A Low-Dose CT Image Restoration Method and System Based on Unsupervised Learning

Abstract

The present invention belongs to the technical field of medical image processing, and relates to a low-dose CT image restoration method and system based on unsupervised learning; it includes obtaining a low-dose CT image and performing feature decomposition using a total variation model to obtain a base layer image and a detail layer image; downsampling the base layer image to obtain a content feature image and a latent texture layer image, respectively performing image enhancement and noise reduction processing on the content feature layer image and the detail layer image to obtain a base layer enhanced image and a detail layer noise-reduced image; segmentally fusing the latent texture layer image, the base layer enhanced image, and the detail layer noise-reduced image to obtain a low-dose CT restored image; the present invention decouples the low-dose CT image, enhances the contrast in the low-frequency region, and denoises the high-frequency region, improving the overall image quality and reducing the doctor's misdiagnosis rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and more specifically, relates to a method and system for low-dose CT image restoration based on unsupervised learning. Background Art

[0002] Computed Tomography (CT) technology has been widely used in clinical screening, diagnostic imaging-guided radiotherapy, etc. due to its advantages such as simple operation and low cost. However, excessive CT dose can induce leukemia, cancer, and other genetic diseases. Therefore, in practical applications, the radiation risk is reduced by lowering the CT dose, and among them, the method of reducing the tube current of the X-ray tube is commonly used clinically to reduce the radiation dose. However, the reduction of the radiation dose will cause serious artifacts and noise pollution in the image, and then problems such as reduced contrast at the edges and corners will occur, resulting in a decline in the quality of CT images and seriously affecting the accuracy of downstream medical image processing tasks and clinical diagnoses. Therefore, researching low-dose CT image restoration algorithms and systems has important clinical significance.

[0003] The relevant research on low-dose CT image restoration by experts in related fields at home and abroad mainly falls into three categories: projection domain methods, iterative methods, and image post-processing methods. Projection domain methods make full use of prior information such as statistical noise distribution and can easily integrate filtering into existing CT systems, but the restored images often lose edges and useful texture details; iterative methods can reconstruct images with a small amount of projection data, but they have a large amount of calculation and a long operation time; image post-processing methods can effectively suppress noise and strip artifacts. Traditional image post-processing methods can effectively suppress noise in low-dose CT images and overcome the disadvantage of unstable performance in previous methods. However, such methods are usually only effective for specific types of noise, and their effects are not ideal when the prior knowledge is inaccurate or the noise is uneven in actual scans. With the development of deep learning, some supervised learning methods can automatically extract useful features in image data, making it easier to adapt to new data to evolve the training model and significantly improving the image quality. However, supervised learning requires training with low-dose CT images and normal-dose images with pixel-level matching. In practical applications, it is impossible for patients to collect low-dose and normal-dose CT images simultaneously, resulting in great difficulty in obtaining paired labeled datasets.

[0004] To solve the training data problem, some unsupervised learning has begun to be applied to low-dose CT image restoration, and these methods are comparable to supervised learning methods. However, they require the assumption that the noise is independent of the signal or only targets specific noise, and when these premises are not met, the effects are not ideal. The present invention aims to complete the restoration task of removing unknown noise and enhancing the image for low-dose CT images without paired datasets, improving the accuracy and efficiency of clinical diagnoses. Summary of the Invention

[0005] The object of the present invention is to overcome the deficiencies of existing low-dose CT image restoration, and provide a low-dose CT image restoration method and system based on unsupervised learning, so as to complete the tasks of removing noise and enhancing contrast of low-dose CT images without paired data labels of low-dose CT images and normal-dose CT images, and realize the restoration of low-dose CT images.

[0006] In the first aspect, a low-dose CT image restoration method based on unsupervised learning is proposed, including:

[0007] S1. Obtain a low-dose CT image and normalize it, and decompose the normalized image using a total variation model to obtain a base layer image and a detail layer image;

[0008] S2. Downsample the base layer image to obtain a content feature layer image and a latent texture layer image, and perform image enhancement on the content feature layer image to obtain a base layer enhanced image;

[0009] S3. Based on a pre-trained denoising processor, perform noise reduction processing on the detail layer image to obtain a detail layer denoised image;

[0010] S4. Segmentally fuse the latent texture layer image, the base layer enhanced image, and the detail layer denoised image to obtain a low-dose CT restored image.

