A CT image noise reduction system and method
By using a generative adversarial network with multiple attention and a joint loss function, the problems of noise and artifacts in low-dose CT image reconstruction are solved, achieving high-quality image reconstruction that meets clinical diagnostic requirements.
Patent Information
- Application Number
- CN202010605214.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-06-29
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2040-06-29
AI Technical Summary
Current CT image reconstruction techniques are unable to reconstruct clinically acceptable CT images under sparse sampling and low-dose conditions, resulting in severe image artifacts and noise, which affects diagnostic results.
A generative adversarial network with multiple attention, including an attention module and an adaptive moment estimation optimizer, is used to map low-dose CT images to normal-dose CT images. The network utilizes high-dimensional and low-dimensional features of the images, as well as local and non-local information, and combines a joint loss function to optimize image quality.
It improves the peak signal-to-noise ratio and structural similarity of low-dose CT images, enhances image detail information, meets the needs of medical diagnosis, eliminates noise and artifacts, and improves image quality.
Smart Images

Figure CN111861910B_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of medical CT imaging technology, and in particular, relates to a CT image noise reduction system and method. Background Art
[0002] Computed Tomography (CT) is a non-invasive imaging method that uses computers and X-rays to obtain tomographic images of the patient's body. It has the advantages of short scanning time, low cost and a wide range of disease monitoring, and is suitable for early screening of diseases and routine physical examinations. However, a large amount of X-ray exposure will have a cumulative effect of radiation dose, greatly increasing 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. The rational application of low-dose CT imaging technology requires that the radiation dose of X-rays to patients be reduced as much as possible while meeting the clinical diagnostic requirements of CT images. Therefore, research and development of CT imaging with higher imaging quality under low-dose conditions has important scientific significance and broad application prospects for the current medical diagnosis field.
[0003] Image Denoising is a professional term in image processing. In the process of digitization and transmission, digital images in reality are often affected by the interference of imaging equipment and external environmental noise, which are called noisy images or noisy images. The process of reducing noise in digital images is called image denoising, sometimes also called image denoising.
[0004] Reducing X-ray radiation during CT imaging will cause a large amount of quantum noise and metal artifacts in the reconstructed image; the amount of data required for normal CT imaging is large, resulting in slow image reconstruction; the long scanning time leads to artifacts caused by the patient's inevitable physiological movement. Summary of the invention
[0005] 1. Technical issues to be solved
[0006] Based on the problem that existing CT image reconstruction technology cannot reconstruct clinically acceptable CT images under sparse sampling and low-dose conditions. Using traditional algorithms to reconstruct sparse low-dose CT images will bring obvious image artifacts and interference information, seriously affecting subsequent clinical diagnosis. The present application provides a CT image denoising system and method.
[0007] 2. Technical solution
[0008] In order to achieve the above-mentioned purpose, the present application provides a CT image denoising system, the system comprising a generative adversarial network, the generative adversarial network is used to achieve mapping between low-dose CT images and normal-dose CT images, and judge the authenticity of the generated images;
[0009] The generative adversarial network includes an attention module and an adaptive moment estimation optimizer; the attention module is used to apply different weights to each channel of the feature map of the image, making full use of the high-dimensional and low-dimensional features of the image as well as local and non-local information;
[0010] The adaptive moment estimation optimizer is used to optimize the generative adversarial network.
[0011] Another implementation method provided by the present application is: the attention module in step 1 is embedded in the generative adversarial network, and the attention module includes a channel attention submodule and a cross self-attention submodule;
[0012] The channel attention submodule is used to assign different weights to different feature maps in the channel direction;
[0013] The cross self-attention submodule is used to improve the utilization of non-local information and can obtain non-local features along the horizontal and vertical directions.
[0014] The channel attention submodule is used to apply different weights to each channel of the image feature map, which can be used to fuse high-dimensional and low-dimensional features; the cross self-attention submodule obtains the attention map through learning, making full use of the local and non-local information of the image
[0015] Another implementation mode provided by the present application is: the generative adversarial network includes a first generator, a second generator, a first discriminator and a second discriminator;
[0016] The first generator is used to complete the low-dose CT image denoising task;
[0017] The second generator is used to complete the noise simulation process from normal-dose CT to low-dose CT;
[0018] The first discriminator is used to encourage the first generator to generate a normal-dose CT image from a low-dose CT image;
[0019] The second discriminator is used to encourage the second generator to generate low-dose CT images from normal-dose CT images.
