WO2022032445A1 - 一种重建神经网络及其应用 - Google Patents

一种重建神经网络及其应用 Download PDF

Info

Publication number
WO2022032445A1
WO2022032445A1 PCT/CN2020/108251 CN2020108251W WO2022032445A1 WO 2022032445 A1 WO2022032445 A1 WO 2022032445A1 CN 2020108251 W CN2020108251 W CN 2020108251W WO 2022032445 A1 WO2022032445 A1 WO 2022032445A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
convolutional neural
domain
reconstructed
image
Prior art date
Application number
PCT/CN2020/108251
Other languages
English (en)
French (fr)
Inventor
郑海荣
李彦明
江洪伟
万丽雯
张其阳
胡战利
Original Assignee
深圳高性能医疗器械国家研究院有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳高性能医疗器械国家研究院有限公司 filed Critical 深圳高性能医疗器械国家研究院有限公司
Priority to PCT/CN2020/108251 priority Critical patent/WO2022032445A1/zh
Publication of WO2022032445A1 publication Critical patent/WO2022032445A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects

Definitions

  • the present application belongs to the technical field of image reconstruction, and in particular relates to a reconstruction neural network and its application.
  • CT computed tomography
  • CT X-ray computed tomography
  • mammography and high-pitch helical CT the data acquisition angle is usually limited by the size of the object being acquired and the flexibility of scanning. These factors will lead to incomplete data in the Radon transform domain, a problem known as the finite angle problem, which presents serious challenges in CT imaging tasks.
  • CT imaging requires the detector and the X-ray light source to rotate around the target to be measured once (360 degrees) to collect data, and then use the Filter-backprojection (FBP) algorithm for reconstruction.
  • FBP reconstruction algorithm is based on the traditional signal processing theory, and the data needs to be complete and enough data collected for one week.
  • this application provides a reconstruction neural network and its application.
  • the application provides a kind of reconstruction neural network, including the first convolutional neural network part, the domain transformation module and the second convolutional neural network part;
  • the first convolutional neural network part is used for the learning of filters and weighted weights under different geometric imaging conditions under different ray metrology conditions;
  • the domain transformation module is used for the forward flow of data from the sinusoidal domain to the image domain, and the back-propagation of gradient errors from the image domain to the sinusoidal domain;
  • the second convolutional neural network part is used to further strengthen the function of the first-stage filtering and weighting network and the processing of artifacts.
  • the first convolutional neural network part is a one-layer or multi-layer structure.
  • the first convolutional neural network part has four layers.
  • the second convolutional neural network part is a one-layer or multi-layer structure.
  • the second convolutional neural network part has 18 layers.
  • the second convolutional neural network part includes 4 residual connections.
  • the first convolutional neural network part is a filter weighting network
  • the second convolutional neural network part is a residual codec network
  • the first convolutional neural network part, the domain transformation module is cascaded with the second convolutional neural network part.
  • the domain transformation module adopts a back-projection transformation algorithm
  • the back-projection transformation algorithm supports forward propagation of data
  • the back-projection transformation algorithm supports back-propagation of errors
  • Another embodiment provided by the present application is: the training of the reconstructed neural network adopts the Adam optimization algorithm.
  • the present application also provides an application of a reconstructed neural network, where the reconstructed neural network is applied to X-ray CT reconstruction, ultrasonic tomography or terahertz tomography.
  • the reconstructed neural network provided in this application is aimed at computed tomography (CT) systems in the medical and industrial fields.
  • the reconstruction neural network provided in this application is a mixed-domain convolutional neural network.
  • the reconstruction neural network provided in this application is used to reduce streak artifacts in CT reconstructed images in the case of limited-angle acquisition scans.
  • the reconstructed neural network provided by this application embeds the analytical algorithm of traditional domain transformation into the network, which can very effectively avoid occupying a huge amount of computing resources.
  • the reconstructed neural network provided by the present application, after the domain transformation network, cascades an encoding and decoding residual network to solve the shortcomings of the FBP algorithm and the FBP mapping network algorithm.
  • the reconstruction neural network provided in this application is a deep neural network spanning two domains for finite-angle CT reconstruction.
