Research On Fundus Image Registration Algorithm Via Deep Learning

Image registration is the operation of transforming the corresponding images into the same coordinate system through geometric space transformation.It is often used as a preprocessing step of information fusion.Fundus images contain rich clinical diagnostic information,and the registration of different fundus images can assist clinical diagnosis: by registering fundus images of different periods,the changes and development of the patient’s condition can be tracked; by registering fundus images of different modal,multimodal information can be combined to assist clinical diagnosis.However,the traditional fundus image registration often requires a large amount of calculation,and the registration parameter reuse rate is low,while the registration method based on deep learning shows the advantages of being fast and can maintain good registration accuracy.The fundus registration algorithm based on deep learning is studied.Specifically,this thesis studies the unimodal registration of color fundus images and the multimodal registration of color fundus image-OCT images respectively:(1)Unimodal color fundus image registration.Aiming at the problem that the dataset is too small in the deep learning-based method for color fundus image registration,a data synthesis method based on image pair metamorphosis decoupling is proposed.This thesis makes full use of the pairing data in the registration task to decouple the lesions and deformations between image pairs and recombine them to expand the dataset.Meanwhile,the loss function suitable for fundus image is selected and the registration model is optimized.Experimental results on the dataset of color fundus images show that the RMSE of our method reaches 0.380,and the DSC reaches 0.790,which are significantly better than the current unsupervised methods.Meanwhile,our data synthesis method can achieve stable improvement in multiple registration models.(2)Multimodal color fundus-OCT images registration.Aiming at the large difference in size and position between the color fundus images and the OCT images and the difficulty of training registration models,this thesis proposes an unsupervised registration method based on OCT vessels.A two-stage registration model is used to simplify the registration process and speed up the registration,while improving the registration effect.In order to take into account large displacement and small deformation,an affine registration network is employed for initial registration firstly,and then a deformable registration network is used for fine registration.During model training,the synthetic unimodal color fundus image is used to pre-train the affine registration network,and the multimodal data is used for further training to get the final registration model parameters.The experimental results on the color fundus-OCT images dataset show that the RMSE of our method reaches 1.334,and the DSC reaches 0.644.The registration result is significantly better than the traditional method,and the registration speed is improved.