WO2022027216A1 - 一种图像降噪方法及其应用 - Google Patents

一种图像降噪方法及其应用 Download PDF

Info

Publication number
WO2022027216A1
WO2022027216A1 PCT/CN2020/106753 CN2020106753W WO2022027216A1 WO 2022027216 A1 WO2022027216 A1 WO 2022027216A1 CN 2020106753 W CN2020106753 W CN 2020106753W WO 2022027216 A1 WO2022027216 A1 WO 2022027216A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
self
noise reduction
network
reduction method
Prior art date
Application number
PCT/CN2020/106753
Other languages
English (en)
French (fr)
Inventor
郑海荣
李彦明
江洪伟
万丽雯
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/CN2020/106753 priority Critical patent/WO2022027216A1/zh
Publication of WO2022027216A1 publication Critical patent/WO2022027216A1/zh

Links

Images

Classifications

    • G06T5/70—
    • G06T5/60—
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Definitions

  • the present application belongs to the technical field of image scanning, and in particular relates to an image noise reduction method and its application.
  • Magnetic Resonance Imaging is a new examination technology based on the principle that atomic nuclei with magnetic distance can produce transitions between energy levels under the action of a magnetic field.
  • the flow situation is of great value in the diagnosis of degenerative diseases.
  • MRI is realized by the action of high-frequency magnetic field outside the body, and the signal is generated by the radiation energy from the substances in the body to the surrounding environment.
  • the imaging process is similar to image reconstruction and CT, but MRI does not rely on external radiation, absorption and reflection, nor does it rely on radioactive substances.
  • the gamma radiation in the body uses the interaction between the external magnetic field and the object to image, and the high-energy magnetic field is harmless to the human body.
  • Magnetic resonance imaging (MRI) scanning is one of the main medical imaging methods used for screening and diagnosis of human organ diseases, and is an effective tool for medical diagnosis.
  • noise generated during image acquisition or transmission can compromise MRI quality and severely reduce the accuracy of these diagnoses.
  • noisy low-quality MRI images can affect the accuracy of automated computer analyses such as classification, segmentation, and registration. Therefore, research on MRI denoising is of great significance for obtaining high-quality MRI output, and has important scientific significance and application prospects in the field of medical diagnosis.
  • the reconstructed MRI image has too much noise, which affects the doctor's diagnosis of the disease.
  • the present application Based on the problem that the reconstructed MRI image is too noisy under the existing fast imaging situation in a specific situation, which affects the doctor's diagnosis of the disease, the present application provides an image noise reduction method and its application.
  • the present application provides an image noise reduction method, the method includes the following steps:
  • Step 1 Build a self-correcting U-net convolutional neural network based on the U-net network
  • Step 2 Use the L 1 norm as the loss function
  • Step 3 Optimize the parameters of the self-correcting U-net convolutional neural network to make the loss function converge
  • Step 4 Take the image with noise as the input of the network, use the image with noise and the corresponding noise-free image as the network label, train the network, and obtain the mapping relationship from the image with noise to the image without noise;
  • Step 5 Denoise the image to be denoised through the trained network to obtain a denoised image.
  • the self-correcting convolutional neural network in the step 1 is an encoding and decoding network similar to the U-net network.
  • the self-correcting convolutional neural network includes a self-correcting convolution layer
  • the self-correcting convolution layer includes a first channel and a second channel.
  • Another implementation manner provided by the present application is to perform convolution and batch normalization processing on the data of the first channel.
  • Another implementation manner provided by the present application is: down-sampling the data of the second channel.
  • sampling rate is 4.
