Improving image registration by correspondence interpolation

Abstract

This paper presents how using a correspondence-based interpolation scheme for 3D image registration improves the registration accuracy. The interpolator takes into account correspondences across slices, which is an advantage, particularly when the volume has thick slices, and where anatomies lie non-parallel to the slice direction. We use our previously presented approach for correspondence-based interpolation and demonstrate results on two different datasets, brain and cardiac MRI. The results are evaluated (i) qualitatively by examination of gradient images and cardiac pig atlases and (ii) quantitatively by registering downsampled brain data using two different interpolators and subsequently applying the deformation fields to the original data. The results show that the interpolator provides better gradient images and a more sharp cardiac atlas. Moreover, it provides better deformation fields on downsampled data, increasing the registration accuracy of original data to 5.8% on average with respect to a standard interpolator.

Cardiac Image Super-Resolution with Global Correspondence Using Multi-Atlas PatchMatch

“…This is a trade-off for improved through-plane resolution. Future work will include exploring the possibility to extend MAPM to patch-based segmentation and to exploit neighboring correspondence [18] to preserve the original SA slice and image self similarity [19] to increase the robustness.…”

MRI Upsampling Using Feature-Based Nonlocal Means Approach

“…In one category of these methods, a single LR image from the subject is available for upsampling. Adjacent slices of the LR image may be registered using a non-rigid method to reconstruct the slices between them [1–6]. These registration-based methods best work when each slice is acquired from a thin slab of the organ as blurred edges between thick slices lead to poor coregistration.…”