A CT image ring artifact suppression method, device, equipment and storage medium

Abstract

The application provides a CT image ring artifact suppression method, device, equipment and storage medium, an initial CT image is acquired, and a data set is made according to the initial CT image; a RAS_UNet network model is constructed, and a loss function is designed for the RAS_UNet network model; then the data set of the initial CT image is input into the RAS_UNet network model, the RAS_UNet network model can suppress the artifact in the initial CT, and a first CT image and a second CT image after artifact suppression processing are output, the removal of the ring artifact is realized, finally, effective information in the first CT image and the second CT image is fused, and a final CT image after fusion is output, details and structures of the final CT image after fusion can be effectively recovered, and the signal-to-noise ratio can be obviously improved.

Description

Technical Field

[0001] This invention relates to the field of computer imaging technology, and in particular to a method, apparatus, device, and storage medium for suppressing ring artifacts in CT images. Background Technology

[0002] As a commonly used non-destructive testing technique in the medical and industrial fields, X-ray computed tomography (CT) imaging can reconstruct the internal information of objects at high resolution, greatly facilitating the interpretation of the internal structure, material, density, and defects of target objects. However, due to manufacturing issues and other problems with the hardware equipment in actual imaging systems, concentric rings with varying brightness centered at a fixed point often appear in the reconstructed tomographic images, known as ring artifacts. The causes of ring artifacts are not singular; for example, detector pixel damage, inconsistent detector channel response, data acquisition system malfunctions, and insufficient X-ray photons can all produce ring artifacts. Therefore, effectively suppressing ring artifacts and improving the image quality of X-ray tomographic images has significant theoretical and practical implications.

[0003] In related technologies, research on ring artifact suppression can be divided into two categories: artifact suppression based on the projection domain and artifact suppression based on the image domain. For example, Raven designed a numerical filter based on Fourier transform in the projection domain, which effectively reduced the impact of ring artifacts on image quality. However, visible artifact traces still exist in the filtered image. Titanenko used compressed sensing theory to correct sine wave data by minimizing the Tikhonov functional. This method effectively suppressed artifacts, and the reconstructed image showed almost no visible ring artifacts, significantly improving image quality. However, this algorithm is highly complex and computationally intensive, and some parameters need to be adjusted according to the image characteristics. Sun et al. performed "interpolation-reconstruction-filtering-reconstruction" on the sine wave, suppressing ring artifacts while preserving image details. The proposed algorithm has a good correction effect on severe ring artifacts caused by large-area detector damage. However, the interpolation and filtering processes in this algorithm change the original information of the image and reduce the signal-to-noise ratio.

[0004] In summary, traditional algorithms suffer from problems such as incomplete artifact removal and reduced image signal-to-noise ratio when it comes to suppressing ring artifacts in CT images. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and storage medium for suppressing ring artifacts in CT images, which solves the defects of incomplete artifact removal and low signal-to-noise ratio in existing technologies for CT images. It achieves the effect of suppressing ring artifacts to the greatest extent while protecting the structural information of CT images, and at the same time significantly improving the signal-to-noise ratio of CT images.

[0006] This invention provides a method for suppressing ring artifacts in CT images, comprising:

[0007] Obtain the dataset of initial CT images;

[0008] Construct the RAS_UNet network model;

[0009] Design a loss function for the RAS_UNet network model;

[0010] The dataset is input into the RAS_UNet network model, which outputs the first CT image and the second CT image after artifact suppression processing.

[0011] The effective information in the first CT image and the second CT image is fused to output the final fused CT image.

[0012] According to the present invention, a method for suppressing ring artifacts in CT images includes a loss function in the image domain and a loss function in the projection domain; the first CT image is a CT image after artifact suppression in the image domain, and the second CT image is a CT image after artifact suppression in the projection domain.

[0013] According to the present invention, a method for suppressing ring artifacts in CT images is provided. The RAS_UNet network model adopts the UNet network as the model framework. The downsampling part of the RAS_UNet network model includes one double convolutional module and four Inception modules for extracting artifact features at different scales. The upsampling part includes four upsampling modules and three residual modules for identifying and processing artifacts, and also includes one convolutional kernel.

[0014] According to a method for suppressing ring artifacts in CT images provided by the present invention, a loss function is designed for the RAS_UNet network model, specifically including:

[0015] The gradient of the initial CT image is calculated simultaneously using forward and backward differencing, and the maximum value of the forward and backward differencing is taken as the gradient of the initial CT image.

[0016]

[0017] Where I represents the initial CT image, (i,j) represents the coordinates of a pixel in the image, G1 represents the image gradient calculated by forward difference, G2 represents the image gradient calculated by backward difference, max represents taking the maximum value, and Gsin is the final image gradient in the sine curve.

