A lung CT image super-resolution method based on improved SRGAN

Abstract

The application discloses a kind of lung CT image super-resolution methods based on improved SRGAN, applied to computer vision technical field, including the following steps: S1, construct low resolution-high resolution image pair, pre-process lung CT image, construct data set;S2, design a kind of multi-scale residual block MS_ResB replaces original residual block in SRGAN;S3, design the perceptual loss function based on Charbonnier loss;S4, build the lung CT image generative adversarial network based on multi-scale residual block;S5, the data set of S1 is constructed to the lung CT image generative adversarial network based on multi-scale residual block built in S4 is trained;S6, by the network model of training obtained in S5, low resolution image is carried out super-resolution, the present application can extract image multi-granularity features, including different levels of global and local features, so as to retain more original and detail information in the generated super-resolution image, help to construct more realistic super-resolution image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and more particularly to a lung CT image super-resolution method based on an improved SRGAN. BACKGROUND

[0002] According to the GLOBOCAN 2020 cancer report published by the World Health Organization, the incidence of lung cancer is 11.4%, second only to breast cancer (incidence rate of 11.7%), and the mortality rate of lung cancer is 18%, ranking the highest. As can be seen, lung cancer is a disease with extremely high incidence and the highest mortality rate in the world. Early diagnosis and treatment of lung diseases can improve the survival rate of lung cancer by at least 30%. Medical lung CT images are important information for doctors to analyze and diagnose the patient's condition. Since the early symptoms of lung disease are not obvious, high-resolution lung CT images can improve the accuracy of diagnosis and have important significance for early diagnosis of the disease. However, medical CT images, including lung CT images, are limited by imaging resolution, image acquisition time, and patient tolerable radiation dose of medical equipment during imaging, and the resolution of the obtained images is limited. The problem of low-resolution CT images is that poor image quality leads to loss of detailed information, which seriously interferes with medical diagnosis. Therefore, improving the quality of CT images and speeding up the reconstruction speed are of particular importance to diagnosis. At present, there are three common methods to improve the resolution of medical CT images:

[0003] 1. Increase the number of sensor pixels or increase the photosensitive area of the sensor, but this method often brings more serious scattering noise and increases the size of the imaging device.

[0004] 2. Change the scanning mode of the imaging device, which requires high precision of medical equipment and increases the radiation dose.

[0005] 3. Super-resolution reconstruction technology. Image resolution refers to the number of pixels in an image, which can directly reflect the information contained in the image. The higher the image resolution, the clearer the appearance of the image, and the more details contained in the unit area, and more information can be obtained from the image. Image super-resolution (SR) refers to recovering a high-resolution (HR) image from a low-resolution (LR) image or image sequence. SR reconstruction has been a research hotspot in the field of computer vision for a long time, especially in the field of SR reconstruction of medical images.

[0006] Under the condition of not changing the device hardware condition and radiation dose, the resolution of the lung CT image is improved by using the super resolution technology, which better assists the clinical diagnosis of medical staff, and has important significance in the field of SR reconstruction of medical images. With the rapid development of deep learning, the super resolution based on generative adversarial network (GAN) has more reliable visual effect, but there are still several problems for medical CT images: (1) the medical CT image dataset is relatively small, the training sample is less due to the difficulty of collection, the precision is not high, and the traditional data enhancement method (flip, translation, etc.) can increase the data quantity to a certain extent, but there are problems of poor effect or large error; (2) the existing super resolution methods (SRCNN, SRGAN, ESRGAN, etc.) are mostly based on natural images, and the natural images themselves contain a large amount of information, while the medical CT images themselves contain single and small amount of information, so this kind of method cannot be directly used for medical images; (3) there are few methods specially for medical CT image super resolution, especially for the super resolution of lung CT images, and most of them use residual network as the backbone network, the residual network model is large, the training time is long, and at the same time, the visual perception quality cannot be improved. Therefore, a medical lung CT image super resolution method which can solve the above problems is urgently needed. SUMMARY

[0007] Therefore, the application provides a lung CT image super resolution method based on an improved SRGAN, which can retain more original and detailed information in the generated super-resolution image, combines the generative adversarial network GAN with the super resolution demand of the lung CT image, and generates a lung CT super-resolution image with high visual perception quality.

