Multiview Diffeomorphic Registration for Motion and Strain Estimation from 3D Ultrasound Sequences

This paper presents a new registration framework for estimating myocardial motion and strain from multiple views of 3D ultrasound sequences. The originality of our approach resides in the estimation of the transformation directly from the multiple views rather than from a single view or a reconstructed compounded sequence. This allows us to exploit all spatio-temporal information available in the input views avoiding occlusions and image fusion errors that could lead to some inconsistencies in the motion quantification result. In addition, by using the original input images, speckle information (which is an important feature for motion estimation and could be blurred out in the fusion process) should remain consistent between temporal image frames. We propose a multiview diffeomorphic registration strategy that enforces smoothness and consistency in the spatio-temporal domain by modeling a continuous 3D+t velocity field as a sum of B-spline kernels. This 3D+t continuous representation allows us to robustly cope with variations in heart rate resulting in different number of images acquired per cardiac cycle for different views. The similarity measure is obtained by extension of a pairwise mean square error metric where a weighting scheme balances the contribution of the different views. We have carried out experiments on synthetic 3D ultrasound images with known ground truth and on in-vivo multiview 3D data sets of two volunteers. It is shown that the inclusion of several views improves the consistency of the strain curves and reduces the number of segments where a non-physiological strain pattern is observed.