A method for metal artifact removal in CT images based on dual-domain progressive enhancement network

Abstract

The present invention discloses a method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network, and relates to the technical field of image data processing. The method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network comprises the following steps: modeling data set processing, processing effect evaluation and judgment, artifact removal evaluation and judgment, and optimization effect evaluation and judgment. The present invention evaluates the effect of data set preprocessing to determine whether to execute multi-level dual-domain progressive denoising network model construction, then evaluates the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to determine whether to execute model optimization, and finally obtains the optimization effect evaluation index after executing model optimization, and determines whether to execute re-acquisition of simulation data sets, thereby achieving more accurate removal of metal artifacts from CT images, and solving the problem of low precision of metal artifact removal from CT images based on a dual-domain progressive enhancement network in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network. Background Art

[0002] With the development of medical imaging technology, computed tomography (CT) technology has become one of the "gold standards" in clinical medicine, especially in emergency, tumor detection and surgical planning. However, the quality of CT images is affected by many factors, and metal artifacts are one of the main challenges. Metal artifacts are artifacts in CT images caused by metal objects (such as artificial joints, dental implants, stents, etc.). The high density of metal materials prevents X-rays from fully penetrating, resulting in strong artifacts in CT images. These artifacts usually appear as banded shadows, stripes or irregular patterns, which seriously affect the quality of CT images.

[0003] In the past, artifact removal in medical images mainly relied on traditional image processing methods (such as filtering, interpolation, reconstruction algorithms, etc.). However, traditional methods are often unable to effectively deal with complex artifact types, especially metal artifacts. In recent years, deep learning methods have become the mainstream in the research of CT metal artifact removal tasks. Such methods are usually based on architectures such as convolutional neural networks and generative adversarial networks. By learning the characteristics of artifacts from a large amount of training data and automatically removing artifacts, the quality of CT images is improved, thereby improving the accuracy of clinical diagnosis and treatment.

[0004] For example, a new method for removing metal artifacts from CT images published with publication number: CN118628599A includes: a first step of preparing a simulated data set to construct a training set of damaged CT images; a second step of constructing a dual-domain network model, including image preprocessing, prior image generation based on wavelet fusion, constructing a projection domain repair network based on mask UNet, constructing an image domain repair network based on improved UNet, and designing an overall loss function of the dual-domain network model; a third step of removing metal artifacts from clinical CT images affected by metal artifacts using the dual-domain network model obtained in the second step, and adjusting the size of the clinical CT images affected by metal artifacts to be consistent with the size of the training images.

[0005] Another example is a metal artifact removal method based on a dual-domain Fourier neural network, which is announced with the announcement number CN114494498B, including: removing metal artifacts from both local and global aspects in the chord diagram domain and the CT image domain respectively through fast Fourier convolution. By making full use of the global receptive field provided by Fourier convolution, the chord diagram damaged by metal can be repaired using long-distance information on the chord diagram, thereby removing metal artifacts from the original data. At the same time, in the image domain, a U-Net containing fast Fourier convolution jump connections is used to eliminate secondary artifacts and global inconsistencies caused by artifact removal. In the process of training the network, a multi-window width loss function is used to optimize the model performance under clinically important window widths.

[0006] However, the above technology has at least the following technical problems:

[0007] Deep learning models rely on a large number of CT images with metal artifacts and their corresponding artifact-free images for training. However, in actual clinical practice, it is relatively difficult to obtain a CT image dataset with metal artifacts. Manually labeling high-quality metal artifact areas requires professional medical knowledge, and the dataset labeling process is cumbersome. The scarcity of training data may lead to insufficient model generalization capabilities.

[0008] It is also necessary to consider that although deep learning methods have made significant progress in artifact removal, while removing metal artifacts, they may also lead to inaccurate processing of details in the image, affecting clinical diagnosis. There is a problem of low accuracy in metal artifact removal in CT images based on the dual-domain progressive enhancement network. Summary of the invention

[0009] The purpose of the present invention is to provide a method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network, which solves the problem of low precision of metal artifact removal from CT images based on a dual-domain progressive enhancement network in the prior art and achieves more accurate removal of metal artifacts from CT images.