  • x i is the pixel value of the image before noise reduction
  • yi is the pixel value of the real image without noise
  • n is the total number of pixel values.
  • step 3 in the optimization in step 3, the Adam optimization algorithm is used to optimize the self-correcting convolutional neural network.
  • the present application also provides an application of an image noise reduction method, wherein the image noise reduction method according to any one of claims 1 to 9 is applied to single photon emission computed tomography, magnetic resonance imaging, low dose CT images or low counts Positron emission tomography.
  • the image denoising method provided by the present application is an under-sampling MRI image denoising method based on self-correction convolution.
  • the image noise reduction method provided in this application is based on processing human brain MR image noise reduction. Considering the fact that the MRI technique itself inherently adds Rician noise to the output image, denoising the MR image before taking the image for further processing yields clearer structural details than existing denoising methods, showing The ability of the present application for noise suppression and structure preservation of MRI images is demonstrated.
  • the image noise reduction method provided by the present application solves the problem of unclear image in some cases of fast MRI imaging, the reconstructed image is relatively noisy.
  • the image noise reduction method provided by the present application uses self-correcting convolution to replace the original convolution layer, which can increase the receptive field of the image without increasing the network parameters, so that the feature map can more accurately represent the features of the image.
  • the receptive field is defined as the area where the convolutional neural network feature can see the input image, in other words the feature output is affected by the pixels in the receptive field area. Therefore, the larger the receptive field, the more representative the feature map representing the image features.
  • the L 1 norm is used to replace the original L 2 norm, so that the image can better retain the edge details, so as to meet the needs of the doctor's diagnosis.
  • the image noise reduction method provided by the present application improves image quality.
  • the image denoising method provided by this application uses a self-correcting convolution to denoise images in the context of deep learning, which increases the receptive field without increasing network parameters, and further improves the quality of denoised images.
  • Fig. 1 is the generator network schematic diagram of the image noise reduction method of the present application
  • Fig. 2 is the schematic diagram of the self-correction convolution process of the present application
  • FIG. 3 is a schematic diagram of the self-calibration block of the present application.
  • the L1 norm is the sum of the absolute values of each parameter, which conforms to the Laplace distribution and is not completely differentiable, and there will be many corners on the image.
  • the L2 norm is the sum of the squares of the individual parameters, which conforms to a Gaussian distribution and is completely differentiable. Compared with L1, the edges and corners on the image are rounded a lot.
  • the present application provides an image noise reduction method, the method includes the following steps:
  • Step 1 Build a self-correcting U-net convolutional neural network based on the U-net network
  • Step 2 Use the L 1 norm as the loss function
  • Step 3 Optimize the parameters of the self-correcting U-net convolutional neural network to make the loss function converge
  • Step 4 Take the image with noise as the input of the network, use the image with noise and the corresponding noise-free image as the network label, train the network, and obtain the mapping relationship from the image with noise to the image without noise;
  • Step 5 Denoise the image to be denoised through the trained network to obtain a denoised image.
  • Step 3 is performed on the parameters in the network constructed in step 1, that is, the parameters are trained, and the parameters are continuously iteratively optimized and changed to achieve a good mapping relationship.