  • the neural network learns the filters and weight coefficients for CT reconstruction in the sinusoidal domain, and learns the removal of artifacts in the image domain.
  • the reconstructed neural network provided by the present application adopts the analytical algorithm to realize the transformation from the sinusoidal domain to the CT image domain, avoiding the occupation of a huge amount of computing resources when the fully connected layer is used to realize the domain transformation. Backpropagation.
  • the second convolutional neural network part adopts the residual and dimensionality reduction structure, which can effectively perform the correction of artifacts and some other potential problems (scattering, noise, etc.).
  • the reconstructed neural network provided by this application can get rid of the drawbacks brought by the existing mapping FBP algorithm.
  • the result can be obtained directly after the back-projection operation of the network.
  • the network in front of the projection is also strictly limited to learning filters and weights, so the inherent flexibility of the neural network itself cannot be effectively utilized.
  • this application fully releases the structure of the previous network of domain transformation, and uses (Convolutional Neural Network) CNN to replace the existing fully connected layer, so that it can freely learn filters and weight parameters.
  • the reconstruction neural network provided by the present application after back-projection, cascades the coding and decoding residual network, which is used to further deal with the problem of stripe artifacts caused by the limited angle reconstruction. Compared with the existing method, it only learns filtering before the domain transformation to reduce Striping artifacts, the effect is very noticeable.
  • Fig. 1 is the reconstruction neural network architecture schematic diagram of the present application
  • FIG. 2 is a schematic diagram of the comparison results of different methods of the present application.
  • the data acquisition angle is usually limited by the size of the object to be acquired and the scanning flexibility, resulting in the inability to acquire complete data.
  • traditional reconstruction algorithms are directly used. , will cause severe streaking artifacts.
  • the complete data here refers to: the data in the range of 180 degrees in the case of parallel beams, and the data in the range of 180 degrees + the fan angle of the fan beam in the case of fan beams. Less than the perfect acquisition angle is the limited angle acquisition mode.
  • the present application provides a reconstruction neural network, including a first convolutional neural network part, a domain transformation module and a second convolutional neural network part;
  • the first convolutional neural network part is used for the learning of filters and weighted weights under different geometric imaging conditions under different ray metrology conditions;
  • the domain transformation module is used for the forward flow of data from the sinusoidal domain to the image domain, and the back-propagation of gradient errors from the image domain to the sinusoidal domain;
  • the second convolutional neural network part is used to further strengthen the function of the first-stage filtering and weighting network and the processing of artifacts.
  • the first convolutional neural network part is a one-layer or multi-layer structure.
  • the first convolutional neural network part has 4 layers.
  • the second convolutional neural network part is a one-layer or multi-layer structure.
  • the second convolutional neural network part has 18 layers.
  • the second convolutional neural network part includes 4 residual connections.
  • the first convolutional neural network part is a filter weighting network
  • the second convolutional neural network part is a residual codec network; the first convolutional neural network part, the domain transformation module and the The second convolutional neural network is cascaded.
  • the domain transformation module adopts a back-projection transformation algorithm, the back-projection transformation algorithm supports the forward propagation of data, and the back-projection transformation algorithm supports the backward propagation of errors.
  • the training of the reconstructed neural network adopts the Adam optimization algorithm.
  • the present application also provides an application of a reconstructed neural network, where the reconstructed neural network is applied to X-ray CT reconstruction, ultrasonic tomography or terahertz tomography.
  • the network consists of three parts: the first part is the convolutional neural network acting in the sine image domain, that is, the first convolutional neural network part, and the second part is the domain transformation operation used to connect the sine domain and the CT image domain to realize the dual domain
