WO2022011690A1 - 一种自监督学习方法及应用 - Google Patents

一种自监督学习方法及应用 Download PDF

Info

Publication number
WO2022011690A1
WO2022011690A1 PCT/CN2020/102732 CN2020102732W WO2022011690A1 WO 2022011690 A1 WO2022011690 A1 WO 2022011690A1 CN 2020102732 W CN2020102732 W CN 2020102732W WO 2022011690 A1 WO2022011690 A1 WO 2022011690A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
self
supervised learning
learning method
dose
Prior art date
Application number
PCT/CN2020/102732
Other languages
English (en)
French (fr)
Inventor
江洪伟
郑海荣
李彦明
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/CN2020/102732 priority Critical patent/WO2022011690A1/zh
Publication of WO2022011690A1 publication Critical patent/WO2022011690A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models

Definitions

  • the present application belongs to the technical field of computed tomography (CT) systems in the medical and industrial fields, and in particular relates to a self-supervised learning method and application.
  • CT computed tomography
  • Computed Tomography is a non-invasive imaging detection method that obtains tomographic images of the patient's body through computers and X-rays. It has the advantages of short scanning time, low cost and a wide range of disease monitoring. for early screening and routine physical examination of the disease. However, a large amount of X-ray exposure will cause the cumulative effect of radiation dose, which will greatly increase the possibility of various diseases, thereby affecting the physiological functions of the human body, destroying human tissues and organs, and even endangering the life safety of patients.
  • CT Computed Tomography
  • CT imaging quality is poor under existing low-dose conditions.
  • the present application Based on the existing problem of poor CT imaging quality under low-dose conditions, the present application provides a self-supervised learning method and application.
  • the present application provides a self-supervised learning method, which comprises the following steps:
  • Step 1 Model the image noise
  • Step 2 generate the approximate target pixel value of the current pixel from the low-dose CT image, and obtain the target pixel point;
  • Step 3 randomly crop an image block from the low-dose CT input image, randomly select N pixels from the image block, and replace the target pixel with the currently selected pixel to obtain the target image;
  • Step 4 Train the network and gradually reach a convergence state.
  • the noise image is composed of a clean image and noise.
  • the current pixel point of the input low-dose CT image is Randomly select pixels in the 5 ⁇ 5 neighborhood as the target pixel of the current pixel.
  • step 3 an image block x j with a size of 64 ⁇ 64 pixels is randomly cropped from the low-dose CT input image.
  • Another embodiment provided by the present application is: in the step 3, the image block is larger than the receptive field of the selected convolutional neural network.
  • N is one tenth of the size of the image block.
  • training the network includes calculating the loss of the N pixel points.
  • the loss function is a mean square error loss function.
  • Another implementation manner provided by the present application is: the network framework adopts the Unet structure as a whole, and the network is optimized by the Adam optimizer.
  • the present application also provides an application of self-supervised learning, in which the self-supervised learning method according to any one of claims 1 to 9 is applied to CT image noise reduction or MRI image noise reduction.
  • the self-supervised learning method provided in this application is a self-supervised learning method for low-dose CT image noise reduction.
  • the self-supervised learning method provided in this application is based on the idea of the self-supervised learning method.
  • the method of this application can generate labels from input data (low dose CT images). Therefore, the training requirements of the neural network can be met, so that the neural network can obtain the ability to denoise CT images under low-dose conditions through learning.
  • the self-supervised learning method provided in this application generates the target image by randomly replacing some pixels in the neighborhood of the input low-dose CT image, thereby breaking through the data requirements of the traditional convolutional neural network for training sets, and abandoning supervision While learning the shortcomings, the powerful feature extraction ability of the convolutional neural network is fully utilized, which greatly improves the CT imaging quality under low-dose conditions.
  • the self-supervised learning method provided in this application utilizes the powerful feature extraction capability of the convolutional neural network, and can train the network end-to-end without manual intervention, and achieve noise reduction.
  • the self-supervised learning method provided in this application does not require normal dose CT images, has low data requirements, and has huge advantages in the absence of clinical paired data.
  • Fig. 1 is the first schematic diagram of the convolutional neural network based on supervised learning of the present application
  • FIG. 2 is a second schematic diagram of the supervised learning-based convolutional neural network of the present application.
  • the encoder part consists of 5 layers of convolution with kernel size of 5 ⁇ 5 and the ReLU activation function
  • the decoder part consists of 5 layers of deconvolution and ReLU activation functions corresponding to the encoder.
  • the encoder The convolution and deconvolution layers corresponding to the decoder use residual connections similar to the residual network, and the final network output obtains the denoised CT image.
  • the final output of the generator is the input image minus the last layer of the generator's convolution output image, which can be obtained Image after denoising;
  • the discriminator part uses a 3 ⁇ 3 ⁇ 3 convolution kernel, LeakyReLU activation function and batch regularization operation, and finally outputs the prediction through the fully connected layer and the Sigmoid activation function.