[0018] First, calculate the row mean of Gsin to obtain Gmean. Then, use Gmean as the weight to weight the gradient of Gsin in the sinusoidal domain.

[0019]

[0020] Among them, WG sin (i,j) represents the gradient weighting coefficients of the sine wave image;

[0021] Weighting gradients within the sliced ​​graph domain:

[0022]

[0023] Where ε is a positive real number to avoid the denominator being 0, R(p) represents a local region window centered at pixel p in the initial CT image, q is the pixel point in the local region window, and g p,q G represents the weighting coefficients of a Gaussian distribution centered at p. img For the image gradient in the slice image, WG img These are the image gradient weighting coefficients. This represents the partial derivative of I along the column direction in the image;

[0024] Using the weighted gradient as a regularization term and the structural similarity coefficient as a fidelity preservation term, the loss functions in the image domain and the projection domain are calculated as follows:

[0025]

[0026] Among them, Loss sin Loss is the loss function for the projection domain. img λ1 and λ2 are the loss functions for the image domain; SSIM is the fidelity term, representing the structural similarity coefficient; λ1 and λ2 ∈ [0,1].

[0027] According to the present invention, a method for suppressing ring artifacts in CT images fuses effective information from a first CT image and a second CT image to output a fused final CT image, specifically including:

[0028] The first CT image and the second CT image are decomposed using a non-subsampled contour wave transform algorithm to obtain multiple sub-images to be fused from the first CT image and multiple sub-images to be fused from the second CT image.

[0029] Calculate the residuals of the pairwise corresponding sub-images to be fused. If the residuals exceed a set residual threshold, the pixel values ​​of the sub-images to be fused with lower contrast are used as the fusion result, and the fused sub-image is output. If the residuals do not exceed the set residual threshold, the pixel values ​​of the sub-images to be fused with higher contrast are used as the fusion result, and the fused sub-image is output.

[0030] The fused sub-images are combined to output the final CT image.

[0031] According to the present invention, a method for suppressing ring artifacts in CT images is provided, wherein the sub-images to be fused include low-frequency sub-band images and high-frequency sub-band images.

[0032] According to the present invention, a method for suppressing ring artifacts in CT images is provided, wherein the residual is the absolute value of the difference between the gray values ​​of the two sub-images to be fused; and the residual threshold is the product of the sum of the gray values ​​of the two sub-images to be fused and the ratio value.

[0033] The present invention also provides a device for suppressing ring artifacts in CT images, comprising:

[0034] The data collection module is used to acquire the dataset of initial CT images;

[0035] The model building module is used to build the RAS_UNet network model;

[0036] The function design module is used to design the loss function for the RAS_UNet network model;

[0037] The artifact processing module is used to input the dataset into the RAS_UNet network model and output the first CT image and the second CT image after artifact suppression processing.

[0038] The image fusion module is used to fuse the effective information in the first CT image and the second CT image, and output the fused final CT image.

[0039] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described methods for suppressing ring artifacts in CT images.

[0040] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CT image ring artifact suppression method as described above.

[0041] This invention provides a method, apparatus, device, and storage medium for suppressing ring artifacts in CT images. The method involves acquiring an initial CT image and creating a dataset based on it; constructing a RAS_UNet network model and designing a loss function for it; inputting the dataset of the initial CT image into the RAS_UNet network model, which suppresses artifacts in the initial CT image and outputs a first and second CT image after artifact suppression processing, thus removing ring artifacts; finally, fusing the effective information from the first and second CT images to output a fused final CT image. The details and structure of the fused final CT image are effectively restored, and the signal-to-noise ratio is significantly improved. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the method for suppressing ring artifacts in CT images provided by the present invention;

[0044] Figure 2 This is a flowchart of the overall process for suppressing ring artifacts in CT images provided by the present invention;

[0045] Figure 3 This is a schematic diagram of the overall structure of the RAS_UNet network model provided by this invention;

[0046] Figure 4 This is a schematic diagram of the Inception module in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the Residual module in an embodiment of the present invention;

[0048] Figure 6 It is a sine wave with bar artifacts in an embodiment of the present invention;

[0049] Figure 7 This is the gradient image in the embodiments of the present invention;

[0050] Figure 8 This is the weighted gradient image in the embodiments of the present invention;

[0051] Figure 9 This is a schematic diagram of the morphology of the annular artifact in an embodiment of the present invention;

[0052] Figure 10 This is a schematic diagram of the annular artifact and edge curve in an embodiment of the present invention;

[0053] Figure 11 This is a schematic diagram of the weighted effect of the annular artifact in an embodiment of the present invention;

[0054] Figure 12 This is a schematic diagram of NSCT multiscale decomposition in an embodiment of the present invention;

[0055] Figure 13 This is a schematic diagram of the non-downsampling pyramid principle in an embodiment of the present invention;

[0056] Figure 14 This is a schematic diagram of NSDFB first-level decomposition and reconstruction in an embodiment of the present invention;

[0057] Figure 15 This is a schematic diagram of NSDFB four-channel directional decomposition and frequency band division in an embodiment of the present invention;

[0058] Figure 16 This is a flowchart of the image fusion algorithm in an embodiment of the present invention;

[0059] Figure 17 This is a schematic diagram of the image fusion effect in an embodiment of the present invention;

[0060] Figure 18 This is a three-dimensional rendering of the specimen in an embodiment of the present invention;

[0061] Figure 19 This is a schematic diagram illustrating the suppression effect of the RAS_UNet model on strip artifacts in a sinusoidal image in an embodiment of the present invention.

