The Research Of MRI Non-rigid Registration Based On Multi-constrained Deep Network

In recent years,deep learning has been widely used in medical image processing,and their combination has changed the way medical staff diagnose diseases.Magnetic resonance imaging(MRI)has become one of the commonly used clinical examination methods with the advantages of non-radiation,high soft tissue resolution and high spatial resolution.However,the non-rigid deformation caused by the unconscious movement of the patient during MRI acquisition process affects the anatomical correspondence between the acquired images.This paper aims to solve the problem of non-rigid deformation between MRI images by using the registration technology based on deep learning.To this end,this paper proposes a registration method based on deep learning to reduce the complexity of registration network and construct the relationship between generating deformation field.Meanwhile,because the intensity changes between images will affect the registration results,a deformation consistency loss is designed and tested on cardiac MRI and breast dynamic contrast-enhanced MRI(DCE-MRI).The work of this paper is mainly divided into the following three parts.(1)Research on deformation registration network based on multiple constraints: To correct the non-rigid deformation among MRIs,this paper proposes a multi-constrained network(MC-Net)for unsupervised MRI registration.The different output information is obtained by exchanging the positions of the moving image and the fixed image in the deformation prediction network,and then combines these outputs to construct similarity constraints,regularization constraints and cycle symmetry constraints between different images.Finally,the loss function formed by these constraints is used to train the registration network.The experimental result of dice similarity coefficient is 0.767,and the result of jacobian determinant is 0.195%.(2)Registration model based on new deformation consistency loss function: To solve the problem that most of the existing deep learning deformation registration networks are trained based on the similarity between images,the registration results are easily affected by intensity changes.A deformation consistency loss is designed to establish constraints between the deformation fields generated at different stages in the multi-module registration network.Experiments have proved that on the basis of cardiac MRI dataset,the results with the loss function are better than those without this loss function in the intensity preserve and the deformation field generation.The result of SSIM increased by 0.013,and the average result of RMSE is 0.00297.(3)Research on DCE-MRI non-rigid registration based on deformation consistency loss function: To solve the lesions distortion caused by intensity variation in DCE-MRI registration process,MC-Net with deformation consistency loss function is used to register breast DCE-MRI.First,this paper combines commonly used data augmentation methods to augment breast DCE-MRI data to solve the problem of insufficient data.Then,the MC-Net combined with deformation consistency loss is trained on the breast augmentation data to obtain the registration model.Finally,the experiment proves that the proposed method can effectively solve the problem of lesion distortion caused by the obvious intensity variation between images during DCE-MRI deformation registration,and improve the accuracy of registration.