  • the present application provides a self-supervised learning method, the method includes the following steps:
  • Step 1 Model the image noise
  • Step 2 generate the approximate target pixel value of the current pixel from the low-dose CT image, and obtain the target pixel point;
  • Step 3 randomly crop an image block from the low-dose CT input image, randomly select N pixels from the image block, and replace the target pixel with the currently selected pixel to obtain the target image;
  • Step 4 Train the network and gradually reach a convergence state.
  • the input means that the low-dose CT image is used as the input of the network, the image obtained after replacing the pixels is used as the target image, the loss between the output of the network and the target image is calculated, and the network is trained.
  • a noisy image can be considered to be a combination of clean images and noise to a certain extent, namely:
  • x represents the noisy image
  • s represents the clean image
  • n represents the noise and artifacts of the image. So simply put, the task of image noise reduction is to separate the noise image x into two parts: s and n, and remove the noise n to obtain a clean image s.
  • Noise reduction methods are usually based on the assumption that the pixels i and j in the clean image s are not statistically independent, namely:
  • the pixel value of any point in the image has a certain relationship with other pixel values in the image.
  • the pixel value of the current pixel can be accurately predicted. . This interdependence between pixels forms the basis of this application.
  • the noise image is composed of a clean image and noise.
  • step 2 from the current pixel point of the input low-dose CT image Randomly select pixels in the 5 ⁇ 5 neighborhood as the target pixel of the current pixel.
  • an image block x j with a size of 64 ⁇ 64 pixels is randomly cropped from the low-dose CT input image.
  • the image block is larger than the receptive field of the selected convolutional neural network.
  • N is one tenth of the size of the image block.
  • training the network in step 4 includes calculating the loss of the N pixels.
  • the loss function is a mean square error loss function.
  • the overall network framework adopts the Unet structure, and the network adopts Adam optimizer for optimization.
  • the low-dose CT noise image is generally regarded as the input of the convolutional neural network, and then the normal-dose CT image is used as the target, and the back-propagation algorithm is used to reduce the noise. Learning the mapping from low-dose CT images to normal CT-dose images, when the network converges, the learned convolutional neural network can realize low-dose CT noise reduction.
  • FIG. 1 The schematic diagram of the noise reduction method based on convolutional neural network is shown in Figure 1.
  • Each predicted pixel value output by the convolutional neural network are obtained from the pixels of the fixed size receptive field area x RF(i) , that is, the set of pixels in the receptive field area affects the predicted pixels.
  • the convolutional neural network can be regarded as the input as the receptive field area x RF (i) , the output is the predicted value of the center pixel of the receptive field function, that is:
  • parameter ⁇ is the learnable parameter of the convolutional neural network.
  • paired training data (x j , s j ) are usually required, where x j and s j represent the jth low-dose CT image and the corresponding normal-dose CT image, respectively, represents the region centered on pixel i in the jth low-dose CT image in the training dataset, represents the pixel point i in the jth normal dose CT image in the training dataset, set input to the convolutional neural network, As the target value, the predicted value and target value of the convolutional neural network are minimized, namely:
  • the neural network parameter ⁇ can gradually meet the requirements of the noise reduction task.
  • Convolutional neural networks based on supervised learning must be trained with paired training data, but it is difficult to obtain paired CT images in the clinical stage. Usually, we can only obtain CT images under low-dose conditions.
  • Self-supervised learning is to generate the target from the input through a certain method, so that the training of the neural network can be carried out normally.
  • This application is based on the idea of self-supervised learning algorithm, in the absence of target pixel value , by some means to generate approximate target pixel values from low-dose CT images
  • the network can thus be trained with the following objective function:
  • the size of the image patch is larger than the receptive field size of our network, and then we randomly select N pixels from the image patch,
  • the size of N is set to one-tenth of the size of the image block, and then randomly select pixels in the 5 ⁇ 5 neighborhood of the selected pixels to replace the currently selected pixels to obtain the target image x j ⁇ N , at this time, the low
  • the dose CT image block x j is used as the input, and the replaced image blocks x j ⁇ N are used as the target.
  • only the loss of these N pixels is calculated, namely:
  • the loss function L uses the mean squared error loss function
  • the overall network framework uses the common Unet structure
  • the entire network is optimized using the Adam optimizer
  • the present application also provides an application of self-supervised learning, in which the self-supervised learning method according to any one of claims 1 to 9 is applied to CT image noise reduction or MRI image noise reduction.
  • MRI images are MRI noise images.
  • it can also be applied to other types of medical image noise reduction.
  • This application is not directed to a specific network structure, has universality, and can be applied to any network structure.
  • This application can implement the denoising task end-to-end, and the network can be trained without paired data.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Processing (AREA)

