CT image super-resolution reconstruction method based on generative adversarial network

Abstract

The invention provides a CT (Computed Tomography) image super-resolution reconstruction method based on a generative adversarial network. The method comprises the following steps of: 1, collecting high-resolution image data and corresponding low-resolution image data; 2, establishing a CycleGAN model comprising a first generator, a second generator, a first discriminator and a second discriminator; a third step of training a CycleGAN model by using the collected high-resolution image data and the corresponding low-resolution image data so as to optimize parameters of the CycleGAN model; and a fourth step of performing image conversion on the to-be-enhanced medical image which is input into the trained CycleGAN model.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and medical technology, and in particular to a method for super-resolution reconstruction of electronic computed tomography (CT) images using a High-to-Low Cycle Generative Adversarial Network (HLCGAN) and a Low-to-High Cycle Generative Adversarial Network (LHCGAN), which can enhance the texture details and clarity of the CT images. Background Art

[0002] The uneven distribution of medical resources in remote areas makes it difficult to effectively improve the medical level in remote areas. Under this trend, residents in remote areas are facing the current situation of urgent need for high-quality medical services. In addition, hospitals in remote areas cannot afford the high cost of purchasing high-configuration CT machines, so they can only choose CT machines with slightly worse imaging effects and cheaper prices. The resolution of the CT images taken by this method is low, which makes it more difficult for doctors to judge the condition.

[0003] Therefore, it is desirable to computationally improve noisy low-resolution CT (LRCT) images acquired under low-dose CT (LDCT) protocols to the quality of high-resolution CT (HRCT) images.

[0004] The main challenges in restoring HRCT images are as follows. First, LRCT images contain different or more complex spatial variations, correlations, and statistical properties than natural images, which limits the super-resolution (SR) imaging performance of traditional methods. Second, during the reconstruction process, the noise in the original projection data is introduced into the image domain, resulting in unique noise and artifact patterns. This makes it difficult for the algorithm to produce high-quality image results. Finally, because the sampling and degradation operations are coupled and ill-conditioned, SR tasks performed using traditional methods cannot effectively restore some detailed features and there is a risk of generating blurred appearances and new artifacts. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a CT image super-resolution reconstruction method based on a generative adversarial network in view of the above-mentioned defects in the prior art, by combining multiple nonlinear SR functional blocks for SR CT (SRCT) imaging, accompanied by a residual module to learn high-frequency details, and then performing adversarial learning in a cyclic manner to generate SRCT images that are better in both perception and quantity.

[0006] According to the present invention, a CT image super-resolution reconstruction method based on a generative adversarial network is provided, comprising:

[0007] The first step: collecting high-resolution image data and corresponding low-resolution image data;

[0008] Step 2: Establish a CycleGAN model including a first generator, a second generator, a first discriminator, and a second discriminator;

[0009] The third step: using the collected high-resolution image data and the corresponding low-resolution image data to train the CycleGAN model to optimize the parameters of the CycleGAN model;

[0010] Step 4: Perform image conversion on the medical image to be enhanced by inputting the trained CycleGAN model.

[0011] Preferably, the third step trains the CycleGAN model using a loss function based on forward and backward cycle consistency to ensure consistency of image content during the conversion process.

[0012] Preferably, the loss function combines adversarial loss, cycle consistency loss, identity loss and joint sparse transformation loss.

[0013] Preferably, the first generator maps the input image from a high-resolution image domain to a low-resolution image domain, performs image conversion from high resolution to low resolution, and the first discriminator evaluates the image quality of the low-resolution image domain generated by the first generator using a first standard; and the second generator maps the input image from a low-resolution image domain to a high-resolution image domain, performs image conversion from low resolution to high resolution, and the second discriminator evaluates the image quality of the high-resolution image domain generated by the second generator using a second standard.

[0014] Preferably, the operation of the first generator mapping the input image from the high-resolution image domain to the low-resolution image domain is reversible, and the operation of the second generator mapping the input image from the low-resolution image domain to the high-resolution image domain is reversible.

[0015] Preferably, in the fourth step, the first generator converts the medical image to be enhanced into an enhanced high-quality image.

[0016] Preferably, in the CycleGAN model, the first generator and the first discriminator constitute a high-resolution-low-resolution generative adversarial network, the second discriminator and the second generator constitute a low-resolution-high-resolution generative adversarial network, and the third step uses unpaired high-resolution data and low-resolution image data to train the high-resolution-low-resolution generative adversarial network to learn to downgrade and downsample the high-resolution images; subsequently, using the paired input and output of the trained high-resolution-low-resolution generative adversarial network, the low-resolution-high-resolution generative adversarial network is trained to perform image super-resolution processing using paired low-resolution and high-resolution images.

