Abstract
The accurate estimation of point correspondences is often required in a wide variety of medical image processing applications including image registration. Numerous point correspondence methods have been proposed, each exhibiting its own characteristics, strengths and weaknesses. This chapter presents a comparative study of four automatic point correspondence methods. The four featured methods are the Automatic Extraction of Corresponding Points approach, the Trimmed Iterated Closest Points scheme, the Correspondence by Sensitivity to Movement technique and the Self-Organizing Maps network. All methods are presented, mainly focusing on their distinct characteristics. An extensive set of dental images, subject to unknown transformations, was employed for the qualitative and quantitative evaluation of the four methods, which was performed in terms of registration accuracy. After assessing all methods, it was deduced that the Self-Organizing Maps approach outperformed in most cases the other three methods in comparison.
TopBackground
There are numerous applications in medical imaging where geometrical registration is performed including the alignment of data between two modalities with anatomical information (CT-CT or CT-MRI), between anatomical atlases and dynamic studies (CT-PET, CT-SPECT, MRI-PET) or between images of the same modality acquired at different times (Maintz & Viergever, 1998).
Several image features may be exploited for the matching process, depending on the modalities used, the specific application and the implementation of the transformation utilized. There are numerous comprehensive surveys of medical image registration (Hajnal, Hill & Hawkes, 2001; Zitova & Flusser 2003), in terms of imaging modalities and employed techniques. The medical image registration methods can be classified into two main categories: image similarity-based methods and point-based methods. In image similarity-based methods, the registration of different images involves the optimization of a quantity measuring the similarity between the images, under constraints imposed by the preferred transformation model (Johnson & Christensen, 2003; Pluim, Maintz & Viergever, 2003). On the other hand, in point-based methods, registration involves the determination of the coordinates of corresponding features/points in different images such as landmark points, ridges or surfaces, and the estimation of a geometrical transformation using these corresponding features (Bookstein, 1997; Pitiot, Thompson & Toga, 2002; Pizer, Fritsch, Yushkevich, Johnson & Chaney, 1996). The corresponding features can be determined either manually or automatically.
In the literature, numerous automatic correspondence methods have been devised (Cao, Pan, Li, Balachandran, Fitzpatrick, Chapman & Dawant, 2004; Chetverikov, Svirko & Stepanov, 2002). According to the general methodology, a correspondence method incorporates two main steps; namely, detection and matching. The detection step aims at extracting salient anatomical points of the visible anatomy and/or geometrical interest points located at the locus of the optimum of some geometric property, such as L-shaped corners, T-shaped junctions and Y-shaped junctions (Laliberte, Gagnon & Sheng, 2003; Likar & Pernus, 1999). After the detection of these interest points, the correspondence between them can be established by the matching step.
Key Terms in this Chapter
Dental Imaging: The representation of dental elements on a film, acquired through specific radiographic protocols
Point Extraction: The automatic or manual process of extracting points of interest from an image
Automatic Point Correspondence: The automatic process of estimating the homologous points on the corresponding image of a set of initial points from the reference image.
Image Registration: The process of matching a corresponding image to a reference image, using suitable transformations
Template Matching: Automatic point correspondence method based on matching regions from the reference image to transformed regions of the corresponding image
Self Organizing Maps: Automatic iterative method for allocating point correspondences, based on Kohonen’s neural network
Iterated Closest Points: Automatic point correspondence method which works by minimizing the average distance from a set of points on the reference image to a set of points on the corresponding image