  • the data flows forward and reverse, and the last one acts on the convolutional neural network in the CT image domain, that is, the second convolutional neural network part.
  • This application proposes a hybrid domain reconstruction neural network to solve the CT reconstruction problem in the limited angle acquisition mode.
  • the network consists of three parts cascaded: a filtering weighting network, a domain transform module and a residual codec network.
  • the overall framework of the network is shown in Figure 1.
  • the upper right part is the overall architecture diagram, and the CNN-A part and the CNN-B part are the detailed internal structure diagram of the neural network.
  • (p, q), (m, n) represent image dimensions.
  • the following numbers indicate the number of feature maps in this layer.
  • Arrows indicate the flow of data.
  • Stride(2, 2) represents the stride of the CNN convolution process.
  • the functional principle of this network can be expressed in mathematical symbols as follows:
  • the acquired sinusoid is represented by y ⁇ R p ⁇ q
  • the target CT image to be reconstructed is represented by x ⁇ R m ⁇ n .
  • the filter function is represented by F
  • the weight matrix is represented by W
  • the domain transformation operation is represented by T bp
  • the encoder is represented by E
  • the decoder is represented by D
  • the function of the network can be represented by the following functions:
  • the main function of the first convolutional neural network part, the CNN-A part, is to learn the filter function F and the weight matrix W from the training set.
  • the domain transform module implements Tab in the formula .
  • the main function of the second convolutional neural network part, the CNN-B part, is to learn the encoder E and the decoder D from the training set.
  • the specific implementation method of the CNN-A part is: using a 4-layer CNN (L1-L4) structure (4 layers are used for description here, which can be 1 layer or more layers).
  • CNN-B has 4 residual connections: the output of the L2 layer and the input of the L17 layer, the output of the L4 layer and the input of the L15 layer, the output of the L6 layer and the input of the L13 layer, the output of the L8 layer and the input of the L11 layer. enter.
  • CNN-B has three dimensionality reduction processes: L3, L5 and L7 layers, and three dimensionality raising processes: L11, L13 and L15 layers.
  • the specific implementation form of the domain transformation module is the back-projection transformation algorithm.
  • the transformation algorithm supports both forward propagation of data and backward propagation of errors.
  • the forward propagation formula of the back projection algorithm is as follows:
  • I(x,y) represents the output feature image of this module.
  • S(t,c) represents the input feature image of this module, which is derived from the CNN-A part.
  • c A(x, y, t) means that the spatial point (x, y) is projected onto the detector according to the current acquisition geometric form, and the position of the projection point on the detector is c; t means the t-th acquisition angle.
  • the forward propagation of the back-projection algorithm is to realize the accumulation of the data collected by the detector c, which has a projection relationship with the spatial point (x, y) under all the collection angles P.
  • the back-projection algorithm error back propagation formula is as follows:
  • Err represents the error returned by the domain transformation module
  • Loss represents the output error of the entire network.
  • the training of the network uses the Adam optimization algorithm, the initial learning rate is 3 ⁇ 10 -5 , and the learning rate decays by 0.98 times after every 1000 steps.
  • the implementation carrier of the present application is program code, and the program code can be implemented and written in any mainstream deep learning framework (Tensorflow, Pytorch, Caffe, etc.).
  • This application can use either a software carrier (the software carrier used in this application), or a dedicated hardware, such as an FPGA, as a carrier. It is completely possible to solidify the trained network into the hardware to realize this application.
  • the present application is mainly aimed at the finite angle problem, because the finite angle problem is difficult to deal with by traditional methods.
  • the present application can also be used for noise removal during reconstruction with low loading parameters (low voltage, low current, etc.).
  • the present application can also be used for CT reconstruction problems in the case of sparse sampling (if a circle is defined to be 360 degrees, and the acquisition of one image data per degree is full sampling, then sparse sampling is similar to the case of 2 degrees apart and one image data is acquired at 4 degrees).
  • the present application is not limited to be used in X-ray CT reconstruction, it can be applied to all fields of reconstruction using tomography theory, such as ultrasonic tomography, terahertz tomography, and the like.