[0010] A method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network comprises the following steps:

[0011] S1, performing data set preprocessing on the acquired simulation data set to obtain a CT image modeling data set; the simulation data set includes a CT image data set and a metal mask data set;

[0012] S2, evaluating the CT image modeling data set to obtain an image processing quality evaluation index, and judging whether to execute the multi-level dual-domain progressive denoising network model construction based on the image processing quality evaluation index; the image processing quality evaluation index is used to quantify the effect of data set preprocessing;

[0013] S3, after executing the multi-level dual-domain progressive denoising network model construction, evaluating the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to obtain an artifact removal conformity evaluation index, and judging whether to perform model optimization based on the artifact removal conformity evaluation index, wherein the artifact removal conformity evaluation index is used to quantitatively evaluate the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts;

[0014] S4, obtaining an optimization effect evaluation index after executing the model optimization, and judging whether to execute re-acquisition of the simulation data set according to the image processing quality evaluation index and the artifact removal compliance evaluation index, wherein the optimization effect evaluation index is used to quantify the effect of the evaluation model optimization.

[0015] Furthermore, the specific steps for obtaining a CT image modeling data set are as follows: performing data set cleaning on the simulation data set to obtain a first CT image modeling data set; performing data set standardization and normalization processing on the first CT image modeling data set to obtain a second CT image modeling data set; performing data set enhancement and data set segmentation on the second CT image modeling data set to obtain a CT image modeling data set.

[0016] Furthermore, the specific steps of evaluating the CT image modeling data set to obtain the image processing quality evaluation index are as follows: obtaining initial evaluation related data before performing data set preprocessing, the initial evaluation related data including the number of missing values ​​of the initial data, the mean of the initial data set, the signal mean of the initial data set and the noise variance of the initial data set; obtaining processing evaluation related data after performing data set preprocessing, the processing evaluation related data including the number of missing values ​​of the processing data, the data set related evaluation value, the Euclidean distance of the data set, the signal mean of the processing data set, the noise variance of the processing data set, and the data set processing time; obtaining a data set quality evaluation value and a preprocessing efficiency evaluation value based on the initial evaluation related data, the processing evaluation related data and the reference processing data obtained from a preset database, the reference processing data including a processing missing reduction rate reference value, a processing noise reference value and a data set processing reference time; combining the data set quality evaluation value, the preprocessing efficiency evaluation value and the preprocessing influencing factor, the data set quality evaluation weight and the preprocessing efficiency evaluation weight obtained from the preset database to obtain the image processing quality evaluation index.

[0017] Furthermore, the method for obtaining the image processing quality evaluation index is as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] In the formula, IQ represents the image processing quality evaluation index, τpf represents the preprocessing impact factor, ε1 represents the data set quality assessment weight, DQ represents the data set quality assessment value, ε2 represents the preprocessing efficiency assessment weight, PQ represents the preprocessing efficiency assessment value, Q c represents the number of missing values ​​in the initial data, Q y represents the number of missing values ​​in the processed data, ΔQ represents the reference value of the missing value reduction rate, μ′ represents the relevant evaluation value of the data set, and μ c represents the mean of the initial data set, D(c,y) represents the Euclidean distance of the data set, μ x,y Represents the mean value of the processed data set signal, μ x,c represents the initial data set signal mean, σ n,y represents the noise variance of the processed data set, σ n,c represents the noise variance of the initial data set, ΔS represents the reference value of the processed noise, and T y represents the dataset processing time, and T′ represents the dataset processing reference time.

[0022] Furthermore, the specific steps for judging whether to execute the construction of the multi-level dual-domain progressive denoising network model based on the image processing quality assessment index are as follows: compare the image processing quality assessment index with the preset processing quality assessment threshold range obtained from the preset database; if the image processing quality assessment index is within the preset processing quality assessment threshold range, then execute the construction of the multi-level dual-domain progressive denoising network model, and at the same time evaluate the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to obtain the artifact removal conformity evaluation index; if the image processing quality assessment index exceeds the preset processing quality assessment threshold range, then do not execute the construction of the multi-level dual-domain progressive denoising network model, and remind the preset personnel to re-preprocess the data set.

[0023] Furthermore, the specific steps of evaluating the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to obtain an artifact removal conformity evaluation index are as follows: obtaining artifact removal related evaluation values ​​after the multi-level dual-domain progressive denoising network model is trained, the artifact removal related evaluation values ​​including artifact peak signal-to-noise ratio, initial peak signal-to-noise ratio, removal edge similarity, removal texture similarity and conformity structure similarity; combining the artifact removal related evaluation values ​​with removal reference data obtained from a preset database to obtain an artifact removal conformity evaluation index, the removal reference data including a similarity reference value and a removal evaluation weight, the similarity reference value including a reference edge similarity, a reference texture similarity and a structure similarity reference value, and the removal evaluation weight including an artifact removal conformity evaluation weight, a detail conformity evaluation weight and a structure similarity evaluation weight.