[0062] Figure 20 This is a schematic diagram illustrating the effect of the RAS_UNet model in this embodiment of the invention on suppressing ring artifacts in tomographic images;

[0063] Figure 21 This is a schematic diagram of the NSCT image fusion effect in an embodiment of the present invention;

[0064] Figure 22 This is a comparison chart of the ring artifact suppression effects of different algorithms in the embodiments of the present invention;

[0065] Figure 23 This is a schematic diagram of the structure of the CT image ring artifact suppression device provided by the present invention;

[0066] Figure 24 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0067] Figure label:

[0068] 231: Data collection module; 232: Model building module; 233: Function design module; 234: Artifact processing module; 235: Image fusion module. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0070] Figure 2 This is a flowchart illustrating the overall process of the CT image annular artifact suppression method provided in this embodiment of the invention. Randon is the Randon transform, and IRadon is the inverse Randon transform. Limited data has always been a significant factor limiting the application of deep learning in the CT field. To train a high-performance annular artifact suppression network model under limited data conditions, this invention selects the UNet model as the network foundation and combines it with the Inception network structure to extract deeper image features. Simultaneously, to prevent the gradient backpropagation effect from weakening as the network depth increases during training, leading to the model getting stuck in a local optimum, this invention introduces the concept of residual connection networks. For the geometric features of annular artifact images, this invention designs specialized loss functions based on the projection domain and image domain respectively, allowing the network to assign more weights to artifact features. Finally, based on the Nonsubsampled Contourlet Transform (NSCT) method, a specialized image fusion rule is designed for annular artifact images to fuse the output images of the projection domain and image domain networks.

[0071] Specifically, embodiments of the present invention provide a method for suppressing ring artifacts in CT images, such as... Figure 1 As shown, the method includes:

[0072] S1. Obtain the dataset of initial CT images.

[0073] In this step, as an optional embodiment of the present invention, a CT scan experiment can be conducted using a self-developed 3D-μCT system to obtain a dataset of real CT images. The hardware of this system mainly includes an X-ray source, a flat panel detector, and a stage. The X-ray source used is the XWT model manufactured by X-RAY WorX GmbH, and its main parameters are shown in Table 1; the detector used is the Perkin Elmer XRD0822, and its main technical parameters are shown in Table 2.

[0074] Maximum tube voltage Maximum tube current Maximum power Minimum focal size 225kV 3.0mA 350W 3μm

[0075] Table 1

[0076] Switching screen CsI ​​(cesium iodide) Pixel area: Total area <![CDATA[204.8×204.8mm 2 ]]> Pixel matrix: Total number 1024×1024 Pixel size 0.2mm Communication methods RJ45 Gigabit Ethernet A / D conversion bit depth 16bit

[0077] Table 2

[0078] S2. Input the dataset of initial CT images into the RAS_UNet network model, and output the first CT image and the second CT image after artifact suppression processing.

[0079] This invention constructs the RAS_UNet network model, as follows: Figure 3 As shown, the RAS_UNet network model includes the following structure:

[0080] The RAS_UNet network model uses the UNet network as its framework. The downsampling part of the RAS_UNet network model uses Inception modules instead of convolutions, while the upsampling part introduces a residual mechanism, ultimately forming RAS_UNet for suppressing ring artifacts in X-ray tomographic images. The downsampling part of the RAS_UNet network model includes one double convolution module and four Inception modules for extracting artifact features at different scales. The upsampling part includes four upsampling modules and three residual modules for identifying and processing artifacts, as well as one convolution kernel. The RAS_UNet network model uses a "skip connection" approach to combine the two parts of image information, preserving the original image information to the greatest extent possible. These two parts of image information refer to the information from the CT images during the downsampling process and the information from the CT images during the upsampling process when the CT image dataset is input into the RAS_UNet network model for training. The two parts of image information here are information from the same CT image at different processing stages. When downsampling the CT image, some low-frequency information is lost, so "skip connections" are used during upsampling to "add back" this lost low-frequency information.

