Research On Multimodal Brain Image Registration Algorithms Based On Deep Learning

Accurate registration of multimodal brain images enables alignment of anatomical and functional images of the brain,resulting in more comprehensive and accurate information.This helps doctors to better understand the development and progression of brain diseases and to provide more accurate diagnosis and treatment options for patients.Multimodal medical image registration plays an important role in the prevention,diagnosis and treatment of brain diseases such as Alzheimer Disease and brain glioma.MRI images of the brain usually have multiple modalities such as T1,T2 and Flair in the actual diagnostic process,where T1-weighted images facilitate the clear presentation of anatomical structures and T2-weighted images reveal the lesions of tissues.By combining the two modalities of T1-weighted images and T2-weighted images,the spatial localization accuracy of anatomical structures and abnormal lesions in the brain can be improved,helping doctors to more accurately assess the size,location and morphology of lesions.In recent years,by introducing deep learning techniques into the field of medical image registration,the speed and accuracy of registration has been improved to a certain extent,but still faces some problems and difficulties.For example,For example,the lack of image boundary information can lead to registration errors in training.In addition,the folding point in registration results causes the model performance to fail to meet the clinical diagnosis requirements.To address the above issues,this paper investigates 2D brain T1 and T2 registration,designs novel model structure and loss function to optimise the model,and validates the optimised model registration performance on publicly available dataset.The innovations of this paper are as follows:(1)To address the registration errors caused by missing image boundaries,a dualconstrained multimodal medical image registration model based on boundary information enhancement is proposed for multimodal brain image registration,named BIE-Net.Firstly,a boundary enhancement network is introduced into the BIE-Net pre-training network to compensate for the missing T1 weighted image boundary information and reduce the registration errors caused by the lack of boundary information.Secondly,the number of deformation field folding points is reduced by designing a new loss function to reduce the prediction error of the images.The experimental results show that compared with advanced methods such as Voxel Morph,BIE-Net can effectively compensate for the T1 image boundary information and reduce the number of deformation field folding points to improve the registration performance of the model.(2)The classical deep learning based model training requires sufficient data support,and when only a few data are available,the performance of the model drops due to overfitting and poor generalization.To address this problem,a multimodal medical image registration model based on few-shot learning named Reverse-Net is proposed.Firstly,a reverse teaching network is designed to generate additional training data using the deformation field,which in turn passes more structural knowledge to the network and improves the ability of the model to extract features.In addition,multiple joint loss functions are introduced to improve the accuracy of registration and the training stability of the network.The experimental results show that Reverse-Net can reduce the number of segmentation labels used and improve the ability of the model to extract features without affecting the performance of the model,which in turn improves the registration accuracy of the model.