Modeling observer performance in low-dose CT assessments

项目总结/文摘

Project Summary/Abstract X-ray computed tomography (CT) has become a mainstay of diagnostic imaging in many areas of medicine because of its ability to render internal structures of the body with high accuracy. This has resulted in a substantial increase in the use of CT imaging in the United States. As a result, there has been sustained interest in dose reduction in CT imaging. However, demonstrating effective dose reduction is challenging. By definition, such techniques seek to retain diagnostic quality with little or no measureable effect on diagnostic performance. Clinical reader studies using receiver operating characteristic (ROC) methodology are the accepted standard for evaluating diagnostic performance effects. However, these studies are expensive and time consuming, and identifying small effects requires prohibitively large sets of readers and cases. This has led to the development of “model-observers” for dose reduction claims at the US Food and Drug Administration (FDA). At this time, at least three dose reduction claims at FDA have used model observer studies to substantiate their claim of CT dose reduction using iterative reconstruction algorithms. The basis for this project is our recognition that such models have had relatively little validation, given the complexity of both the human visual system and the images being evaluated. We propose an in-depth characterization of human observer responses in tasks related to dose reduction in CT. The purpose of this research is to develop and validate a model (or models) of observer performance for use in assessments of image reconstruction for CT dose reduction. For a model observer to be of use in this area, it must be accepted as a reasonable predictor of human-observer performance for some range of relevant tasks. This motivates our general approach, and many specifics of our research plan. Our plan is to collect an initial set of psychophysical data, use this data to develop our model, and then predict performance in CT reconstructions from simulations at a variety of doses. We then collect the psychophysical data on these images to quantify predictive accuracy and to compare it to the accuracy of other models. Specific Aim 1 involves the collection of psychophysical data in tasks with noise statistics similar to CT dose assessments. Specific Aim 2 seeks to develop models of task performance by fitting model parameter for several candidate models to the data from Aim 1. Specific Aim 3 proposes a prospective prediction of observer performance in a new set of psychophysical data from images that have been reconstructed using modern iterative methods. At the conclusion of the project period, we expect to have a better understanding of how observers perform difficult localization and discrimination tasks in noisy CT images, and how this process in influenced by the dose associated with images.