[0024] Furthermore, the specific steps for determining whether to perform model optimization based on the artifact removal compliance evaluation index are as follows: compare the artifact removal compliance evaluation index with a preset artifact compliance threshold range obtained from a preset database; if the artifact removal compliance evaluation index is within the preset artifact compliance threshold range, model optimization is not performed; if the artifact removal compliance evaluation index exceeds the preset performance compliance threshold range, model optimization is performed, and at the same time, an optimization effect evaluation index after performing model optimization is obtained.

[0025] Furthermore, the specific steps for obtaining the optimization effect evaluation index after the execution model is optimized are as follows: obtain the initial optimization effect evaluation related values ​​before the execution model is optimized, the initial optimization effect evaluation related values ​​include the initial artifact removal compliance evaluation index, the initial computational complexity and the initial convergence speed; obtain the optimization evaluation related values ​​after the execution model is optimized, the optimization evaluation related values ​​include the optimization artifact removal compliance evaluation index, the optimization computational complexity and the optimization convergence speed; perform a difference operation on the initial convergence speed and the optimized convergence speed to obtain the optimized convergence speed change, and obtain the optimization effect evaluation index according to the initial optimization effect evaluation related values, the optimization evaluation related values, the optimized convergence speed change, the complexity reference change obtained from the preset database and the optimization evaluation weights, the optimization evaluation weights include the artifact optimization evaluation weights, the convergence speed optimization evaluation weights and the computational complexity optimization evaluation weights.

[0026] The present invention has the following technical effects or advantages:

[0027] 1. The present invention evaluates the effect of data set preprocessing to determine whether to execute the multi-level dual-domain progressive denoising network model construction, then evaluates the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to determine whether to execute model optimization, and finally determines whether to re-acquire the simulation data set, thereby realizing the evaluation and optimization of the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts, thereby achieving more accurate removal of metal artifacts in CT images, and effectively solving the problem of low accuracy of metal artifact removal in CT images based on dual-domain progressive enhancement network in the prior art.

[0028] 2. The present invention obtains artifact removal compliance evaluation values ​​after training the multi-level dual-domain progressive denoising network model, and then combines the artifact removal compliance evaluation values ​​with the removal reference data obtained from a preset database to obtain an artifact removal compliance evaluation index, thereby realizing a numerical evaluation of the conformity of the multi-level dual-domain progressive denoising network model to remove artifacts, thereby realizing a more accurate evaluation of the conformity of the multi-level dual-domain progressive denoising network model to remove artifacts.

[0029] 3. The present invention obtains the initial optimization effect evaluation related value before executing the model optimization, and simultaneously obtains the optimization evaluation related value after executing the model optimization, and then performs a difference operation on the initial convergence speed and the optimized convergence speed to obtain the optimized convergence speed change. Finally, the optimization effect evaluation index is obtained according to the initial optimization effect evaluation related value, the optimized evaluation related value, the optimized convergence speed change, the complexity reference change obtained from the preset database and the optimized evaluation weight, thereby realizing the numerical evaluation of the model optimization effect, and further realizing a more accurate evaluation of the model optimization effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flowchart of a method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network provided by an embodiment of the present invention;

[0031] Figure 2 A schematic diagram showing changes in the artifact removal compliance evaluation index according to the removal edge similarity provided by an embodiment of the present invention;

[0032] Figure 3 A schematic diagram of how the artifact removal compliance evaluation index varies with the removal texture similarity provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The embodiment of the present invention solves the problem of low precision of metal artifact removal of CT images based on a dual-domain progressive enhancement network in the prior art by providing a method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network. A CT image modeling dataset is obtained by performing dataset preprocessing on an acquired simulation dataset, and then the CT image modeling dataset is evaluated to obtain an image processing quality evaluation index. Based on the image processing quality evaluation index, it is determined whether to execute multi-level dual-domain progressive denoising network model construction, and then the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts is evaluated to obtain an artifact removal conformity evaluation index. It is determined whether to execute model optimization and obtain an optimization effect evaluation index after executing model optimization. Finally, it is determined whether to execute re-acquisition of the simulation dataset based on the image processing quality evaluation index and the artifact removal conformity evaluation index, thereby achieving more accurate removal of metal artifacts from CT images.

