Non-rigid Registration of 3D Ultrasound Images Using Model-Based Segmentation

Non-rigid image registration of large 3D volumes depends heavily on image size. Voxel-based approaches for registration of large data sets are computationally intensive and take a lot of processing time, usually exponentially proportional to their resolution. In this paper, we present a method for non-rigid image registration that is independent of the image size. It can be applied to large data sets without sacrificing performance. The method is based on computing a thin plate spline (TPS) transformation between corresponding control points derived from a rapid and highly accurate model-based segmentation. The segmentation process does not depend on the image size and defines accurate point-based correspondence. Those properties were used to estimate an approximating TPS between a source and a target volumes and to deform the target to align with the source. The method was applied to image sequences of pediatric 3D ultrasound data. Registration of 3D ultrasound images is challenging because of poor-defined image gradients, artifacts, low contrast, noise and speckle. Quantitative and qualitative results indicate that the method is capable of registering large volumes acquired throughout the cardiac cycle from both intra- and inter-patient ultra- sound sequences. More importantly, the actual registration does not depend on image intensity and can be applied to multi-modal applications.

1. Introduction

Image registration is a computational method for establishing point-by-point correspondence between two data sets [5]. The registration process typically relies on corresponding image points known as control points or landmarks. These points are used to estimate a transformation function that deforms one image to align with the other. Non-rigid registration uses non-rigid transformation functions such as polynomials, splines, or multi- quadrics [6], [2], [3]. These could be formulated as either approximating or interpolating functions that require the solution of a system of equations. The non-rigid registration is usually achieved in an iterative fashion by minimizing a cost or energy function that represents the geometric and/or intensity differences between the two data sets. The performance of the registration algorithm depends on the quality of the corresponding landmarks and the image resolution in terms of accuracy and speed. In voxel-based non-rigid image registration, the cost function is estimated by optimizing an intensity-based similarity metric, such as mutual information or normalized cross correlation. These techniques could be accurate but are time consuming and impractical for many interventional procedures. Furthermore, the optimization process could be trapped in false local maxima, resulting in large registration errors.