Non-Rigid Multimodal Medical Image Registration using Optical Flow and Gradient Orientation

Optical flow models are widely used for different image registration applications due to their accuracy and fast computation. Major disadvantages to overcome for medical image registration are large deformations and inaccurate regularisation at discontinuities, which cannot be modelled accurately with quadratic regularisers, and an intensity dependent energy term, which does not allow for images of different modalities. In this work we present a multi-level framework utilising multiple warps, which succeeds in estimating larger deformations. We introduce a non-quadratic penalty function, for a better modelling of discontinuities, that are caused by sliding motion of ribs against the lungs during respiration. Our algorithm is extended to multimodal image registration tasks by maximising the local alignment of the image intensity gradient orientation. We demonstrate the findings on synthetic 3D CT data and clinical CT-CT images as well as on CT-MRI data. Quantitative evaluation using the Dice coefficient shows improvements of our new approach for single-modal data for the interface between lungs and ribs compared to a commonly used parametric free form deformations (FFD) method and equally good results for multimodal data.