[0081] UNet is a symmetric, end-to-end, unsupervised convolutional neural network model. Originally proposed for medical image segmentation, it has achieved significant results in medical image segmentation since its introduction in 2015, gradually becoming a benchmark model for medical image segmentation. Due to the limited image data, this invention is based on the UNet network structure, utilizing its symmetric network layer "skip connection" mechanism to accelerate network training curve convergence while maintaining image signal-to-noise ratio. Ring artifacts in X-ray CT images have a simple structure, generally appearing as concentric rings in the image domain and as single-pixel-width straight lines in the projection domain. UNet is a lightweight model with a shallow network depth. Experiments show that directly using the UNet structure cannot accurately identify artifact features; while suppressing ring artifacts, the network introduces additional random artifacts, leading to image distortion. To extract deeper artifact features, this invention introduces... Figure 4 The Inception module shown replaces the convolutional module in downsampling. By combining multiple convolutional kernels, the Inception module has a sparse structure, which can generate denser data, improving network performance while ensuring efficient resource utilization.

[0082] The introduction of the Inception module deepens the network, giving it a stronger ability to extract deep feature structures. However, it also increases the likelihood of gradient vanishing, making the network prone to getting trapped in local optima during training. To mitigate the vanishing problem, this invention introduces techniques such as gradient vanishing into the deeper layers of the network. Figure 5 The Residual module shown increases the network width while optimizing gradient backpropagation.

[0083] When training the RAS_UNet network model, the loss function reflects the accuracy of the model's prediction results, quantitatively describing the difference between the predicted and theoretical values. A suitable loss function can accelerate the training of the network model and improve its performance. Based on the characteristics of ring artifacts in different domains, this invention designs targeted loss functions for the RAS_UNet network model, namely, a loss function in the image domain and a loss function in the projection domain.

[0084] Specifically, such as Figure 6 As shown, in the projection domain, the ring artifact appears as strip-shaped artifacts with varying brightness, and the width of the lines is one pixel. Compared to an ideal sine wave, the image structure of a sine wave with artifacts differs only at the straight lines; the rest of the image is identical. Therefore, in training the network model, this invention aims to focus the network's attention as much as possible on the straight lines.

[0085] To highlight the strip artifacts, this invention transfers the image to the gradient domain. Since the width of the strip artifacts is one pixel, to ensure that the structure of the strip artifacts remains unchanged in the gradient domain (the width is still one pixel), as shown in formula (1), this invention simultaneously uses forward and backward differencing to calculate the gradient of the initial CT image, as follows: Figure 7 As shown, the maximum value of the forward difference and the backward difference is finally taken as the gradient of the image.

[0086]

[0087] In formula (1), I represents the initial CT image; G1 represents the image gradient calculated by forward difference; G2 represents the image gradient calculated by backward difference; max represents taking the maximum value; G sin This represents the final image gradient in the sine curve.

[0088] Formula (1) improves the contrast of strip artifacts while also highlighting the structure and details of the image. If the loss error is calculated directly in the gradient domain, the network will blur the structure and details of the image while removing strip artifacts, and introduce new artifacts. To protect the original information of the image, this invention uses Formula (2) to weight the image gradient based on Formula (1).

[0089]

[0090] Where WGsin(i,j) represents the gradient weighting coefficient of the sine curve.

[0091] Depend on Figure 7 It can be seen that, relative to the structure and details of the image, the gradient of the strip artifacts exhibits a regularity, that is, the straight-line regions where the strip artifacts are located all have relatively large gradient values. Based on this characteristic, formula (2) first applies G... sin Calculate the mean by row to obtain G. mean , and then with G mean For weights on G sin Gradients are weighted in the sinusoidal domain to highlight bar artifacts while suppressing image structure and detail. The weighted result is as follows: Figure 8 As shown.

[0092] Compared to the projection domain, the characteristics of ring artifacts in the image domain are more complex, such as... Figure 9 As shown, it is represented by a ring with a width of several pixels.

[0093] By analyzing the grayscale curve of the image, it was found that, Figure 10 As shown, the grayscale curve of the annular artifact exhibits a roof-like shape, while the grayscale curves of details and edge contours in the image exhibit a step-like shape. This difference also exists in the gradient domain, and the pixel values ​​of the annular artifact alternate between positive and negative. Figure 10In the image, (a) is a roof edge; (b) is a step edge.

[0094] To highlight the characteristics of the ring artifact, the gradient is weighted in the sliced ​​map domain using the concept of relative variation, according to formula (3):

[0095]

[0096] In formula (3), ε is taken as a very small positive real number to avoid the denominator being 0; R(p) represents a local region window centered on pixel p in the image; q is the pixel in this local window; g p,q G represents the weighting coefficients of a Gaussian distribution centered at p. img For the image gradient in the slice image, WG img These are the image gradient weighting coefficients. This represents the partial derivative of image I along the q direction (q represents the column direction in the image).

