Research On Image Registration Of Thoracic Aorta

In thoracic aortic disease surgery,doctors can only see X-ray images that do not show vascular structures,so we need to combine preoperative CT and intraoperative X-ray images showing aortic vessels.To this end,this paper proposes a deep neural network registration algorithm based on branch codec structure to provide navigation for surgery.The main work and innovations of this paper are as follows:(1)In this paper,a real-time thoracic aortic preoperative CT slice group and Digitally Reconstructed Radiograph(DRR)method are used to create a superimposed DRR image database based on cast transformation.-This database was used as a training set and test set for deep neural networks to achieve pose estimation of aortic DRR images.This database is created by the ray tracing-based DRR method.The simulated X-rays are projected on the CT slice group and contain more than 50,000 images after spatial geometric transformation.(2)A deep neural network Inception-branch based on the codec branch structure is proposed to estimate the pose of the DRR image of the thoracic aorta.The encoder part of the network is composed of Inception structure,which has good depth and width,and has strong ability to deal with complex problems.It can extract useful features in complex image internal structure and complete regression training.The decoder portion of the network is a branch structure.Since the displacement transformation and the angular transformation in the image transformation are different in the image,the displacement transformation and the angular transformation are separately decoded and calculated to reduce mutual interference between them.Finally,different loss functions are defined for the calculation of displacement and angle to improve the accuracy of parameter regression.Using the created superimposed DRR image database training,experiments and analysis including different regressions,decoders,encoders,cropping methods and various branching structures were carried out,and finally an optimal neural network structure was found.In terms of displacement error and angle error,the network added to the branch decoding structure is more than 30%higher than the network without branch decoding.(3)The deep neural network Inception-branch proposed in this paper is used to register the preoperative CT of the thoracic aorta with the real intraoperative X-ray image,and achieve high precision in visual perception and objective index evaluation.The validity and superiority of the proposed algorithm are proved.In this paper,all the experiments are based on data collected from real operations.The time is reduced from more than 3 minutes to about 2 seconds.At the same time,from the precision point of view,the deep neural network with branch decoding structure is better than the deep neural network without branch decoding structure,which reduces the error by 30%,which proves the superiority of the proposed algorithm.From the registration results,the aortic vascular structure in the DRR image can be clearly and completely superimposed on the intraoperative X-ray image,which reduces the root mean square error by 2.1%.