Time Continuous Tracking and Segmentation of Cardiovascular Magnetic Resonance Images Using Multidimensional Dynamic Programming

Publisher Website
Google Scholar
Abstract
In this article, we propose a semiautomatic method for time-continuous contour detection in all phases of the cardiac cycle in magnetic resonance sequences. The method is based on multidimensional dynamic programming. After shape parameterization, cost hypercubes are filled with image-feature derived cost function values. Using multidimensional dynamic programming, an optimal path is sought through the sequence of hypercubes. Constraints can be imposed by setting limits to the parameter changes between subsequent hypercubes. Quantitative evaluation was performed on 20 subjects. Average border positioning error over all slices, all phases and all studies, was 1.77 ± 0.57 mm for epicardial and 1.86 ± 0.59 mm for endocardial contours. The average error in end-diastolic and end-systolic volumes over all studies was small: 4.24 ± 4.62 mL and −4.36 ± 4.26 mL, respectively. The average error in ejection fraction was 4.82 ± 3.01%. The reported results compare favorable to the best-reported results in recent literature, underlining the potential of this method for application in daily clinical practice.