[0017] Preferably, the CT image super-resolution reconstruction method based on the generative adversarial network also includes a fifth step: performing model evaluation on the trained CycleGAN model.

[0018] Preferably, the first generator and the second generator use the same convolutional neural network structure but have different convolutional neural network parameters.

[0019] Preferably, the acquired images include MR images and CT images, the first step further includes an image preprocessing operation, and the fourth step further includes an image postprocessing operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] A more complete understanding of the present invention and its attendant advantages and features will be more readily appreciated by reference to the following detailed description taken in conjunction with the accompanying drawings, in which:

[0021] Figure 1 The overall flow chart of the CT image super-resolution reconstruction method based on the generative adversarial network according to the preferred embodiment of the present invention is schematically shown.

[0022] Figure 2 A schematic diagram of a CycleGAN model according to a preferred embodiment of the present invention is schematically shown.

[0023] Figure 3 The original CT image and the image after super-resolution reconstruction by the model are shown.

[0024] Figure 4 The original CT image and the SR-CT image of the model are shown.

[0025] Figure 5 It is shown that SR-CT images show clearer capillary details than LRCT images.

[0026] Figure 6 and Figure 7 The Gaussian convolution process is schematically shown.

[0027] It should be noted that the drawings are used to illustrate the present invention, rather than to limit the present invention. Note that the drawings showing the structures may not be drawn to scale. In addition, in the drawings, the same or similar elements are marked with the same or similar reference numerals. DETAILED DESCRIPTION

[0028] In order to make the contents of the present invention clearer and easier to understand, the contents of the present invention are described in detail below in conjunction with specific embodiments and drawings.

[0029] Figure 1 The overall flow chart of the CT image super-resolution reconstruction method based on the generative adversarial network according to the preferred embodiment of the present invention is schematically shown.

[0030] like Figure 1 As shown, the CT image super-resolution reconstruction method based on the generative adversarial network according to the preferred embodiment of the present invention includes:

[0031] The first step S1: collect high-resolution image data and corresponding low-resolution image data.

[0032] Preferably, in the first step S1, after collecting the high-resolution image data and the corresponding low-resolution image data, data preprocessing is further performed on the high-resolution image data and the low-resolution image data.

[0033] Specifically, before medical CT image enhancement, appropriate CT image data is selected as the input of the model. This includes selecting representative and diverse CT images from various scanned parts (for example: head CT, chest CT, spine CT, abdomen and pelvic CT, etc.). After the CT image is selected, a series of preprocessing operations are required to adapt to the requirements of the model. For example, the image needs to be cropped or resized first to ensure the consistency of the input. Then, in order to improve the training effect and stability of the model, the image is also normalized so that its pixel values ​​fall within a standard range. In addition, some data enhancement techniques such as rotation, flipping and scaling can also be applied to enrich the diversity of the data and improve the generalization ability of the model.

[0034] The second step S2: establish a first generator G AB , the second generator G BA And the first discriminator D A , the second discriminator D B CycleGAN model, such as Figure 2 shown.

[0035] Among them, the first generator G AB Map the input image from the high-resolution image domain to the low-resolution image domain, and perform the image conversion from high resolution to low resolution. The first discriminator D AThe image quality of the low-resolution image domain generated by the first generator is evaluated using the first criterion; and the second generator G BA Map the input image from the low-resolution image domain to the high-resolution image domain, and perform the image conversion from low resolution to high resolution. The second discriminator D B The image quality of the high-resolution image domain generated by the second generator is evaluated using the second standard. The first standard and the second standard can be set based on specific image quality requirements.

[0036] CycleGAN is an unsupervised method that aims to learn MR-to-CT mapping using unpaired MR and CT images, i.e., multiple MR and CT images from different subjects without deformation alignment. In this stage, the CycleGAN model will be initialized and built.

[0037] The third step S3: training the CycleGAN model using the collected high-resolution image data and the corresponding low-resolution image data to optimize the parameters of the CycleGAN model;

[0038] Preferably, the CycleGAN model is trained using a loss function based on forward and backward cycle consistency to ensure the consistency of image content during the conversion process.