[0011] Furthermore, to avoid the amplification of the noise level by image enhancement, the present invention uses the principle of the total variation model to decompose the image after normalizing the low-dose CT image to obtain a base layer image and a detail layer image, realizing the decoupling of image enhancement and noise suppression. The base layer image retains the basic information of the low-dose CT image, and the detail layer image retains the details and noise information of the low-dose CT image; step S1 specifically includes:

[0012] S11. Normalize the low-dose CT image to obtain a normalized image, and decompose the normalized image using a total variation model to obtain a base layer image;

[0013] S12. Perform a difference operation on the normalized image and the base layer image to obtain a detail layer image, expressed as:

[0014] I det (x,y) = I(x,y) - I base (x,y);

[0015] where, I(x,y) represents the normalized image, I base (x,y) represents the base layer image, I det (x,y) represents the detail layer image.

[0016] Furthermore, based on the principle of total variation feature decomposition, the minimum energy functional expression of the present invention is proposed, which is expressed as:

[0017]

[0018] where I base (x, y) represents the base layer image, I(x, y) represents the normalized image, λ is the regularization parameter, and its calculation formula is:

[0019]

[0020]

[0021] M and N respectively represent the width and height of the normalized image, * represents the convolution operation, and T s is the convolution matrix.

[0022] Furthermore, the base layer image contains a large amount of background information and potential texture features. The background information directly affects the image contrast, and the texture features can help the detail layer recovery. Therefore, the present invention uses the Sobel convolution operator A to perform convolution operation on the content feature layer image to obtain the horizontal edge enhanced image of the content feature layer image, uses the convolution operator B to perform convolution operation on the content feature layer image to obtain the vertical edge enhanced image of the content feature layer image, and fuses the horizontal edge enhanced image and the vertical edge enhanced image to obtain the base layer enhanced image I base1 , and its specific process is as follows:

[0023] S21. Continuously use three convolutional layers with a stride of 2 to extract the content feature layer image from the base layer image;

[0024] S22. Use a convolutional layer with a stride of 2 to obtain the potential texture layer image from the content feature layer image;

[0025] S23. Perform image enhancement on the content feature layer image, including:

[0026] Perform convolution operation on the content feature layer image using the Sobel convolution operator A to obtain the horizontal edge enhanced image. The horizontal edge enhanced image and the Sobel operator A are respectively expressed as:

[0027] G x = I base1 (x, y) * A;

[0028]

[0029] Perform convolution operation on the content feature layer image using the Sobel convolution operator B to obtain the vertical edge enhanced image. The vertical edge enhanced image and the Sobel operator B are respectively expressed as:

[0030] G y = I base1 (x,y) * B;

[0031]

[0032] Fuse the horizontally edge-enhanced image and the vertically edge-enhanced image to obtain the base layer enhanced image, denoted as:

[0033]

[0034] where I base1 (x,y) represents the content feature layer image, and I' base1 (x,y) is the base layer enhanced image.

[0035] Furthermore, the training process of the pre-trained denoising processor is as follows:

[0036] Obtain the low-dose CT image and normalize it, and decompose it using the total variation model to obtain the detail layer image;

[0037] Downsample the detail layer image to obtain a low-resolution image, and construct a group of sub-images similar to it based on the low-resolution image;

[0038] Train the denoising processor using the low-dose CT image and the group of sub-images;

[0039] where the group of sub-images includes multiple similar sub-images, which are respectively used to replace the clear low-dose CT image and the noisy low-dose CT image corresponding to the low-resolution image.

[0040] Furthermore, during the denoising process of the detail layer image, by minimizing the following loss function to achieve the denoising effect, and no prior knowledge of the noise is required in this process. The loss function used is:

[0041]

[0042] where N s represents the number of images, and I det1 = s i + n i represents the low-resolution image, represents the sub-image similar to the low-resolution image.

[0043] Furthermore, add the latent texture layer image and the detail layer denoised image, feed the addition result to the ResNet network for upsampling to obtain the first feature layer image; fuse the first feature layer image and the base layer enhanced image, and feed the fusion result to the ResNet network for upsampling to obtain the restored CT image.

[0044] In a second aspect, based on the method of the first aspect, a low-dose CT image restoration system based on unsupervised learning is provided, including:

[0045] An acquisition module, configured to acquire low-dose CT images;

[0046] A decoupling module, configured to perform feature decoupling on the low-dose CT images by using a total variation model to obtain a base layer image and a detail layer image;

[0047] A downsampling module, configured to downsample the base layer image to obtain a content feature layer image and a latent texture layer image; and downsample the detail layer image to obtain a low-resolution image;

[0048] An enhancement module, configured to enhance the horizontal edges and vertical edges of the content feature layer image by using two Sobel operators respectively, and fuse the two enhanced images to obtain a base layer enhanced image;

[0049] A denoising module, configured to perform similarity denoising processing on the detail layer image to obtain a detail layer denoised image;

[0050] A fusion module, configured to segmentally fuse the base layer enhanced image, the latent texture layer image, and the detail layer denoised image to obtain a low-dose CT restored image.

