A method and device for super-resolution denoising of low-dose CT images

Abstract

The present invention discloses a low-dose CT image super-resolution denoising method, comprising: obtaining a low-dose CT image; constructing a training system based on a generative adversarial network, including a generator and a discriminator, the generator including a feature extraction module, at least one lightweight and efficient pixel attention module, at least one multi-scale detail context module and a fusion module, the fusion module is used to fuse the denoising feature vector and the feature vector closely related to the context details to obtain a predicted high-dose CT image; constructing a total loss function of the training system, using the total loss function to train the training system, at the end of the training, the generator with determined parameters constitutes a CT image super-resolution denoising model; when applied, the low-dose CT image is input into the CT image super-resolution denoising model, and a super-resolution denoised CT image is obtained by calculation. The method can suppress the noise of low-dose CT images and improve the resolution of low-dose CT images.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning, and specifically relates to a low-dose CT image super-resolution denoising method and device. Background Art

[0002] Computed Tomography (CT) imaging is also known as computer tomography. It uses a computer to process many combined x-rays to measure specific areas of the cross-section generated from different angles to scan the object, allowing the user to see the inside of the object without cutting. Since CT imaging technology is a cross-sectional imaging, it can reconstruct images and display tissues or organs in any direction, showing the lesions more comprehensively to prevent omissions; it has high density resolution, and can also display subtle lesions with density changes, and can clarify the nature of the lesions; in addition, CT has the advantages of being non-invasive and fast imaging, and has become a widely used and highly safe medical diagnostic technology.

[0003] With the continuous development of CT, CT diagnosis plays an increasingly important role in early disease screening. However, high radiation is generated during CT scanning, which can cause harm to the human body. Current medical research shows that 1.5%-2% of tumors may be caused by the high radiation dose of CT. Especially for patients with high-risk diseases and infants, high-dose CT diagnosis is no longer applicable.

[0004] Due to the high radiation problem of high-dose CT diagnosis, the concept of low-dose CT (LDCT) was proposed. While ensuring that other scanning parameters remain unchanged, the radiation dose of radiation CT is reduced by reducing the tube current. Because the structure of the lungs and other tissues and organs is very different, with a large amount of air and low density, a CT scan with a dose 75%-90% lower than the conventional dose can form a more satisfactory image. Although low-dose CT can reduce the damage caused by radiation, this will cause the projected data to be contaminated. Low-dose CT lung images will not only have obvious noise, but also strip artifacts. When the CT image contains a small amount of noise, image domain technology is usually used to obtain good image quality. However, for CT images with severe noise, these noises even completely cover the structural information. Image domain technology will sacrifice the structure of the image while eliminating the noise. The interference of these noises will reduce the probability of lung cancer being screened out, and will interfere with doctors' judgment of the disease to a certain extent, which is not conducive to clinical diagnosis. Therefore, it is of great significance to denoise and restore low-dose CT lung images.

[0005] Although traditional noise removal algorithms have achieved good results in restoring low-dose CT lung images, due to the influence of block noise, while suppressing noise and artifacts, it also causes loss of edge details.

[0006] At present, the mainstream methods for high-quality reconstruction of low-dose CT can be divided into projection domain filtering method, iterative method and image domain method. The projection domain filtering method mainly uses the known noise distribution information to design the corresponding filtering function, improves the quality of the projection image in the projection domain, and thus improves the quality of the reconstructed image, such as the FDK algorithm; the projection domain filtering method is prone to cause the loss of some structures in the projection image, which causes the reconstructed image to also lose some structural information. The iterative method mainly uses the statistical characteristics of the projection data and the prior information of the reconstructed image to design the corresponding optimization function, and obtains high-quality reconstructed images by optimizing the objective function during the iteration process, such as the joint algebraic iteration technology; the iterative method is usually time-consuming, and the hyperparameters need to be adjusted based on experience. Image domain methods are divided into two categories: one is the traditional method, such as Block-Matching and 3D filtering (BM3D), dictionary learning, etc.; the other is the image domain deep learning technology, such as residual codec convolutional neural network; however, the image domain method has little potential to reduce the X-ray dose while ensuring image quality. Neither the traditional image domain method nor the image domain deep learning method can restore the image details that have been lost in the reconstructed image. Summary of the invention

[0007] The present invention provides a low-dose CT image super-resolution denoising method, by which the low-dose CT image noise can be suppressed and the low-dose CT image resolution can be improved.