[0034] The technical solution in the embodiment of the present invention is to solve the problem of low precision of metal artifact removal in CT images based on the dual-domain progressive enhancement network. The overall idea is as follows:

[0035] The effect of data set preprocessing is evaluated to determine whether to execute the multi-level dual-domain progressive denoising network model construction. Then, the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts is evaluated to determine whether to execute model optimization. Finally, the optimization effect evaluation index after model optimization is obtained, and it is determined whether to re-acquire the simulation data set, thereby achieving more accurate removal of metal artifacts in CT images.

[0036] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0037] like Figure 1 As shown, it is a flowchart of a method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network provided by an embodiment of the present invention, and the method includes the following steps: S1, modeling data set processing: performing data set preprocessing on the acquired simulation data set to obtain a CT image modeling data set, and the simulation data set includes a CT image data set and a metal mask data set; S2, processing effect evaluation judgment: evaluating the CT image modeling data set to obtain an image processing quality evaluation index, and judging whether to execute a multi-level dual-domain progressive denoising network model construction based on the image processing quality evaluation index, and the image processing quality evaluation index is used to quantify the effect of data set preprocessing; S3, removing artifacts Evaluation and judgment: After executing the construction of the multi-level dual-domain progressive denoising network model, the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts is evaluated to obtain the artifact removal conformity evaluation index. Based on the artifact removal conformity evaluation index, it is judged whether to execute model optimization. The artifact removal conformity evaluation index is used to quantitatively evaluate the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts; S4, optimization effect evaluation and judgment: obtain the optimization effect evaluation index after executing model optimization, and judge whether to execute re-acquisition of simulation data set according to the image processing quality evaluation index and the artifact removal conformity evaluation index. The optimization effect evaluation index is used to quantitatively evaluate the effect of model optimization.

[0038] It should be understood that simulation data sets are obtained through simulation software such as TOPAS (Tool for Open Parallel Analysis of Simulations). CT image data sets refer to medical image sets obtained through simulated or actual CT scanners, and metal mask data sets refer to the effects of metal objects (such as metal implants, prostheses, teeth, titanium alloy brackets, etc.) on CT images during simulated or actual CT scanning.

[0039] It should be added that the specific steps to obtain the CT image modeling data set are as follows: perform data set cleaning on the simulation data set to obtain the first CT image modeling data set; perform data set standardization and normalization on the first CT image modeling data set to obtain the second CT image modeling data set; perform data set enhancement and data set segmentation on the second CT image modeling data set to obtain the CT image modeling data set.

[0040] Specifically, the Python library (OpenCV) is used to clean the data set. The simulation data set is checked for duplicates and deleted, and the missing data in the simulation data set is processed by interpolation. The mean of the first CT image modeling data set is scaled to the range of 0 to 1 through standardization and normalization functions (such as scikit-learn). The second CT image modeling data set is enhanced through the ImageDataGenerator in Keras, and then the data set is segmented through image segmentation technology (such as Canny edge detection) to obtain the CT image modeling data set.

[0041] The embodiment of the present invention obtains a CT image modeling dataset by performing dataset preprocessing on the acquired simulation dataset, thereby improving the reliability of the CT image modeling dataset, and further improving the quality of metal artifact removal in CT images based on a dual-domain progressive enhancement network.

[0042] Furthermore, the specific steps of evaluating the CT image modeling dataset to obtain the image processing quality evaluation index are as follows: obtaining initial evaluation related data before performing dataset preprocessing, the initial evaluation related data include the number of missing values ​​of the initial data, the mean of the initial dataset, the signal mean of the initial dataset, and the noise variance of the initial dataset; obtaining processing evaluation related data after performing dataset preprocessing, the processing evaluation related data include the number of missing values ​​of the processing data, the dataset related evaluation value, the Euclidean distance of the dataset, the signal mean of the processing dataset, the noise variance of the processing dataset, and the processing time of the dataset; obtaining the dataset quality evaluation value (i.e., DQ in the image processing quality evaluation index) and the preprocessing efficiency evaluation value (i.e., PQ in the image processing quality evaluation index) based on the initial evaluation related data, the processing evaluation related data, and the reference processing data obtained from the preset database, the reference processing data include the processing missing reduction rate reference value, the processing noise reference value, and the dataset processing reference time; combining the dataset quality evaluation value, the preprocessing efficiency evaluation value, the preprocessing influencing factor, the dataset quality evaluation weight, and the preprocessing efficiency evaluation weight obtained from the preset database to obtain the image processing quality evaluation index.

