Research On Key Technologies Of Human PET Image Registration

Positron emission computed tomography(PET)technology is an image scanning technology that injects radionuclides labeled with a substance into the human body and reflects the metabolic activities of life through the accumulation of the substance in the human body.PET images of human body are of great significance to the exploration of systemic diseases,focal metabolism and focal metastasis.However,the scanning time of dynamic PET image is too long and the patient’s local limb movement,even breathing and heartbeat will lead to the deviation of human PET image.Therefore,it is necessary to make spatial alignment of human PET image,which is register the human PET image.At present,human body PET image registration faces two problems:on the one hand,when there is a large offset of local limbs,the registration effect of most registration methods on limbs is not good.On the other hand,due to the huge size of PET image itself and the transformation of human body,the amount of parameters of the model is too large to be trained.This paper combines the ideas of human modeling,point cloud registration and neural network and The main research contents are as follows:(1)Aiming at the problem of hinge registration with large offset of local limbs in human PET image,a hinge registration method of human PET image based on point cloud mapping is proposed in this paper.Firstly,the human body is modeled by combining the parametric human model and human skin modeling method,and the point cloud is determined and extracted by combining the Laplace mapping method,then the hinge registration is carried out on the point cloud image.Finally combined with the interpolation method,The point cloud transformation field obtained from point cloud registration is transformed into a smooth pixel transformation field.This method is tested on the real 3D human body PET image data set provided by Henan Provincial People’s Hospital and the control experiment includes two classical traditional registration methods,three registration methods based on deep learning and a hinge registration method.The results show that,this method completes the articulated registration task of human PET image and the registration accuracy reaches 93%.(2)Due to the large size of 3D human PET scanning data,the training parameters,computation of convolutional neural network are huge and the registration time is too long,a symmetric lightweight depth convolution network registration method is proposed to solve these problems in the registration process.Methods the similarity between images is maximized in the differential homeomorphic mapping space and the forward transform and inverse transform are estimated symmetrically to ensure the topology and global smoothness of the transform.At the same time,the ordinary convolution in the convolution neural network is replaced by ghost convolution and the model is pruned to realize the effective compression of the model.Results this method is tested on the real PET scanning 3D data set provided by Henan Provincial People’s Hospital and the latest registration algorithm is selected as the control experiment.The results show that this method can greatly maintain the topology and smoothness of the original image while the dice evaluation reaches 97%,greatly reduce the amount of network calculation and parameters and accelerate the registration speed of PET scanning 3D data.Conclusion the symmetric lightweight depth convolution network registration method proposed in this paper can significantly reduce the demand for computing resources and speed up the registration while maintaining a high level of registration effect.