Abstract
The main aim of the breast screening programme (BSP) is to provide early detection of breast lesions permitting efficient treatment. At present the BSP is technically run using conventional screen film (S/F) X-ray mammography, which suffers from a number of inherent limitations. Most of these limitations could be overcome by the use of digital mammography technology. Such a technology has to, however, show image performance that is as good as or better than that for S/F systems. The performance of a digital mammography system can be quantified and compared in a number of different ways. These include contrast detail and contrast-to-noise ratio (CNR) measurements. In practice, various limitations and problems have, however, become apparent in applying these measurements. For the CNR measurements the main problems are the presence of the uncorrected heel effect and the effect of the size of region of interest (ROI). The contrast detail analysis, using human observers, is time consuming and suffers from the presence of significant inter-observer error. These can be solved by using an "automatic observer". One limitation of using the automated approach is that the relationship between automated and human observer scoring was not fully explored across the wide variety of systems and circumstances encountered in practice. An alternative approach would be to predict contrast detail response from measurements of modulation transfer function (MTF) and detection quantum efficiency (DQE) using a model of the imaging process. In this thesis, the procedure of measuring CNR was revised and the use of an automatic approach for contrast detail measurements was further examined, using different modalities of digital mammography. The contrast detail performance was then analysed across a range of doses for a wide range of clinically used digital mammography systems, using the human results predicted from the automated measurements. MTF and DQE were also measured for the detectors used in these systems. A simple signal-matched noise-integration model was then adopted to theoretically predict the contrast detail response of these systems. The most remarkable findings of this thesis are as follows: the use of multiple small ROIs led to CNR results that were essentially the same as if a heel effect correction had been applied; the automated measurements can be used to predict the threshold contrast for a typical observer; an encouragingly good level of agreement was found between the experimental contrast detail data and theoretical predictions. Finally, image performance of promising hybrid pixel semiconductor detectors, not commercially available, was also evaluated with the aid of Monte Carlo simulation, for application to digital mammography.