Unsupervised Registration Model And Registration Algorithms Evaluation For Lung CT

Registration of lung CT images is an important research topic in non-rigid medical image registration.By registering high-resolution CT images with different respiratory phases,it is possible to master the changing rule of lung tissue of patients during breathing movement.This has practical significance for predicting the movement of lung tissue,tracking radiotherapy for lung tumors,and diagnosing chronic obstructive pulmonary disease.In recent years,due to the potential to meet the instantaneous and high-precision requirements of clinical applications,deep learning has begun to be applied to certain rigid or non-rigid medical image registration.This paper explores suitable models for lung CT registration based on deep learning.First,our models take unlabeled 3D image pairs as input,transform the deep convolutional neural network as a function of parameter sharing.The displacement field is obtained by the Networks.Then we use spatial transformation and displacement field to warp the float ing image.The similarity between fixed image and warped image and the regularization term of the displacement field are the objective functions to obtain the optimal parameters.The registration process is unsupervised.We design two models with different depths.The registration effects of different optimization methods and different convolution kernel sizes under the two models are explored.The research is conducted on public data sets and self-obtained data sets.Experiments proved that designs with Adam optimizer and deeper levels,and using larger convolution kernels when obtaining the displacement field have higher accuracy and stronger robustness.Different designs run dozens even hundreds of times faster than iterative-based approaches.In addition,the evaluation of registration algorithms,as another major development direction of the registration technology,has no uniform gold standard.Therefore,this paper proposes a comprehensive evaluation framework,including visualization results,image quality,lung shape,global and local displacement fields,and running time,to achieve different aspects comparison of the algorithm.Six widely used registration algorithms were selected in the study,and the public lung CT image data set was used to verify the comprehensive evaluation ability of the process.The results show that the proposed evaluation framework can effectively compare the registration algorithms.