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Graphics (AREA)
  • Geometry (AREA)
  • Software Systems (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

本申请属于图像重建技术领域，特别是涉及一种重建神经网络及其应用。当投影数据覆盖的扫描角度范围小于180度(平行束层析成像)时，在重建的图像中会出现严重的伪影，因此，在有限角度采集条件下将无法实现高质量的图像重建。本申请提供了一种重建神经网络，包括第一卷积神经网络部、域变换模块和第二卷积神经网络部；所述第一卷积神经网络部，用于不同射线计量情况下滤波器与不同几何成像条件下加权权重的学习；所述域变换模块，用于数据从正弦域到图像域的正向流动，以及梯度误差从图像域到正弦域的反向传播；所述第二卷积神经网络部，用于进一步加强第一级滤波加权网络功能，以及伪影的处理。在图像域学习伪影的去除。

Description

一种重建神经网络及其应用 技术领域
本申请属于图像重建技术领域，特别是涉及一种重建神经网络及其应用。
背景技术
计算机辅助的X射线扫描在对大脑疾病和损伤的诊断中起着革命性的作用。传统的X光检查最多只能产生一幅大脑投影图像，这样的影像分辨力不高。为了解决这个问题，研究者设计了计算机断层扫描(computedtomography，简称CT)。CT是以X射线从多个方向沿着头部某一选定断层层面进行照射，测定透过的X射线量，数字化后经过计算机算出该层面组织各个单位容积的吸收系数，然后重建图像的一种技术。这是一种图质好、诊断价值高而又无创伤、无痛苦、无危险的诊断方法。它使我们能够在任何深度或任何角度重建脑的各种层面结构。CT能够显示出脑创伤后遗症、损伤、脑瘤和其他大脑病灶的位置，这样，也就可以通过CT来诊断一个人行为变化在脑水平上的病因。
X射线计算机断层扫描(CT)可以产生清晰，高质量的图像，并且在临床诊断和手术图像引导中起着非常重要的作用。CT可以帮助放射科医生识别和诊断传染病，肌肉骨骼疾病，心血管疾病，外伤甚至某些种类的癌症。但是，在CT的一些实际应用中，例如乳腺X线摄影和大螺距螺旋CT，数据采集角度通常受被采集对象尺寸和扫描灵活性的限制。这些因素将导致Radon变换域中的数据不完整，这个问题称为有限角度问题，这在CT成像任务中提出了严峻的挑战。通常情况下，CT成像需要探测器与X射线光源围绕待测目标旋转一周(360度)采集数据，之后采用Filter-backprojection(FBP)算法进行重建。FBP重建算法基于传统信号处理理论，需要数据是完备的采集一周的足够多的数据。
当投影数据覆盖的扫描角度范围小于180度(平行束层析成像)时，在重建的图像中会出现严重的伪影，因此，在有限角度采集条件下将无法实现高质量的图像重建。
发明内容
1.要解决的技术问题
基于当投影数据覆盖的扫描角度范围小于180度(平行束层析成像)时，在重建的图像中会出现严重的伪影，因此，在有限角度采集条件下将无法实现高质量的图像重建的问题，本申请提供了一种重建神经网络及其应用。
2.技术方案
为了达到上述的目的，本申请提供了一种重建神经网络，包括第一卷积神经网络部、域 变换模块和第二卷积神经网络部；
所述第一卷积神经网络部，用于不同射线计量情况下滤波器与不同几何成像条件下加权权重的学习；
所述域变换模块，用于数据从正弦域到图像域的正向流动，以及梯度误差从图像域到正弦域的反向传播；
所述第二卷积神经网络部，用于进一步加强第一级滤波加权网络功能，以及伪影的处理。
本申请提供的另一种实施方式为：所述第一卷积神经网络部为一层或者多层结构。
本申请提供的另一种实施方式为：所述第一卷积神经网络部为4层。
本申请提供的另一种实施方式为：所述第二卷积神经网络部为一层或者多层结构。
本申请提供的另一种实施方式为：所述第二卷积神经网络部为18层。
本申请提供的另一种实施方式为：所述第二卷积神经网络部包括4个残差连接。
本申请提供的另一种实施方式为：所述第一卷积神经网络部为滤波加权网络，所述第二卷积神经网络部为残差编码解码器网络；所述第一卷积神经网络部、所述域变换模块与所述第二卷积神经网络部级联。
