Xin Wang PhD, Presenter: Nothing to Disclose

Purpose/Objective(s):Two-dimensional (2D) to 2D matching of the kV setup and digitally reconstructed radiography (DRR) images is a popular patient setup technique for image-guided radiotherapy (IGRT). Mutual information based methods were developed to co-register these two types of images on commercial linear accelerators[1]. However, additional operator involvement is often needed to correct the mismatch of mutual information due to the movement, shape, intensity changes between the two images. The shape invariant feature transform (SIFT) method has been used to detect the shape features in the overlapping region of the projection digital photographs [2]. In this abstract we present our work on applying the SIFT method to the image co-registration of DRR and kV setup images. The hypothesis of this study is that the SIFT method can be used to detect the horizontal and vertical shifts between the kV setup and DRR images. Materials/Methods: The kV setup and DRR images are of very different contrast and intensity ranges. Without a preset cue or marker system, it is hard to find intensity and shape matching image features. We modified the SIFT method to estimate the imaging shifts and scaling factors of the kV setup image with reference to the DRR. Specifically, selection of good matching points is essential to the success of the feature identification. Because of the high misidentification rate, centers of the kV setup and DRR images were read from the DICOM header as a reference. The matching points were selected in the corresponding regions of the image, for example, from left quadrant of both images. After that, a rigid transformation was estimated to account for the horizontal and vertical shifts and scaling factor between the two images. The kV setup image was then transformed and overlaid on the DRR image. At this stage, we performed a visual check of the matching accuracy between these two images. The rotation between the two images was ignored considering the clinical reality.Results:We first selected a pair of anterior-to-posterior pelvic kV setup and DRR images from a patient as an experimental set and achieved a registration accuracy comparable to the visual check. We then introduced artificially arranged vertical and horizontal shifts between the two images. The linear regression analysis of the introduced vs. detected shifts showed a slope of 1.15 and 0.98 with an R2 (sample correlation coefficient) of 0.89 and 0.99 for the horizontal and vertical shifts, respectively. Conclusions: We proved in principle that the feature-based image co-registration is viable for images with completely different contrast and intensity modes. We are currently refining the feature selection criterion of the SIFT method to further improve of the co-registration accuracy.References1.Kessler, M.L., Br J Radiol, 2006. 79 Spec No 1: p. S99-108.2.Brown, M.a.D.G.L., Int. J. Comput. Vision, 2007. 74(1): p. 59-73.

Wang, X, Automatic Feature-based Co-Registration of Digitally Reconstructed Radiography (DRR) and kV Setup Images.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14043551.html