Image Registration In Digital Subtraction Angiography Technique

In current clinic practice, Digital Subtraction Angiography technique is an unparallel visualization tool for the blood vascular structures. How to further improve its imaging function and quality to provide more powerful diagnostic value is still an active research topic in the field of medical imaging. In this thesis, image registration problem in DSA technique is discussed and a new image registration approach is developed. Focusing on the registration framework, following works are carried out in this thesis: Firstly, the vascular centerlines in DSA images are extracted by multiscale Gabor filters. The flexible frequency bands and enhancement effects,are fully utilized to enhance centerlines of the blood vessels in various size and greatly reduce the background noise. The non maximum suppression algorithm is taken to get the local maximum points on the response map. Then, vessel centerlines are extracted by a hysteresis-thresholding segmentation. The evenly chosen small number of points on the centerlines forms the control points set. Secondly, according to the local dissimilarity of the DSA image pairs (live image and mask image), a synthetic template pairs, which are strictly registered, are used to test the robustness of two widely used similarity measurements: combined invariant moments (affine invariant, blur invariant and contrast invariant) and Mutual Information. Test results demonstrate the combined invariant moments are too sensitive towards the local dissimilarity to correctly calculate the displacement of the control points on the vessel part. While the Mutual Information can be comparatively resistant to the local difference, and also be tolerant to local deformation within a small range, which is robust enough as for the DSA images. Thus the mutual Information is chosen as the similarity measurement to search for the control point displacements. Finally, the thin-plate spline (TPS) interpolation function is utilized to conduct image warping towards the mask image based on the corresponding control point sets. In order to reduce the influence from the displacements mistakes, a smoothed TPS function is taken, which can control the warping and smooth degree by adjusting the smooth parameter. Also, a weighted interpolation method is given to make the warping result more accurate and complete. Experiments carried on several types of DSA data set demonstrate that the proposed algorithms can remove most of the artifacts and noisy with low computational cost, the vascular structures can therefore be better visualized.