Myocardium tracking via matching distributions.

OBJECTIVE

The goal of this study is to investigate automatic myocardium tracking in cardiac Magnetic Resonance (MR) sequences using global distribution matching via level-set curve evolution. Rather than relying on the pixelwise information as in existing approaches, distribution matching compares intensity distributions, and consequently, is well-suited to the myocardium tracking problem.

MATERIALS AND METHODS

Starting from a manual segmentation of the first frame, two curves are evolved in order to recover the endocardium (inner myocardium boundary) and the epicardium (outer myocardium boundary) in all the frames. For each curve, the evolution equation is sought following the maximization of a functional containing two terms: (1) a distribution matching term measuring the similarity between the non-parametric intensity distributions sampled from inside and outside the curve to the model distributions of the corresponding regions estimated from the previous frame; (2) a gradient term for smoothing the curve and biasing it toward high gradient of intensity. The Bhattacharyya coefficient is used as a similarity measure between distributions. The functional maximization is obtained by the Euler-Lagrange ascent equation of curve evolution, and efficiently implemented via level-set. The performance of the proposed distribution matching was quantitatively evaluated by comparisons with independent manual segmentations approved by an experienced cardiologist. The method was applied to ten 2D mid-cavity MR sequences corresponding to ten different subjects.

RESULTS

Although neither shape prior knowledge nor curve coupling were used, quantitative evaluation demonstrated that the results were consistent with manual segmentations. The proposed method compares well with existing methods. The algorithm also yields a satisfying reproducibility.

CONCLUSION

Distribution matching leads to a myocardium tracking which is more flexible and applicable than existing methods because the algorithm uses only the current data, i.e., does not require a training, and consequently, the solution is not bounded to some shape/intensity prior information learned from of a finite training set.