[0008] A low-dose CT image super-resolution denoising method comprises the following steps:

[0009] (1) obtaining a real high-dose CT image and processing the real high-dose CT image into a low-dose CT image;

[0010] (2) constructing a training system based on a generative adversarial network, including a generator and a discriminator, the generator including a first feature extraction module, at least one lightweight and efficient pixel attention module, at least one multi-scale detail context module and a fusion module, the first feature extraction module is used to extract a shallow feature vector from an input low-dose CT image, the lightweight and efficient pixel attention module is used to denoise the input shallow feature vector and output a denoised feature vector, the multi-scale detail context module is used to expand the receptive field of the input shallow feature vector and output a context-detail closely connected feature vector, and the fusion module is used to fuse the denoised feature vector and the context-detail closely connected feature vector to obtain a predicted high-dose CT image;

[0011] (3) Constructing a total loss function of the training system, outputting first discriminant information through the discriminator based on the real high-dose CT image, outputting the predicted high-dose CT image through the generator based on the low-dose CT image, and outputting second discriminant information through the discriminator based on the predicted high-dose CT image. The first loss function of the discriminator is constructed based on the expectation of the first discriminant information and the second discriminant information;

[0012] Constructing a second loss function of the discriminator according to the expectation of the second discriminant information and the first discriminant information;

[0013] Constructing a total loss function of the training system according to the mean square error between the predicted high-dose CT image and the real high-dose CT image data, the mean square error between the first discriminant information and the second discriminant information, the first loss function, and the second loss function;

[0014] (4) The training system is trained using the total loss function. At the end of the training, the generator with the determined parameters constitutes a CT image super-resolution denoising model;

[0015] (5) When applied, the low-dose CT image is input into the CT image super-resolution denoising model, and a super-resolution denoised CT image is obtained through calculation.

[0016] By introducing a light weight pixel attention block (LPAB), we can efficiently extract prominent target features and suppress the noise of surrounding environment features on target features. By introducing a multiscale detail context block (MDCB), we can expand the receptive field. By obtaining feature information of different sizes, we can better extract the context details and dependencies of target features, and fuse them with the denoised target features to enhance the target feature information and obtain a denoised super-resolution CT image.

[0017] The first feature extraction module uses at least one convolution layer, that is, uses the convolution layer to extract the low-dose CT image to obtain a shallow feature vector.

[0018] The lightweight and efficient pixel attention module includes a first channel, a second channel and a first fusion unit. The first channel includes a first branch and a second branch. Each branch uses at least one convolution to denoise the input shallow feature vector. The denoising results of the two branches are processed by a first activation function to obtain a first channel denoising result. The second channel uses at least one convolution to denoise the input shallow feature vector to obtain a second channel denoising result. The first fusion unit fuses the input first channel denoising result and the second channel denoising result to obtain a denoised feature vector.

[0019] The first branch uses at least one 1×1 convolution to realize denoising of the input shallow feature vector, and the second branch uses at least one 1×1 convolution and at least one 3×3 convolution to realize denoising of the shallow feature vector. The denoising results of the two branches are processed by the first activation function to obtain the first channel denoising result.

[0020] The first fusion unit includes a first splicing module and a lightweight module. The first splicing module is used to splice the input first channel denoising result and the second channel denoising result, and output a first splicing feature vector. The lightweight module uses at least one 1×1 convolution layer to achieve lightweight quantization of the input first splicing feature vector to obtain a denoising feature vector.

[0021] By performing multi-channel convolution on shallow features, it is possible to denoise shallow features and extract target features at the same time, thereby improving efficiency while highlighting target features. The 1×1 convolution kernel is used to reduce the parameters of the convolution operation and to lightweight the network.

[0022] The multi-scale detail context module includes a second feature extraction module, a first normalization layer, a second activation function, a second normalization layer and a third activation function. The second feature extraction module uses multiple dilated convolutions of different scales to realize multi-scale feature extraction of shallow feature vectors. The extraction results of the second feature extraction module are processed by the first normalization layer and the second activation function to obtain a multi-scale feature vector. The second fusion unit is used to fuse the multi-scale feature vectors. The fused results are processed by the second normalization layer and the third activation function to obtain a context detail closely related feature vector.

