Deep Learning-Based Study For Brain Magnetic Resonance Image Registration

Brain magnetic resonance image registration is a critical step in various neuroimaging studies,such as those of Alzheimer’s disease,neuropsychiatric disorders and attention deficit and hyperactivity disorder.In addition,high-performance registration can significantly improve the correctness of the analysis and thus provide a solid foundation for neuroimaging studies.Conventional image registration algorithms usually adjust parameters for different image pairs in an iterative optimization manner,which makes them computationally intensive and time-consuming,and thus not suitable for clinical applications.In recently,deep learning-based image registration algorithms have been proven to have higher registration performance and computational efficiency,and thus become the mainstream of current research.However,there are still some shortcomings in existing studies that need to be further investigated,such as large deformation due to individual subject differences or disease progression,insufficient characterization ability of deep learning-based image registration networks,and the effectiveness of registration algorithms on downstream tasks.Based on the above background and challenges,this paper focuses on the design of registration networks based on unsupervised learning strategies and proposes two deep learning-based brain magnetic resonance image registration algorithm as follows:(1)Multi-resolution transformer-based deformable registrationTo address the problem of large deformation caused by individual differences,this paper proposes a multi-resolution Transformer registration network(MTReg-Net),which considers both long-range spatial relations and local detail information in images via Transformer and convolutional neural networks,and the multi-resolution framework is used to optimize the deformation field in a coarse-to-fine manner to better align image regions with large deformations.In this paper,we performed extensive quantitative and qualitative evaluations on four neuroimaging public datasets and demonstrated that MTReg-Net outperforms existing registration algorithms in terms of alignment accuracy.In addition,statistical analysis is performed on all brain anatomical structures,demonstrating that MTReg-Net can effectively align large deformation regions compared to other registration algorithms.(2)Multi-scale complexity-aware convolutional neural network for deformable registrationTo further consider the large deformations caused by individual differences or disease progression,the insufficient representation ability of the registration network,and the effectiveness of the registration algorithm on downstream tasks,this paper designs a multiscale complexity-aware registration network(MSCAReg-Net),which progressively improves the estimation of deformation via a feature transformation stage and a complexityaware stage.Specifically,U-Net is used in the feature transformation stage to realize the transformation from image features to deformation magnitude features,followed by the design of a multi-scale complexity-aware module(MSCA-Module)in the complexityaware stage to perceive and further optimize deformations with different complexity.In addition,a feature calibration module(FC-Module)and a feature aggregation module(FAModule)are integrated in U-Net to enhance the deformation representation capability of UNet and further consolidate the recognition capability of MSCA-Module in terms of deformation complexity.The experimental results show that MSCAReg-Net outperforms existing registration algorithms in terms of registration accuracy,especially SyN which has the best performance.Notably,the ablation experimental analysis confirms the effectiveness of MSCA-Module in terms of perceiving deformations of different complexity and proves the effectiveness of FC-Module and FA-Module in enhancing the deformation representation capability.Furthermore,the evaluation results on downstream tasks demonstrate the potential of MSCAReg-Net for large applications.Besides,the evaluation results on downstream tasks validate that MSCAReg-Net has a large application potential.