本申请提供的另一种实施方式为：域变换模块采用反投影变换算法，所述反投影变换算法支持数据的正向传播，所述反投影变换算法支持误差的反向传播。
本申请提供的另一种实施方式为：所述重建神经网络的训练采用Adam优化算法。
本申请还提供一种重建神经网络的应用，将所述重建神经网络应用于X射线CT重建、超声波断层成像或者太赫兹断层成像。
3.有益效果
与现有技术相比，本申请提供的一种重建神经网络的有益效果在于：
本申请提供的重建神经网络，针对医学和工业领域计算机断层扫描(CT)系统。
本申请提供的重建神经网络，为一种混合域卷积神经网络。
本申请提供的重建神经网络，用于减少有限角度采集扫描情况下CT重建图像中的条纹伪影。
本申请提供的重建神经网络，将传统域变换的解析算法嵌入到网络中，可以非常有效的避免占用巨量的计算资源。
本申请提供的重建神经网络，在域变换网络后面级联上编码解码残差网络，解决FBP算法以及FBP映射网络算法的缺点。
本申请提供的重建神经网络，为一种跨越两个域的深度神经网络用于有限角度CT重建。 神经网络在正弦域学习CT重建时的滤波器与权重系数，在图像域学习伪影的去除。
本申请提供的重建神经网络，采用解析算法实现正弦域到CT图像域的变换，避免采用全连接层实现域变换时对巨量计算资源的占用，本申请算法支持数据的正向传播与误差的反向传播。
本申请提供的重建神经网络，第二卷积神经网络部采用残差与降维结构，可以有效进行伪影校正与其他一些潜在问题(散射、噪声等)的校正。
本申请提供的重建神经网络，摆脱了现有的映射FBP算法带来的弊端，其网络的反投影操作后就直接得到结果，无法对有限角度重建条状伪影问题的进行处理，且其反投影前面的网络也严格限制为学习滤波器与权重，所以神经网络自身固有的灵活性得不到有效发挥。本申请针对现有方法的不足，充分放开域变换前面的网络的结构，采用(Convolutional Neural Network)CNN取代现有的全连接层，使其自由的学习滤波器与权重参数。
本申请提供的重建神经网络，在反投影之后级联编码解码残差网络，用于进一步处理有限角度重建带来的条状伪影问题，比起现有方法只在域变换之前学习滤波来降低条状伪影，效果非常显著。
附图说明
图1是本申请的重建神经网络架构示意图；
图2是本申请的不同方法对比结果示意图。
具体实施方式
在下文中，将参考附图对本申请的具体实施例进行详细地描述，依照这些详细的描述，所属领域技术人员能够清楚地理解本申请，并能够实施本申请。在不违背本申请原理的情况下，各个不同的实施例中的特征可以进行组合以获得新的实施方式，或者替代某些实施例中的某些特征，获得其它优选的实施方式。
在CT的一些实际应用中，例如乳腺X线摄影和大螺距螺旋CT，数据采集角度通常受被采集对象尺寸和扫描灵活性的限制，导致无法采集到完备的数据，此时直接采用传统重建算法，将导致严重的条状伪影。此处完备数据是指：在平行束时180度范围的数据，扇形束时为180度+扇形束扇角范围的数据。小于完备采集角度即是有限角度采集模式。
参见图1～2，本申请提供一种重建神经网络，包括第一卷积神经网络部、域变换模块和第二卷积神经网络部；
所述第一卷积神经网络部，用于不同射线计量情况下滤波器与不同几何成像条件下加权权重的学习；
所述域变换模块，用于数据从正弦域到图像域的正向流动，以及梯度误差从图像域到正弦域的反向传播；
所述第二卷积神经网络部，用于进一步加强第一级滤波加权网络功能，以及伪影的处理。
进一步地，所述第一卷积神经网络部为一层或者多层结构。
进一步地，所述第一卷积神经网络部为4层。
进一步地，所述第二卷积神经网络部为一层或者多层结构。
进一步地，所述第二卷积神经网络部为18层。
进一步地，所述第二卷积神经网络部包括4个残差连接。
进一步地，所述第一卷积神经网络部为滤波加权网络，所述第二卷积神经网络部为残差编码解码器网络；所述第一卷积神经网络部、所述域变换模块与所述第二卷积神经网络部级联。
进一步地，所述域变换模块采用反投影变换算法，所述反投影变换算法支持数据的正向传播，所述反投影变换算法支持误差的反向传播。
进一步地，所述重建神经网络的训练采用Adam优化算法。
本申请还提供一种重建神经网络的应用，将所述重建神经网络应用于X射线CT重建、超声波断层成像或者太赫兹断层成像。
该网络包括三个部分：第一部分是作用于正弦图域中的卷积神经网络即第一卷积神经网络部，第二部分是域变换操作用来连接正弦域与CT图像域，实现双域数据正反向流动，最后一个作用于CT图像域中的卷积神经网络即第二卷积神经网络部。