[0020] Another implementation manner provided by the present application is: the first generator includes a feature extraction unit, an image reconstruction unit and a residual connection unit, and the residual connection unit includes a mean filter; the second generator includes a feature extraction unit, an image reconstruction unit and a residual connection unit, and the residual connection unit includes a mean filter.
[0021] Another implementation method provided by the present application is: the feature extraction submodule is composed of 12 groups of 3×3 convolutions and LeakyReLU activation functions, and the output of each layer of convolution operations is merged along the channel direction at the end, and then the channel attention submodule is used to autonomously assign weights to each channel.
[0022] Another implementation manner provided by the present application is: the first discriminator is composed of 6 groups of convolutions and LearkyReLU activation functions, wherein the size of the convolution kernel is 3×3; the second discriminator is composed of 6 groups of convolutions and LearkyReLU activation functions, wherein the size of the convolution kernel is 3×3.
[0023] Another implementation manner provided by the present application is: it also includes a joint loss function module, and the joint loss function module is used to further improve the image quality.
[0024] Another implementation manner provided by the present application is that the joint loss function module includes an adversarial loss submodule, a cycle consistency loss submodule and a structural recovery loss submodule.
[0025] The present application also provides a CT image noise reduction method, the method comprising:
[0026] 1) Processing an image using the CT image noise reduction system according to any one of claims 1 to 7;
[0027] 2) extracting image patches from the low-dose CT image dataset as input and extracting corresponding image patches from the normal-dose CT image dataset as reference;
[0028] 3) Train the generative adversarial network and gradually reach a convergence state.
[0029] Another implementation method provided by the present application is: applying different weights to each channel of the feature map of the image includes performing a global average pooling operation on the input feature map to obtain a 1×1×C vector, and then using a 1×1 convolution operation to compress and restore the vector along the channel direction. After two 1×1 convolution operations and then passing through a sigmoid function, the required weight vector can be obtained. Finally, the input feature map is multiplied by the weight vector to obtain the final output.
[0030] 3. Beneficial effects
[0031] Compared with the prior art, the CT image denoising system and method provided by the present application has the following beneficial effects:
[0032] The CT image noise reduction system provided in the present application is a computer tomography (CT) system in the medical and industrial fields.
[0033] The CT image denoising method provided in this application realizes CT image denoising based on a multi-attention cycle consistency generative adversarial network, which improves the image peak signal-to-noise ratio and structural similarity while enhancing the image detail information, thereby obtaining a CT image that better meets diagnostic needs.
[0034] The CT image denoising method provided in the present application is to solve the problem of poor CT imaging quality and many noise artifacts under low-dose conditions.
[0035] The CT image denoising system provided in this application is a cycle-consistent generative adversarial network based on multiple attentions to improve the quality of low-dose CT imaging. The attention mechanism can greatly improve the reuse of low-dimensional and high-dimensional information and the fusion of local and non-local information, thereby enhancing the performance of traditional convolution operations and being able to eliminate noise and artifacts of low-dose CT to a large extent.
[0036] The CT image denoising method provided in this application specifically designs a joint loss function to improve the quality of CT images, and further ensures that the generated CT images meet the medical diagnosis requirements by combining multiple loss functions.
[0037] The CT image denoising system provided in this application is based on a multiple attention mechanism, strives to extract image features more effectively, starts from high-dimensional and low-dimensional features as well as local and non-local information, and is implemented through two different attention mechanisms, thereby greatly improving the detail expression of the generated CT image.
[0038] The general CT denoising method uses a single loss function and cannot guarantee the quality of the generated image. The CT image denoising system provided in this application effectively guarantees the quality of the output image by combining multiple loss functions.
[0039] The CT image denoising system provided in the present application adds a residual connection with a mean filter, which effectively improves the convergence speed of the network and improves the training efficiency of the network. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 It is a schematic diagram of the channel attention submodule structure of this application;
[0041] Figure 2 This is a schematic diagram of the cross self-attention submodule structure of this application;
[0042] Figure 3 is a schematic diagram of the first generator structure of this application;
[0043] Figure 4 It is a structural schematic diagram of the feature extraction unit of the present application;
[0044] Figure 5is a schematic diagram of the structure of the image reconstruction unit of the present application;
[0045] Figure 6 is a schematic diagram of the structure of the first discriminator of the present application;
[0046] Figure 7 It is a schematic diagram of the generative adversarial network of this application;
[0047] Figure 8 It is a schematic diagram comparing the results of different methods of this application. DETAILED DESCRIPTION
[0048] Hereinafter, the specific embodiments of the present application will be described in detail with reference to the accompanying drawings. According to these detailed descriptions, those skilled in the art can clearly understand the present application and can implement the present application. Without violating the principles of the present application, the features in the various embodiments can be combined to obtain new implementations, or certain features in certain embodiments can be replaced to obtain other preferred implementations.