  • the self-correcting convolutional neural network in the step 1 is an encoding and decoding network similar to the U-net network.
  • the self-correcting convolutional neural network is an encoding and decoding network similar to the U-net network: the image undergoes a layer of convolution, a layer of self-correction convolution, pooling, a layer of self-correction convolution, pooling, and a layer of Self-correcting convolution, pooling, self-correcting convolution, self-correcting convolution, bilinear interpolation upsampling, self-correcting convolution, bilinear interpolation upsampling, self-correcting convolution, bilinear interpolation upsampling, self-correcting convolution Corrected convolution, convolution.
  • the number of convolution kernels are: 1, 16, 16, 32, 32, 64, 64, 128, 128, 128, 64, 64, 32, 32, 16, 16, 1.
  • the sizes of the images are: 128 ⁇ 128, 64 ⁇ 64, 32 ⁇ 32, 16 ⁇ 16, 32 ⁇ 32, 64 ⁇ 64, 128 ⁇ 128.
  • the network is trained by taking the noisy MRI images as the input of the network and the corresponding noise-free MRI images as the network labels.
  • Noise reduction is performed from the noisy MR image through the trained network to obtain a denoised image conforming to the doctor's diagnosis.
  • the self-correcting convolutional neural network includes a self-correcting convolution layer, and the self-correcting convolution layer includes a first channel and a second channel.
  • sampling rate is 4.
  • the self-correcting convolution structure in Figure 1 is shown in Figure 2.
  • the number of channels is evenly divided into the upper channel, that is, the second channel, and the lower channel, that is, the first channel.
  • the above data is processed by the self-correction module, and the data of the lower channel is merged with the above data after convolution and batch normalization operations, and the data of the same size as the input data is obtained as input.
  • the network structure of the self-correction block in Fig. 2 is shown in Fig. 3, and the dimension of the input data is C/2 ⁇ H ⁇ W.
  • the sampling rate is r, which is set to 4 in this application.
  • the dimension of the data after sampling is: C/2 ⁇ H/r ⁇ W/r, then convolution and batch normalization operations are performed, and then upsampling with a sampling rate of r is performed to restore the dimension of the data, and then it is performed with the input data. Adding, the result obtained goes through the sigmoid activation function.
  • the result of the convolution and batch normalization operations is multiplied by the previous results, and the result obtained is subjected to the convolution and batch normalization operations to obtain the final self-corrected convolution result.
  • the L 1 norm in the step 2 is:
  • x i is the pixel value of the image before noise reduction
  • yi is the pixel value of the real image without noise
  • n is the total number of pixel values.
  • the present application also provides an application of an image noise reduction method, wherein the image noise reduction method according to any one of claims 1 to 9 is applied to single photon emission computed tomography, magnetic resonance imaging, low dose CT images or low counts Positron emission tomography.
  • Step 1 Build a self-correcting convolutional neural network framework. This application replaces the original convolution module with the self-correcting convolution shown in Figure 2 on the basis of the U-net network;
  • Step 2 Use the L1 norm between the noise reduction result obtained after the low-quality magnetic resonance image in the paired data passes through the network designed in Step 1 and the high-quality as the loss function;
  • Step 3 Use the optimizer to optimize the loss function in step 2, iteratively optimize the parameters in the network designed in step 1, and finally make the loss function in step 2 converge;
  • Step 4 Give the network noise and noise-free data.
  • the noisy data passes through the network to obtain the L1 norm between the denoised data and the noise-free data as the loss function.
  • Use the optimizer to optimize it to change Step 1 Design the parameters in the network, and finally get the mapping of the network parameters in step 1;
  • Step 5 Pass the network parameters of the mapping relationship trained in Step 4 to the test image, and finally obtain a denoised image.