[0008] In order to achieve the above purpose, the application provides the following technical scheme:

[0009] A lung CT image super resolution method based on an improved SRGAN, comprising the following steps:

[0010] S1, constructing a low resolution-high resolution image pair, preprocessing the lung CT image, and constructing a dataset;

[0011] S2, designing a multi-scale residual block MS_ResB to replace the original residual block in SRGAN;

[0012] S3, designing a perception loss function based on Charbonnier loss;

[0013] S4, building a lung CT image generative adversarial network based on the multi-scale residual block;

[0014] S5, training the lung image generative adversarial network based on the multi-scale residual block built in S4 by the data set built in S1;

[0015] S6, super-resolution of the low-resolution image by the network model trained in S5.

[0016] Preferably, the construction of the low-resolution-high-resolution image pair in S1 is specifically: dividing the lung CT image into a training set and a test set, performing down-sampling algorithm processing on the images in the training set to obtain a low-resolution image I LR , and the low-resolution image and the real high-resolution image I HR in the original training set constitute a low-resolution-high-resolution image pair.

[0017] Preferably, the multi-scale residual block designed in S2 includes 2 1*1 convolution layers and 6 3*3 convolution layers, and the 1*1 convolution layer is used to reduce the parameter amount, and the 3*3 convolution layer is used to extract image features.

[0018] Preferably, the lung CT image generative adversarial network based on the multi-scale residual block in S4 includes a generator Gnet and a discriminator Dnet, the generator Gnet includes: 3 convolution blocks, 5 multi-scale residual blocks, and 1 up-sampling module; wherein the first generator convolution block is used for shallow feature extraction of the image, the 5 multi-scale residual blocks and the second generator convolution block are used for high-level feature extraction of the image, the up-sampling module is used to restore the size of the feature map to the size of the target super-resolution image, and the third convolution block is used for super-resolution image reconstruction; the discriminator Dnet includes 8 discriminator convolution blocks, 2 fully connected layers, and 2 activation functions; the 8 discriminator convolution blocks respectively include an activation function, wherein the second, fourth, sixth and eighth discriminator convolution blocks have a step of 2, which functions to reduce the size of the feature map while reducing the calculation amount, and the first, third, fifth and seventh discriminator convolution blocks have a step of 1; the first fully connected layer has a dimension of 1024, and the second fully connected layer has a dimension of 1; the last layer activation function adopts a sigmoid activation function, and the remaining activation functions all select gelu as the activation function.

[0019] Preferably, S5 is specifically:

[0020] S51, initializing the generator Gnet, the discriminator Dnet and the optimizer;

[0021] S52, inputting the low-resolution image I LR into the generator Gnet to generate a high-resolution image I SR ;

[0022] S53, respectively inputting the high-resolution image I SRand real high-resolution image I HR In the input discriminator Dnet, the probability value of determining the input image as the real high-resolution image I HR ;

[0023] S54, calculate the loss of the discriminator Dnet, and the discriminator loss is back propagated to the input side, and the gradient value of the discriminator Dnet is calculated according to the discriminator loss;

[0024] S55, calculate the loss of the generator Gnet, and the generator loss is back propagated to the input side, and the gradient value of the generator Gnet is calculated according to the generator loss;

[0025] S56, update the parameters of the discriminator Dnet and the generator Gnet in turn, the parameters are bias and weight.

[0026] Preferably, the S56 further comprises: S57, test the S56 multi-scale residual block based lung CT image generative adversarial network by the test set.

