrate the completed system's performance against human observers for accreditation and for manufacturing quality control of standard mammographic phantom images. The following methods and procedures were followed to achieve the goals of the research: (1) human variabilities in the American College of Radiology accreditation process were simulated by observer studies involving 30 medical physicists and these were compared to the same number of diagnostic radiologists and untrained control group of observers; (2) current digitization technologies were presented and performance test procedures were developed; three devices were tested which represented commercially available high, intermediate and low-end contrast and spatial resolution capabilities; (3) optimal image processing schemes were applied and tested which performed low, intermediate and high-level computer vision tasks; and (4) the completed system's performance was tested against human observers for accreditation and for manufacturing quality control of standard mammographic phantom images. The results from application of the procedures were as follows: (1) the simulated American College of Radiology mammography accreditation program phantom evaluation process demonstrated human observer variabilities which might affect consistent accreditation; (2) the device characterization and testing demonstrated that high-resolution film digitization necessary for these purposes may be adequately performed by a 2048 by 2048 by 12-bit cooled CCD camera; (3) Fourier domain template matching, combined with derivative filters, provided salient localization of the test objects in the standard mammographic phantom; and (4) the high level classification decision was adequately modeled by a Bayesian classifier using threshold contrast as measured from the target observer group. (Abstract shortened by UMI.).