We propose to quantitatively assess image quality in magnetic resonance imaging (MRI) and use it to evaluate and optimize the almost unlimited number of ways to acquire and reconstruct an image. In fast MR, one can increase time to improve image quality in a manner analogous to the way that one can increase dose in x-ray and nuclear imaging. An informed trade-off requires a quantitative measure of image quality, heretofore not used in MR imaging. We will focus on important variables in fast MR imaging, including those associated with acquisition sampling and reconstruction. As done in x-ray and nuclear imaging, we will use human observer and model detection of isolated pathologies to assess image quality. However, detectability is insufficient for many reasons, e.g, because multiple image features are important and because the goal might be guidance of a needle for intervention to a readily seen target. As a second, and probably more useful, method for evaluating image quality in fast MR, we will use a perceptual difference computer model (PDM) developed in our laboratory that determines the visual difference between a high quality reference image and a more quickly obtained, but possibly degraded, MR image. A low PDM score indicates similarity of the images, and there is a threshold, below which, one cannot tell the difference between the two. We hypothesize that these two quantitative image quality approaches will be useful for assessing and optimizing MR images. A computer method such as PDM for assessing image quality is significant given the large parameter space and numbers of images; e.g., a spiral study can easily generate thousands of different images, precluding a conventional human-observer ROC analysis. We will demonstrate the method on selected, important design issues in spiral, radial, and combined acquisition technique (CAT) fast imaging methods. When this research is successfully completed, we will have established useful tools for evaluating and optimizing the almost unlimited number of MR imaging methods.
Showing the most recent 10 out of 18 publications