[0039] Specifically, for example, in the CycleGAN model, the first generator G AB and the first discriminator D A Composed of high-resolution-low-resolution generative adversarial network, the second discriminator D B and the second generator G BA A low-resolution-high-resolution generative adversarial network is formed, and in the third step, the high-resolution-low-resolution generative adversarial network is trained to learn to downgrade and downsample the high-resolution images using unpaired high-resolution data and low-resolution image data; subsequently, the paired input and output of the trained high-resolution-low-resolution generative adversarial network is used to train the low-resolution-high-resolution generative adversarial network to perform image super-resolution processing using paired low-resolution and high-resolution images.

[0040] The goal of the model training phase is to optimize the parameters of the model so that it can complete high-quality image conversion tasks. Moreover, in the absence of paired training samples, CycleGAN is trained through a special loss function that takes into account the forward and reverse cycle consistency to ensure that the consistency of image content is maintained during the conversion process. Preferably, the training process will continue for multiple cycles until the performance of the model reaches a satisfactory level or meets a predetermined stopping criterion.

[0041] The fourth step S4: performing image conversion on the medical image to be enhanced which is input into the trained CycleGAN model.

[0042] In the fourth step S4, the first generator G AB Convert the medical image to be enhanced into an enhanced high-quality image.

[0043] The fourth step may further include image post-processing operations, such as image denormalization, so as to obtain an output that better meets actual application requirements.

[0044] Preferably, if Figure 1 As shown, the CT image super-resolution reconstruction method based on the generative adversarial network according to the preferred embodiment of the present invention may also include:

[0045] The fifth step S5: performing model evaluation on the trained CycleGAN model.

[0046] Model evaluation is a key step to verify the model effect. Objective indicators such as PSNR (peak signal-to-noise ratio) and SSIM (structural similarity index) can be used to evaluate the quality of generated images. In addition, medical experts can also be invited to evaluate the enhanced images to obtain professional opinions and feedback on image quality and usability.

[0047] After the model has been fully trained and evaluated, it can be deployed in actual clinical diagnosis or research scenarios. During the application process, feedback from users and medical experts can still be collected to understand the performance and effect of the model in actual applications. These feedbacks will serve as the basis for further optimization and improvement of the model.

[0048] The network of the present invention can effectively improve the quality of CT low-resolution images.

[0049] Specific examples of the present invention are described below.

[0050] (1) Problem modeling: Let a∈A be the input LRCT image and matrix b∈B be the output HRCT image. The traditional ill-posed linear super-resolution problem can be expressed as:

[0051] a=PHb+n

[0052] Where P represents the downsampling matrix, H represents the system blur matrix, and n represents noise and other interference. The goal of the present invention is to improve the noisy LDCT images obtained under the condition of simple medical equipment to HRCT images through computational methods. The main challenges of restoring HRCT images can be listed as follows:

[0053] 1. Compared with natural images, LDCT images contain different or more complex spatial variations, correlations, and statistical properties, which limits the super-resolution imaging performance of traditional methods.

[0054] 2. During the reconstruction process, the noise in the original projection data is introduced into the image domain, resulting in unique noise and artifact patterns. This makes it difficult for the algorithm to produce high image quality results.

[0055] 3. Since the sampling and degradation operations are coupled and ill-posed, traditional methods cannot perform super-resolution tasks on edge levels, cannot effectively restore some details, and there is a risk of generating blurred appearance and new artifacts.

[0056] To address these limitations, the present invention proposes an advanced neural network that performs SR-CT imaging by combining multiple nonlinear super-resolution functional blocks and uses a residual module to learn high-frequency details, followed by adversarial learning in a cyclic manner to generate perceptually and quantitatively superior SR-CT images.

[0057] (2) CycleGAN model: Current deep learning-based algorithms use a feed-forward convolutional neural network (CNN) to learn a nonlinear mapping parameterized by θ, where θ = {AB, BA}, which can be used to transform the observed model It is expressed as:

[0058]

[0059] Among them G θ (a) represents the generator matrix and θ = {AB, BA}, in order to obtain a reliable A suitable loss function must be specified to encourage G θ (a) Generate an SR image based on the training sample:

[0060]