实施例
本申请提出了一种混合域重建神经网络用于解决有限角度采集模式下的CT重建问题，网络由三部分级联而成：滤波加权网络、域变换模块与残差编码解码器网络。网络的整体框架如图1所示。
右上部分为总体架构图，CNN-A部分与CNN-B部分为神经网络的详细内部结构图。图示中(p，q)，(m，n)表示图像维度。后面的数字表示本层的特征图的数目。箭头表示数据的流向。Stride(2，2)表示CNN卷积过程的步进。
本网络的功能原理可以采用数学符号表示如下：有限角度采集模式下，采集到的正弦图采用y∈R p×q表示，待重建的目标CT图像采用x∈R m×n表示。滤波函数采用F表示，权重矩阵采用W表示，域变换操作采用T bp表示，编码器采用E表示，解码器采用D表示，网络的功能可以用如下函数表示：
Figure PCTCN2020108251-appb-000001
其中
Figure PCTCN2020108251-appb-000002
第一卷积神经网络部即CNN-A部分的主要功能是从训练集中学习滤波函数F和权重矩阵W。域变换模块实现公式中的T ab。第二卷积神经网络部即CNN-B部分的主要功能是从训练集中学习编码器E与解码器D。
CNN-A部分具体实现方式是：采用4层CNN(L1-L4)结构(本处采用4层进行说明，可以是1层，也可以是更多层)。
CNN-B部分具体实现方式是：采用18层CNN(L1-18)结构(本处采用18层进行说明，可以是1层，也可以是更多层)。CNN-B具有4个残差连接：分别连接L2层的输出与L17层的输入，L4层的输出与L15层的输入，L6层的输出与L13层的输入，L8层的输出与L11层的输入。CNN-B具有三次降维过程：L3层、L5层和L7层，三次升维过程：L11、L13和L15层。
域变换模块具体实现形式是采用的反投影变换算法，变换算法即支持数据的正向传播，也支持误差的反向传播。反投影算法正向传播公式如下：
Figure PCTCN2020108251-appb-000003
其中，I(x,y)表示本模块的输出特征图像。S(t,c)表示本模块的输入特征图像，其来源于CNN-A部分。c＝A(x,y,t)表示根据当前采集几何形式，将空间点(x,y)投影到探测器上，投影点在探测器上的位置为c；t表示第t个采集角度。反投影算法正向传播就是实现将所有采集角度P下的与空间点(x,y)具有投影关系的探测器c点采集到的数据累加起来。反投影算法误差反向传播公式如下：
Figure PCTCN2020108251-appb-000004
其中，
Figure PCTCN2020108251-appb-000005
Err表示域变换模块回传得到的误差，Loss表示整个网络的输出端误差。
网络的训练使用Adam优化算法，初始学习率为3×10 -5，每1000步后学习率衰减0.98倍。本申请的实施载体是程序代码，程序代码可以在任何主流深度学习框架(Tensorflow、Pytorch、Caffe等)中实施编写。
如图2所示，随着有限角度的减少，其他方法条状伪影越来越严重，而本申请对条状伪 影的抑制作用非常明显。
从实验结果可以看出在有限角度为120度时，本申请明显好于其他方法，本申请非常适合处理有限角重建问题中的伪影问题。
本申请既可以采用软件载体(本申请采用的软件载体)，也可以采用专用的硬件作为载体，比如FPGA等。完全可以将训练好的网络固化到硬件里面实现本申请。
本申请主要针对有限角问题，是因为有限角问题采用传统方法很难处理。本申请同样也可以用于低加载参数(低电压，低电流等)重建时的噪声去除。本申请同样可以用于稀疏采样(如果定义一周360度，每一度采集一张图像数据为全采样，那么稀疏采样就是类似间隔2度，4度采集一张图像数据)情况下的CT重建问题。
本申请不限于用于X射线CT重建中，其可以适用于所有采用断层成像理论进行重建的领域中，比如超声波断层成像，太赫兹断层成像等。
尽管在上文中参考特定的实施例对本申请进行了描述，但是所属领域技术人员应当理解，在本申请公开的原理和范围内，可以针对本申请公开的配置和细节做出许多修改。本申请的保护范围由所附的权利要求来确定，并且权利要求意在涵盖权利要求中技术特征的等同物文字意义或范围所包含的全部修改。