Abstract

一种图像降噪方法及其应用，该方法包括：构建自校正卷积神经网络；将L1范数作为损失函数；对自校正卷积神经网络进行优化；将带有噪声的图像作为网络的输入，将带有噪声的图像和对应的无噪声图像作为网络标签，对网络进行训练，得到由带噪声的图像到无噪声的图像映射关系；将需要降噪的图像通过训练好的网络进行降噪得到降噪后的图像。该方法可以解决图像不清楚的问题。

Description

一种图像降噪方法及其应用 技术领域
本申请属于图像扫描技术领域，特别是涉及一种图像降噪方法及其应用。
背景技术
磁共振成像(MRI)是根据有磁距的原子核在磁场作用下，能产生能级间的跃迁的原理而采用的一项新检查技术，MRI有助于检查癫痫患者脑的能量状态和脑血流情况，对变性病诊断价值很大。MRI是通过体外高频磁场作用，由体内物质向周围环境辐射能量产生信号实现的，成像过程与图像重建和CT相近，只是MRI既不靠外界的辐射、吸收与反射，也不靠放射性物质在体内的γ辐射，而是利用外磁场和物体的相互作用来成像，高能磁场对人体无害。
磁共振成像(MRI)扫描是用于人体器官疾病的筛查和诊断的主要医学成像方法之一，是用于医学诊断的有效工具。但是，在图像采集或传输过程中产生的噪声会损害MRI质量，并严重降低这些诊断的准确性。嘈杂的低质量MRI图像会影响自动计算机分析的准确性，例如分类，分割和配准。因此，对MRI去噪的研究对于获得高质量的MRI输出具有重要意义,同时对于医疗诊断领域具有重要的科学意义和应用前景。
现有的快速成像情况下，重建的MRI图像噪声过大，影响医生诊断病情。
发明内容
1.要解决的技术问题
基于在特定情况下，现有的快速成像情况下，重建的MRI图像噪声过大，影响医生诊断病情的问题，本申请提供了一种图像降噪方法及其应用。
2.技术方案
为了达到上述的目的，本申请提供了一种图像降噪方法，所述方法包括如下步骤：
步骤1：构建基于U-net网络基础上的自校正U-net卷积神经网络；
步骤2：将L 1范数作为损失函数；
步骤3：对自校正U-net卷积神经网络的参数进行优化，使得损失函数收敛；
步骤4：将带有噪声的图像作为网络的输入，将带有噪声的图像和对应的无噪声图像作为网络标签，对网络进行训练，得到由带噪声的图像到无噪声的图像映射关系；
步骤5：将需要降噪的图像通过训练好的网络进行降噪得到降噪后的图像。
本申请提供的另一种实施方式为：所述步骤1中所述自校正卷积神经网络为类似于U-net网络的编码解码网络。
本申请提供的另一种实施方式为：所述步骤1中自校正卷积神经网络包括自校正卷积层，所述自校正卷积层包括第一通道和第二通道。
本申请提供的另一种实施方式为：对所述第一通道的数据进行卷积和批归一化处理。
本申请提供的另一种实施方式为：对所述第二通道的数据进行下采样。
本申请提供的另一种实施方式为：所述采样率为4。
本申请提供的另一种实施方式为：所述步骤2中L 1范数为：
Figure PCTCN2020106753-appb-000001
其中，x i为降噪之前的图像的像素值，y i为无噪声的真实图像的像素值，n为像素值的总个数。
本申请提供的另一种实施方式为：所述步骤3中优化时采用Adam优化算法对所述自校正卷积神经网络进行优化。
本申请还提供一种图像降噪方法的应用，将权利要求1～9中任一项所述的图像降噪方法应用于单光子发射计算机断层成像、磁共振成像、低剂量CT图像或者低计数正电子发射型断层成像。
3.有益效果
与现有技术相比，本申请提供的一种图像降噪方法的有益效果在于：
本申请提供的图像降噪方法，为一种基于自校正卷积的欠采样MRI图像降噪方法。
本申请提供的图像降噪方法，基于处理人脑MR图像降噪。考虑到MRI技术本身会固有地将Rician噪声添加到输出图像这一事实，在将图像摄取进行进一步处理之前，对MR图像进行降噪，获得比现有降噪方法更加清晰的结构细节信息，展示了本申请对MRI图像的噪声抑制和结构保存的能力。
本申请提供的图像降噪方法，针对某些快速MRI成像情况下，重建的图像噪声比较大，解决图像不清楚的问题。
本申请提供的图像降噪方法，使用自校正卷积代替原始的卷积层，能在不增加网络参数的情况下增加图像的感受野，使得特征图更加准确的表示图像的特征。感受野被定义为卷积神经网络特征所能看到输入图像的区域，换句话说特征输出受感受野区域内的像素点的影响。所以，感受野越大，表述图像特征的特征图越具有代表性。
本申请提供的图像降噪方法，在损失函数中，使用L 1范数代替原始的L 2范数，使图像能更好的保留边缘细节，以此来满足医生诊断的需求。
本申请提供的图像降噪方法，提高图像质量。
本申请提供的图像降噪方法，在深度学习的背景下降噪图像，使用了一种自校正卷积，在不增加网络参数的同时增加感受野，使降噪的图像质量得到进一步的提升。
