Research On 2D-3D Registration Algorithm Of Double X-ray Images Based On Motion Decomposition Model

With the development of medical technology,the image-guided radiotherapy has gradually become one of the most important methods for cancer treatment.In order to ensure the accuracy of radiotherapy during the treatment,the pre-operative 3D CT images and the intra-operative 2D X-ray images are often used for 2D-3D registration to estimate the patient's current posture.However,because the patient's spatial motion is difficult to be fully reflected by a single Xray image,many iterations are required in the optimization.In each iteration,the computation burden is very heavy to generate Digitally Reconstructed Radiograph(DRR).Thus,the 2D-3D registration algorithm is not fast enough in the clinical practice.To address these issues,the following two works have been carried out in this thesis.(1)A 2D-3D registration method of the double X-ray images is proposed based on a 2D pre-registration process.First,the rigid transformation of CT is substituted by the inverse rigid transformation of the ray source and DRR,when the DRR is iteratively updated in 2D-3D registration.Moreover,every pixel of DRR is calculated by a core of the Graphics Processing Unit(GPU),respectively.Therefore,the computational complexity of DRR generation is effectively reduced,and the calculation of ray tracing is fully parallelized.Furthermore,the CT motion is decomposed into the orthogonal planes,according to the geometric relationship between the coordinate system of X-ray images and the coordinate system of CT.A new 2D-2D approximately rigid registration is performed on the orthogonal planes.Finally,the result of 2D-2D registration is used to guide the optimization of 2D-3D registration.The simulations on a skull phantom CT and a truncated chest CT demonstrate that,the registration time is significantly reduced and the precision is also improved for the proposed method.(2)A 2D-3D registration method of double X-ray images is proposed based on Convolutional Neural Networks(CNN).First,the 2D-3D registration problem is considered as a mapping from the residual image between the intra-operative X-ray images and DRRs to the set-up error of patients.At the same time,the registration problem is simplified by decomposing the CT rigid transform parameters into the in-plane parameter estimation and the out-of-plane parameter estimation.Moreover,a new regression network is proposed to derive the set-up error by the parallel and series connection of three independent neural network models.All models are trained and tested on hundreds of thousands of labeled DRRs and simulated X-ray images.Because the iterative optimization is not necessary in computing the set-up error via the trained network models,the speed of registration is close to real-time in the experiments.