[0043] The method to obtain the image processing quality evaluation index is as follows:

[0044] ;

[0045] ;

[0046] ;

[0047] In the formula, IQ represents the image processing quality evaluation index, τ pf represents the preprocessing impact factor, ε1 represents the data set quality assessment weight, DQ represents the data set quality assessment value, ε2 represents the preprocessing efficiency assessment weight, PQ represents the preprocessing efficiency assessment value, Q c represents the number of missing values ​​in the initial data, Q y represents the number of missing values ​​in the processed data, ΔQ represents the reference value of the missing value reduction rate, μ′ represents the relevant evaluation value of the data set, and μ c represents the mean of the initial data set, D(c,y) represents the Euclidean distance of the data set, μ x,y Represents the mean value of the processed data set signal, μ x,c represents the initial data set signal mean, σ n,y represents the noise variance of the processed data set, σ n,c represents the noise variance of the initial data set, ΔS represents the reference value of the processed noise, and T y represents the dataset processing time, and T′ represents the dataset processing reference time.

[0048] In this embodiment, the number of missing values ​​of the initial data, the mean of the initial data set, the mean of the initial data set signal, and the noise variance of the initial data set are obtained through a Python library (such as pandas); the number of missing values ​​of the processed data, the signal mean of the processed data set, the signal mean of the processed data set, and the noise variance of the processed data set are obtained through a Python library (such as NumPy), where the data set related evaluation value is the absolute value of the difference between the signal mean of the processed data set and 0.5; the Euclidean distance of the data set is obtained through a Python library (such as SciPy), and the data set processing time is obtained through a Python library (such as time).

[0049] Among them, the processing missing reduction rate reference value is represented by summing and averaging the collected historical processing missing reduction rates, the processing noise reference value is represented by summing and averaging the collected historical signal-to-noise ratios, and the data set processing reference time is represented by summing and averaging the collected historical data set processing times.

[0050] The preprocessing influence factor is used to describe the degree of influence of the quality of the data set preprocessing influence factor on the image processing quality evaluation index. Specifically, the preprocessing influence factor is the influence factor corresponding to the preprocessing preset in the preset database, which represents the numerical value of the influence of the quality of the data set preprocessing influence factor on the image processing quality evaluation index. When used, the preset preprocessing influence factor can be directly obtained from the preset database, and the corresponding relationship can be a preset mapping relationship. For example, the quality of the data set preprocessing influence factor and the influence factor corresponding to the preset preprocessing in the preset database form a mapping set, and the quality of the real-time data set preprocessing influence factor is input into the mapping set to obtain the corresponding influence factor. The mapping relationship can be one-to-one or many-to-one. In the embodiment of the present invention, the value range is [0, 1].

[0051] The data set quality assessment weight represents the degree of influence of the quality of data set preprocessing on the image processing quality assessment index, and is a preset value in a preset database. When in use, the weight corresponding to the preset data set quality assessment can be directly obtained from the preset database, and the corresponding relationship can be a preset mapping relationship. For example, the quality of data set preprocessing and the weight corresponding to the preset data set quality assessment in the preset database form a mapping set, and the quality of real-time data set preprocessing is input into the mapping set to obtain the corresponding weight. The mapping relationship can be one-to-one or many-to-one. In the embodiment of the present invention, the value range is [0, 1].

[0052] In the embodiment of the present invention, the sum of the data set quality assessment weight and the preprocessing efficiency assessment weight is 1, and the preprocessing efficiency assessment weight quantifies the influence of the efficiency of data set preprocessing on the image processing quality assessment index.