[0061] in represents the loss function, (a i , b i ), paired LRCT and HRCT images for training. To address the limitations, the cyclic SR-CT model of the present invention is as follows Figure 2 As shown. Generator G AB Responsible for converting LRCT image A into HRCT image B, while G BA Responsible for converting the image of domain B to the image of domain A. Generator G AB The goal is to generate realistic B domain images, and the generator G BA The goal is to generate realistic images in domain A. The two generators can use the same or similar convolutional neural network structure, but have different convolutional neural network parameters. AThe goal of the discriminator D is to evaluate the image quality of the high-resolution image domain produced by the generator to distinguish the generated B-domain images from the real A-domain images. B The goal is to evaluate the image quality of the low-resolution image domain generated by the generator to distinguish the generated A domain images from the real B domain images. AB , the cycle consistency loss ensures that G AB (G BA (A))≈A, that is, the mapping from domain A to domain B and then to domain A is reversible; for the generator G BA , the cycle consistency loss ensures that G BA (G AB (B))≈B, that is, the mapping from domain B to domain A and then to domain B is reversible. Therefore, the following optimization problem can be proposed:

[0062]

[0063] in

[0064]

[0065] in, Represents the expectation operator; and They represent the probability that the input dataset belongs to real data and generated data, respectively. In order to strengthen the mapping between the source domain and the target domain and standardize the training process, the network proposed in this paper combines four types of loss functions: adversarial loss; cycle consistency loss; identity loss; and joint sparse transformation loss.

[0066] (3) Adversarial loss: For edge matching, an adversarial loss is used to force the generated images to follow the empirical distribution in the source and target domains. To improve the training quality, the Wasserstein distance with gradient penalty can be used instead of the negative log-likelihood. Therefore, we obtain the adversarial objective for G:

[0067]

[0068] in Indicates the parameters Find the gradient, represents uniform sampling of pairs of G(a) and b along a straight line, ||·|| 2 Indicates L 2 norm. The first two terms involve the Wasserstein estimate, the third term penalizes the deviation of the gradient norm of its input from 1, and λ is the regularization parameter. Similar adversarial loss It is defined for edge matching in the reverse direction.

[0069] (4) Cycle consistency loss: Although adversarial training can be used well for edge matching, using adversarial loss alone cannot ensure that the learned function can successfully transform the source input into the target output. AB (G BA The cycle consistency between (a)) and a, the cycle consistency loss can be expressed as:

[0070]

[0071] Among them, ||·|| 1 Indicates L 1 norm. Since the cycle consistency loss encourages G BA (G AB (a))≈a and G AB (G BA (b))≈b, which are called forward cycle consistency and backward cycle consistency, respectively. The domain adaptation mapping refers to the cycle-reconstruction mapping. In fact, it imposes the constraint of sharing the latent space to encourage the preservation of the source content during the cycle-reconstruction mapping. In other words, cycle consistency forces the latent code to deviate from the prior distribution in the cycle-reconstruction mapping. In addition, cycle consistency can help prevent the degradation phenomenon in adversarial learning.

[0072] (5) Identity loss: Since HR images should be refined versions of their LR counterparts, it is necessary to use identity loss to regularize the training process. 2 Compared with the loss, L 1 The loss does not overly penalize large differences, nor tolerate small errors between the estimated and target images. Therefore, in this case, L 1 Losses are used first to mitigate L 2 Loss limitation. In addition, L 1 Loss also enjoys L 2 The same fast convergence speed of loss. The formula of identity loss is as follows:

[0073]

[0074] In the bidirectional mapping, G AB (b)(or G BA The size of (a)) is the same as the size of b (or a).

[0075] (6) Joint Sparse Transformation Loss: Total Variation (TV) has shown state-of-the-art performance in promoting image sparsity and reducing noise in piecewise constant images. In order to express the sparsity of the image, the present invention formulates a nonlinear total variation based loss according to the joint constraint, as follows:

[0076]

[0077] where ε is a scaling factor. The above constrained minimization combines two parts: the first part is used to sparse the reconstructed image and reduce obvious artifacts, and the second part is to minimize the difference image bG AB (x) helps preserve anatomical features. Essentially, these two parts need to be jointly minimized under bidirectional constraints. In the present invention, the control parameter ε is set to 0.5. In the case of ε = 1, L JST (G AB ) is considered a traditional TV loss.

[0078] (7) Overall objective function: During the training process, the network proposed in the present invention is fine-tuned in an end-to-end manner to minimize the following overall objective function:

[0079]

[0080] where λ 1 ,λ 2 ,λ 3 is a parameter used to balance different penalties.