[0027] Preferably, the discriminator Dnet loss function in the S54 is:

[0028]

[0029] The loss function of the generator Gnet in the S55 is:

[0030] Gnet_loss = cbn_loss + a·adv_loss + β·VGG_loss

[0031] In the formula, cbn_loss represents Charbonnier loss, adv_loss represents adversarial loss, VGG_loss represents VGG loss, and a and β are weight parameters of adversarial loss and VGG loss in the generator loss; The calculation method of cbn_loss is as follows:

[0032]

[0033] In the formula, ε = 1 × 10 -6 ;

[0034] The calculation formula of adv_loss is as follows:

[0035]

[0036] In the formula, represents a generator network with a parameter θ G , θ G = {W 1:L ; b 1:L} represents the parameters of each layer in the network. Represents a parameter θ G The discriminator network, θ D ={W 1:L b 1:L} represents the parameters of each layer in the network;

[0037] The VGG loss is calculated as follows: The high-resolution image I... SR and true high-resolution images I HR The I values ​​obtained by inputting the VGG19 network respectively SR Feature maps and I HR The VGG loss is obtained by calculating the Euclidean distance between the two feature maps, as shown below:

[0038]

[0039] In the formula, W i,j and H i,j This represents the width and height of the feature map in the VGG network, where x takes values ​​in the range [1, W]. i,j ], the range of y is [1, H i,j ], This represents the feature map output after the input image passes through the j-th convolutional layer (after the activation function) following the i-th max pooling layer in the VGG19 network.

[0040] Preferably, the optimizer in S51 is the ADAM optimization algorithm, and the initial learning rate lr is 1×10. -3 The momentum parameters are [β1, β2] = [0.9, 0.999].

[0041] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a lung CT image super-resolution method based on an improved SRGAN, which has the following beneficial effects:

[0042] 1. To address the issue of limited sample size in medical images, the multi-scale residual block MS_ResB designed in this invention can extract multi-granularity features from images, including global and local features at different levels, so as to retain more original and detailed information in the generated super-resolution image, which helps to construct a more realistic super-resolution image;

[0043] 2. The skip connection structure in the generator network Gnet of the present invention sequentially transmits shallow features (including edges, textures, etc.) to each layer in the network, which effectively transmits and preserves the detailed information of the image, and helps to recover a super-resolution image with less noise.

[0044] 3. The multi-scale residual block MS_ResB of the present application removes the BN layer and only has convolution layers, which can improve performance while reducing model complexity, and can also improve the problem of excessive smoothing of super-resolution results;

[0045] 4. The perception loss function based on Charbonnier loss designed in the present application can better handle outliers and improve performance, and solves the problems of excessive smoothing of super-resolution images, lack of high-frequency details, and lack of visual authenticity of generated super-resolution images caused by the original loss function. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0047] Figure 1 The flowchart of the present application;

[0048] Figure 2 The schematic diagram of the multi-scale residual block of the present application;

[0049] Figure 3 The schematic diagram of the generator Gnet of the present application;

[0050] Figure 4 The schematic diagram of the discriminator Dnet of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] The present application discloses a lung CT image super-resolution method based on improved SRGAN, as shown in Figure 1 The method comprises the following steps:

[0053] The method comprises the following steps: Figure 1

[0054] S1, construct a low-resolution-high-resolution image pair, pre-process the lung CT image, and construct a data set;

[0055] ​S2, design a multi-scale residual block MS_ResB (Multi-scale Resblock) to replace the original residual block in SRGAN;

[0056] S3, design a perception loss function based on Charbonnier loss;

[0057] S4, build a lung CT image generative adversarial network based on a multi-scale residual block;

[0058] S5, train the lung image generative adversarial network based on the multi-scale residual block built in S4 by using the dataset constructed in S1;

[0059] S6, perform super-resolution on a low-resolution image by using the network model trained in S5.

[0060] In another embodiment of the application, the low-resolution-high-resolution image pair constructed in S1 is specifically: the lung CT image is divided into a training set and a test set, the images in the training set are processed by a downsampling algorithm to obtain a low-resolution image I LR , and the low-resolution image and the real high-resolution image I HR in the original training set constitute a low-resolution-high-resolution image pair.

[0061] In another embodiment of the application, as shown in Figure 2 , the multi-scale residual block designed in S2 includes two 1*1 convolution layers and six 3*3 convolution layers, the 1*1 convolution layer is used to reduce the parameter quantity, and the 3*3 convolution layer is used to extract image features, and the residual block only includes convolution operation and does not include other operations such as normalization.