Abstract

一种自监督学习方法及应用，包括：对图像噪声建模；从低剂量CT图像中产生当前像素的近似目标像素值，获取目标像素点；随机从低剂量CT输入图像中裁剪出图像块，从图像块中随机选取N个像素点，将目标像素点替换掉当前选取的像素点得到目标图像；训练网络，逐步达到收敛状态。无需人工干预即可端到端地进行网络的训练，并实现降噪。

Description

一种自监督学习方法及应用 技术领域
本申请属于医学和工业领域计算机断层扫描(CT)系统技术领域，特别是涉及一种自监督学习方法及应用。
背景技术
计算机断层扫描(Computed Tomography，CT)，是一种通过计算机和X射线来获取病人躯体断层图像的非侵入式影像学检测方法，它具有扫描时间短，费用低廉和疾病监测范围广等优点，适用于疾病的早期筛查和常规性体检。然而，大量的X射线照射会出现辐射剂量的累计效应，大幅度增加各种疾病发生的可能性，进而影响人体生理机能，破坏人体组织器官，甚至危害到患者的生命安全。
合理应用低剂量CT成像技术需要在满足CT图像的临床诊断要求下，同时尽可能的降低X射线对患者的辐射剂量，因此，研究和开发低剂量条件下成像质量更高的CT成像，对于目前的医疗诊断领域都有着重要的科学意义和广阔的应用前景。但目前在临床阶段，获取成对的CT图像(低剂量CT图像和与之对应的正常剂量CT图像)较为困难，在仅有低剂量CT图像的情况下，如何使用神经网络来学习到低剂量CT图像到正常剂量CT图像的映射在临床应用领域有着巨大的发展前景。
由于在CT成像时降低X射线的辐射会导致重建图像产生大量量子噪声和金属伪影；正常CT成像需采集的数据量较大，导致图像重建速度慢；扫描时间长，病人人体生理机能运动导致图像伪影；基于神经网络的方法必须使用成对数据进行训练，在临床阶段获取成对数据较为困难，这大大降低了临床应用的可能性。
现有的低剂量条件下CT成像质量较差。
发明内容
1.要解决的技术问题
基于现有的低剂量条件下CT成像质量较差的问题，本申请提供了一种自监督学习方法及应用。
2.技术方案
为了达到上述的目的，本申请提供了一种自监督学习方法，所述方法包括如下步骤：
步骤1：对图像噪声建模；
步骤2：从低剂量CT图像中产生当前像素的近似目标像素值，获取目标像素点；
步骤3：随机从低剂量CT输入图像中裁剪出图像块，从图像块中随机选取N个像素点，将目标像素点替换掉当前选取的像素点得到目标图像；
步骤4：训练网络，逐步达到收敛状态。
本申请提供的另一种实施方式为：所述步骤1中噪声图像由干净图像和噪声共同组成。
本申请提供的另一种实施方式为：所述步骤2中从输入的低剂量CT图像当前像素点
Figure PCTCN2020102732-appb-000001
的5×5邻域内随机选取像素点作为当前像素点的目标像素点。
本申请提供的另一种实施方式为：所述步骤3中随机从低剂量CT输入图像中裁剪出64×64像素大小的图像块x j。
本申请提供的另一种实施方式为：所述步骤3中图像块大于选取卷积神经网络的感受野。
本申请提供的另一种实施方式为：所述步骤3中N为图像块大小的十分之一。
本申请提供的另一种实施方式为：所述步骤4中训练网络包括计算所述N个像素点的损失。
本申请提供的另一种实施方式为：所述损失函数为均方误差损失函数。
本申请提供的另一种实施方式为：所述网络框架整体采用Unet结构，所述网络采用Adam优化器进行优化。
本申请还提供一种自监督学习的应用，将所述权利要求1～9中任一项所述的自监督学习方法应用于CT图像降噪或者MRI图像降噪。
3.有益效果
与现有技术相比，本申请提供的一种自监督学习方法的有益效果在于：
本申请提供的自监督学习方法，为一种用于低剂量CT图像降噪的自监督学习方法。
本申请提供的自监督学习方法，是鉴于自监督学习方法的思想，在没有标签(正常剂量CT图像)的情况下，通过本申请的方法，可以从输入数据(低剂量CT图像)中生成标签值，从而得到满足神经网络的训练要求，使得神经网络通过学习获得对低剂量条件下CT图像降噪的能力。