[0097] When calculating the weighted gradient, it is worth noting that... The calculation of the gradient of an image should be performed in order to utilize the characteristic of "alternating positive and negative gray values". The line containing the difference direction should pass through the center of the annular artifact when calculating the gradient.

[0098] In an optional embodiment of the present invention, to verify the effect of the designed weighted gradient, a simple program was written using Matlab R2018b software, and tests were conducted using real X-ray tomographic images. The test results are as follows: Figure 11 As shown in the figure, where a is the image containing the ring artifact, b is the gradient image of a, and c is the weighted gradient image. It can be seen that after gradient weighting, the characteristics of the ring artifact are more obvious.

[0099] To remove ring artifacts while preserving image information as much as possible, the weighted gradient designed in this invention is used as a regularization term, and the structural similarity coefficient is used as a fidelity preservation term. The final loss function is shown in formula (4):

[0100]

[0101] In formula (4), Loss sin and Loss img λ1 and λ2 are the loss functions for the projection domain and the image domain, respectively; SSIM is the fidelity term, representing the structural similarity coefficient; λ1 and λ2 ∈ [0,1] are used to adjust the weights between the regularization term and the fidelity term.

[0102] In step S2, the dataset of the initial CT images is input into the RAS_UNet network model. The RAS_UNet network model can perform artifact suppression processing on the initial CT images in the image domain and the projection domain respectively, and output the first CT image after artifact suppression in the image domain and the second CT image after artifact suppression in the projection domain.

[0103] S3. Fuse the effective information from the first CT image and the second CT image, and output the final fused CT image.

[0104] To maximally suppress ring artifacts and preserve image information, it is necessary to fuse the effective information from artifact suppression results in the image domain and projection domain into a single image. The quality of image details should be the primary goal, and multi-scale image decomposition provides a method for image processing at different scales. This invention decomposes the dual-domain images (i.e., the first CT image and the second CT image) into different image scales, and then fuses the information at different scales according to a designed fusion rule to enrich image details and improve image contrast. This invention employs an image fusion method based on Nonsubsampled Contourlet Transform (NSCT) decomposition.

[0105] Specifically, step S3 includes the following steps:

[0106] S31. The first CT image and the second CT image are decomposed using a non-subsampled contourlet transform algorithm to obtain multiple sub-images to be fused from the first CT image and multiple sub-images to be fused from the second CT image. The sub-images to be fused include low-frequency sub-band images and high-frequency sub-band images.

[0107] Specifically, the NSCT decomposition process is as follows: Figure 12 As shown, the Nonsubsampled Pyramid (NSP) is responsible for multi-resolution decomposition of the image, and the Nonsubsampled Directional Filter Bank (NSDFB) is responsible for multi-directional decomposition. Each NSP decomposition stage generates a high-frequency sub-band and a low-frequency sub-band. Then, the high-frequency sub-band from each NSP stage is decomposed using NSDFB in multiple directions, combining singularities distributed in the same direction into a single coefficient, thus obtaining sub-band images of different scales and directions.

[0108] Figure 13Figure (a) shows a schematic diagram of the first-order decomposition and reconstruction of NSP, where H0(z) and H1(z) are the low-pass and high-pass decomposition filters, respectively, and G0(z) and G1(z) are the low-pass and high-pass reconstruction filters, respectively. The decomposition iteration structure of NSP is as follows: Figure 13 As shown in (b), this filter bank does not downsample the image. The next-level nonsubsampled filter bank (NSFB) is formed by performing 2×2 row and column sampling on the previous-level NSFB.

[0109] Figure 14 This diagram illustrates the first-level decomposition and reconstruction using NSDFB. U0(z) and U1(z) are the decomposition filters, and V0(z) and V1(z) are the synthesis filters. The basic module of NSDFB is a dual-channel non-subsampled filter bank. Through its sector filter bank and quadrant filter bank, the image can be divided into four sub-bands. This process avoids image resampling operations and possesses translation invariance. Figure 15 The general process of NSDFB four-channel directional decomposition and frequency band division is given. k The outputs correspond to respectively Figure 15 The k (k = 0, 1, 2, 3) directions in the image. If a sub-band image at a certain scale is decomposed into k-level directions, 2k directional sub-band images of the same size as the original input image can be obtained. After J-level NSCT decomposition, one low-frequency sub-band image and... A bandpass direction sub-band image, where k j Let be the directional decomposition series at scale j.

[0110] S32. Calculate the residuals of the pairwise corresponding sub-images to be fused. If the residuals exceed a set residual threshold, the pixel values ​​of the sub-images to be fused with lower contrast are used as the fusion result, and the fused sub-image is output. If the residuals do not exceed the set residual threshold, the pixel values ​​of the sub-images to be fused with higher contrast are used as the fusion result, and the fused sub-image is output.