[0081] (8) Supervised learning using CGAN: When we have access to paired datasets, we can train our model to solve the SR-CT problem in a supervised manner. Given training paired data from the true joint distribution, i.e., (a, b) ~ P data (A, B), the supervision loss can be defined as follows:

[0082]

[0083]

[0084]

[0085] Figure 3The left side is the original CT image, and the right side is the image after super-resolution reconstruction using this model. In the super-resolution reconstructed computed tomography image on the right, the detailed anatomical structure is significantly enhanced. The density of the lung parenchyma and its fine structures, such as alveoli and tiny bronchi, are clearly visible, providing the necessary image resolution for the detection of early lesions. The branching structures of the trachea and bronchi are accurately displayed, allowing detailed observation of their secondary branches. In addition, the vascular structures of the pulmonary arteries and veins are also clearly displayed in the image, which helps to evaluate vascular lesions. The outline of the heart and the identification of large blood vessels, such as the aorta and superior vena cava, are also more accurate due to super-resolution technology, which is of great significance for the diagnosis of vascular diseases. The structure of the chest wall, including the details of the ribs and thoracic vertebrae, is carefully depicted, which is crucial for the evaluation of bone lesions. If there is abnormal effusion or gas in the chest cavity, this technology can more clearly show these pathological features. At the same time, the layers of soft tissue and subcutaneous tissue are enhanced, providing more information for the diagnosis of soft tissue lesions.

[0086] Figure 4 The image shows a large round mass in the medium (central part of the chest). The left side is the original CT image, and the right side is the SR-CT image of this model. This set of CT images shows a cross-section of the central area of ​​the chest. It can be seen from the figure that the SR-CT image can show that the mass has smooth boundaries, which may indicate that it is benign in nature, such as thymoma, bronchial cyst or germ cell tumor. In addition, the noise of LRCT is removed, which can make it easier for doctors to judge.

[0087] Figure 5 In the figure, SR-CT images show clearer capillary details than LRCT images. This means that the microvascular structure of the lungs is better visualized in SR-CT images, which is very important for evaluating the integrity of blood vessels and detecting early vascular lesions. In the SR-CT image on the right, the branching structure of the bronchi and blood vessels is more detailed, which can help doctors identify tiny abnormalities. SR-CT images provide higher contrast and brightness, making diagnostic imaging clearer and reducing the risk of misdiagnosis or missed diagnosis due to poor image quality. In summary, super-resolution reconstruction technology has demonstrated its important value in improving CT image quality, reducing the required radiation dose, and facilitating clinical decision-making. This progress has far-reaching implications for the early diagnosis and treatment planning of oncology, infectious diseases, and lung diseases.

[0088] The VGG19 model can be used. VGG19 is a variant of VGGNet. Its name comes from its 19-level structure, which is arranged in order, including 16 convolutional layers and 3 fully connected layers. Compared with VGG16, it adds three convolutional layers, making it more expressive. Each convolutional layer uses a 3×3 convolution kernel for convolution, and each fully connected layer contains 4096 neurons. The input image size of the network is 224×224×3. After five pooling layers and 16 convolutional layers, a 7×7×512 feature image is output. After three fully connected layers, the posterior probabilities of 1000 categories are finally output. Then, after a soft-max function, the largest probability result is selected as its posterior probability. The overall model gradually reduces the information in the length and width directions of an image, but gradually increases the information in the channel, converting the information in the pixel space into semantic information. The main feature of VGG19 is the use of very small convolution kernels and very deep network structures. This allows it to process larger images with a smaller memory footprint and excel at identifying objects in images.

[0089] In order to solve the problem that the generated image pays too much attention to edge details and causes the loss of key central content, this paper uses edge smoothing to retain more complete central content. This method is similar to the effect of human eye observation, which clarifies the key central area and blurs the edges. The following function Gaussian function is used to filter the image

[0090]

[0091] where σ 2 Represents the variance of x and y. The value of the Gaussian function is high in the center area and very small in other edge areas. It can be well applied to image blur processing to smooth the image and reduce detail information. By convolving the image with the Gaussian kernel, the high-frequency part is smoothed, thereby achieving a blurring effect.

[0092] The Gaussian convolution process is as follows Figure 6 and Figure 7 As shown in the figure, the steps of using a 3×3 Gaussian convolution kernel to perform Gaussian filtering on an image are as follows: 1. Normalize the Gaussian convolution kernel to effectively improve the stability of the convolution network and reduce the occurrence of problems such as gradient vanishing or gradient exploding. 2. Traverse each pixel in the image, and for each pixel, consider a 3×3 neighborhood centered on the pixel. Multiply the pixel value in the neighborhood with the weight of the corresponding position in the Gaussian convolution kernel, and accumulate the product results to obtain the weighted sum. Finally, we can get Figure 7As shown in the figure, it is obvious that the distribution of the image after convolution with the Gaussian convolution kernel is close to the two-dimensional Gaussian distribution. Therefore, the Gaussian filtering method can be used to obtain an image with prominent central content and blurred edge content, thereby reducing the interference of edge noise on the generation of cartoon images.