[0062] In another embodiment of the application, the lung CT image generative adversarial network based on the multi-scale residual block in S4 includes a generator Gnet and a discriminator Dnet, the generator Gnet includes: three generator convolution blocks, five multi-scale residual blocks, and one upsampling module; wherein the first generator convolution block is used for shallow feature extraction of the image, the five multi-scale residual blocks and the second generator convolution block are used for high-level feature extraction of the image, the upsampling module is used for restoring the size of the feature map to the size of the target super-resolution image, and the third convolution block is used for super-resolution image reconstruction; as Figure 3As shown, the image extracts shallow features through the first convolutional block, then extracts high-level features through 5 multi-scale residual blocks and the second generator convolutional block, and after passing through the upsampling module, the size is restored to the size of the target super-resolution image, and finally the resolution image is reconstructed through the third generator convolutional block. Among them, the 5 multi-scale residual blocks, the second generator convolutional block and the upsampling module form a skip connection, which sequentially transmits the shallow features (including edges, textures, etc.) to each layer in the network, well transmitting and maintaining the detail information of the image, which helps to restore the super-resolution image with less noise.

[0063] The discriminator Dnet includes 8 discriminator convolutional blocks, 2 fully connected layers, and 2 activation functions; as shown in the formula (1), the 8 discriminator convolutional blocks each include an activation function, wherein the second, fourth, sixth and eighth discriminator convolutional blocks have a step size of 2, which functions to reduce the size of the feature map while reducing the calculation amount, and the first, third, fifth and seventh discriminator convolutional blocks have a step size of 1; the first fully connected layer has a dimension of 1024, and the second fully connected layer has a dimension of 1; the last layer activation function adopts a sigmoid activation function, and the remaining activation functions all select gelu as the activation function. Figure 4

[0064] In another embodiment of the present application, S5 is specifically:

[0065] S51, initializing the generator Gnet, the discriminator Dnet and the optimizer;

[0066] S52, inputting a low-resolution image I LR into the generator Gnet to generate a high-resolution image I SR .

[0067] S53, inputting the high-resolution image I SR and the real high-resolution image I HR into the discriminator Dnet respectively, to determine the probability value of the input image being the real high-resolution image I HR .

[0068] S54, calculating the loss of the discriminator Dnet, and performing back propagation of the discriminator loss to the input side, and calculating the gradient value of the discriminator Dnet according to the discriminator loss;

[0069] S55, calculating the loss of the generator Gnet, and performing back propagation of the generator loss to the input side, and calculating the gradient value of the generator Gnet according to the generator loss;

[0070] S56, sequentially updating the parameters of the discriminator Dnet and the generator Gnet, and the parameters are bias and weight.

[0071] ​In another embodiment of the present application, the discriminator Dnet loss function in S54 is:

[0072]

[0073] The loss function of the generator Gnet in S55 is:

[0074] Gnet_loss = cbn_loss + a · adv_loss + β · VGG_loss

[0075] where cbn_loss represents Charbonnier loss, adv_loss represents adversarial loss, VGG_loss represents VGG loss, and a and β are weight parameters of the adversarial loss and the VGG loss in the generator loss. The calculation method of cbn_loss is as follows:

[0076]

[0077] where ε = 1 × 10 -6 ;

[0078] The calculation formula of adv_loss is as follows:

[0079]

[0080] where G represents a generator network with parameters θ G , θ G = {W 1:L ; b 1:L} represents the parameters of each layer in the network, and D represents a discriminator network with parameters θ G , θ D = {W 1:L ; b 1:L} represents the parameters of each layer in the network.

[0081] The calculation method of VGG loss is as follows: high-resolution image I SR and real high-resolution image I HR are input into the VGG19 network to obtain I SR feature map and I HR feature map, and the Euclidean distance between the two feature maps is calculated to obtain the VGG loss.

[0082]

[0083] where W i,j and H i,j represent the width and height of the feature map in the VGG network, and x takes values in the range of [1, W i,j], y is in the range of [1, H i,j ], represents a feature map output after the input image passes through the VGG19. represents a feature map output after the input image passes through the jth convolutional layer after the ith max-pooling layer in the VGG19 network (after the activation function).

