3D echocardiography has brought added value to ultrasound imaging in clinical practice. In comparison to its 2D counterpart, 3D echo provides more anatomical information on the spatial relationship between structures. It assists in the diagnostic process and helps in training new echocardiographers. Its real time capability makes 3D echo a suitable imaging modality for intraoperative monitoring [11].

One of the inherent limitations of 3D echo is its dependence on the incidence angle [8]. Signal dropouts often occurs in the presence of strongly reflective structures. Transesophageal Echocardiography (TEE) and Intracardiac Echocardiography (ICE) probes that have a clear access to the heart during minimally invasive procedures have limited field of view due to their small aperture size.

There have been attempts to overcome these limitations using post-processing techniques, aiming at both improvement of image quality as well as extension of field of view. Such approaches are mainly focused on compounding information from several acquisitions. When combining image data from several acquisitions, they must be both registered and fused. Registration is performed to optimally transform the images into one common coordinate system while fusion combines the redundant data for optimal visual. Some studies focused more on the accuracy of the registration [2, 13, 14], while others at the improvement of image quality through fusion [1, 9, 16]. This paper focuses on the registration part.

Some studies utilize tracker information for registration [2, 16]. Either electromagnetic or optical trackers are commonly used. For electromagnetic tracking, there is a risk of disturbance in the electromagnetic field when ferro magnetic material is present. In the case of optical tracker, a clear line of sight from the optical marker to the sensor has to be kept.

Image based registration methods can be grouped into voxel based and feature based methods. Many available registration methods rely on optimization of a cost or similarity function that is computed directly from the voxel intensity values. Some of the known functions are: sum of absolute differences, sum of squared differences, normalized cross correlation (NCC) and mutual information [1, 12, 15]. Voxel based methods are to some extent dependent on pose initialization. Some works used tracker information for initial registration then refined it using a voxel based method [2, 16]. Other approaches also perform optimization but on extracted features from the image instead of directly on the intensity values. For example, Ni et al. [13] utilized 3D SIFT which is a scale invariant and robust gradient based feature descriptor to perform registration on 3D ultrasound of the liver to build panoramic volumes. Similarly, Schneider et al. [14] used features based on Laplacian-of-Gaussian filtering to register images of a porcine heart in a water tank while Grau & Noble [9] used monogenic signal analysis.

In this work, a new approach is proposed. The proposed registration is based on signal decomposition into a polynomial basis using the optical flow method by Farnebäck [5]. Farnebäck optic flow was chosen for its ability to handle potential uncertainty in the images. In contrast to the voxel based methods that optimize global cost function by applying a series of transformations, the Farnebäck method attracts the floating image to the optimum position and orientation that matched the reference image based on the analytical solution of the optic flow problem. This approach was taken to allow for robust and fast registration in images with very small frame-to-frame transformations as well as larger transformations.

The general overview of the proposed registration method can be seen in Fig. 1.