Using prior knowledge of human anatomy to constrain MR image acquisition and reconstruction: half k-space and full k-space techniques.

In this note, we demonstrate how to utilize prior knowledge of human cranial anatomy to constrain full k-space and half k-space acquisition and reconstruction of 128 times 128 images. We used a database of magnetic resonance head images to derive new basis functions which represent the most important features of the head. the "training" images were also used to derive formulas for reconstructing head images from a subset of the usual 128 phase-encoded signals and to determine the optimal k-space locations of those signal measurements. We used this algorithm, called Feature-Recognizing MRI, to reconstruct 128 times 128 head images from 50-60% of the signals filling the full k-space. Furthermore, we combined the algorithm with a conventional half k-space technique to create 128 times 128 images from only 60% of the 80 signals required by the usual unconstrained half k-space imaging. Thus, the prior knowledge represented by the image database, together with a half k-space technique, made it possible to construct accurate magnetic resonance images from only 30-40% of the complete set of 128 signals. In other words, a database of head images was used to devise a 1/3 k-space method for imaging the head.