Abstract:
To improve the tagged cardiac magnetic resonance (CMR) image analysis, we propose a 3D (2D space + 1D time) energy minimization framework, based on learning first- and second-order visual appearance models from voxel intensities. The former model approximates the marginal empirical distribution of intensities with two linear combinations of discrete Gaussians (LCDG). The second-order model considers an image of a sample from a translation–rotation invariant 3D Markov–Gibbs random field (MGRF) with multiple pairwise spatiotemporal interactions within and between adjacent temporal frames. Abilities of the framework to accurately recover noise-corrupted strain slopes were experimentally evaluated and validated on 3D geometric phantoms and independently on in vivo data. In multiple noise and motion conditions, the proposed method outperformed comparative image filtering in restoring strain curves and reliably improved HARP strain tracking during the entirety of the cardiac cycle. According to these results, our framework can augment popular spectral domain techniques, such as HARP, by optimizing the spectral domain characteristics and thereby providing more reliable estimates of strain parameters.
