|
The radiation dose delivered to the population by diagnostic medical imaging devices is continuously increasing, with x-ray computed tomography (CT) and nuclear medicine being the main contributors. Adjusting the dose only makes sense when it can be associated with an objective image quality metric. Moreover, most of the modern imaging modalities generate a substantial quantity of 3-dimensional (3D) data that obliges the radiologists to change how they read images: Radiologists often flip through the image slices in order to make the pathology "pop out of the image", and while they look at certain regions of the image they might miss clinically relevant structures in the visual periphery. These crucial and unique aspects of diagnostic decisions with 3D data sets are not taken into account by the image quality tools presently used by the medical physicists.:Background The goal of this proposal is to provide metrics of medical image quality that allow prediction of the trade-off between dose and diagnostic accuracy of radiologists in 3D CT images. We propose to develop and test several types of model observers in two practical applications where dose is of particular concern, (1) imaging lung nodules in chest CT, which are a relatively simple case that will be used in developing and validating model observers and (2) hepatic metastases, which are much more difficult to detect and need CT optimization.:Aim and objectives We will begin by studying the problem of lung-nodule detection in CT. This is a case for which dose reduction has already been implemented in CT, but it serves the purpose of having a clinically relevant set of data that can be used to validate the model observers. The detection of lung nodules has the property that a 3D movement through the slices is helpful in distinguishing nodules from normal anatomy. The models developed will then be extended to the case of hepatic metastases. Detection of metastases will be examined with human observers (radiologists and trained non-clinical observers) first with 2-Alternative forced choice (2-AFC) experiments and then, in order to be closer to the clinical reality, with Localization ROC (LROC) experiments. Both of these are well established psychophysical techniques for evaluating reader performance. We will compare perceptual performance as well as search strategies including dwell-time at different locations, number of re-fixations for a given location, total number of eye movements and sequence of eye movement fixations. We propose the following approach: (1) pooling the 2D model responses across the different image slices; (2) computing the response on the entire 3D image at once in a manner analogous to the 2D calculation; (3) taking into account the interaction between the speed of scrolling performed through the slices and the temporal response of the human visual system. :Materials and methods If successful, the proposal will result in a model observer that could be used for CT images and possibly be extended to other uses of CT and other 3D modalities (PET, SPECT, tomosynthesis). If we make the analogy of results obtained with the use of recent iterative reconstruction algorithms, this work has the potential to reduce the dose delivered to the many patients undergoing diagnostic CT examinations typically by up to 40% on average.:Expected results and importance
|