| Dynamic (time dependent) imaging protocols have the potential to yield valuable physiologic information and therefore allow for mechanism-based diagnosis. One method for extracting this information is to apply to the image data mathematical models that quantify the kinetic behavior of the physiologic system. If a kinetic model is fit to the dynamic data of every pixel within an image, then the resulting model parameters can be redisplayed as a parametric image. This in turn provides the parameter values and spatial distributions, thereby preserving the image's inherent heterogeneity information. While parametric image analysis is gaining utility in the assessment of a variety of pathologies, it can be confounded by noisy data (poor signal to noise ratios) at the individual pixel level. We postulated that population-based modeling techniques would be useful when modeling each pixel's noisy dynamic data. To venture into the population-based modeling domain, we chose to implement Bayesian modeling techniques, which we hypothesized would improve the accuracy of the parameter estimates in the presence of poor signal to noise, while not altering the raw imaging data. The Bayesian method was evaluated using simulation and phantom studies, where it performed well compared with other methods. We then illustrated the use of this method by applying it to two different pathologies, namely the assessment of tumor vasculature in an animal cancer model and the evaluation of the synovium in rheumatoid arthritis patients. In both the simulations and applications, the Bayesian method performed well, showed improved parameter accuracy over other methods, and regularized the parameter estimates without overly smoothing the parametric images. |