本申请提供的自监督学习方法中图像先验信息的使用将有助于稀疏角度低剂量CT图像的重建，从而在大幅度降低辐射剂量的前提下获得高质量的CT图像。
本申请提供的自监督学习方法，通过在输入的低剂量CT图像中对部分像素进行邻域随机替换来产生目标图像，从而突破了传统卷积神经网络对于训练集成对数据的要求，在摒弃监督学习缺点的同时，充分利用了卷积神经网络强大的特征提取能力，大大提高了低剂量条件下CT成像质量。
本申请提供的自监督学习方法，利用了卷积神经网络强大的特征提取能力，无需人工干预即可端到端地进行网络的训练，并实现降噪。
本申请提供的自监督学习方法，无需正常剂量CT图像，对数据要求低，在临床缺乏成对数据的情况下，有着巨大的优势。
附图说明
图1是本申请的基于监督学习的卷积神经网络第一示意图；
图2是本申请的基于监督学习的卷积神经网络第二示意图。
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。
Hu Chen等人于2017年在IEEE Transactions on Medical Imaging期刊上发表文章“Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network(RED-CNN).”，首次将编码-解码结构的深度残差卷积神经网络应用于低剂量CT图像质量改善问题，使得低剂量条件下获取到的CT图像结构更加清晰。其中，编码器部分由5层卷积核大小为5×5的卷积加ReLU激活函数组成，解码器部分由5层与编码器相对应的反卷积加ReLU激活函数组成，此外，编码器与解码器相对应的卷积与反卷积层使用类似于残差网络的残差连接，最终网络输出得到降噪后的CT图像。
Jelmer M.Wolterink等人于2017年在IEEE Transactions on Medical Imaging期刊上发表文章“Generative Adversarial Networks for Noise Reduction in Low-Dose CT”，成功将生成对抗网络(GAN)应用于低剂量CT成像领域，其中生成器使用的卷积核大小为3×3×3，卷积核数量由最开始的32个逐步增加到64个，最终增加到128个，去除了池化操作，所有卷积层后均使用LeakyReLU激活函数提训练稳定性，此外，为保证生成器学习到的是低剂量CT图像中的噪声部分，生成器的最终输出为输入图像减去生成器最后一层卷积输出图像，即可得到降噪后图像；判别器部分使用了3×3×3大小的卷积核，LeakyReLU激活函数以及批正则化操作，最终通过全连接层以及Sigmoid激活函数输出预测。
参见图1～2，本申请提供一种自监督学习方法，所述方法包括如下步骤：
步骤1：对图像噪声建模；
步骤2：从低剂量CT图像中产生当前像素的近似目标像素值，获取目标像素点；
步骤3：随机从低剂量CT输入图像中裁剪出图像块，从图像块中随机选取N个像素点，将目标像素点替换掉当前选取的像素点得到目标图像；
步骤4：训练网络，逐步达到收敛状态。
输入的意思是低剂量CT图像用作网络的输入，替换像素点后得到的图像作为目标图像，计算网络的输出和目标图像之间的损失，训练网络。
噪声图像在一定程度上可以被认为是干净图像和噪声共同组成，即：
x＝s+n       (1)
其中，x表示噪声图像，s表示干净图像，而n就表示图像的噪声和伪影等。所以简单地来说，图像降噪任务就是讲噪声图像x分离为两个部分：s和n，去除掉噪声n从而得到干净图像s。
降噪方法通常是基于干净图像s中的像素点i和j之间在统计上不是独立的假设，即：
p(s i|s j)≠p(s i)        (2)