[0084] In another embodiment of the application, the optimizer in S51 is an ADAM optimization algorithm, and the initial learning rate lr is 1x10 -3 , and the momentum parameters [β1, β2] = [0.9, 0.999].

[0085] In another embodiment of the application, after S56, there is further S57, testing the multi-scale residual block-based lung CT image generative adversarial network of S56 by using the test set, in particular:

[0086] The TCGA-LUAD lung cancer lung CT dataset in the American tumor genome atlas (TCGA) database is used as dataset 1, and the lung CT image dataset DeepLesion released by the latest clinical center of the National Institutes of Health (NIH) is used as dataset 2. 418 lung CT images are randomly selected from the TCGA-LUAD lung cancer lung CT dataset as a training set, 52 are randomly selected as a test set, and 470 lung CT images are randomly selected from DeepLesion as a training set, and 30 are randomly selected as a test set. In the test, the batchsize is set to 4, an image region with a size of 256x256 is randomly cropped from a single image for training, the Adam optimization algorithm is used, the initial learning rate is 10-3, and the training epoch is 50. In this embodiment, experiments are performed when the up-sampling factor is 4, and the peak signal to noise ratio PSNR (peak signal to noise ratio, unit: dB, the larger the value, the smaller the distortion) and the structural similarity SSIM (structural similarity, the larger the value, the smaller the image distortion) two evaluation indexes are selected to objectively evaluate the performance of the algorithm.

[0087] The test results (scale=4) are shown in Table 1 and Table 2. Among them, × represents not using the part, √ represents using the part; the best result is shown in bold; Table 1 is a comparison of test results on dataset 1, and Table 2 is a comparison of test results on dataset 2. As can be seen from the data in the first and third rows of Table 1, the reconstructed charboniner-based perceptual loss is very effective, and the network trained using the perceptual loss function of the application achieves the best effect whether it is the optimal value of PSNR and SSIM or the average value of 50 epochs. The results in the second and third rows of Table 1 show that, under the same other experimental conditions, the optimal PSNR, SSIM and average value of 50 epochs are all optimal, indicating that the effect of the multi-scale residual block MS_ResB proposed in the application is better than that of the residual block in the original SRGAN network. Table 2 is a comparison of results on dataset 2, and the analysis method is the same as Table 1. As can be seen from Table 2, the loss function reconstructed in the application and the designed residual block MS_ResB make the network optimal in the three indicators of optimal PSNR, SSIM and average PSNR / SSIM. In summary, the results in Tables 1 and 2 fully demonstrate the role and effect of the residual block MS_ResB and the charboniner-based perceptual loss proposed in the application, and the super-resolution network of the lung CT image constructed based on the two has a small distortion in super-resolution results and more details, which is closer to the real image. The experiment effectively shows from the objective quality that the method of the application helps to construct more realistic and detailed high-resolution lung CT images.

[0088] Table 1: Comparison of results on dataset 1 (scale=4)

[0089]

[0090] Table 2: Comparison of results on dataset 2 (scale=4)

[0091]

[0092] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0093] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lung CT image super-resolution method based on improved SRGAN, characterized in that, The method comprises the following steps: S1, constructing a low-resolution-high-resolution image pair, preprocessing the lung CT image, and constructing a data set; S2, designing a multi-scale residual block MS_ResB to replace the original residual block in SRGAN; S3, designing a perception loss function based on Charbonnier loss; S4, building a lung CT image generative adversarial network based on a multi-scale residual block; S5, training the lung image generative adversarial network based on the multi-scale residual block built in S4 through the data set constructed in S1; S6, performing super-resolution on the low-resolution image through the network model trained in S5. The multi-scale residual block designed in S2 comprises two 1*1 convolution layers and six 3*3 convolution layers, the 1*1 convolution layer is used to reduce the parameter quantity, and the 3*3 convolution layer is used to extract image features; the multi-scale residual block only contains convolution operation and does not contain other operations; The lung CT image generative adversarial network based on the multi-scale residual block in S4 comprises a generator Gnet and a discriminator Dnet, the generator Gnet comprises three convolution blocks, five multi-scale residual blocks, and one up-sampling module; Wherein the first generator convolution block is used for shallow feature extraction of the image, the five multi-scale residual blocks and the second generator convolution block are used for high-level feature extraction of the image, the up-sampling module is used for restoring the size of the feature map to the size of the target super-resolution image, and the third generator convolution block is used for super-resolution image reconstruction; the convolution kernel size of all convolution blocks in the discriminator Dnet is 3*3, wherein the second, fourth, sixth and eighth convolution layers have a step size of 2, which reduces the size of the feature map and reduces the calculation amount, and the first, third, fifth and seventh convolution layers have a step size of 1; the first full connection layer has a dimension of 1024, and the second full connection layer has a dimension of 1; the last layer adopts a sigmoid activation function, and the activation layers connected to the remaining convolution blocks all select gelu as the activation function; the five multi-scale residual blocks, the second generator convolution block and the up-sampling module form a skip connection structure, and the shallow features are sequentially transmitted to each layer in the network.

