Research On Medical Image Registration Algorithm Based On SURF Improvement

Medical image registration means that images come from different time, conditions, and equipments through a space transformation to get alignment with another medical image. Currently, image registration has been applied into remote sensing image processing and computer vision fields, especially becoming a research hotspot in medicine. During medical diagnosis,2D image information can be provided for patients by MRI, CT and other equipments, which cannot provide the internal structure information visually. Image 3D reconstruction can exactly show the architectural feature and space position of body organs by 3D image being reconstructed from 2D image. Image registration is key point.This thesis studies the image registration based on feature. Feature extraction and matching is the key of featured image registration. SURF algorithm has been applied into feature extraction and matching stage. It has certain robustness to image translation, rotation and scaling, and speeds up the operation with the function of integral image and box filter. But it extracts lesser feature points during medical image matching, and causes deficiency in amount of feature matching, which cannot reach the requirements. Also it is prone to error matching in feature matching stage, with great influence to the follow-up in registration accuracy and poor effect of affine transformation. Through above, this thesis proposed the improved SURF algorithm. In feature extraction stage, it sets up affine camera model to simulate image sequence in different perspectives and scale space combined with box filter for testing feature points to determine main direction and generates 64D descriptor. Finally these points map into original image to increase the testing quantity. In feature point matching stage, it uses the advanced GTM (Graph Transform Matching) algorithm and sets up GTM image of matching points in two images. Calculate the difference value of adjacent matrix by GTM image, and delete error matching pairs for optimization and accuracy improvementThe thesis both adopts preliminary and fine registration ways. In preliminary registration period, it uses rigid body transformation model to make image transformation. Owing to the low accuracy by solving with rigid body transformation model, it utilizes GA (Genetic Algorithm) to optimize matrix parameters and gives preliminary registration to images. In fine registration period, it uses thin plate spline functions to do fine registration for preliminary registration images and improves the accuracy. It proves the superiority of SURF algorithm improvement by testing and analyzing the common medical images of translation, rotation and affine.