This work reports an investigation in which localization-receiver-operating-characteristics (LROC) studies employing hybrid images were used. Hybrid images are normal Ga-67 scans modified by the addition of Monte Carlo simulated lesions. After determining a target image contrast using human observers, we conduct pilot LROC studies to determine the optimal parameters for the reconstruction methods. Then we perform a human observer LROC comparison study using the optimal parameters obtained in the pilot studies to determine the relative impact on detection accuracy of the various corrections introduced. We study the effectiveness of attenuation compensation, scatter compensation, and detector resolution compensation strategies used with the RBI reconstruction method and compare it to FBP reconstruction. The relative ranking of the test strategies agreed in most cases with those of previous studies that employed simulated projections of digital anthropomorphic phantoms, thus confirming the findings of those studies. In this work, we specifically focus on the lesion-detection performance of two channelized models, the channelized non-prewhitening observer and the channelized Hotelling observer. They are used in the context of generating performance curves for parameter optimization tasks. We chose channelized models, as they are based on the phenomenon of frequency-selective-contrast-sensitivity exhibited by the human visual system and demonstrated by various psycho-physical studies. Working with these models, however, requires various amounts of knowledge regarding the lesion-free background data, which in the context of hybrid data, is either not known, or difficult to estimate. Various methods are explored that approximate this background data in a manner that is useful for the model observers. We specifically explore boot-strap methods, K-neighbor methods and the wavelets linear enhancement method to generate this approximation. Of the various methods studied, the wavelet method was found to produce the best results. We perform statistical tests to assess the significance of the performance curves obtained. Wavelets can also be used for feature-selective non-linear contrast enhancement of images. This may be especially useful for lesion-detection tasks. The image sets used in our studies are filtered on reconstruction using a 3-D Gaussian filter. This reduces reconstruction noise and enhances lesion definition by bringing in counts from neighboring planes into the plane of interest. 3-D low-pass filtering, however, can introduce considerable masking of the lesion due to blurring of neighboring anatomical structures. Non-linear enhancement using wavelets can undo this blurring effect. This is possible since wavelets employ a multi-scale decomposition scheme. Using this methodology, it is possible to enhance the features of interest while de-emphasizing others. We explore the possibility of lesion contrast enhancement using wavelet filters and conduct human observer LROC studies to assess its impact on detection accuracy. We then propose an internal noise mechanism for the model observer and study its impact using the CNPW model observer. Within the framework of model observers, an internal noise mechanism has been suggested to match the models performance to that of human performance. This is because, model observers typically out-perform humans in detection tasks, as, humans face uncertainty due to choice in the detection process which is absent in a noiseless model. A noise model can simulate this choice and in this work we implement it by chaining a probability prior, that encapsulates the models degree of uncertainty, to a Bernoulli random event generator, that implements the models choice. To achieve this, we define a data-dependent test-statistic/representative variable which is then used to sample the prior. Various methods of defining the parameters of the prior are explored and importantly, methods that use the test variable’s estimated distribution. We see that this leads to different shades of noise-models, beginning from the passive model, that merely degrades the performance curves leaving the dynamics intact, to more active ones, that modify the dynamics as well. A theoretical expression is also derived for the average amount of noise that the mechanism introduces in the observer. Finally, we show that the inclusion of a noise mechanism can yield mean performance curves with greater statistical significance provided we simulate multiple noise instances. (Abstract shortened by UMI.)
Identifer | oai:union.ndltd.org:UMASS/oai:scholarworks.umass.edu:dissertations-5281 |
Date | 01 January 2008 |
Creators | Pereira, Nicholas F |
Publisher | ScholarWorks@UMass Amherst |
Source Sets | University of Massachusetts, Amherst |
Language | English |
Detected Language | English |
Type | text |
Source | Doctoral Dissertations Available from Proquest |
Page generated in 0.0023 seconds