Research On Registration And Correction Of Breathing Motion Image Based On Convolutional Neural Network

Early screening and preventive diagnosis of cancer are particularly critical based on the high morbidity and mortality of cancer.Respiratory imaging equipment such as PET\CT can improve image support for the diagnosis of thoracic and abdominal tumors and assist doctors in diagnosis.However,during PET\CT imaging,respiratory motion can be generated due to limitations of scanning speed,collection efficiency and involuntary respiratory movements.Artifacts,which reduce image quality,cause trouble for doctors in diagnosis.Current respiratory motion artifact correction is still a challenge.For this reason,In this paper,a method of registration and correction of respiratory motion images based on Convolution Neural Network(CNN)is studied for the correction of respiratory motion artifacts.This paper mainly includes the following two aspects.(1)The traditional registration technologies are to align two or more images by iteratively optimizing the objective function,but the methods have a large amount of computation.Recent researches have shown that CNN of deep learning can be applied to image registration.At present,some studies have proposed using predefined registration examples to train registration model,but it is not easy to obtain predefined registration examples.For this reason,we construct an unsupervised and deformable deep learning image registration framework.The registration framework is mainly composed of 3D CNN,Spatial Transformer Networks(STN),etc.It can transform the pixel of one image into the corresponding position of another image,thus achieving deformable registration.(2)In this paper,two artifact correction methods are experimented on simulated PET/CT 3D geometric phantom data and pixel phantom data respectively.The first method is to use registration model to correct the artifact image directly;the second method is to use registration model to align different temporal frames,and then to fuse the registered temporal frames.In this paper,the artifact registration methods achieve Dice similarity coefficient(Dice)of82.12% and 83.76% on the simulated PET/CT 3D geometric phantom data and pixel phantom data respectively.In terms of operational efficiency,registration data of one case cost 5+0.5s and 11+0.5s respectively.The experimental results show the effectiveness and instantaneity of theproposed methods.