[0023] The dilated convolutions of different scales are used to expand the receptive field and obtain feature information of different sizes, thereby expanding the receptive range of the features, aggregating more detailed feature information, and better extracting the context details and dependencies of the target features through fusion.

[0024] The first loss function loss_d1 is:

[0025] loss_d1 = σ(D(y)-E[D(G(x)])

[0026] The second loss function loss_d2 is:

[0027] loss_d2 = σ(D(G(x))-E[D(y)])

[0028] The total loss function loss_d3 of the training system is:

[0029]

[0030] Where x is a low-dose CT image, y is a real high-dose CT image, σ() is a Sigmoid function, G(x) is the predicted high-dose CT image output by the generator of the low-dose CT image, D(y) is the first discriminant information output by the discriminator of the real high-dose CT image, D(G(x)) is the second discriminant information output by the discriminator of the predicted high-dose CT image, E[D(y)] is the expectation of the first discriminant information, E[D(G(x)] is the expectation of the second discriminant information, α, β and γ are weight coefficients, the generator D() is trained by the VGG network model, y i is the i-th real high-dose CT image, x i is the i-th low-dose CT image.

[0031] If the actual high-dose CT image is more realistic than the predicted high-dose CT image, the first loss function loss_d1 is 1; if the predicted high-dose CT image is more realistic than the actual high-dose CT image, the second loss function loss_d2 is 0.

[0032] When training the training system, the optimizer uses Adam, the initial learning rate is set to 0.001, and then decays by 0.1 times every 50 epochs.

[0033] The fusion module includes a second splicing module and an upsampling module. The second splicing module is used to splice the input denoising feature vector and the context detail closely related feature vector to output a second splicing feature vector. The upsampling module is used to enlarge the pixels of the input second splicing feature vector to obtain a predicted high-dose CT image.

[0034] A device for super-resolution denoising of low-dose CT images based on a generative adversarial network, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that a CT image super-resolution denoising model constructed by a method for super-resolution denoising of low-dose CT images using a generative adversarial network is used in the computer memory;

[0035] When the computer processor executes the computer program, the following steps are implemented:

[0036] The low-dose CT image is input into the CT image super-resolution denoising model, and the super-resolution denoised CT image is obtained through calculation.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention provides a lightweight and efficient block based on a pixel attention scheme (light weight pixel attention block, LPAB), which improves the expressive power of convolution and efficiently denoises low-dose CT images.

[0039] 2. The present invention provides a multiscale detail context block (MDCB) that can increase the receptive field and better extract the interdependence of detail context. It can extract the information of low-dose computed tomography (LDCT) images at different scale factors and obtain the relationship and dependency with the target information.

[0040] 3. The present invention fuses the denoised feature vector obtained by the pixel attention module with the context-detail closely connected feature vector obtained by the multi-scale detail context module through an upsampling module, expands the pixels, and finally obtains a super-resolution, denoised CT image. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of the structure of a low-dose CT image super-resolution denoising method provided by an embodiment of the present invention, wherein k represents the convolution kernel size, n represents the output channel size, s represents the step size, and p represents the padding number;

[0042] Figure 2 A schematic diagram of the structure of a multi-scale detail context module provided in an embodiment of the present invention;

[0043] Figure 3 A comparison chart of the reconstruction visualization results of different denoising methods provided in the embodiments of the present invention;

[0044] Figure 4 A CT image restored by a low-dose CT image super-resolution denoising method provided by an embodiment of the present invention, a is a low-resolution noisy low-dose CT lung image, and b is the restored CT image. DETAILED DESCRIPTION

[0045] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0046] A flow chart of a low-dose CT image super-resolution denoising method is shown in Figure 1 As shown, the specific steps are as follows:

[0047] Step 1: Collect real high-dose CT lung images, and degenerate them into low-dose CT lung images through image preprocessing methods. Each high-dose CT lung image and the corresponding low-dose CT lung image form a sample.

[0048] Step 2: Input low-dose CT images and use two parallel channels to extract feature details at different levels from the low-dose CT images.

[0049] The specific process is:

[0050] Step 2.1: Collect low-dose CT images. Extract the shallow feature H of the low-dose CT image through a 3×3 convolution. The shallow feature H can be set as shown in formula (1).

[0051] H=f shallow (I NLR ) (1)

[0052] Among them, f shallow is the 3×3 convolution function for shallow learning, I NLR It is a low-resolution CT image with noise (Noise and Low-resolution), and NLR is low resolution and noise.

