Feature‐recognizing MRI

Abstract

We describe a new theory of MR imaging that utilizes prior information in the form of a set of “training” images thought to be similar to the “unknown” objects to be scanned. First, the training images are processed to find an orthonormal series representation of these images that is more convergent than the usual Fourier series. The coefficients in this new series can be calculated from a subset of the phase‐encoded signals needed to construct the Fourier image representation. The characteristics of the training images also determine exactly which phase‐encoded signals should be measured in order to minimize error in the image reconstruction. The optimal phase‐encodings are usually scattered nonuniformly in k‐space. Good results were obtained when this theory was applied to imaging data from simulated objects and to experimental data from phantom scans. This theory provides the basis for developing efficient scanning and image reconstruction techniques that are “tailored” to each body part or to particular disease states.