[0053] Specific assumptions: preprocessing impact factor is 0.9, data set quality assessment weight is 0.5, preprocessing efficiency assessment weight is 0.5, the number of missing values ​​in the initial data is 100, the reference value of the missing reduction rate is 0.2, the data set correlation assessment value is 0.1, the initial data set mean is 0.8, the processing data set signal mean is 60, the initial data set signal mean is 55, the processing data set noise variance is 20, the initial data set noise variance is 25, the processing noise reference value is 5, and the data set processing reference time is 100 minutes. The image processing quality assessment index can be calculated by the above method, and the statistical table of the change of the image processing quality assessment index is shown in Table 1:

[0054] Table 1 Statistics of changes in image processing quality evaluation index

[0055]

[0056] It can be seen from the first and second groups of data in Table 1 that as the number of missing values ​​in the processed data and the Euclidean distance of the data set decrease, the data set quality assessment value increases, and the image processing quality assessment index increases accordingly; in addition, as in the second and third groups of data, as the data set processing time decreases, the preprocessing efficiency assessment value increases, and the image processing quality assessment index increases accordingly.

[0057] The image processing quality evaluation index is used to quantify the effect of data set preprocessing, where changes in the data set quality evaluation value and the preprocessing efficiency evaluation value directly lead to changes in the image processing quality evaluation index. Specifically, as the data set quality evaluation value and the preprocessing efficiency evaluation value increase, the image processing quality evaluation index increases. In addition, the image processing quality evaluation index includes parameters from multiple aspects, and each parameter is related to each other and does not exist independently. For example, the Euclidean distance of the data set is used to measure the difference between the simulation data set and the CT image modeling data set. As the number of missing values ​​in the processed data decreases, it indicates that the difference between the simulation data set and the CT image modeling data set decreases, that is, the Euclidean distance of the data set decreases. At the same time, as the number of missing values ​​in the processed data increases, it indicates that a longer data set processing time is required, that is, the data set processing time increases, which in turn causes the data set noise variance to increase. Therefore, through a quantitative approach, the correlation and mutual influence between various factors are considered, and the image processing quality evaluation index is obtained through a comprehensive analysis. The numerical evaluation of the data set preprocessing effect is realized, and the data set preprocessing effect is judged through the numerical evaluation, thereby achieving a more accurate evaluation of the data set preprocessing effect.

[0058] Furthermore, the specific steps for judging whether to execute the construction of the multi-level dual-domain progressive denoising network model based on the image processing quality assessment index are as follows: compare the image processing quality assessment index with the preset processing quality assessment threshold range obtained from the preset database; if the image processing quality assessment index is within the preset processing quality assessment threshold range, then execute the construction of the multi-level dual-domain progressive denoising network model, and at the same time evaluate the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to obtain the artifact removal conformity evaluation index; if the image processing quality assessment index exceeds the preset processing quality assessment threshold range, then do not execute the construction of the multi-level dual-domain progressive denoising network model, and remind the preset personnel to re-preprocess the data set.

[0059] In this embodiment, the preset processing quality assessment threshold range is set by professionals according to standards in the field. For example, the preset processing quality assessment threshold range is set to 0.7 to 1. By making judgments based on the preset processing quality assessment threshold range, a more accurate judgment of the data set preprocessing effect is achieved.

[0060] Furthermore, the specific steps of constructing a multi-level dual-domain progressive denoising network model are as follows: A1, constructing an attention embedded sine domain restoration network, which includes embedding an input layer, embedding a hidden layer, and embedding an output layer; A2, constructing an image domain restoration network, which includes restoring an input layer, restoring a hidden layer, and restoring an output layer; A3, constructing an image detail enhancement network, which includes a parallel expansion fusion block and a pixel-level attention block.

[0061] In an embodiment of the present invention, an attention embedding sinusoidal domain restoration network is constructed to minimize the impact on the overall geometric structure by correcting only the areas affected by metal in the sinusoidal map without modifying other areas; the weight map is applied to the feature map by element-by-element multiplication, so that the model can pay more attention to severe artifact areas, thereby improving artifact removal and image reconstruction.

[0062] An image domain restoration network is constructed. In order to further remove residual metal artifacts in the image domain, the image domain restoration network applies U-Net to further remove metal artifacts in the image domain through residual learning.

[0063] Construct an image detail enhancement network, which is used to restore and reconstruct detailed information in the image, ensuring that the final output not only reduces artifacts but also retains clear and rich details. The image detail enhancement network consists of an initial convolutional layer, three enhanced pixel-level residual attention blocks, and a final convolutional layer, all of which are connected by a global skip strategy. The expressive power of the model is enhanced by integrating features from different layers while retaining the original information from the input layer.