附图说明
图1是本申请的图像降噪方法的生成器网络示意图；
图2是本申请的自校正卷积过程示意图；
图3是本申请的自校正块示意图。
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。
L1范数是各个参数绝对值之和，符合拉普拉斯分布，是不完全可微的，表现在图像上会有很多角出现。
L2范数是各个参数的平方和，符合高斯分布，是完全可微的。和L1相比，图像上的棱角被圆滑了很多。
参见图1～3，本申请提供一种图像降噪方法，所述方法包括如下步骤：
步骤1：构建基于U-net网络基础上的自校正U-net卷积神经网络；
步骤2：将L 1范数作为损失函数；
步骤3：对自校正U-net卷积神经网络的参数进行优化，使得损失函数收敛；
步骤4：将带有噪声的图像作为网络的输入，将带有噪声的图像和对应的无噪声图像作为网络标签，对网络进行训练，得到由带噪声的图像到无噪声的图像映射关系；
步骤5：将需要降噪的图像通过训练好的网络进行降噪得到降噪后的图像。
对步骤1所构建的网络中的参数进行步骤3，即对其参数进行训练，不断的迭代优化改变其参数达到好的映射关系。
进一步地，所述步骤1中所述自校正卷积神经网络为类似于U-net网络的编码解码网络。
该自校正卷积神经网络是一种类似于U-net网络的编码解码网络：图像经过一层卷积，一层自校正卷积、池化、一层自校正卷积、池化、一层自校正卷积、池化、自校正卷积、自校正卷积、双线性插值上采样、自校正卷积、双线性插值上采样、自校正卷积、双线性插值上采样、自校正卷积、卷积。卷积核个数分别为：1，16，16，32，32，64，64，128，128，128，64，64，32，32，16，16，1.图像的大小分别为：128×128，64×64，32×32，16×16， 32×32，64×64，128×128。
将带有噪声的MRI图像作为网络的输入和对应的无噪声的MRI图像作为网络标签，对网络进行训练。
训练网络，得到由带噪声的MRI图像到无噪声的MRI图像映射关系G。
从带噪声的MR图像通过训练好的网络进行降噪得到符合医生诊断的降噪后的图像。
进一步地，所述步骤1中自校正卷积神经网络包括自校正卷积层，所述自校正卷积层包括第一通道和第二通道。
进一步地，对所述第一通道的数据进行卷积和批归一化处理。
进一步地，对所述第二通道的数据进行下采样。
进一步地，所述采样率为4。
图1中的自校正卷积结构图如图2所示。对于输入的数据将通道数平均分为上通道即第二通道和下通道即第一通道。上面的数据进行自校正模块的处理，下面通道的数据经过卷积和批归一化操作之后与上面的数据进行合并，得到与输入数据尺寸大小一样的数据作为输入。
图2中的自校正块网络结构如图3所示，输入数据的维度为C/2×H×W。对于上面的数据进行下采样，采样率为r，在本申请中设置为4。采样后的数据维度为：C/2×H/r×W/r,之后做卷积和批归一化操作，之后进行采样率为r的上采样恢复数据的维度，之后与输入的数据进行相加，得到的结果经过Sigmoid激活函数。
对于下面的数据经过卷积和批归一化操作的结果与之前上面的结果相乘，得到的结果经过卷积和批归一化操作的得到最终的自校正卷积结果。
进一步地，所述步骤2中L 1范数为：
Figure PCTCN2020106753-appb-000002
其中，x i为降噪之前的图像的像素值，y i为无噪声的真实图像的像素值，n为像素值的总个数。在本自校正卷积的U-net网络框架中，为了使图像细节更加尖锐，降噪后的图像保留细节边缘。
进一步地，所述步骤3中优化时采用Adam优化算法对所述自校正卷积神经网络进行优化。
本申请还提供一种图像降噪方法的应用，将权利要求1～9中任一项所述的图像降噪方法应用于单光子发射计算机断层成像、磁共振成像、低剂量CT图像或者低计数正电子发射型断层成像。
实施例
步骤一：搭建自校正卷积神经网络框架，本申请在U-net网路的基础上将原始的卷积模块换成图2所示的自校正卷积；
步骤二：将配对的数据中低质量磁共振图像经过步骤一中所设计的网络后得到的降噪结果与高质量之间的L1范数作为损失函数；
步骤三：对步骤二中的损失函数使用优化器进行优化，迭代的优化步骤一中所设计网路中的参数，最后使得步骤二中的损失函数达到收敛；
步骤四：给网络带噪声和无噪声的数据，带噪声的数据经过网络得到降噪后的数据与无噪声的数据之间的L1范数作为损失函数使用优化器对其进行优化来改变步骤一设计网络中的参数，最后得到步骤一中的网络参数的映射；
步骤五：对测试的图像经过步骤四所训练好的映射关系的网络参数，最终得到降噪后的图像。
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 (9)

