Bilateral filter regularized accelerated Demons for improved discontinuity preserving registration.

The classical accelerated Demons algorithm uses Gaussian smoothing to penalize oscillatory motion in the displacement fields during registration. This well known method uses the L2 norm for regularization. Whereas the L2 norm is known for producing well behaving smooth deformation fields it cannot properly deal with discontinuities often seen in the deformation field as the regularizer cannot differentiate between discontinuities and smooth part of motion field. In this paper we propose replacement the Gaussian filter of the accelerated Demons with a bilateral filter. In contrast the bilateral filter not only uses information from displacement field but also from the image intensities. In this way we can smooth the motion field depending on image content as opposed to the classical Gaussian filtering. By proper adjustment of two tunable parameters one can obtain more realistic deformations in a case of discontinuity. The proposed approach was tested on 2D and 3D datasets and showed significant improvements in the Target Registration Error (TRE) for the well known POPI dataset. Despite the increased computational complexity, the improved registration result is justified in particular abdominal data sets where discontinuities often appear due to sliding organ motion.

Intensity-based elastic registration incorporating anisotropic landmark errors and rotational information.

PURPOSE: Thin-plate splines (TPS) represent an effective tool for estimating the deformation that warps one set of landmarks to another based on the physical equivalent of thin metal sheets. In the original formulation, data used to estimate the deformation field are restricted to landmark locations only and thus does not allow to incorporate information about the rotation of the image around the landmark. It furthermore assumes that landmark positions are known exactly which is not the case in real world applications. These localization inaccuracies are propagated to the entire deformation field as each landmark has a global influence. We propose to use a TPS approximation method that incorporates anisotropic landmark errors and rotational information and integrate it into a hierarchical elastic registration framework (HERA). The improvement of the registration performance has been evaluated. METHODS: The proposed TPS approximation scheme integrates anisotropic landmark errors with rotational information of the landmarks. The anisotropic landmark errors are represented by their covariance matrices estimated directly from the image data as a minimal stochastic localization error, i.e. the Cramér-Rao bound. The rotational attribute of each landmark is characterized by an additional angular landmark, thus doubling the number of landmarks in the TPS model. This allows the TPS approximation to better cope up with local deformations. RESULTS: We integrated the proposed TPS approach into the HERA registration framework and applied it to register 161 image pairs from a digital mammogram database. Experiments showed that the mean squared error using the proposed TPS approximation was superior to pure TPS interpolation. On artificially deformed breast images HERA, with the proposed TPS approximation, performed significantly better than the state-of-the-art registration method presented by Rueckert. CONCLUSION: The TPS approximation approach proposed in this publication allows to incorporate anisotropic landmark errors as well as rotational information. The integration of the method into an intensity-based hierarchical non-rigid registration framework is straightforward and improved the registration quality significantly.