Visual-Search Model Observers in Assessing CT Acquisition Parameters

Medical images are commonly used by radiologists for diagnostic purposes. With the ever-evolving technology in the medical imaging systems’ technology, radiologists and medical experts face numerous choices between different imaging modalities, acquisition protocols, and reconstruction algorithms. Therefore, there is a growing need for developing a reliable image quality assessment metric to help choose and optimize the appropriate system’s parameters. A state-of-the-art approach for assessment of medical image quality is to quantify image quality by measuring the performance of the radiologists when performing a clinical task of interest. This approach is known as task-based image quality assessment. In a task-based image quality assessment, an observer such as a radiologist is to read an ensemble of images that are produced using different imaging systems or using the same system but with different parameters. An example of a task can be lesion detection. Then the performance of the observer is quantified using a figure of merit and the system or parameter(s) that yields the better performance is chosen. Note that employing a radiologist to read images is costly and time-consuming; therefore, a great deal of research has been dedicated to developing mathematical model observers to be used instead of humans in such studies. In this dissertation, we aimed to develop a novel model observer called the visual-search (VS) observer that could be used as a surrogate to human observers. To evaluate the effectiveness of our VS observer, we used simulated (CT) images to optimize CT acquisition parameters. For our study, we used a basic acquisition setting, i.e., parallel-beam geometry and a filtered-backprojection reconstruction. We aimed to evaluate the effect of the number of projection angles and the number of detectors on the performance of our observer. A varying range of these parameters were tested under different tasks. The tasks were chosen to be either lesion detection or lesion localization. We also implemented the channelized Hotelling observer that is used often in studies where the purpose is the optimization of acquisition parameter and compare its performance to that of our visual search observer. We also performed human observer studies to use as the gold standard in measuring image quality. Our results showed that sampling parameters affected visual search observer performance in a similar manner to that of human. The channelized Hotelling observer, on the other hand, was not affected by the variations in the angular sampling. Our results indicated that the visual search framework is an effective modification of the channelized Hotelling observer for assessing the effects of acquisition parameters on human-observer performance in our simulated CT images. We also extended our visual search observer to the projection space as projection studies are often considered for acquisition parameter optimization. This approach allowed us to study the effect of the same CT acquisition parameter i.e., the angular sampling, without the complications associated with the reconstruction process. We compared the performance of our observer on projection data vs. filtered-back projection reconstruction. The reconstructions were generated using different filters. Although projection performance was always higher compared to the reconstruction regardless of the filter, the projection performance correlated with that of the reconstruction under varying values of the angular sampling. Our projection analysis using visual search observer was the first of its kind and showed promising results in assessing the effect of angular sampling using CT simulated projections.