Landmark-Based Elastic Matching of 3D Medical Images

We describe an approach for 3D image registration based on a set of corresponding anatomical point landmarks and approximating thin-plate splines. This approach is an extension of the original interpolating thin-plate spline approach and has been introduced in [7] to cope with landmark localization errors. The extension is important for practical applications since landmark extraction is always prone to error. Additionally, we are investigating 3D di erential operators for semi-automatic localization of anatomical point landmarks. 1 Elastic Image Matching In neurosurgery and radiotherapy planning it is important to either register images from di erent modalities, e.g. CT (X-ray Computed Tomography) and MR (Magnetic Resonance) images, or to match images to atlas representations. If only rigid transformations were applied, then the accuracy of the resulting match often is not satisfactory w.r.t. clinical requirements. In general, nonrigid or elastic transformations are required to cope with the variations between the data sets. This contribution is concerned with elastic matching of medical image data based on a set of corresponding anatomical point landmarks. Previous work on this topic has concentrated on i) selecting the corresponding landmarks manually and on ii) using an interpolating transformation model (e.g., Bookstein [2], Evans et al. [4]). The basic approach draws upon thin-plate splines and is computationally e cient, robust, and general w.r.t. di erent types of images and atlases. Also, the approach is well-suited for user-interaction which is important in clinical scenaria. However, an interpolation scheme forces the corresponding landmarks to exactly match each other. The underlying assumption is that the landmark positions are known exactly. In real applications, however, the localization of landmarks is always prone to error. This is true for interactive as well as for automatic landmark localization.