[0051] Advantages of the present invention:

[0052] The present invention provides a low-dose CT image restoration method based on unsupervised learning. For unpaired low-dose CT images, a total variation decomposition method is used to separate the base layer and the detail layer of the LDCT images, realizing the decoupling of contrast enhancement and denoising, and restoring the quality of low-dose CT images.

[0053] Existing deep learning LDCT image restoration algorithms can only target specific noises and rely on prior knowledge for denoising, while the specific noises do not fully reflect the actual CT device noise distribution. The present invention uses feature similarity to remove noises, so that it still has good performance when the prior knowledge is inaccurate or the noises are uneven. When deep learning removes noise suppression, problems such as reduced contrast of contours and corners, and even over-smoothing of lesion areas will occur, affecting the diagnosis of related diseases by clinicians. Therefore, the present invention decouples low-dose CT images, enhances the contrast of low-frequency content features, denoises the high-frequency region, and uses a segmented fusion image method to obtain a low-dose CT restored image, improving the overall image quality and reducing the doctor misdiagnosis rate. Description of the Drawings

[0054] Figure 1 It is a flowchart of the low-dose CT image restoration method of the present invention;

[0055] Figure 2The low-dose CT image restoration process according to an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of the denoising principle of the detail layer image of the present invention; Specific implementation manner

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0058] Existing supervised learning training datasets require pixel-level paired low-dose and normal-dose CT images. Obtaining such a dataset requires the patient to undergo two different-dose CT scans simultaneously, which is unreasonable in clinical applications. Even if the patient is scanned with two doses simultaneously, it may be impossible to achieve complete pixel registration between the two due to reasons such as patient breathing and instrument equipment. Therefore, the present invention proposes an unsupervised learning-based low-dose CT image restoration method and system to complete the low-dose CT restoration task without a pixel-level paired dataset.

[0059] An unsupervised learning-based low-dose CT image restoration method, as Figures 1 - 3 shown, includes:

[0060] S1. Obtain a low-dose CT image and normalize it, and decompose the normalized image using a total variation model to obtain a base layer image and a detail layer image;

[0061] S2. Downsample the base layer image to obtain a content feature layer image and a latent texture layer image, and perform image enhancement on the content feature layer image to obtain a base layer enhanced image;

[0062] S3. Based on a pre-trained denoising processor, perform noise reduction processing on the detail layer image to obtain a detail layer denoised image;

[0063] S4. Segmentally fuse the latent texture layer image, the base layer enhanced image, and the detail layer denoised image to obtain a low-dose CT restored image.

[0064] In one embodiment, as Figure 2 shown, experiments are carried out using 1 / 4-dose abdominal CT data of the 2016 NIH AAPM-Mayo Clinic Low-Dose CT Grand Challenge. The image normalization is performed to increase the detail discernibility, and the total variation model is used to process the normalized abdominal CT data to achieve the decoupling of the base layer enhancement and the detail layer denoising in subsequent tasks, specifically including:

[0065] S11. Normalize the low-dose abdominal CT image to obtain the normalized image I(x,y), and decompose the normalized image using the total variation model to obtain the base layer image;

[0066] Preferably, establish the total variation model: Assume that I(x,y) is disturbed by Gaussian noise n(x,y) with a mean of 0 and a variance of σ 2 The image noise model is

[0067] I(x,y) = I 0 (x,y) + n(x,y), (x,y) ∈ Ω;

[0068] Generally, the total variation of the noisy image I(x,y) is significantly greater than the total variation of the original clean image I 0 (x,y). Therefore, the purpose of image denoising can be achieved by minimizing the total variation. Minimizing the total variation of the image can be understood as minimizing the overall gradient magnitude. Because when the overall gradient magnitude of the image is smaller, it means that the gray level change of the image is smaller, the overall performance is smoother, and it also represents that the influence of noise on the image is smaller. The gradient magnitude here refers to the direction in which the gray level of the image pixel changes fastest. The minimum energy functional of the total variation model is expressed as:

[0069]

[0070] where (I,n) ∈ BV×L2 indicates that in this TV model, the original image I belongs to the bounded variation BV space and the noise n belongs to the L 2 space, and λ is the Lagrange multiplier.