[0053] Step 2.2: Adopting the strategy of two parallel channels, a mapping is constructed according to the output features corresponding to the input features. The mapping consists of 12 lightweight and efficient pixel attention modules (LPAB) and 10 multi-scale detail context modules (MDCB). In order to prevent the residual learning of each channel from decreasing with the increase of the network, as shown in Equation (2) and Equation (3), where f LPAB represents LPAB, k represents the number of LPAB, L out Represents the output of LPAB, L k-1 It is the k-1th level LPAB.

[0054]

[0055]

[0056] Among them, the specific steps to construct LPAB and improve the expressiveness of convolution are as follows:

[0057] The lightweight and efficient pixel attention module includes a first channel, a second channel and a first fusion unit. The first channel includes a first branch and a second branch. The shallow feature vector is denoised by two 1×1 convolutions of the first branch to realize the denoising of the input shallow feature vector. At the same time, the target feature vector is obtained by denoising the shallow feature vector through 1×1 convolution and 3×3 convolution of the second branch. The denoising results of the two branches are processed by the first activation function Sigmold to obtain the first channel denoising result. The second channel uses 1×1 convolution and 3×3 convolution to realize the denoising of the input shallow feature vector to obtain the second channel denoising result, and ensures the same feature vector size as the first channel denoising result. The first fusion unit splices the input first channel denoising result and the second channel denoising result, and the output first spliced ​​feature vector is subjected to 1×1 convolution to realize the lightweight quantization of the input first spliced ​​feature vector to obtain the denoised feature vector.

[0058] Step 2.3: Given the second channel containing MDCB, as in equations (4) and (5), where f MDCB represents MDCB, n represents the level of MDCB, M out Indicates the output of MDCB, M n-1 It is the n-1th level MDCB.

[0059]

[0060]

[0061] Among them, the specific steps of constructing MDCB are as follows Figure 2 As shown, the specific steps are as follows:

[0062] MDCB includes a second feature extraction module, a first batch normalization layer (BN) and a second activation function (PRelu), a second fusion unit, a second batch normalization layer (BN) and a third activation function (PRelu). The second feature extraction module uses dilated convolutions with dilation scales of 1, 2 and 3 respectively to realize multi-scale feature extraction of shallow feature vectors. The extraction results of the second feature extraction module are processed by the first batch normalization layer (BN) and the second activation function (PRelu) to obtain multi-scale feature vectors K3D1, K3D2, and K3D3, wherein K is the size of the convolution kernel and D is the size of the hole convolution kernel. The second fusion unit is used to fuse the multi-scale feature vectors. The fused result is subjected to 1×1 convolution, and then processed by the second batch normalization layer (BN) and the third activation function (PRelu) to obtain a feature vector closely connected with context details.

[0063] Step 2.4: The first fusion module includes a splicing module and an upsampling module (Upsclae). The second splicing module is used to splice the input denoising feature vector and the feature vector closely related to the context details, and output the second spliced ​​feature vector. The upsampling module is used to expand the pixels of the output second spliced ​​feature vector to obtain a predicted high-dose CT image.

[0064] As shown in formula (6), where f Upscale represents the upsampling module, f Concat is the second splicing module, I CHR For low-resolution noisy CT images (Noise and Low-resolution) CHR is high-resolution and clear, and the Pixelshuffle upsampling module is used to expand the pixels.

[0065] I CHR =f Upscale (f Concat (M out +L out )) (6)

[0066] The specific steps to build the upsampling module are as follows:

[0067] After the second concatenated feature vector is convolved with a specification of k3n256s1p1 (convolution kernel is 3, channel is 25, step size and padding number are both 1), the convolution result is input into Pixelshuffle with a resolution magnification of 2 (scale=2) through the fourth activation function, and then convolved with k3n256s1p1 to obtain the predicted high-dose CT image.

[0068] Step 3: Construct the total loss function of the training system. According to the real high-dose CT image, the discriminator outputs the first discriminant information, the low-dose CT image outputs the predicted high-dose CT image through the generator, and the predicted high-dose CT image outputs the second discriminant information through the discriminator. The expectation of the first discriminant information and the second discriminant information constructs the first loss function of the discriminator.