  1. 一种图像降噪方法，其特征在于：所述方法包括如下步骤：
    步骤1：构建自校正卷积神经网络；
    步骤2：将L 1范数作为损失函数；
    步骤3：对自校正卷积神经网络进行优化,使得损失函数收敛；
    步骤4：将带有噪声的图像作为网络的输入，将带有噪声的图像和对应的无噪声图像作为网络标签，对网络进行训练，得到由带噪声的图像到无噪声的图像映射关系；
    步骤5：将需要降噪的图像通过训练好的网络进行降噪得到降噪后的图像。
  2. 如权利要求1所述的图像降噪方法，其特征在于：所述步骤1中所述自校正卷积神经网络为基于U-net网络基础上的自校正U-net卷积神经网络。
  3. 如权利要求1所述的图像降噪方法，其特征在于：所述步骤1中自校正卷积神经网络包括自校正卷积层，所述自校正卷积层包括第一通道和第二通道。
  4. 如权利要求3所述的图像降噪方法，其特征在于：对所述第一通道的数据进行卷积和批归一化处理。
  5. 如权利要求3所述的图像降噪方法，其特征在于：对所述第二通道的数据进行下采样。
  6. 如权利要求5所述的图像降噪方法，其特征在于：所述采样率为4。
  7. 如权利要求1所述的图像降噪方法，其特征在于：所述步骤2中L 1范数为：
    Figure PCTCN2020106753-appb-100001
    其中，x i为降噪之前的图像的像素值，y i为无噪声的真实图像的像素值，n为像素值的总个数。
  8. 如权利要求1所述的图像降噪方法，其特征在于：所述步骤3中优化时采用Adam优化算法对所述自校正卷积神经网络进行优化。
  9. 一种图像降噪方法的应用，其特征在于：将权利要求1～8中任一项所述的图像降噪方法应用于单光子发射计算机断层成像、磁共振成像、低剂量CT图像或者低计数正电子发射型断层成像。
PCT/CN2020/106753 2020-08-04 2020-08-04 一种图像降噪方法及其应用 WO2022027216A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/106753 WO2022027216A1 (zh) 2020-08-04 2020-08-04 一种图像降噪方法及其应用