[0111] Compared to the original images, the first and second CT images processed by the RAS_UNet model have their ring artifacts suppressed, but the details and structure of the images are inevitably smoothed, resulting in a decrease in image contrast. To improve image contrast and sharpen the image contour information, this invention combines the information from both CT images and uses local residuals as the fusion rule when fusing the effective information from the first and second CT images.

[0112] The implementation steps of the algorithm in this step are as follows: Figure 16As shown, G1 and G2 represent the local contrast of the image, (i,j) represents the coordinates of the pixel in the image, Th represents the gray threshold (i.e., the residual threshold mentioned above), Res represents the gray difference of the image (i.e., the residual mentioned above), and ratio is a ratio value. Ratio is a value determined manually based on the degree of gray fluctuation in the image region, and it is recommended to take a value range of 0.04-0.125.

[0113] As analyzed above, edges and details in an image have wider contours compared to ring artifacts. Image fusion aims to improve image contrast while suppressing ring artifacts. Therefore, during fusion, we compare the local information of two corresponding sub-images to be fused, i.e., we compare their grayscale values. If the difference in local information (i.e., the image grayscale difference) exceeds a set threshold (i.e., the grayscale threshold), an artifact is considered to exist. When fusing images, the pixel values ​​of the sub-image with lower contrast are directly used as the fusion result, and the fused sub-image is output to suppress artifacts. If the difference in local information (i.e., the image grayscale difference) does not exceed the set threshold (i.e., the grayscale threshold), the pixel values ​​of the sub-image with higher contrast are used as the fusion result, and the fused sub-image is output to improve image contrast.

[0114] S33. Combine the fused sub-images to output the final CT image.

[0115] For example, the first CT image and the second CT image are decomposed to produce their own high-frequency sub-images and low-frequency sub-images. The high-frequency sub-images of the first CT image and the high-frequency sub-images of the second CT image are fused together, and the low-frequency sub-images of the first CT image and the low-frequency sub-images of the second CT image are fused together to form two fused sub-images. Finally, the two fused sub-images are combined to form the final CT image.

[0116] To verify the effectiveness of the designed fusion rule, we conducted image fusion experiments using tomographic images with and without ring artifacts as input. The experimental results are as follows: Figure 17 As shown, (a) is the original image containing ring artifacts, (b) is the standard image without artifacts, and (c) is the image fused from (a) and (b). We objectively evaluate the similarity between the fused image and the standard image using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Mean Squared Error (MSE). The results show that the PSNR, SSIM, and MSE of the fused images are 33.48, 0.963, and 8.9×e⁻⁴, respectively.

[0117] This invention focuses on the suppression of ring artifacts in real tomographic images. To train and test the performance of the aforementioned RAS_UNet network model, we acquired three sets of tomographic images using the 3D-μCT system described above. The visualization structures of the three specimens are as follows: Figure 18 As shown, (a) is a soil sample; (b) is a frozen soil sample; and (c) is a stop valve sample.

[0118] To train a network model for suppressing ring artifacts, this invention, based on CT reconstruction principles and using Matlab 2018 software, incorporates random ring artifacts into CT tomographic images using Radon forward and inverse transforms, generating three training and test sets. The original data is then used as the standard images for network training. The main parameters of the datasets are shown in Table 3.

[0119]

[0120] Table 3

[0121] As shown in Table 3, when preparing the dataset, the ratio of the training set to the test set was set to 4:1. Additionally, when loading the dataset, random geometric transformations were applied to the images to improve network training performance. After multiple experiments, the optimal training parameters for the network were determined, as shown in Table 4.

[0122]

[0123] Table 4

[0124] Using the parameters shown in Table 4, we trained the network model for both sinusoidal and tomographic images. Figure 19 This demonstrates the effectiveness of the RAS_UNet model in suppressing stripe artifacts in sinusoidal images. Figure 20 The effectiveness of the RAS_UNet model in suppressing ring artifacts in tomographic images is demonstrated.

[0125] Depend on Figure 19 and Figure 20 As can be seen, after processing by the RAS_UNet network model, artifacts are effectively removed in both the projection domain and the image domain. Finally, the dual-domain images are fused using the NSCT fusion algorithm, and the fusion result is as follows. Figure 21As shown in Table 5, this invention compares the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Metric (SSIM), and Mean Squared Error (MSE) of the images before and after fusion. Table 5 shows that after fusion, both PSNR and SSIM are improved, while MSE is significantly reduced, indicating that the details and structure of the image are effectively preserved through dual-domain image fusion.