也就是说图像中的任意一点像素值与图像中其他像素值都存在一定关联，换句话说，通过观察当前像素点周围像素值的大小和分布可以对当前像素点的像素值做出准确的预测。这种像素之间的相互依赖关系成为了本申请的基础。
进一步地，所述步骤1中噪声图像由干净图像和噪声共同组成。
进一步地，所述步骤2中从输入的低剂量CT图像当前像素点
Figure PCTCN2020102732-appb-000002
的5×5邻域内随机选取像素点作为当前像素点的目标像素点。
进一步地，所述步骤3中随机从低剂量CT输入图像中裁剪出64×64像素大小的图像块x j。
进一步地，所述步骤3中图像块大于选取卷积神经网络的感受野。
进一步地，所述步骤3中N为图像块大小的十分之一。
进一步地，所述步骤4中训练网络包括计算所述N个像素点的损失。
进一步地，所述损失函数为均方误差损失函数。
进一步地，所述网络框架整体采用Unet结构，所述网络采用Adam优化器进行优化。
在使用监督学习方法实现低剂量CT降噪任务时，一般情况下都是将低剂量CT噪声图像视为卷积神经网络的输入，再将正常剂量CT图像当作目标，通过反向传播算法来学习低剂量CT图像到正常CT剂量图像的映射，当网络收敛后，学习完成的卷积神经网络即可实现低剂量CT降噪。
基于卷积神经网络降噪方法示意图如图1所示，卷积神经网络输出的每一个预测像素值
Figure PCTCN2020102732-appb-000003
都是由固定大小感受野区域x RF(i)的像素来获得，即影响预测像素的是感受野区域内的像素集合，此时，卷积神经网络可以被视为输入为感受野区域x RF(i)，输出为感受野中心像素的预测值
Figure PCTCN2020102732-appb-000004
的函数，即：
Figure PCTCN2020102732-appb-000005
其中参数θ是卷积神经网络的可学习参数。
在监督学习中，通常需要成对的训练数据(x j,s j)，其中x j和s j分别表示第j张低剂量CT图像和与之对应的正常剂量CT图像，
Figure PCTCN2020102732-appb-000006
表示训练数据集中的第j张低剂量CT图像中以像素i为中心的区域，
Figure PCTCN2020102732-appb-000007
表示训练数据集中的第j张正常剂量CT图像中的像素点i，将
Figure PCTCN2020102732-appb-000008
输入到卷积神经网络，将
Figure PCTCN2020102732-appb-000009
作为目标值，最小化卷积神经网络的预测值和目标值，即：
Figure PCTCN2020102732-appb-000010
其中：
Figure PCTCN2020102732-appb-000011
通过最小化上式，即可使得神经网络参数θ逐步满足降噪任务要求。
基于监督学习的卷积神经网络必须使用成对的训练数据来训练，但在临床阶段获取成对的CT图像较为困难，通常情况下我们只能在低剂量条件下获取CT图像。
自监督学习就是通过一定方法从输入中生成目标，从而使得神经网络的训练可以正常进行。本申请就是以自监督学习算法思想为基础，在没有目标像素值
Figure PCTCN2020102732-appb-000012
的情况下，通过一定手段从低剂量CT图像中产生近似的目标像素值
Figure PCTCN2020102732-appb-000013
从而可以使用下面的目标函数对网络进行训练：
Figure PCTCN2020102732-appb-000014
如何从低剂量CT图像中产生当前像素的近似目标像素值
Figure PCTCN2020102732-appb-000015
是本申请的重点。前面已经提到，图像像素之间在统计上不是独立的，也就是说像素之间存在一定的关联性，而且两个像素点在空间位置上越相邻，它们之间的关联性越高，通俗点说，在图像低频区域，相邻像素点与当前像素点的像素值差异较小，而CT图像存在大量的低频区域。所以考虑在训练阶段可以从输入的低剂量CT图像当前像素点
