Automated Tracking of Deformable Objects Based on Non-Rigid Registration of Cardiac Images

We developed an automated technique for non-rigid image registration as a basis for tracking the heart in contrast-enhanced cardiac magnetic resonance (CMR) image sequences. The goal of the present work was to validate this technique against conventional manual analysis. Our approach is based on a multi-scale extension of the normalized 2D cross-correlation algorithm in combination with level-set methods. This technique was tested on short-axis CMR (Philips 1.5T) image sequences obtained in 11 patients at the mid level of the left ventricle during first pass of a gadolinium bolus. Myocardial identification required around 5s for a 60-frame sequence. To validate the technique, myocardial boundaries were manually traced on all frames by an experienced interpreter. Comparison between automatically registered and manually traced boundaries was performed by computing Hausdorff distance (2.1 +/- 1.4px), mean absolute distance (0.9 +/- 0.7px), root mean square distance (1.0 +/- 0.8px) and Dice coefficient (0.8 +/- 0.1). These results indicate that the proposed technique allows fast and accurate non-rigid image registration, and may thus be successfully used for deformable object tracking in cardiac image sequences.