Claims (10)

  1. 一种重建神经网络，其特征在于：包括第一卷积神经网络部、域变换模块和第二卷积神经网络部；
    所述第一卷积神经网络部，用于不同射线计量情况下滤波器与不同几何成像条件下加权权重的学习；
    所述域变换模块，用于数据从正弦域到图像域的正向流动，以及梯度误差从图像域到正弦域的反向传播；
    所述第二卷积神经网络部，用于进一步加强第一级滤波加权网络功能，以及伪影的处理。
  2. 如权利要求1所述的重建神经网络，其特征在于：所述第一卷积神经网络部为一层或者多层结构。
  3. 如权利要求2所述的重建神经网络，其特征在于：所述第一卷积神经网络部为4层。
  4. 如权利要求1所述的重建神经网络，其特征在于：所述第二卷积神经网络部为一层或者多层结构。
  5. 如权利要求4所述的重建神经网络，其特征在于：所述第二卷积神经网络部为18层。
  6. 如权利要求5所述的重建神经网络，其特征在于：所述第二卷积神经网络部包括4个残差连接。
  7. 如权利要求1～6中任一项所述的重建神经网络，其特征在于：所述第一卷积神经网络部为滤波加权网络，所述第二卷积神经网络部为残差编码解码器网络；所述第一卷积神经网络部、所述域变换模块与所述第二卷积神经网络部级联。
  8. 如权利要求7所述的重建神经网络，其特征在于：所述域变换模块采用反投影变换算法，所述反投影变换算法支持数据的正向传播，所述反投影变换算法支持误差的反向传播。
  9. 如权利要求8所述的重建神经网络，其特征在于：所述重建神经网络的训练采用Adam优化算法。
  10. 一种重建神经网络的应用，其特征在于：将权利要求1～9中任一项所述的重建神经网络应用于X射线CT重建、超声波断层成像或者太赫兹断层成像。
PCT/CN2020/108251 2020-08-10 2020-08-10 一种重建神经网络及其应用 WO2022032445A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/108251 WO2022032445A1 (zh) 2020-08-10 2020-08-10 一种重建神经网络及其应用

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/108251 WO2022032445A1 (zh) 2020-08-10 2020-08-10 一种重建神经网络及其应用

Publications (1)

Publication Number Publication Date
WO2022032445A1 true WO2022032445A1 (zh) 2022-02-17

Family

ID=80247471

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/108251 WO2022032445A1 (zh) 2020-08-10 2020-08-10 一种重建神经网络及其应用

Country Status (1)