[0126]

[0127] Table 5

[0128] Furthermore, in an optional embodiment of the present invention, to highlight the superiority of the CT image ring artifact suppression method provided by the present invention, we used specimen 1 (rock and soil) as the test object and conducted comparative verification using three methods. Two of these algorithms achieved ring artifact suppression from the projection domain and the image domain, respectively (hereinafter referred to as Algo1 and Algo2 to represent these two algorithms), and the other was a deep learning method using UNet as the network structure. Figure 22 The artifact suppression effects of four methods are shown, where (a) is the original input image; (b) is the processing result of Algo1 algorithm; (c) is the processing result of Algo2 algorithm; (d) is the processing result of UNet network model; (e) is the processing result of the CT image ring artifact suppression method provided by the present invention; and (f) is the standard image without ring artifacts.

[0129] As can be seen from the above processing results, the CT image ring artifact suppression method provided by the present invention can solve the defects of incomplete artifact removal and low signal-to-noise ratio in related technologies, and achieve the effect of suppressing ring artifacts to the greatest extent while protecting the structural information of CT images, and significantly improving the signal-to-noise ratio of CT images.

[0130] The CT image ring artifact suppression device provided by the present invention is described below. The CT image ring artifact suppression device described below can be referred to in correspondence with the CT image ring artifact suppression method described above.

[0131] like Figure 23 As shown, the CT image ring artifact suppression device provided by the present invention includes:

[0132] Data collection module 231 is used to acquire a dataset of initial CT images;

[0133] Model building module 232 is used to build the RAS_UNet network model;

[0134] Function design module 233 is used to design loss functions for the RAS_UNet network model;

[0135] The artifact processing module 234 is used to input the dataset into the RAS_UNet network model and output the first CT image and the second CT image after artifact suppression processing.

[0136] The image fusion module 235 is used to fuse the effective information in the first CT image and the second CT image and output the fused final CT image.

[0137] The CT image ring artifact suppression device provided by the present invention can be used to perform the above-mentioned CT image ring artifact suppression method to solve the defects of incomplete removal of artifacts and low signal-to-noise ratio in related technologies. It achieves the effect of suppressing ring artifacts to the greatest extent while protecting the structural information of CT images, and at the same time significantly improving the signal-to-noise ratio of CT images.

[0138] Figure 24 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 24 As shown, the electronic device may include: a processor 2410, a communication interface 2420, a memory 2430, and a communication bus 2440, wherein the processor 2410, the communication interface 2420, and the memory 2430 communicate with each other through the communication bus 2440. The processor 2410 can call logical instructions in the memory 2430 to execute a CT image ring artifact suppression method, which includes:

[0139] Obtain the dataset of initial CT images;

[0140] Construct the RAS_UNet network model;

[0141] Design a loss function for the RAS_UNet network model;

[0142] The dataset is input into the RAS_UNet network model, which outputs the first CT image and the second CT image after artifact suppression processing.

[0143] The effective information in the first CT image and the second CT image is fused to output the final fused CT image.

[0144] Furthermore, the logical instructions in the aforementioned memory 2430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0145] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to perform the CT image ring artifact suppression method provided by the above methods, the method comprising:

[0146] Obtain the dataset of initial CT images;

[0147] Construct the RAS_UNet network model;

[0148] Design a loss function for the RAS_UNet network model;

[0149] The dataset is input into the RAS_UNet network model, which outputs the first CT image and the second CT image after artifact suppression processing.

[0150] The effective information in the first CT image and the second CT image is fused to output the final fused CT image.

[0151] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CT image ring artifact suppression method provided by the methods described above, the method comprising:

[0152] Obtain the dataset of initial CT images;

[0153] Construct the RAS_UNet network model;

[0154] Design a loss function for the RAS_UNet network model;

[0155] The dataset is input into the RAS_UNet network model, which outputs the first CT image and the second CT image after artifact suppression processing.

[0156] The effective information in the first CT image and the second CT image is fused to output the final fused CT image.

[0157] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A CT image ring artifact suppression method characterized by, include: Obtain the dataset of initial CT images; Construct the RAS_UNet network model; Design a loss function for the RAS_UNet network model; The dataset is input into the RAS_UNet network model, which outputs the first CT image and the second CT image after artifact suppression processing. The effective information in the first CT image and the second CT image is fused to output the final fused CT image; The loss function includes a loss function in the image domain and a loss function in the projection domain; the first CT image is a CT image after artifact suppression in the image domain, and the second CT image is a CT image after artifact suppression in the projection domain; The RAS_UNet network model uses the UNet network as its model framework. The downsampling part of the RAS_UNet network model includes one double convolutional module and four Inception modules for extracting artifact features at different scales. The upsampling part includes four upsampling modules and three residual modules for identifying and processing artifacts, and also includes one convolutional kernel. Design a loss function for the RAS_UNet network model, specifically including: The gradient of the initial CT image is calculated simultaneously using forward and backward differencing, and the maximum value of the forward and backward differencing is taken as the gradient of the initial CT image. ； Where I represents the initial CT image, (i,j) represents the coordinates of a pixel in the image, G1 represents the image gradient calculated by forward difference, G2 represents the image gradient calculated by backward difference, max represents taking the maximum value, and Gsin is the final image gradient in the sine curve. First, calculate the row mean of Gsin to obtain Gmean. Then, use Gmean as the weight to weight the gradient of Gsin in the sinusoidal domain. ； wherein G sin (i,j) represents a sinogram gradient weighting factor; Weighting gradients within the sliced ​​graph domain: ； Where ε is a positive real number to avoid the denominator being 0, R(p) represents a local region window centered at pixel p in the initial CT image, q is the pixel point in the local region window, and g p,q G represents the weighting coefficients of a Gaussian distribution centered at p. img For the image gradient in the slice image, WG img These are the image gradient weighting coefficients, ( )q represents the partial derivative of I along the column direction in the image; Using the weighted gradient as a regularization term and the structural similarity coefficient as a fidelity preservation term, the loss functions in the image domain and the projection domain are calculated as follows: ； Among them, Loss sin Loss is the loss function for the projection domain. img λ1 and λ2 are the loss functions for the image domain; SSIM is the fidelity term, representing the structural similarity coefficient; λ1 and λ2 ∈ [0,1].