[0064] Specifically, the CT image modeling dataset is taken as input, the loss function threshold and the maximum number of training times are set, and the training parameters are initialized; the linear interpolation corrected sinusoidal graph and the sine domain artifact weight component are spliced ​​in the channel dimension and input into the attention embedding sine domain restoration network, and the sine domain corrected sinusoidal graph is output; the corrected sinusoidal graph and the linear interpolation corrected sinusoidal graph are input into the filtered back projection to obtain the sine domain corrected image and the linear interpolation corrected image; the sine domain corrected image and the linear interpolation corrected image are input into the image domain restoration network, and the image domain corrected image is output; the image domain corrected image is input into the image detail enhancement network, and the final reconstructed image is output.

[0065] The embodiment of the present invention improves the reliability of removing metal artifacts from CT images by executing a multi-level dual-domain progressive denoising network model construction, thereby achieving more accurate removal of metal artifacts from CT images.

[0066] Furthermore, the specific steps of evaluating the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to obtain an artifact removal conformity evaluation index are as follows: obtaining artifact removal related evaluation values ​​after training the multi-level dual-domain progressive denoising network model, the artifact removal related evaluation values ​​including artifact peak signal-to-noise ratio, initial peak signal-to-noise ratio, removal edge similarity, removal texture similarity and conformity structure similarity; combining the artifact removal related evaluation values ​​with the removal reference data obtained from a preset database to obtain an artifact removal conformity evaluation index, the removal reference data including a similarity reference value and a removal evaluation weight, the similarity reference value including a reference edge similarity, a reference texture similarity and a structure similarity reference value, and the removal evaluation weight including an artifact removal conformity evaluation weight, a detail conformity evaluation weight and a structure similarity evaluation weight.

[0067] The method to obtain the artifact removal compliance evaluation index is as follows:

[0068] ;

[0069] Wherein, AR represents the artifact removal compliance assessment index, δ1 represents the artifact removal compliance assessment weight, PSNR(w,k) represents the artifact peak signal-to-noise ratio, PSNR(c,k) represents the initial peak signal-to-noise ratio, δ2 represents the detail compliance assessment weight, B represents the removed edge similarity, ΔB represents the reference edge similarity, H represents the removed texture similarity, ΔH represents the reference texture similarity, δ3 represents the structural similarity assessment weight, SSIM(w,c) represents the compliance structure similarity, and ΔX represents the structural similarity reference value.

[0070] In an embodiment of the present invention, the artifact peak signal-to-noise ratio and the initial peak signal-to-noise ratio are obtained through MATLAB (such as the psnr() function), the removed edge similarity and the removed texture similarity are obtained through a Python library (such as OpenCV), and the conforming structure similarity is obtained through MATLAB (such as the ssim() function).

[0071] Among them, the result of summing and averaging the collected historical edge similarities represents the reference edge similarity, the result of summing and averaging the collected historical texture similarities represents the reference texture similarity, and the result of summing and averaging the collected historical structure similarities represents the structure similarity reference value.

[0072] The artifact removal compliance evaluation weight is used to describe the influence of the artifact removal compliance situation on the artifact removal compliance evaluation index. In the embodiment of the present invention, the sum of the artifact removal compliance evaluation weight, the detail compliance evaluation weight and the structural similarity evaluation weight is 1.

[0073] The detail compliance assessment weight indicates the degree of influence of the detail compliance in artifact removal on the artifact removal compliance assessment index, and is a preset value in a preset database. When in use, the weight corresponding to the preset detail compliance assessment can be directly obtained from the preset database, and the corresponding relationship can be a preset mapping relationship. For example, the detail compliance in artifact removal and the weight corresponding to the preset detail compliance assessment in the preset database form a mapping set, and the detail compliance in real-time artifact removal is input into the mapping set to obtain the corresponding weight. The mapping relationship can be a one-to-one correspondence or a many-to-one relationship. In the embodiment of the present invention, the value range is [0, 1].

[0074] The structural similarity evaluation weight is a numerical value of the degree of influence of the structural similarity of quantitative artifact removal preset in the preset database on the artifact removal compliance evaluation index. When used, the weight corresponding to the preset structural similarity evaluation can be directly obtained from the preset database, and the corresponding relationship can be a pre-set mapping relationship. For example, the structural similarity of artifact removal and the weight corresponding to the structural similarity evaluation preset in the preset database form a mapping set, and the structural similarity of real-time artifact removal is input into the mapping set to obtain the corresponding weight. The mapping relationship can be a one-to-one correspondence or a many-to-one relationship. In the embodiment of the present invention, the value range is [0, 1].