[0069] Constructing a second loss function of the discriminator according to the expectation of the second discriminant information and the first discriminant information;

[0070] Constructing a total loss function of the training system according to the mean square error between the predicted high-dose CT image and the real high-dose CT image data, the mean square error between the first discriminant information and the second discriminant information, the first loss function, and the second loss function;

[0071] Specifically, the total loss function constructed is:

[0072] The first loss function loss_d1 is:

[0073] loss_d1 = σ(D(y)-E[D(G(x)])

[0074] The second loss function loss_d2 is:

[0075] loss_d2 = σ(D(G(x))-E[D(y)])

[0076] The total loss function loss_d3 of the training system is:

[0077]

[0078] Where x is a low-dose CT image, y is a real high-dose CT image, σ() is a Sigmoid function, G(x) is the predicted high-dose CT image output by the generator of the low-dose CT image, D(y) is the first discriminant information output by the discriminator of the real high-dose CT image, D(G(x)) is the second discriminant information output by the discriminator of the predicted high-dose CT image, E[D(y)] is the expectation of the first discriminant information, E[D(G(x)] is the expectation of the second discriminant information, α, β and γ are weight coefficients, the generator D() is trained by the VGG network model, y i is the i-th real high-dose CT image, x i is the i-th low-dose CT image.

[0079] If the actual high-dose CT image is more realistic than the predicted high-dose CT image, the first loss function loss_d1 is 1; if the predicted high-dose CT image is more realistic than the actual high-dose CT image, the second loss function loss_d2 is 0.

[0080] The training system is trained using the total loss function. At the end of the training, the generator with determined parameters constitutes a CT image super-resolution denoising model.

[0081] Specific experimental examples

[0082] (1) Select experimental data

[0083] We used a real clinical dataset from the 2016 NIH-AAPMMayo Clinic Low-Dose CT Grand Challenge to train our model. In the experiment, 1500 CT images were randomly selected as training data and 200 CT images were used as test data. In addition, three real LDCT images were selected from the COVID-CT dataset to evaluate the robustness of the model.

[0084] (2) Experimental results

[0085] According to the steps in the above-mentioned lightweight parallel generative adversarial network based on pixel attention low-dose CT image super-resolution denoising method, a multi-scale generative adversarial network for super-resolution and denoising of low-dose CT images is trained. After the model is constructed, it is fine-tuned by loading the generator with the highest accuracy in the training model, and then the model performance is verified with the images in the validation set to obtain the super-resolution and denoising model CT-LSDGAN for low-dose CT images.

[0086] Figure 3 As shown, taking one picture in the validation set as an example, with different In order to evaluate the performance of CT-LSDGAN, we quantitatively compare our method with the most advanced restoration methods, namely BM3D+BICUBIC, SRGAN, EDSR, VDSR, and ESRGAN. Figure 3 As shown, our CT-LSDGAN obtains better metrics than other methods at 2x scale.

[0087] Figure 4 As shown in the figure, taking the three pictures in the validation set as an example, a is a high-dose CT lung image; b is the model COVID-CT restoration image effect diagram. It can be seen intuitively that COVID-CT has excellent super-resolution and denoising performance.

[0088] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A low-dose CT image super-resolution denoising method, characterized in that: The following steps are involved: (1) obtaining a real high-dose CT image and processing the real high-dose CT image into a low-dose CT image; (2) constructing a training system based on a generative adversarial network, including a generator and a discriminator, the generator including a first feature extraction module, at least one lightweight and efficient pixel attention module, at least one multi-scale detail context module and a fusion module, the first feature extraction module is used to extract a shallow feature vector from an input low-dose CT image, the lightweight and efficient pixel attention module is used to denoise the input shallow feature vector and output a denoised feature vector, the multi-scale detail context module is used to expand the receptive field of the input shallow feature vector and output a context-detail closely connected feature vector, and the fusion module is used to fuse the denoised feature vector and the context-detail closely connected feature vector to obtain a predicted high-dose CT image; (3) constructing a total loss function of the training system, outputting first discriminant information through the discriminator according to the real high-dose CT image, outputting the predicted high-dose CT image through the generator according to the low-dose CT image, and outputting second discriminant information through the discriminator according to the predicted high-dose CT image, and constructing the first loss function of the discriminator according to the expectation of the first discriminant information and the second discriminant information; Constructing a second loss function of the discriminator according to the expectation of the second discriminant information and the first discriminant information; Constructing a total loss function of the training system according to the mean square error between the predicted high-dose CT image and the real high-dose CT image data, the mean square error between the first discriminant information and the second discriminant information, the first loss function, and the second loss function; (4) The training system is trained using the total loss function. At the end of the training, the generator with the determined parameters constitutes a CT image super-resolution denoising model; (5) When applied, the low-dose CT image is input into the CT image super-resolution denoising model, and a super-resolution denoised CT image is obtained through calculation; The first loss function loss_d1 is: loss_d1 = σ(D(y)-E[D(G(x)]) The second loss function loss_d2 is: loss_d2 = σ(D(G(x))-E[D(y)]) The total loss function loss_d3 of the training system is: Where x is a low-dose CT image, y is a real high-dose CT image, σ() is a Sigmoid function, G(x) is the predicted high-dose CT image output by the generator of the low-dose CT image, D(y) is the first discriminant information output by the discriminator of the real high-dose CT image, D(G(x)) is the second discriminant information output by the discriminator of the predicted high-dose CT image, E[D(y)] is the expectation of the first discriminant information, E[D(G(x)] is the expectation of the second discriminant information, α, β and γ are weight coefficients, the generator D() is trained by the VGG network model, y i is the i-th real high-dose CT image, x i is the i-th low-dose CT image.