[0049] In 2017, Elmer M. Wolterink et al. published an article titled "Generative Adversarial Networks for Noise Reduction in Low-Dose CT" in the journal IEEE Transactions on Medical Imaging, successfully applying generative adversarial networks (GANs) to the field of low-dose CT imaging. The convolution kernel size used by the generator was 3×3×3, and the number of convolution kernels gradually increased from the initial 32 to 64, and finally to 128. The pooling operation was removed, and the LeakyReLU activation function was used after all convolution layers to improve training stability. In addition, to ensure that the generator learns the noise part of the low-dose CT image, the final output of the generator is the input image minus the output image of the last layer of convolution of the generator, to obtain the denoised image; the discriminator part uses a convolution kernel of size 3×3×3, a LeakyReLU activation function, and a batch normalization operation, and finally outputs the prediction through a fully connected layer and a Sigmoid activation function.
[0050] In 2018, Yang et al. published an article titled "Low-dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss" in the journal IEEE transactions on medical imaging. They added perceptual loss to the generative adversarial network, making the generated images have more detailed information and better visual effects. The generator part of the network consists of 8 convolution operations and ReLU activation functions, and the discriminator consists of 6 convolution operations and ReLU activation functions. The generator and discriminator are trained simultaneously, and the discriminator can promote the generator to generate images that meet the requirements better.
[0051] See also Figures 1 to 8 , the present application provides a CT image denoising system, the system comprising a generative adversarial network, the generative adversarial network is used to achieve mapping between low-dose CT images and normal-dose CT images, and judge the authenticity of the generated images;
[0052] The generative adversarial network includes an attention module and an adaptive moment estimation optimizer; the attention module is used to apply different weights to each channel of the feature map of the image, making full use of the high-dimensional and low-dimensional features of the image as well as local and non-local information;
[0053] The adaptive moment estimation optimizer is used to optimize the generative adversarial network.
[0054] The generative adversarial network is the overall structure of the network, which includes a generator and a discriminator. The generator learns the mapping between low-dose CT images and normal CT images, and the discriminator learns to determine whether the input image is a real image. The so-called adversarial refers to the mutual confrontation between the generator and the discriminator. The generator learns the characteristic distribution of real data, and the discriminator distinguishes the true from the false from the real data and the data generated by the generator. The generator hopes that the generated data will deceive the discriminator as much as possible, while the discriminator hopes to be able to identify the data generated by the generator, thus forming an adversarial relationship. The two constantly compete in generation and confrontation, learn together, and gradually reach a Nash equilibrium. Finally, the data generated by the generator is sufficient to be indistinguishable from the real, making it impossible for the discriminator to distinguish the true from the false; the attention module is a sub-module embedded in the generator and the discriminator, which is used to improve the performance of the generator and the discriminator; the Adam optimizer is a gradient update method used during backpropagation, which can ensure the normal network training and effectively improve the network convergence speed.
[0055] Furthermore,: the attention module in step 1 is embedded in the generative adversarial network, and the attention module includes a channel attention submodule and a cross self-attention submodule;
[0056] The channel attention submodule is used to assign different weights to different feature maps in the channel direction;
[0057] The cross self-attention submodule is used to improve the utilization of non-local information and can obtain non-local features along the horizontal and vertical directions.
[0058] The attention module contains two different attention mechanisms: channel attention and cross self-attention. Traditional convolutional neural networks obtain higher-dimensional information of images by continuously superimposing convolution operations, but they often lack sufficient flexibility in utilizing feature information from low dimensions. Among them, channel attention mainly assigns different weights to different feature maps in the channel direction, which can make more full use of feature information from low dimensions and high dimensions. The weight is not determined manually, but obtained by the network through learning, which further increases the autonomy of the network. Channel attention is like Figure 1 shown.
