【作者】 王彬；

【导师】 曹秒；

【作者基本信息】 长春理工大学 ， 生物医学工程， 2023， 硕士

【摘要】 医学图像配准在病灶定位、疾病诊断、手术导航、放射治疗等领域发挥着极其重要的作用,实现高精度、高效率的配准对于辅助临床诊疗具有积极意义。传统配准算法迭代速度慢、普适性差,难以达到临床领域实时配准的目标要求。近年随着深度学习的快速发展,卷积神经网络在图像配准领域得到广泛使用。基于深度学习的图像配准主要分为基于图像对的相似性度量以及直接利用深度回归网络预测变形场两类方法,前者容易受到数据集中异常值的影响,且对配准参数估计进行迭代计算或进行非线性配准时处理时效性差,后者利用深度神经回归网络直接对转换参数进行预测的配准速度快、精度高。本文遵循直接预测变形场的配准方式,对基于深度学习的单模态MR脑图像配准技术进行研究,主要研究内容如下:1)实验数据集预处理。首先去除对配准工作不具有实际意义的组织干扰信息,通过对MR脑部图进行脑提取和重采样,保留图像配准感兴趣的区域。其次对大小不一、灰度分布范围不均的影像数据进行剪裁和归一化操作,在缩放体素后进行仿射对齐。最后采取变形的方式实现数据增强,提高模型的鲁棒性和泛化能力。2)提出基于分布式反馈网络的无监督可变形图像配准算法。首先为提高网络特征能力,搭建基于U-Net的网络作为配准框架内部反馈模块。其次为增强配准图像对特征信息的交流与充分融合,通过级联的方式完成反馈模块的层次连接,完成配准子网络的构建。最后为促进网络内部的高阶变形信息修正低层次表征,利用多次迭代子网络得到分布式反馈配准网络FIR-Net,以获取具有强大表示的变形场空间向量,提高配准的精度。3)提出反馈结合注意力机制的无监督可变形配准算法。首先针对U-Net与反馈机制结合时导致损失下降时的不稳定问题,在反馈模块中引入通道与空间注意力机制,促使网络学习远程依赖关系,提高梯度下降在远距离输出时的稳定性。其次为降低影像解剖结构高相似性对精配准的难度,在FIR-Net内部级联极化自注意力机制得到空间反馈配准网络SF-Net,以降低高低阶信息融合中特征图谱之间语义壁垒,提升模型学习获取重点变形信息的能力。最后在三个脑部MR图像数据集上进行实验并通过量化与可视化分析,证明算法在医学图像配准中的可行性。更多还原

【Abstract】 Image registration has great application value in the field of computer and medicine.However,the traditional registration algorithm has slow iteration speed,poor universality and low registration accuracy,which is difficult to meet the requirements of real-time registration in clinical field.In recent years,with the rapid development of deep learning,convolutional neural networks have been widely used in the field of image registration.Image registration based on deep learning is mainly divided into two methods: similarity measurement based on image pairs and direct use of deep regression network to predict deformation field.The former method is easily affected by outliers in data sets,and the processing timeliness is poor when iterative calculation of registration parameter estimation or nonlinear registration is carried out.The latter method uses deep neural regression network to directly predict the conversion parameters.The registration speed is fast and the accuracy is high.This paper follows the registration mode of directly predicting the deformation field,and studies the single-mode MR brain image registration technology based on deep learning.The main contents are as follows:1)The data set used in the experiment is preprocessed.In order to remove different tissue information and interference information that are not of practical significance to the registration work of this paper,such as skull and neck,firstly,the brain extraction and resampling of MR brain map are carried out,leaving only the region of interest in image registration.Secondly,the image data of different sizes and uneven gray distribution range are cropped and normalized,and the affine alignment is carried out after scaling the voxel.This paper also adopts the method of deformation to realize data enhancement and improve the robustness and generalization ability of the model.2)An unsupervised deformable image registration algorithm based on distributed feedback network is proposed.In this paper,the framework of the registration network is improved,and the matching U-Net feedback medium network is built.The distributed feedback registration network FIR-Net is obtained by iteratively cascading the U-Net network generation and feedback.By increasing the number of cascades and iterations of feedback blocks in FIR-Net,the high-order deformation information is used to correct the low-level representation,so that the high-level and low-level information of MR feature maps can be fully integrated.Finally,a powerful high-level representation is extracted for obtaining the spatial vector of the deformation field,and the deformation information of higher-level registration image pairs is obtained to improve the registration accuracy.3)An unsupervised deformable registration algorithm combining feedback and attention mechanism is proposed.In order to improve the instability of the loss function combined with U-Net and feedback mechanism during network training,the feedback module is integrated with the channel and spatial attention mechanism,so that the network can learn the remote dependence and improve the stability of gradient descent in longdistance output.In addition,in order to alleviate the difficulty of high similarity of image anatomical structure to fine registration,the spatial feedback registration network SF-Net is obtained by cascading the polarization self-attention mechanism within the feedback network,which reduces the semantic barrier between feature maps in high and low order information fusion,and promotes the ability of network learning to obtain key deformation information,and further improves the registration accuracy.In this paper,sufficient experiments are carried out on three brain MR image datasets and quantitative and visual analysis are carried out to prove the feasibility of the proposed algorithm in medical image registration tasks.更多还原