2. The low-dose CT image super-resolution denoising method according to claim 1, characterized in that: The first feature extraction module uses at least one convolution layer, that is, uses the convolution layer to extract the low-dose CT image to obtain a shallow feature vector.

3. The low-dose CT image super-resolution denoising method according to claim 1, characterized in that: The lightweight and efficient pixel attention module includes a first channel, a second channel and a first fusion unit. The first channel includes a first branch and a second branch. Each branch uses at least one convolution to denoise the input shallow feature vector. The denoising results of the two branches are processed by a first activation function to obtain a first channel denoising result. The second channel uses at least one convolution to denoise the input shallow feature vector to obtain a second channel denoising result. The first fusion unit fuses the input first channel denoising result and the second channel denoising result to obtain a denoised feature vector.

4. The low-dose CT image super-resolution denoising method according to claim 3, characterized in that: The first branch uses at least one 1×1 convolution to realize denoising of the input shallow feature vector, and the second branch uses at least one 1×1 convolution and at least one 3×3 convolution to realize denoising of the shallow feature vector. The denoising results of the two branches are processed by the first activation function to obtain the first channel denoising result.

5. The low-dose CT image super-resolution denoising method according to claim 3, characterized in that: The first fusion unit includes a first splicing module and a lightweight module. The first splicing module is used to splice the input first channel denoising result and the second channel denoising result, and output a first splicing feature vector. The lightweight module uses at least one 1×1 convolution layer to achieve lightweight quantization of the input first splicing feature vector to obtain a denoising feature vector.

6. The low-dose CT image super-resolution denoising method according to claim 1, characterized in that: The multi-scale detail context module includes a second feature extraction module, a first normalization layer, a second activation function, a second normalization layer and a third activation function. The second feature extraction module uses multiple dilated convolutions of different scales to realize multi-scale feature extraction of shallow feature vectors. The extraction results of the second feature extraction module are processed by the first normalization layer and the second activation function to obtain a multi-scale feature vector. The second fusion unit is used to fuse the multi-scale feature vectors. The fused results are processed by the second normalization layer and the third activation function to obtain a context detail closely related feature vector.

7. The low-dose CT image super-resolution denoising method according to claim 1, characterized in that: When training the training system, the optimizer uses Adam, the initial learning rate is set to 0.001, and then decays by 0.1 times every 50 epochs.

8. The low-dose CT image super-resolution denoising method according to claim 1, characterized in that: The fusion module includes a second splicing module and an upsampling module. The second splicing module is used to splice the input denoising feature vector and the context detail closely related feature vector to output a second splicing feature vector. The upsampling module is used to enlarge the pixels of the input second splicing feature vector to obtain a predicted high-dose CT image.

9. A low-dose CT image super-resolution denoising device, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that: The computer memory contains a CT image super-resolution denoising model constructed by the low-dose CT image super-resolution denoising method according to any one of claims 1 to 8; When the computer processor executes the computer program, the following steps are implemented: The low-dose CT image is input into the CT image super-resolution denoising model, and the super-resolution denoised CT image is obtained through calculation.

Images