[0059] like Figure 1 As shown in the figure, the input of the channel attention submodule is a feature map of size H×W×C. In order to apply different weights to each channel, a 1×1×C weight vector is required. The value on this vector is the weight of different channels. To obtain the weight vector, first perform a global average pooling operation on the input feature map to obtain a 1×1×C vector, and then use a 1×1 convolution operation to compress and restore the vector along the channel direction to better achieve information fusion between channels. After two 1×1 convolution operations and then a sigmoid function, the required weight vector can be obtained. Finally, the input feature map is multiplied by the weight vector to obtain the final output. The channel attention operation can be expressed by the following formula:
[0060] s=f(W U δ(W D G(z))) (1)
[0061] Among them, z represents the input feature map of size H×W×C, G represents the global average pooling operation, and W D and W U represents two 1×1 convolution operations, δ and f represent ReLU and sigmoid activation functions, respectively.
[0062] In addition, the traditional convolution operation only performs cross-correlation calculation in the local area, which greatly reduces the utilization of non-local information. The non-local features of the fused image are conducive to improving the quality of the final generated image. In order to improve the utilization of non-local information, a cross self-attention submodule is designed to obtain non-local features in the horizontal and vertical directions. This module first calculates the horizontal and vertical correlation of the current pixel position, and then superimposes the cross self-attention submodule twice to obtain the global correlation, thereby effectively utilizing the global non-local feature information.
[0063] As shown in the figure, assuming that the input of the module is a feature map of size C×H×W, it is obtained by three 1×1 convolutions. and The number of channels of f(x) and g(x) is C It should be lower than h(x). The compression channel is to promote the fusion of information between channels. Assume that the current pixel x i ∈x is the pixel at the i-th position in the feature map, It is a vector composed of pixel values along the channel direction corresponding to f(x) and the current pixel position. g(x) is a vector of pixel values in the horizontal and vertical directions corresponding to the current pixel position, then:
[0064]
[0065] Then v′(x) is first expanded by 1×1 convolution, and then the output attention map is calculated by softmax, and then the obtained attention map is multiplied by the corresponding element of h(x):
[0066]
[0067] Finally, the output is multiplied by the learnable parameter γ and added to the input to get the final output:
[0068] o final (x) = γ·o(x) + x (4)
[0069] The cross-self-attention submodule described above only calculates the correlation in the vertical and horizontal directions. When the cross-self-attention submodule is superimposed twice, the global correlation can be indirectly calculated, which greatly reduces the amount of computation required to directly calculate the global correlation.
[0070] Further, the generative adversarial network includes a first generator, a second generator, a first discriminator and a second discriminator;
[0071] The first generator is used to complete the low-dose CT image denoising task;
[0072] The second generator is used to complete the noise simulation process from normal-dose CT to low-dose CT;
[0073] The first discriminator is used to encourage the first generator to generate a normal-dose CT image from a low-dose CT image;
[0074] The second discriminator is used to encourage the second generator to generate low-dose CT images from normal-dose CT images.
[0075] Furthermore, the first generator includes a feature extraction unit, an image reconstruction unit and a residual connection unit, and the residual connection unit includes a mean filter; the second generator includes a feature extraction unit, an image reconstruction unit and a residual connection unit, and the residual connection unit includes a mean filter.
[0076] The generative adversarial network involved in this application is mainly composed of a generator and a discriminator, wherein the generator mainly realizes the mapping between low-dose CT images and normal-dose CT images, and the discriminator is used to judge the authenticity of the images generated by the generator. The two are trained synchronously and progress together, gradually making the images generated by the generator indistinguishable from the real ones. The generator is mainly composed of two parts: a feature extraction unit and an image reconstruction unit. The input is extracted from the image through a feature extraction network, and then the extracted features are reconstructed through an image reconstruction network. At the same time, the input is output through a residual connection unit with a mean filter. The residual connection unit is to solve the gradient vanishing and gradient exploding problems in training. At the same time, in order to promote information transfer between networks, the mean filter can further improve the convergence speed of the network.
[0077] Furthermore, the feature extraction submodule consists of 12 groups of 3×3 convolutions and LeakyReLU activation functions, and the output of each layer of convolution operations is merged along the channel direction at the end, and then the channel attention submodule assigns weights to each channel autonomously, thereby more effectively utilizing feature information from high and low dimensions.
[0078] The image reconstruction unit first extracts features through three convolution operations of different sizes, among which 1×1 convolution can be used for channel compression, reducing the number of parameters and improving information fusion between channels. A cross self-attention submodule is added at the end of the three branches to improve the utilization of non-local features. Finally, the feature maps obtained by the three branches are merged along the channel direction, and the final output is obtained through three sets of convolution operations.
