Non-rigid Medical Image Registration Based On Deep Learning

Medical image registration is an important task in the field of medical image processing,which is of great significance for image fusion,tumor detection and other clinical work.Image registration aims to find the optimal spatial transformation that maps one image to another.The traditional registration methods iteratively optimize the objective function of each image pair to find the spatial transformation,which has some disadvantages such as long registration time and large amount of calculation.In recent years,with the extensive application of deep learning in medical image research,deep learning-based image registration has become a popular research subject.Although the supervised deep learning-based registration algorithm has improved both the speed and accuracy,it is still difficult to obtain reliable label for supervised training,which leads to the development of unsupervised deep learning-based registration methods in order to overcome the difficulty of obtaining label.This thesis mainly studies the unsupervised deep learning-based registration method to solve the non-rigid registration problem of 3D medical images.This thesis proposes three different unsupervised registration methods,all of which use convolutional neural networks to directly predict the deformation field between image pairs,and obtain registered images after interpolation.Compared with the traditional registration method,it has the advantages of completing the registration without iteration,and can achieve accurate and fast registration.(1)This thesis firstly proposes an unsupervised end-to-end registration method LNet based on 3D convolutional neural network.This method back-propagates image similarity and gradient information to train the network.In LNet,a parallel convolution kernel block LKB is designed.The LKB decomposes the ordinary large convolution kernel,which can enlarge the receptive field and further improve the performance for large deformations estimation.The registration experiment results on the public brain dataset and the cardiac dataset prove that the registration accuracy of LNet is comparable to the state-of-the-art methods,and the registration speed is faster.(2)Considering that only relying on image similarity and gradient information to train the network,the constraints of unsupervised registration are insufficient.On the basis of LNet,this thesis proposes a cycle-consistent image registration framework CIRNet.CIRNet proposes a new registration optimization goal:warp the moving image to a fixed image,and then warp the registration result back to the moving image.The original moving image could be obtained.To achieve this goal,CIRNet introduces a cycle-consistency constraint,that is m→ G(m)→ F(G(m))≈m.Cycle-consistency constraint is more robust to changes in grayscale of image pairs.Quantitative experiments in the public brain dataset and cardiac dataset demonstrate that the registration accuracy of CIRNet is better than the traditional algorithms including SyN and BSpline and other deep learning registration algorithms.Cycle-registration framework and cycle-consistency constraint can significantly improve the accuracy of unsupervised registration algorithms.(3)Based on the cycle-consistent image registration framework CIRNet,this thesis proposes a dual-channel registration network CIRNet-2ch for multimodal medical image registration,which provides a new idea for unsupervised multimodal image registration.This method uses a two-channel network to separately extract image features of different modals and then fuse image features,cycle-consistency is used to constraint multimodal image registration.In this paper,the clinical cardiac CT-MR images are used to evaluate the multimodal registration algorithm,and the heart contour points are marked to compare the results of different registration methods.Experiments show that the proposed method CIRNet-2ch achieves the best results.