2D/3D Multi-Modal Medical Image Registration Based On Iterative Regression

Image-guided surgery(IGS)increases the field of vision through visualization to reduce the risk of tissue damage and improve the accuracy and targeting of lesion location.It needs to obtain the patient’s preoperative three-dimention(3D)images and intraoperative two-dimensional(2D)images.3D images can provide more space information so as to determine the lesion location and plan the surgical procedure.2D images can provide intraoperative real-time information so as to track the surgical instruments and adjust the space position in real time.Therefore,the key of IGS technology is to seek the spatial position transformation relationship between preoperative 3D image and intraoperative 2D image,namely 2D/3D image registration.In IGS system,registration is maintained throughout the treatment process,so improving the accuracy,timeliness and robustness of image registration is a problem to be solved in current research.The 2D/3D registration can be abstracts as an optimization problem including similarity measure function design and optimization strategy selection.In the thesis,we mainly studied the 2D/3D registration algorithm based on the influence of different image information on the accuracy,timeliness and robustness of registration,and focused on the design of similarity measure based on intensity information and feature information.Then,an improved algorithm is proposed.The algorithms proposed in the thesis all are based on iterative optimization to obtain the optimal value.The specific research work is as follows:(1)The dissertation researches 2D/3D medical image registration algorithm based on intensity,and realizes the registration based on normalized cross correlation.Aiming at the problem that the normalized cross-correlation algorithm has single information and small convergence domain,the gaussian Laplace operator is introduced as a new similarity measure to add edge information and internal details.Aiming at the low efficiency of normalized cross correlation algorithm,a multi-resolution strategy is used to improve the registration accuracy and efficiency.(2)Aiming at the problem of the normalized cross-correlation algorithm,the Sobel operator is introduced as a new similarity measure to add angle information of gradient direction,thus increasing the rotation sensitivity.Aiming at the low efficiency of normalized cross correlation algorithm.The multi-resolution strategy is also used.Finally,the purpose to improve the registration precision,registration efficiency and convergence domain is achieved.(3)The 2D/3D medical image registration algorithm based on features is studied.Statistical information is extracted from the histogram of the gradient direction as the feature.Moreover the 2D/3D registration based on the weighted histogram of the gradient direction is realized.Aiming at the problem that the weighted histogram of gradient direction is only suitable for image registration with small foreground and large background and insensitive to translation transformation,the weighted spatial histogram of gradient direction is proposed as a new similarity measure.Statistical features are extracted by two-dimensional spatial histogram in the direction of gradient,which breaks the limitation of the algorithm and broadens the applicable scene under the premise of ensuring accuracy.The proposed algorithm is experimented based on CT images and synthetic X-ray images.Compared with existing algorithms,the experiment results show that the proposed algorithm can improve the accuracy and efficiency,and reduce the sensitivity of initial values.