[0071] Specifically, in the total variation model, according to the total variation feature decomposition principle, the minimized energy functional expression is proposed, which is expressed as:

[0072]

[0073] where I base represents the low-frequency structure layer, that is, the base layer image, I(x,y) represents the normalized image after normalizing the low-dose abdominal CT image. The first term (I base (x,y) - I(x,y) 2 ) in the above formula is the data fidelity term, which makes I base (x,y) have the main structure extracted from the normalized image I. The second term is the regularization term, which ensures I base(x, y) is as smooth as possible locally. Among them, the regularization parameter λ (i.e., the Lagrange multiplier) is used as the conditional variable to limit the total variation, so as to effectively control the level of structural roughness. Doubling the regularization term results in a stronger constraint on the high-frequency part, which can better retain the image details and remove the redundant noise.

[0074] The gradient descent method is used to solve the minimization of the energy functional expression, obtaining the corresponding equation:

[0075]

[0076] Introducing the time step into this equation and combining it with the gradient descent method, the time evolution formula is obtained:

[0077]

[0078] |Ω| represents the area of the complete image I region.

[0079] From the above formula, the calculation formula for the regularization parameter is proposed as:

[0080]

[0081]

[0082] Among them, M and N respectively represent the width and height of the original image, * represents the convolution operation, T s is the convolution matrix. Using the convolution matrix to convolve the normalized image is to ensure that the diffusion coefficient is anisotropic diffusion, so that while obtaining the base layer image, noise can be better eliminated, and the edges and details of the image can also be protected;

[0083] S12. Perform a difference operation on the normalized image and the base layer image to obtain the detail layer image, expressed as:

[0084] I det (x, y) = I(x, y) - I base (x, y);

[0085] Among them, I(x, y) represents the normalized image, I base (x, y) represents the base layer image, I det (x, y) represents the detail layer image.

[0086] Preferably, the base layer image contains a large amount of background information and potential texture features. The background information directly affects the image contrast, and the texture features can help the detail layer to be restored. Therefore, in the present invention, the Sobel convolution operator A is used to perform a convolution operation on the content feature layer image to obtain a horizontal edge enhanced image of the content feature layer image, and the convolution operator B is used to perform a convolution operation on the content feature layer image to obtain a vertical edge enhanced image of the content feature layer image. The horizontal edge enhanced image and the vertical edge enhanced image are fused to obtain a base layer enhanced image. The specific process is as follows:

[0087] S21. Let the size of I(x, y) be H×W. I(x, y) passes through three convolutional layers with a stride of 2 to extract the content feature layer image from the base layer image Then use a convolutional layer with a stride of 2 to obtain the potential texture layer image

[0088] S22. Perform image enhancement on the content feature layer image I base1 including:

[0089] Perform a convolution operation on the content feature layer image using the Sobel convolution operator A to obtain a horizontal edge enhanced image. The horizontal edge enhanced image and the Sobel operator A are respectively represented as:

[0090] G x =I base1 (x,y)*A;

[0091]

[0092] Perform a convolution operation on the content feature layer image using the Sobel convolution operator B to obtain a vertical edge enhanced image. The vertical edge enhanced image and the Sobel operator B are respectively represented as:

[0093] G y =I base1 (x,y)*B;

[0094]

[0095] Fuse the horizontal edge enhanced image and the vertical edge enhanced image to obtain a base layer enhanced image, which is represented as:

[0096]

[0097] wherein, I′ base1 (x,y) is the base layer enhanced image.

[0098] Preferably, during the training process of the pre-trained denoising processor in step S3, the detail layer image I det is downsampled to obtain a low-resolution image Perform similarity denoising on the low-resolution image I det1 That is, construct a set of sub-images similar to a low-resolution image, and use this set of similar sub-images to replace the clear low-dose CT image and the noisy low-dose CT image corresponding to the detail layer image. The similarity denoising principle adopted in this embodiment is as follows Figure 3 As shown, perform a difference operation on the input image (low-dose CT image, sub-image replacing the noisy low-dose CT image) and the constructed target image (sub-image replacing the clear low-dose CT image), and eliminate the dissimilar parts, thereby suppressing the noise and artifacts of the low-dose CT image. This denoising process does not require the normal-dose CT image corresponding to the low-dose CT image as a training sample.

