This proposal develops task-based methods for evaluating and optimizing acquisition, reconstruction, and processing techniques for gated myocardial perfusion SPECT. Gated SPECT yields a series of volume images of the left ventricle, each representing one portion of the cardiac cycle. Several clinical tasks may be performed with the resulting images, including visual observation of motion properties (the detection task). For example, physicians will frequently view animations to assess whether the left ventricular wall appears to move normally. However, reconstruction methods designed for gated SPECT have never been evaluated on such a task, largely due to a lack of appropriate tools. The initial phase of the research involves the development of new tools, including a realistic mathematical phantom that can model a continuum of cardiac motions and an interactive software application for presenting motion images to observers in a controlled experiment. These tools will be used in a series of human observer studies meant to emulate the clinical task of detecting abnormal cardiac motion from animated displays of gated SPECT reconstructions. The initial observer studies will focus on optimizing acquisition parameters, such as the number of gated frames and the presentation frame rate, and on determining whether accurate modeling of physical effects, like attenuation or scatter, has an effect on the detection task. Further observer studies will be used to optimize smoothing parameters for 4D reconstruction and processing methods, and compare these methods to determine if one offers improved performance in the detection task. Lastly, the research proposes to develop a set of computer-based observers, so as to reduce the need for time-consuming human observer studies in the future. The computer-based observers, will incorporate the Channelized Hotelling Observer concept, extending it to the spatiotemporal domain, and will be bases on psychophysical properties of the human visual system. Computer-based observers will be trained and evaluated on the same image detection tasks as in the human observer studies to determine the computer-based observer most likely to emulate human task performance.
Showing the most recent 10 out of 11 publications