Figure PCTCN2020102732-appb-000016
的5×5邻域内随机选取像素点作为当前像素点的目标像素点，即：
Figure PCTCN2020102732-appb-000017
知道了如何获取目标像素点，接下来就可以对整个训练流程加以描述。
首先，随机从低剂量CT输入图像中裁剪出64×64像素大小的图像块x j，该图像块的大小大于我们选取网络的感受野大小，随后我们从图像块中随机选取N个像素点，其中N的大小设置为图像块大小的十分之一，再在选取的像素点5×5邻域内随机选择像素点替换掉当前选取的像素点得到目标图像x j～N，此时，将低剂量CT图像块x j作为输入，替换后的图像块x j～N作为目标，在训练时仅计算这N个像素点的损失，即：
Figure PCTCN2020102732-appb-000018
损失函数L使用均方误差损失函数；
网络框架整体使用常见Unet结构；
整个网络使用Adam优化器来优化；
从低剂量CT图像数据集中提取图像块作为输入，并经由输入产生目标；
训练网络，逐步达到收敛状态。
本申请还提供一种自监督学习的应用，将所述权利要求1～9中任一项所述的自监督学习方法应用于CT图像降噪或者MRI图像降噪。MRI图像的话就MRI噪声图像。当然也可应用于其他类型医学图像降噪。
本申请不针对某一特定网络结构，具有普适性，可应用于任意网络结构。
本申请可以端到端地实现降噪任务，无需成对数据即可对网络进行训练。
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. 一种自监督学习方法，其特征在于：所述方法包括如下步骤：
    步骤1：对图像噪声建模；
    步骤2：从所述图像中产生当前像素的近似目标像素值，获取目标像素点；
    步骤3：随机从所述图像中裁剪出图像块，从图像块中随机选取N个像素点，将目标像素点替换掉当前选取的像素点得到目标图像；
    步骤4：训练网络，逐步达到收敛状态。
  2. 如权利要求1所述的自监督学习方法，其特征在于：所述步骤1中噪声图像由干净图像和噪声共同组成。
  3. 如权利要求1所述的自监督学习方法，其特征在于：所述步骤2中从输入的低剂量CT图像当前像素点
    Figure PCTCN2020102732-appb-100001
    的5×5邻域内随机选取像素点作为当前像素点的目标像素点。
  4. 如权利要求1所述的自监督学习方法，其特征在于：所述步骤3中随机从低剂量CT输入图像中裁剪出64×64像素大小的图像块x j。
  5. 如权利要求4所述的自监督学习方法，其特征在于：所述步骤3中图像块大于选取卷积神经网络的感受野。
  6. 如权利要求1所述的自监督学习方法，其特征在于：所述步骤3中N为图像块大小的十分之一。
  7. 如权利要求1所述的自监督学习方法，其特征在于：所述步骤4中训练网络包括计算所述N个像素点的损失。
  8. 如权利要求7所述的自监督学习方法，其特征在于：所述损失函数为均方误差损失函数。
  9. 如权利要求1～8中任一项所述的自监督学习方法，其特征在于：所述网络框架整体采用Unet结构，所述网络采用Adam优化器进行优化。
  10. 一种自监督学习的应用，其特征在于：将所述权利要求1～9中任一项所述的自监督学习方法应用于CT图像降噪或者MRI图像降噪。
PCT/CN2020/102732 2020-07-17 2020-07-17 一种自监督学习方法及应用 WO2022011690A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/102732 WO2022011690A1 (zh) 2020-07-17 2020-07-17 一种自监督学习方法及应用