2. The CT image ring artifact suppression method of claim 1, characterized by, The effective information from the first CT image and the second CT image is fused to output the final fused CT image, specifically including: The first CT image and the second CT image are decomposed using a non-subsampled contour wave transform algorithm to obtain multiple sub-images to be fused from the first CT image and multiple sub-images to be fused from the second CT image. Calculate the residuals of the pairwise corresponding sub-images to be fused. If the residuals exceed a set residual threshold, the pixel values ​​of the sub-images to be fused with lower contrast are used as the fusion result, and the fused sub-image is output. If the residuals do not exceed the set residual threshold, the pixel values ​​of the sub-images to be fused with higher contrast are used as the fusion result, and the fused sub-image is output. The fused sub-images are combined to output the final CT image.

3. The CT image ring artifact suppression method of claim 2, characterized by, The sub-images to be fused include low-frequency sub-band images and high-frequency sub-band images.

4. The CT image ring artifact suppression method of claim 2, characterized by, The residual is the absolute value of the difference between the gray values ​​of the two sub-images to be fused; the residual threshold is the product of the sum of the gray values ​​of the two sub-images to be fused and the ratio value.

5. A CT image ring artifact suppression apparatus, characterized by comprising: include: The data collection module is used to acquire the dataset of initial CT images; The model building module is used to build the RAS_UNet network model; The function design module is used to design the loss function for the RAS_UNet network model; The artifact processing module is used to input the dataset into the RAS_UNet network model and output the first CT image and the second CT image after artifact suppression processing. The image fusion module is used to fuse the effective information in the first CT image and the second CT image, and output the fused final CT image; The loss function includes a loss function in the image domain and a loss function in the projection domain; the first CT image is a CT image after artifact suppression in the image domain, and the second CT image is a CT image after artifact suppression in the projection domain; The RAS_UNet network model uses the UNet network as its model framework. The downsampling part of the RAS_UNet network model includes one double convolutional module and four Inception modules for extracting artifact features at different scales. The upsampling part includes four upsampling modules and three residual modules for identifying and processing artifacts, and also includes one convolutional kernel. Design a loss function for the RAS_UNet network model, specifically including: The gradient of the initial CT image is calculated simultaneously using forward and backward differencing, and the maximum value of the forward and backward differencing is taken as the gradient of the initial CT image. ； Where I represents the initial CT image, (i,j) represents the coordinates of a pixel in the image, G1 represents the image gradient calculated by forward difference, G2 represents the image gradient calculated by backward difference, max represents taking the maximum value, and Gsin is the final image gradient in the sine curve. First, calculate the row mean of Gsin to obtain Gmean. Then, use Gmean as the weight to weight the gradient of Gsin in the sinusoidal domain. ； wherein G sin (i,j) represents a sinogram gradient weighting factor; Weighting gradients within the sliced ​​graph domain: ； Where ε is a positive real number to avoid the denominator being 0, R(p) represents a local region window centered at pixel p in the initial CT image, q is the pixel point in the local region window, and g p,q G represents the weighting coefficients of a Gaussian distribution centered at p. img For the image gradient in the slice image, WG img These are the image gradient weighting coefficients, ( )q represents the partial derivative of I along the column direction in the image; Using the weighted gradient as a regularization term and the structural similarity coefficient as a fidelity preservation term, the loss functions in the image domain and the projection domain are calculated as follows: ； wherein Loss sin is a loss function of the projection domain, Loss img is a loss function of the image domain; SSIM is a fidelity term representing a structural similarity coefficient; λ1, λ2∈[0,1].

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the CT image ring artifact suppression method as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the CT image ring artifact suppression method as described in any one of claims 1 to 4.

Images