Ordinarily, the lack of higher k-space values causes a blurring artifact in the reconstructed images in which sharp discontinuities create spatial oscillations into nearby regions. In some cases - such as lipids in the scalp - these discontinuities con interfere with low-level signal measurements in nearby regions. If the general location of these bright signals is known, we can postprocess the image to """"""""steer"""""""" the oscillations in a different direction. We have developed a simple approach to solving this problem. By specifying the location of the bright interference, the phase response of a filter is optimized to reduce the oscillations in the region of interest. At present, out technique has been verified in 1-D simulations. It holds promise for significantly reducing the oscillations caused by nearby bright objects. The computational complexity - once the filter has been designed - is on the same order as the original reconstruction from k-space samples. We are currently extending the method to 2-D and more general interference patterns. We investigated the use of a local variance model of the image to indicate a priori knowledge of discontinuities in the original. The model separately represented variations in the horizontal direction and the vertical direction. We assumed that an accurate model could be derived from a standard MR scout image of the object. The local variance model was incorporated into the reconstruction by penalizing roughness in a way inversely related to the local variance of the scout image. In this way, the reconstruction formula was made to be shift variant, allowing for the possibility of spectral extrapolation from the observed k-space samples. Our experiments confirmed that this is a very effective form of image model for the type of prior information that is available from an anatomic scout image. While higher resolution of fine details was somewhat limited, the reconstruction was dramatically sharper near boundaries in the image so that blurring of edges was minimal.?
Showing the most recent 10 out of 19 publications
