Tracking of Tagged MR Images by Bayesian Analysis of a Network of Quads

Automatic tracking of tagged MR image sequences is done frame-by-frame. For each frame, a quadrilateral (quad) detector is run over the image to give a set of “potential quads”. A likelihood function is specified for the detection of potential quads from an image. Quads are picked from the set of potential quads to form a “quilt”. Quads are present where a grid structure is apparent in the image. A prior is specified to govern how the quads should be joined up to form the quilt. The prior for the quilt (i) encourages quads to be close to their positions predicted from the last frame, (ii) encourages neighbouring quads to be close to each other, (iii) discourages intersecting quads, (iv) avoids “tears” in the quilt, and (v) encourages connectedness of quads. With the likelihood and prior densities, a Bayesian analysis is carried out using the Markov Chain Monte Carlo method on the posterior density to give an estimate of the posterior mode. Email: fdelman,john,[email protected] Post: Department of Statistics, University of Leeds, Leeds, LS2 9JT, U.K..