2. The lung CT image super-resolution method based on improved SRGAN according to claim 1, characterized in that, The low-resolution-high-resolution image pair constructed in the S1 is specifically: dividing the lung CT image into a training set and a test set, performing a down-sampling algorithm on the image in the training set to obtain a low-resolution image I LR , and the low-resolution image and the real high-resolution image I HR in the original training set constitute a low-resolution-high-resolution image pair.

3. The lung CT image super-resolution method based on improved SRGAN according to claim 2, characterized in that, The S5 specifically comprises: S51, initializing the generator Gnet, the discriminator Dnet and the optimizer; S52, inputting a low-resolution image I LR to the generator Gnet, generating a high-resolution image I SR ; S53, respectively, the high-resolution image I SR and the real high-resolution image I HR into the discriminator Dnet, determines a probability value that the input image is the real high-resolution image I HR . S54, calculating the loss of the discriminator Dnet, and performing back propagation of the discriminator loss to the input side, and calculating the gradient value of the discriminator Dnet according to the discriminator loss; S55, calculating the loss of the generator Gnet, and performing back propagation of the generator loss to the input side, and calculating the gradient value of the generator Gnet according to the generator loss; S56, sequentially updating the parameters of the discriminator Dnet and the generator Gnet, the parameters being bias and weight.

4. The lung CT image super-resolution method based on improved SRGAN of claim 3, characterized in that, The S56 further comprises S57, testing the lung CT image generative adversarial network based on the multi-scale residual block in S56 through the test set.

5. The lung CT image super-resolution method based on improved SRGAN according to claim 3, characterized in that, The loss function of the discriminator Dnet in S54 is: The loss function of the generator Gnet in S55 is: Gnet_loss = cbn_loss + a * adv_loss + b * VGG_loss In the formula, cbn_loss represents Charbonnier loss, adv_loss represents adversarial loss, VGG_loss represents VGG loss, and a and b are weight parameters of the adversarial loss and the VGG loss in the generator loss; the calculation method of cbn_loss is as follows: where ε = 1 x 10 -6 ; The calculation formula of adv_loss is as follows: In the formula, Represents a parameter θ G The generator network, θ G ={W 1:L b 1:L } represents the parameters of each layer in the network. Represents a parameter θ G The discriminator network, θ D ={W 1:L b 1:L } represents the parameters of each layer in the network; The VGG loss is calculated in the following manner: the high-resolution image I SR and the real high-resolution image I HR are respectively input into the VGG19 network to obtain I SR feature maps and I HR feature maps, and the Euclidean distance between the two feature maps is calculated to obtain the VGG loss, which is specifically represented as follows: wherein W i,j and H i,j represent the width and height of the feature map in the VGG network, x takes values in the range [1, W i,j ] and y takes values in the range [1, H i,j ], represents the output of the jth convolutional layer after the ith max-pooling layer in the VGG19 network.

6. The lung CT image super-resolution method based on improved SRGAN according to claim 3, characterized in that, The optimizer in S51 is ADAM optimization algorithm, the initial learning rate lr is 1x10 -3 , and the momentum parameter [β1, β2] = [0.9, 0.999].

Images