[0079] Furthermore, the first discriminator is composed of 6 groups of convolution and LearkyReLU activation functions, where the size of the convolution kernel is 3×3; the second discriminator is composed of 6 groups of convolution and LearkyReLU activation functions, where the size of the convolution kernel is 3×3. The number and step size of the convolution kernels are shown in the figure, where n represents the number of convolution kernels, s represents the step size of the convolution kernel, and the network ends with two convolution layers, and the final output is to determine whether the input image is real.
[0080] Furthermore, a joint loss function module is included, and the joint loss function module is used to further improve the image quality.
[0081] Furthermore, the joint loss function module includes an adversarial loss submodule, a cycle consistency loss submodule and a structure recovery loss submodule.
[0082] Before designing the joint loss function, we first explain the overall architecture of the cycle-consistent generative adversarial network. As shown in the figure, the architecture mainly consists of two pairs of generators (G ab and G ba ) and the discriminator (D a and D b ), discriminator D b Encourage the generator G ab From low-dose CT images Generate normal dose CT images Discriminator D a Encourage the generator G ba From normal dose CT images Generate low-dose CT images The generator G ab It can complete the low-dose CT image denoising task, the generator G ba It can complete the noise simulation process from normal dose CT to low dose CT. In the test phase, this application mainly uses the trained generator G ab Realize image noise reduction function.
[0083] A single loss function has certain limitations in image generation. In order to improve the quality of generated images, it is necessary to use a joint loss to further improve the image quality. The joint loss function can be expressed as follows:
[0084]
[0085] Among them, λ1 and λ2 are weight coefficients respectively.
[0086] First of all, the whole network is based on the idea of generating adversarial networks, so adversarial loss is one of its core loss functions, and its objective function is:
[0087]
[0088] in,
[0089]
[0090]
[0091] In order to alleviate the gradient vanishing and mode collapse problems of the generative adversarial network during training, the least squares adversarial loss function is used:
[0092]
[0093]
[0094] Where E(·) represents the expectation, α and β represent the input and target data, and P * represents data distribution, a, b and c represent artificially set hyperparameters, where a=0, b=1, c=1.
[0095] In the cycle consistency generative adversarial network, in addition to the adversarial loss, cycle consistency is also added. Using only the adversarial loss to train the function cannot guarantee that the generated image details can meet the requirements, so the cycle consistency loss is added to further constrain the output:
[0096]
[0097] in, represents the L1 norm.
[0098] The adversarial loss and the cycle consistency loss cooperate with each other to constrain each other and ensure that the output result is as close to the truth as possible, but it is difficult to guarantee the peak signal-to-noise ratio and structural similarity of the output image. In order to further improve the image quality, the structural recovery loss is added, which can further improve the image peak signal-to-noise ratio and structural similarity. First of all, past experience shows that although the L2 loss function can improve the peak signal-to-noise ratio of the image to a certain extent, the L2 loss will make the image smooth and lose details, so the L1 loss function is used:
[0099]
[0100] Structural similarity measures the similarity between two images in terms of brightness, contrast, and structure. The calculation formula is:
[0101]
[0102] Where μ and σ represent the mean and standard deviation of the image respectively, C1=(k1L) 2 and C2=(k2L) 2are two small constants to avoid the denominator being 0, where L represents the maximum pixel value of the image. The closer the structural similarity is to 1, the higher the similarity between the two images. At the same time, the gradient descent method is usually used in network training, so the following loss function is designed:
[0103]
[0104] Combining L1 and L SSIM We can get:
[0105] L str =μL1+(1-μ)L SSIM (15)
[0106] The present application also provides a CT image noise reduction method, the method comprising:
[0107] 1) Processing an image using the CT image noise reduction system according to any one of claims 1 to 7;
[0108] 2) extracting image patches from the low-dose CT image dataset as input and extracting corresponding image patches from the normal-dose CT image dataset as reference;
[0109] 3) Train the generative adversarial network and gradually reach a convergence state.
[0110] Furthermore, applying different weights to each channel of the feature map of the image includes performing a global average pooling operation on the input feature map to obtain a 1×1×C vector, and then using a 1×1 convolution operation to compress and restore the vector along the channel direction. After two 1×1 convolution operations and then passing through a sigmoid function, the required weight vector can be obtained. Finally, the input feature map is multiplied by the weight vector to obtain the final output.
