Research On Non-Rigid 3D Brain Image Registration Algorithm Based On Unsupervised

As part of medical image processing,medical image registration plays an essential role in preoperative planning,intraoperative guidance and postoperative health recovery in clinical medicine.Although traditional registration methods are well established,they are time-consuming and extremely depend on doctors’ professional skills.As the variety and quantity of medical images becomes more complex and varied,the work of the medical staff becomes more difficult,with missed and misdiagnosed cases occurring.In recent years,deep learning algorithms have come into the limelight.Research into deep learning registration algorithms is now becoming increasingly hot.Although,deep learning-based registration algorithms have improved substantially in terms of efficiency and precision,there are still some problems with some complex and large-shaped anatomical structures.Therefore,this thesis focuses on building a class of unsupervised and non-rigid registration methods based on Voxel Morph networks to achieve accurate and fast end-to-end registration using deep learning networks to complete the registration task directly.The details are as follows:(1)Build the backbone feature extraction network Voxel Morph,consisting of the design of the convolutional layer,pooling layer and activation function.The unimodal loss function,multimodal loss function and deformation field regularization loss function are designed for unimodal and multimodal registration problems respectively.The original data are pre-processed,including conversion of2 D images to 3D images,resampling,and normalisation.Finally,the traditional iterative registration and deep learning techniques are fused to build an affine preregistration model to pre-align the images,reduce the number of parameters in the network training process,speed up training and improve the precision of registration.(2)To address the problem of unimodal brain MRI registration,this thesis designs a novel unsupervised non-rigid unimodal registration network,"DIT-IVNet".Firstly,this thesis cascades the Transformer to the bottom of the Voxel Morph downsampling,which has a stronger feature extraction effect in training compared to the original Voxel Morph.Secondly,we designe a lightweight multi-scale feature extraction module that cascades into the downsampling phase of the backbone network,sampling convolutional kernels at different scales to extract features.At the same time,the fixed-size patch method in the original Vision Transformer is improved to an adaptive Depatch method,which prevents the missing image information in learning and effectively reduces the training parameters.Finally,comparative,ablation,and generalization experiments are set up in this thesis to demonstrate the superiority and generalization of the designed algorithm.(3)For the multimodal brain MRI registration problem,the feature extraction capability and learning capability of the registration network are relatively high,considering the huge differences in image pixel information between different modalities.In this thesis,the attention mechanism module is cascaded in the downsampling phase of the Voxel Morph network,thus effectively preventing the missing image information during the training process.At the same time,a lightweight residual network designe in this thesis is cascaded to enhance the learning capability of Voxel Morph,thus deepening the network number and enhancing the learning capability.We compare the better registration methods to date,validate the superiority of the multimodal registration algorithm in this thesis,and test it on different brain datasets to verify the generalization capability of the network.