Country Link
WO (1) WO2022032445A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109509235A (zh) * 2018-11-12 2019-03-22 深圳先进技术研究院 Ct图像的重建方法、装置、设备及存储介质
CN110047113A (zh) * 2017-12-29 2019-07-23 清华大学 神经网络训练方法和设备、图像处理方法和设备和存储介质
CN110211194A (zh) * 2019-05-21 2019-09-06 武汉理工大学 一种基于深度学习去除稀疏角度ct成像伪影的方法
CN110660123A (zh) * 2018-06-29 2020-01-07 清华大学 基于神经网络的三维ct图像重建方法和设备以及存储介质
US20200027252A1 (en) * 2018-06-04 2020-01-23 Korea Advanced Institute Of Science And Technology Method for processing multi-directional x-ray computed tomography image using artificial neural network and apparatus therefor
CN110728729A (zh) * 2019-09-29 2020-01-24 天津大学 一种基于注意机制的无监督ct投影域数据恢复方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110047113A (zh) * 2017-12-29 2019-07-23 清华大学 神经网络训练方法和设备、图像处理方法和设备和存储介质
US20200027252A1 (en) * 2018-06-04 2020-01-23 Korea Advanced Institute Of Science And Technology Method for processing multi-directional x-ray computed tomography image using artificial neural network and apparatus therefor
CN110660123A (zh) * 2018-06-29 2020-01-07 清华大学 基于神经网络的三维ct图像重建方法和设备以及存储介质
CN109509235A (zh) * 2018-11-12 2019-03-22 深圳先进技术研究院 Ct图像的重建方法、装置、设备及存储介质
CN110211194A (zh) * 2019-05-21 2019-09-06 武汉理工大学 一种基于深度学习去除稀疏角度ct成像伪影的方法
CN110728729A (zh) * 2019-09-29 2020-01-24 天津大学 一种基于注意机制的无监督ct投影域数据恢复方法

Similar Documents

Publication Publication Date Title
JP6855223B2 (ja) 医用画像処理装置、ｘ線コンピュータ断層撮像装置及び医用画像処理方法
Yu et al. Data consistency based rigid motion artifact reduction in fan-beam CT
Yuan et al. SIPID: A deep learning framework for sinogram interpolation and image denoising in low-dose CT reconstruction
Yu et al. Data consistency based translational motion artifact reduction in fan-beam CT
WO2022110530A1 (zh) 基于spect数据采样与噪声特性的断层图像重建方法
JP5590548B2 (ja) Ｘ線ｃｔ画像処理方法，ｘ線ｃｔプログラムおよび該プログラムが搭載されたｘ線ｃｔ装置
AU2019271915A1 (en) Method and system for motion correction in CT imaging
JP3987024B2 (ja) 横方向のフィルタリング処理を用いたトモシンセシス画像を強調する方法及びシステム
Zhang et al. PET image reconstruction using a cascading back-projection neural network
Fournié et al. CT field of view extension using combined channels extension and deep learning methods
CN115984394A (zh) 一种联合先验图像及卷积稀疏网络的低剂量ct重建方法
Zhang et al. DREAM-Net: Deep residual error iterative minimization network for sparse-view CT reconstruction
Ketola et al. Deep learning-based sinogram extension method for interior computed tomography
CN107886554B (zh) 流数据的重构
CN113344876A (zh) 一种ct和cbct间可变形配准方法
WO2022032445A1 (zh) 一种重建神经网络及其应用
CN111950705A (zh) 一种重建神经网络及其应用
Ma et al. Low dose CT reconstruction assisted by an image manifold prior
Manhart et al. Fast dynamic reconstruction algorithm with joint bilateral filtering for perfusion C-arm CT
US20220375038A1 (en) Systems and methods for computed tomography image denoising with a bias-reducing loss function
Slavine et al. Phantom and preclinical studies for image improvement in clinical CT
Hsieh et al. An algorithm to estimate the object support in truncated images
KR102225747B1 (ko) 단일 뉴럴 네트워크를 이용한 다중 관심영역의 내부 단층촬영 영상 처리 방법 및 그 장치
Wang et al. Improving the quality of sparse-view cone-beam computed tomography via reconstruction-friendly interpolation network
Nuyts et al. Iterative FDK reconstruction for helical CT of the head with rigid motion compensation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20948944

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 04.07.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20948944

Country of ref document: EP

Kind code of ref document: A1