[0099] Specifically, use the low-dose CT image, the sub-image replacing the noisy low-dose CT image, and the sub-image replacing the clear low-dose CT image as training samples, continuously eliminate the dissimilar parts (noise), train to obtain a denoising processor, and use the denoising processor to obtain the detail layer denoised image.

[0100] In one embodiment, use the low-resolution image I det1 = s i + n i to construct a set of similar sub-images, which are respectively denoted as and where s i represents the clean low-dose CT image, n i and n' i are different noises, δ i is the difference between clean low-dose CT images. Use the sub-image to replace the clear low-dose CT image corresponding to the low-resolution image I det1 = s i + n i and use the sub-image to replace the noisy low-dose CT image corresponding to the detail layer image I det1 = s i + n i

[0101] In the algorithm for image restoration based on supervised learning, the paired low-dose and normal-dose CT images are used for training, and the loss function used is expressed as:

[0102]

[0103] ​However, in practical applications, it is impossible for patients to collect low-dose and normal-dose CT images simultaneously, which makes it very difficult to obtain a paired labeled dataset. When the paired images are unavailable, the loss function of supervised learning will be restricted. In the noise removal process of the present invention, the following loss function is minimized to achieve the denoising effect. In this process, prior knowledge of noise is not required. Let θ s be the vector of network parameters optimized with the constructed similar data pairs:

[0104]

[0105] where N s is the number of detail layer images. When and E[δ i |s i +n i = 0, θ s is infinitely close to θ c , that is, when all pixel noises are considered independent of each other, the similarity-based denoising method is effective. Under correlated noise, the following experiment is conducted in this example to prove that the above conditions are met and correlated noise can be removed.

[0106] This example uses the high-frequency images of low-dose CT, that is, the detail layer images, including 4,800 slices of 512×512. Zero-mean Gaussian noise with a standard deviation of 25 is used. The following formula is used to estimate E[δ i |s i +n i = 0:

[0107]

[0108] and are a set of randomly constructed similar image pairs for the given detail layer images. M is set to 4,800×500 = 2.4×10 6 , that is, each image is reused 500 times. The experimental result is E[δ i |s i +n i ∈ [-0.0014, 0.0013], where s i is normalized to [0, 1], and the estimated value is close to 0, meeting the constraint conditions.

[0109] Denoising network and its parameter settings. In the denoising experiment of the high-frequency region of low-dose CT images, the present invention uses a three-layer U-Net as the denoising network, with a depth of 3, a kernel size of 3, and the last layer is a linear activation function. A batch normalization layer is added before each activation function.

[0110] Specifically, in this example, Adam is used as the optimizer, the learning rate is set to 0.0005, the maximum number of iterations is 300, the batch size is set to 32, and the momentum parameters are 0.1 and 0.99 respectively. The entire training process runs under the pytorh framework, and the server uses Intel Xeon E5-2680 v4 (2.4GHz / 14C), 256GB of memory, and a 14.4TB mechanical hard drive.

[0111] Preferably, the process of segmentally fusing the latent texture layer image, the base layer enhanced image, and the detail layer denoised image to obtain a low-dose CT image includes:

[0112] First, add the latent texture layer image and the detail layer denoised image, feed the addition result into a ResNet network with two layers of "conv + Leaky-ReLu (lrelu)" and perform upsampling to obtain a first feature layer image Fuse the first feature layer image and the base layer enhanced image enhanced by the convolution operator, feed the fusion result into a ResNet network with two layers of "conv + Leaky-ReLu (lrelu)", and perform upsampling to obtain the restored CT image.

[0113] In one embodiment, a low-dose CT image restoration system based on unsupervised learning is provided, including:

[0114] An acquisition module for acquiring a low-dose CT image and normalizing it;

[0115] A decoupling module for decoupling the features of the low-dose CT image using a total variation model to obtain a base layer image and a detail layer image;

[0116] A downsampling module for downsampling the base layer image to obtain a content feature layer image and a latent texture layer image; and downsampling the detail layer image to obtain a low-resolution image;

[0117] An enhancement module for enhancing the horizontal edges and vertical edges of the content feature layer image using two Sobel operators respectively, and fusing the two enhanced images to obtain a base layer enhanced image;

[0118] A denoising module for performing similarity denoising processing on the detail layer image to obtain a detail layer denoised image;

[0119] A fusion module for segmentally fusing the base layer enhanced image, the latent texture layer image, and the detail layer denoised image to obtain a low-dose CT restored image.