[0093] In summary, the present invention provides a CT image super-resolution reconstruction method based on an advanced neural network using a generative adversarial network. By combining multiple nonlinear super-resolution reconstruction functional blocks, super-resolution reconstruction imaging of CT images is performed, accompanied by a residual module to learn high-frequency details, and then the trained HLCGAN and LHCGAN are used to generate SRCT images that are better in both perception and quantity.

[0094] In addition, it should be noted that, unless otherwise specified, the terms "first", "second", "third", etc. in the specification are only used to distinguish the various components, elements, steps, etc. in the specification, and are not used to indicate the logical relationship or sequential relationship between the various components, elements, steps, etc.

[0095] It is to be understood that, although the present invention has been disclosed as a preferred embodiment, the above embodiment is not intended to limit the present invention. For any person skilled in the art, without departing from the scope of the technical solution of the present invention, the technical content disclosed above can be used to make many possible changes and modifications to the technical solution of the present invention, or modified into equivalent embodiments of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A CT image super-resolution reconstruction method based on generative adversarial network, characterized in that include: The first step: collecting high-resolution image data and corresponding low-resolution image data; Step 2: Establish a CycleGAN model including a first generator, a second generator, a first discriminator, and a second discriminator; The third step: using the collected high-resolution image data and the corresponding low-resolution image data to train the CycleGAN model to optimize the parameters of the CycleGAN model; Step 4: Perform image conversion on the medical image to be enhanced by inputting the trained CycleGAN model.

2. The CT image super-resolution reconstruction method based on generative adversarial network according to claim 1 is characterized in that: The third step trains the CycleGAN model using a loss function based on forward and backward cycle consistency to ensure the consistency of image content during the conversion process.

3. The CT image super-resolution reconstruction method based on generative adversarial network according to claim 2 is characterized in that: The loss function combines adversarial loss, cycle consistency loss, identity loss and joint sparse transformation loss.

4. The CT image super-resolution reconstruction method based on a generative adversarial network according to claim 1 or 2, characterized in that: The first generator maps the input image from the high-resolution image domain to the low-resolution image domain, performs image conversion from high resolution to low resolution, and the first discriminator uses the first standard to evaluate the image quality of the low-resolution image domain generated by the first generator; and the second generator maps the input image from the low-resolution image domain to the high-resolution image domain, performs image conversion from low resolution to high resolution, and the second discriminator uses the second standard to evaluate the image quality of the high-resolution image domain generated by the second generator.

5. The method for super-resolution reconstruction of CT images based on generative adversarial networks according to claim 4, characterized in that: The operation of the first generator mapping the input image from the high-resolution image domain to the low-resolution image domain is reversible, and the operation of the second generator mapping the input image from the low-resolution image domain to the high-resolution image domain is reversible.

6. The CT image super-resolution reconstruction method based on generative adversarial network according to claim 1 or 2, characterized in that: In the fourth step, the first generator converts the medical image to be enhanced into an enhanced high-quality image.

7. The CT image super-resolution reconstruction method based on generative adversarial network according to claim 1 or 2, characterized in that: In the CycleGAN model, the first generator and the first discriminator form a high-resolution-low-resolution generative adversarial network, the second discriminator and the second generator form a low-resolution-high-resolution generative adversarial network, and the third step uses unpaired high-resolution data and low-resolution image data to train the high-resolution-low-resolution generative adversarial network to learn to downgrade and downsample high-resolution images; subsequently, using the paired input and output of the trained high-resolution-low-resolution generative adversarial network, the low-resolution-high-resolution generative adversarial network is trained to perform image super-resolution processing using paired low-resolution and high-resolution images.

8. The CT image super-resolution reconstruction method based on generative adversarial network according to claim 1 or 2, characterized in that: The CT image super-resolution reconstruction method based on the generative adversarial network also includes a fifth step: performing model evaluation on the trained CycleGAN model.

9. The CT image super-resolution reconstruction method based on generative adversarial network according to claim 1 or 2, characterized in that: The first generator and the second generator use the same convolutional neural network structure, but have different convolutional neural network parameters.

10. The CT image super-resolution reconstruction method based on generative adversarial network according to claim 1 or 2, characterized in that: The acquired images include MR images and CT images, the first step further includes an image preprocessing operation, and the fourth step further includes an image postprocessing operation.

Images