[0111] from Figure 8 It can be seen that the method of the present application can effectively improve the peak signal-to-noise ratio and structural similarity of the image, and at the same time, can restore the image detail information to a certain extent.
[0112] In addition to being used in CT image denoising, this method can be applied to other types of medical image denoising; in addition to being used in denoising, this method can also be applied to the field of image super-resolution after appropriate modifications; the attention mechanism can be considered as a plug-and-play module that can be added to any traditional convolutional neural network workflow to improve the performance of the network.
[0113] Although the present application is described above with reference to specific embodiments, it should be understood by those skilled in the art that many modifications may be made to the configurations and details disclosed in the present application within the principles and scope disclosed in the present application. The scope of protection of the present application is determined by the appended claims, and the claims are intended to cover all modifications included in the equivalent literal meaning or scope of the technical features in the claims.
Claims
1. A CT image noise reduction system, characterized in that: The system includes a generative adversarial network, which is used to achieve mapping between low-dose CT images and normal-dose CT images and determine the authenticity of generated images; The generative adversarial network includes an attention module and an adaptive moment estimation optimizer; the attention module is used to apply different weights to each channel of the feature map of the image, making full use of the high-dimensional and low-dimensional features of the image as well as local and non-local information; The attention module is embedded in the generative adversarial network, and the attention module includes a channel attention submodule and a cross self-attention submodule; the channel attention submodule is used to assign different weights to different feature maps in the channel direction; the cross self-attention submodule is used to improve the utilization of non-local information and can obtain non-local features along the horizontal and vertical directions; the adaptive moment estimation optimizer is used to optimize the generative adversarial network.
2. The CT image noise reduction system according to claim 1, characterized in that: The generative adversarial network includes a first generator, a second generator, a first discriminator and a second discriminator; The first generator is used to complete the low-dose CT image denoising task; The second generator is used to complete the noise simulation process from normal-dose CT to low-dose CT; The first discriminator is used to encourage the first generator to generate a normal-dose CT image from a low-dose CT image; The second discriminator is used to encourage the second generator to generate low-dose CT images from normal-dose CT images.
3. The CT image noise reduction system according to claim 2, characterized in that: The first generator includes a feature extraction unit, an image reconstruction unit and a residual connection unit, and the residual connection unit includes a mean filter; the second generator includes a feature extraction unit, an image reconstruction unit and a residual connection unit, and the residual connection unit includes a mean filter.
4. The CT image noise reduction system according to claim 3, characterized in that: The feature extraction unit consists of 12 groups of 3×3 convolutions and LeakyReLU activation functions, and the outputs of each layer of convolution operations are merged along the channel direction at the end, and then weights are autonomously assigned to each channel through the channel attention submodule.
5. The CT image noise reduction system according to claim 2, characterized in that: The first discriminator consists of 6 groups of convolutions and LearkyReLU activation functions, where the size of the convolution kernel is 3×3; the second discriminator consists of 6 groups of convolutions and LearkyReLU activation functions, where the size of the convolution kernel is 3×3.
6. The CT image noise reduction system according to claim 1, characterized in that: It also includes a joint loss function module, which is used to further improve the image quality.
7. The CT image noise reduction system according to claim 6, characterized in that: The joint loss function module includes an adversarial loss submodule, a cycle consistency loss submodule and a structure recovery loss submodule.
8. A CT image noise reduction method, characterized in that: The method comprises: 1) Processing an image using the CT image noise reduction system according to any one of claims 1 to 7; 2) Extract image patches from the low-dose CT image dataset as input, and extract corresponding image patches from the normal-dose CT image dataset as reference; 3) Train the generative adversarial network and gradually reach convergence.
9. The CT image denoising method according to claim 8, characterized in that: The method of applying different weights to each channel of the feature map of the image includes performing a global average pooling operation on the input feature map to obtain 1×1×C The vector is then compressed and restored along the channel direction using a 1×1 convolution operation. After two 1×1 convolution operations, the vector is then sigmoid The function can get the required weight vector, and finally multiply the input feature map with the weight vector to get the final output.
Citation Information
Patent Citations
Multi-band image feature level fusion method based on attention mechanism generative adversarial network
CN110555458A
Image denoising method based on channel attention mechanism and feature pyramid
CN110766632A
Low-dose CT image restoration and denoising method
CN110930318A
Image multi-style conversion method based on latent variable feature generation
CN110992252A
CT super-resolution GAN constrained by the identical, residual and cycle learning ensemble (GAN-circle)
US20200111194A1