[0120] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A low-dose CT image restoration method based on unsupervised learning, characterized in that, it includes: S1. Obtain a low-dose CT image and normalize it, and decompose the normalized image using a total variation model to obtain a base layer image and a detail layer image; Step S1 specifically includes: S11. Normalize the low-dose CT image to obtain a normalized image, and decompose the normalized image using a total variation model to obtain a base layer image; S12. Perform a difference operation on the normalized image and the base layer image to obtain a detail layer image, expressed as: I det (x, y) = I(x, y) - I base (x, y); Among them, I(x, y) represents the normalized image, and I base (x, y) represents the base layer image, and I det (x, y) represents the detail layer image; The minimized energy functional expression proposed according to the total variation principle in the total variation model is: Among them, I base (x, y) represents the base layer image, I(x, y) represents the normalized image, λ is the regularization parameter, and the calculation formula for the regularization parameter is: M and N respectively represent the width and height of the normalized image, * represents the convolution operation, and T s is the convolution matrix; S2. Downsample the base layer image to obtain a content feature layer image and a potential texture layer image, and perform image enhancement on the content feature layer image to obtain an enhanced base layer image; The processing process of the base layer image: S21. Continuously use three convolutional layers with a stride of 2 to extract a content feature layer image from the base layer image; S22. Use a convolutional layer with a stride of 2 to obtain a potential texture layer image from the content feature layer image; S23. Perform image enhancement on the content feature layer image, including: Perform a convolution operation on the content feature layer image using the Sobel convolution operator A to obtain a horizontal edge enhanced image. The horizontal edge enhanced image and the Sobel operator A are respectively expressed as: G x = I base1 (x, y) * A; Perform a convolution operation on the content feature layer image using the Sobel convolution operator B to obtain a vertical edge enhanced image. The vertical edge enhanced image and the Sobel operator B are respectively expressed as: G y = I base1 (x,y)*B; Fuse the horizontal edge enhanced image and the vertical edge enhanced image to obtain an enhanced base layer image, expressed as: Among them, I base1 (x, y) represents the content feature layer image, I b ′ ase1 (x, y) is the enhanced image of the base layer; S3. Based on a pre-trained denoising processor, perform denoising processing on the detail layer image to obtain a denoised detail layer image; During the denoising process of the detail layer image, the loss function used is: Among them, N s is the number of detail layer images, I det1 = s i + n i represents the low-resolution image, and represents the similar sub-images of the low-resolution image; s i represents the clean low-dose CT image, n i and n i ′ are different noises, and δ i is the difference between clean low-dose CT images; S4. Segmentally fuse the potential texture layer image, the enhanced base layer image, and the denoised detail layer image to obtain a low-dose CT restored image.

2. A low-dose CT image restoration method based on unsupervised learning according to claim 1, characterized in that, the training process of the pre-trained denoising processor is: Obtain a low-dose CT image and normalize it, and decompose it using a total variation model to obtain a detail layer image; Downsample the detail layer image to obtain a low-resolution image, and construct a group of sub-images similar to it based on the low-resolution image; Use the low-dose CT image and the group of sub-images to train the denoising processor; wherein, the group of sub-images includes multiple similar sub-images, which are respectively used to replace the clear low-dose CT image and the noisy low-dose CT image corresponding to the low-resolution image.

3. A low-dose CT image restoration method based on unsupervised learning according to claim 1, characterized in that, the image fusion process in step S4 is: Add the potential texture layer image and the denoised detail layer image, and feed the added result to the ResNet network for upsampling to obtain a first feature layer image; Fuse the first feature layer image and the enhanced base layer image, and feed the fused result to the ResNet network for upsampling to obtain the restored CT image.

4. An unsupervised learning-based low-dose CT image restoration system applying the unsupervised learning-based low-dose CT image restoration method according to any one of claims 1-3, characterized in that, it includes: an acquisition module, configured to acquire a low-dose CT image and normalize it; a decoupling module, configured to perform feature decoupling on the low-dose CT image by using a total variation model to obtain a base layer image and a detail layer image; a downsampling module, configured to downsample the base layer image to obtain a content feature layer image and a latent texture layer image; and downsample the detail layer image to obtain a low-resolution image; an enhancement module, configured to enhance the horizontal edges and vertical edges of the content feature layer image by using two Sobel operators respectively, and fuse the two enhanced images to obtain a base layer enhanced image; a denoising module, configured to perform similarity denoising processing on the detail layer image to obtain a detail layer denoised image; a fusion module, configured to segmentally fuse the base layer enhanced image, the latent texture layer image, and the detail layer denoised image to obtain a low-dose CT restored image.

Images