Research On Lung CT Image Registration Algorithm Based On Multi-scale Fully Convolutional Neural Network

Lung cancer is one of the major diseases that threaten human health in recent years.As a key technology for adjuvant treatment of lung cancer,image registration has been widely used in lung tumor radiation therapy.At present,the development of deep learning technology has made the application prospect of medical image registration broader,but the existing algorithms still have many problems such as poor registration effect and low operation efficiency in lung CT image registration.In view of the current technical defects,this paper studies the non-rigid registration algorithm of lung CT images based on multiscale fully convolutional neural network.The main research results are as follows:Aiming at the problems that existing deep learning image registration algorithms need supervised information,low registration accuracy of large deformation images and poor real-time registration,this paper proposes an unsupervised multi-scale parallel fully convolutional neural network for non-rigid registration of lung CT images.algorithm.The algorithm uses a multi-scale parallel down-sampling module to reduce the image size and obtains a multi-scale low-resolution feature map;the pyramid dilated convolution module is used to extract image information from multi-scale features to improve the receptive field of the model.The algorithm adjusts the emphasis of the network on different deformation features through the adaptive channel attention module,so as to solve the problem that the model is biased towards larger deformation features and has poor registration effect for small deformation;at the same time,the smoothness constraint is added to the loss function to improve the deformation image.and improve the stability and generalization of the algorithm model by augmenting the training sample data.At the same time,the algorithm can perform end-to-end image registration without any supervision information.In the DIRlab and Creatis dataset tests,the target registration errors of the proposed algorithm are1.71 mm and 1.50 mm,respectively,while the one-time iteration target registration errors of the fully convolutional neural network algorithm are 2.83 mm and 2.01 mm,respectively.The experimental results show that the target registration error of the algorithm in this paper is significantly smaller than that of the full convolutional neural network algorithm,and the generalization and stability of the algorithm are also good.Real-time.Aiming at the problems of large network model,slow model training speed,low convergence efficiency,and easy overfitting during training in large-scale networks,this paper proposes a multi-scale full convolution based on joint training.A neural network-based non-rigid image registration algorithm for lung CT.The algorithm uses unsupervised joint training to decompose a large fully convolutional neural network into three small fully convolutional networks of different scales.Computational resources improve training efficiency;at the same time,the miniaturized fully convolutional subnetwork adopts double convolution module for feature extraction,which reduces the number of network layers and solves the problem of easy overfitting of the model;finally,a joint loss function is used to compare the three parameters.A fully convolutional subnetwork is weighted to calculate the image difference loss,and the weight of each scale subnetwork is balanced to improve the stability and accuracy of the algorithm.In the DIR-lab and Creatis dataset tests,the target registration errors of the algorithm are 1.56 mm and1.30 mm,respectively,and the computational complexity of the model is greatly reduced,and the training convergence speed is significantly improved.Experiments show that the multi-scale fully convolutional neural network algorithm based on joint training has better registration effect,and can obtain deformation models with higher registration accuracy while consuming less computing resources and better real-time performance.