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/102732 WO2022011690A1 (zh) 2020-07-17 2020-07-17 一种自监督学习方法及应用

Publications (1)

Publication Number Publication Date
WO2022011690A1 true WO2022011690A1 (zh) 2022-01-20

Family

ID=79554443

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/102732 WO2022011690A1 (zh) 2020-07-17 2020-07-17 一种自监督学习方法及应用

Country Status (1)

Country Link
WO (1) WO2022011690A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114648814A (zh) * 2022-02-25 2022-06-21 北京百度网讯科技有限公司 人脸活体检测方法及模型的训练方法、装置、设备及介质
CN114936979A (zh) * 2022-06-07 2022-08-23 中南大学 一种模型训练方法、图像去噪方法、装置、设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109035169A (zh) * 2018-07-19 2018-12-18 西安交通大学 一种无监督/半监督ct图像重建深度网络训练方法
WO2019147767A1 (en) * 2018-01-24 2019-08-01 Rensselaer Polytechnic Institute 3-d convolutional autoencoder for low-dose ct via transfer learning from a 2-d trained network
CN110599420A (zh) * 2019-05-27 2019-12-20 首都师范大学 基于深度学习的ct图像分块重建方法及系统
CN111260055A (zh) * 2020-01-13 2020-06-09 腾讯科技（深圳）有限公司 基于三维图像识别的模型训练方法、存储介质和设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019147767A1 (en) * 2018-01-24 2019-08-01 Rensselaer Polytechnic Institute 3-d convolutional autoencoder for low-dose ct via transfer learning from a 2-d trained network
CN109035169A (zh) * 2018-07-19 2018-12-18 西安交通大学 一种无监督/半监督ct图像重建深度网络训练方法
CN110599420A (zh) * 2019-05-27 2019-12-20 首都师范大学 基于深度学习的ct图像分块重建方法及系统
CN111260055A (zh) * 2020-01-13 2020-06-09 腾讯科技（深圳）有限公司 基于三维图像识别的模型训练方法、存储介质和设备

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114648814A (zh) * 2022-02-25 2022-06-21 北京百度网讯科技有限公司 人脸活体检测方法及模型的训练方法、装置、设备及介质
CN114936979A (zh) * 2022-06-07 2022-08-23 中南大学 一种模型训练方法、图像去噪方法、装置、设备及存储介质

Similar Documents

Publication Publication Date Title
WO2021077997A1 (zh) 图像去噪的多生成器生成对抗网络学习方法
WO2021159948A1 (zh) 一种基于深度学习的低剂量pet三维重建方法
Yuan et al. SIPID: A deep learning framework for sinogram interpolation and image denoising in low-dose CT reconstruction
CN110728729B (zh) 一种基于注意机制的无监督ct投影域数据恢复方法
WO2022011690A1 (zh) 一种自监督学习方法及应用
Ko et al. Rigid and non-rigid motion artifact reduction in X-ray CT using attention module
CN112258642B (zh) 基于深度学习的低剂量pet数据三维迭代更新重建方法
Hou et al. CT image quality enhancement via a dual-channel neural network with jointing denoising and super-resolution
Li et al. Learning non-local perfusion textures for high-quality computed tomography perfusion imaging
CN111860836A (zh) 一种自监督学习方法及应用
Li et al. A comprehensive survey on deep learning techniques in CT image quality improvement
WO2022094911A1 (zh) 一种共享权重的双区域生成对抗网络及其图像生成方法
WO2022027216A1 (zh) 一种图像降噪方法及其应用
Liang et al. A self-supervised deep learning network for low-dose CT reconstruction
Du et al. X-ray CT image denoising with MINF: A modularized iterative network framework for data from multiple dose levels
CN112419175A (zh) 一种共享权重的双区域生成对抗网络及其图像生成方法
Xu et al. Metal artifact reduction for oral and maxillofacial computed tomography images by a generative adversarial network
CN113436118B (zh) 一种基于多尺度卷积编码网络的低剂量ct图像复原方法
Ma et al. Low dose CT reconstruction assisted by an image manifold prior
Zhang et al. NOWNUNM: nonlocal weighted nuclear norm minimization for sparse-sampling CT reconstruction
Chen et al. Dual-domain modulation for high-performance multi-geometry low-dose CT image reconstruction
CN104899904B (zh) 一种在低辐射剂量下进行血流参数图像的输出方法
Park et al. Unpaired-paired learning for shading correction in cone-beam computed tomography
KR20220073156A (ko) 딥 러닝 기반의 제한각도 컴퓨터 단층촬영상 재구성 시스템
Hein et al. Ring Artifact Correction in Photon-Counting Spectral CT Using a Convolutional Neural Network With Spectral Loss

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: 20945361

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20945361

Country of ref document: EP

Kind code of ref document: A1