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/106753 WO2022027216A1 (zh) 2020-08-04 2020-08-04 一种图像降噪方法及其应用

Publications (1)

Publication Number Publication Date
WO2022027216A1 true WO2022027216A1 (zh) 2022-02-10

Family

ID=80119236

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/106753 WO2022027216A1 (zh) 2020-08-04 2020-08-04 一种图像降噪方法及其应用

Country Status (1)

Country Link
WO (1) WO2022027216A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11540798B2 (en) 2019-08-30 2023-01-03 The Research Foundation For The State University Of New York Dilated convolutional neural network system and method for positron emission tomography (PET) image denoising
CN117372261A (zh) * 2023-12-04 2024-01-09 无锡日联科技股份有限公司 基于卷积神经网络的分辨率重建方法、装置、设备及介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108765291A (zh) * 2018-05-29 2018-11-06 天津大学 基于稠密神经网络和双参数损失函数的超分辨率重建方法
CN109978778A (zh) * 2019-03-06 2019-07-05 浙江工业大学 基于残差学习的卷积神经网络医学ct图像去噪方法
CN110349103A (zh) * 2019-07-01 2019-10-18 昆明理工大学 一种基于深度神经网络和跳跃连接的无干净标签图像去噪方法
CN110378845A (zh) * 2019-06-17 2019-10-25 杭州电子科技大学 一种基于卷积神经网络的极端条件下的图像修复方法
CN110503616A (zh) * 2019-08-28 2019-11-26 上海海事大学 一种应用于图片去噪的生成式网络
CN111310903A (zh) * 2020-02-24 2020-06-19 清华大学 基于卷积神经网络的三维单分子定位系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108765291A (zh) * 2018-05-29 2018-11-06 天津大学 基于稠密神经网络和双参数损失函数的超分辨率重建方法
CN109978778A (zh) * 2019-03-06 2019-07-05 浙江工业大学 基于残差学习的卷积神经网络医学ct图像去噪方法
CN110378845A (zh) * 2019-06-17 2019-10-25 杭州电子科技大学 一种基于卷积神经网络的极端条件下的图像修复方法
CN110349103A (zh) * 2019-07-01 2019-10-18 昆明理工大学 一种基于深度神经网络和跳跃连接的无干净标签图像去噪方法
CN110503616A (zh) * 2019-08-28 2019-11-26 上海海事大学 一种应用于图片去噪的生成式网络
CN111310903A (zh) * 2020-02-24 2020-06-19 清华大学 基于卷积神经网络的三维单分子定位系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LIU JIANG-JIANG; HOU QIBIN; CHENG MING-MING; WANG CHANGHU; FENG JIASHI: "Improving Convolutional Networks With Self-Calibrated Convolutions", 2020 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), IEEE, 13 June 2020 (2020-06-13), pages 10093 - 10102, XP033804900, DOI: 10.1109/CVPR42600.2020.01011 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11540798B2 (en) 2019-08-30 2023-01-03 The Research Foundation For The State University Of New York Dilated convolutional neural network system and method for positron emission tomography (PET) image denoising
CN117372261A (zh) * 2023-12-04 2024-01-09 无锡日联科技股份有限公司 基于卷积神经网络的分辨率重建方法、装置、设备及介质
CN117372261B (zh) * 2023-12-04 2024-02-27 无锡日联科技股份有限公司 基于卷积神经网络的分辨率重建方法、装置、设备及介质

Similar Documents

Publication Publication Date Title
CN108898642B (zh) 一种基于卷积神经网络的稀疏角度ct成像方法
CN111325686B (zh) 一种基于深度学习的低剂量pet三维重建方法
WO2022047625A1 (zh) 一种图像处理方法、系统和计算机存储介质
Kuanar et al. Low dose abdominal CT image reconstruction: An unsupervised learning based approach
CN107481297B (zh) 一种基于卷积神经网络的ct图像重建方法
CN111882503A (zh) 一种图像降噪方法及其应用
US9002134B2 (en) Multi-scale image normalization and enhancement
CN111429379B (zh) 一种基于自监督学习的低剂量ct图像去噪方法及系统
CN112258415B (zh) 一种基于生成对抗网络的胸部x光片超分辨率和去噪方法
Yang et al. Low-dose CT denoising via sinogram inner-structure transformer
WO2022110530A1 (zh) 基于spect数据采样与噪声特性的断层图像重建方法
WO2022027216A1 (zh) 一种图像降噪方法及其应用
CN110070510A (zh) 一种基于vgg-19提取特征的cnn医学图像降噪方法
Hou et al. CT image quality enhancement via a dual-channel neural network with jointing denoising and super-resolution
CN116091636A (zh) 基于双域增强的x射线差分相衬成像不完备数据重建方法
Li et al. Learning non-local perfusion textures for high-quality computed tomography perfusion imaging
Feng et al. Dual residual convolutional neural network (DRCNN) for low-dose CT imaging
Xia et al. Deep residual neural network based image enhancement algorithm for low dose CT images
Zhu et al. STEDNet: Swin transformer‐based encoder–decoder network for noise reduction in low‐dose CT
Chan et al. An attention-based deep convolutional neural network for ultra-sparse-view CT reconstruction
Liu et al. Cascade resunet with noise power spectrum loss for low dose ct imaging
CN113822796B (zh) 基于改进的图像金字塔的多模态脑部影像配准方法
Jin et al. Low-dose CT image restoration based on noise prior regression network
CN114926383A (zh) 基于细节增强分解模型的医学图像融合方法
Mahmoud et al. Variant Wasserstein Generative Adversarial Network Applied on Low Dose CT Image Denoising.

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: 20948675

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 17.07.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20948675

Country of ref document: EP

Kind code of ref document: A1