[0075] Figure 2 , Figure 3 A schematic diagram of the change of the artifact removal compliance evaluation index provided in an embodiment of the present invention, wherein the following settings are made: the artifact removal compliance evaluation weight is 0.5, the detail compliance evaluation weight is 0.3, the structural similarity evaluation weight is 0.2, the artifact peak signal-to-noise ratio is 35 decibels, the initial peak signal-to-noise ratio is 30 decibels, the reference edge similarity is 1, the reference texture similarity is 1, the compliance structure similarity is 0.9, and the structural similarity reference value is 1; wherein Figure 2 Schematic diagram of the change of artifact removal compliance evaluation index with the removal edge similarity. Specifically, the removal texture similarity is set to 0.8. As the removal edge similarity increases, the artifact removal compliance evaluation index increases accordingly. Figure 3 Schematic diagram of the change of artifact removal compliance evaluation index with the removal texture similarity. Specifically, the removal edge similarity is set to 0.5. As the removal texture similarity increases, the artifact removal compliance evaluation index also increases.

[0076] The artifact removal compliance evaluation index is used to quantitatively evaluate the compliance of the multi-level dual-domain progressive denoising network model in removing artifacts. Specifically, the artifact removal compliance evaluation index includes multiple parameters, and each parameter is related to each other and does not exist independently. For example, with the increase of edge removal similarity and texture removal similarity, it indicates that the image edge and texture retention increases during the artifact removal process, that is, the compliance structure similarity increases accordingly. In addition, with the loss of edges and textures, the compliance structure similarity decreases, which in turn leads to an increase in the artifact peak signal-to-noise ratio. Therefore, through a quantitative approach, the correlation and mutual influence between various factors are considered, and the artifact removal compliance evaluation index is obtained through a comprehensive analysis, which realizes the numerical evaluation of the compliance degree of the multi-level dual-domain progressive denoising network model in removing artifacts. The multi-level dual-domain progressive denoising network model is judged through numerical evaluation to remove artifacts. The accuracy of the evaluation of the compliance degree of the multi-level dual-domain progressive denoising network model in removing artifacts is improved.

[0077] Furthermore, the specific steps for determining whether to perform model optimization based on the artifact removal compliance evaluation index are as follows: compare the artifact removal compliance evaluation index with a preset artifact compliance threshold range obtained from a preset database; if the artifact removal compliance evaluation index is within the preset artifact compliance threshold range, model optimization is not performed; if the artifact removal compliance evaluation index exceeds the preset performance compliance threshold range, model optimization is performed, and at the same time, an optimization effect evaluation index after performing model optimization is obtained.

[0078] In the embodiments of the present invention, specifically, the model optimization achieves model data level optimization by reducing the number of features through dimensionality reduction methods such as principal component analysis; hyperparameter optimization is achieved through automated hyperparameter optimization tools such as Optuna; and model evaluation optimization is achieved through Matplotlib visualization of evaluation results.

[0079] Among them, the preset artifact compliance threshold range is set by professionals according to standards in the field. For example, the artifact peak signal-to-noise ratio range (generally 0 decibels to 50 decibels), the initial peak signal-to-noise ratio range (generally 0 decibels to 50 decibels), the edge removal similarity range (generally 0.7 to 1), the texture removal similarity range (generally 0.7 to 1), and the compliance structure similarity range (generally 0.7 to 1) are set to determine the preset optimization effect threshold range. By combining the preset artifact compliance threshold range for judgment and optimization, the degree of compliance with artifact removal of the multi-level dual-domain progressive denoising network model is improved.

[0080] Furthermore, the specific steps for obtaining the optimization effect evaluation index after the execution model is optimized are as follows: obtain the initial optimization effect evaluation related values ​​before the execution model is optimized, the initial optimization effect evaluation related values ​​include the initial artifact removal compliance evaluation index, the initial computational complexity and the initial convergence speed; obtain the optimization evaluation related values ​​after the execution model is optimized, the optimization evaluation related values ​​include the optimization artifact removal compliance evaluation index, the optimization computational complexity and the optimization convergence speed; perform a difference operation on the initial convergence speed and the optimized convergence speed to obtain the optimized convergence speed change (i.e., ΔV in the optimization effect evaluation index), and obtain the optimization effect evaluation index according to the initial optimization effect evaluation related values, the optimization evaluation related values, the optimized convergence speed change, the complexity reference change obtained from the preset database and the optimization evaluation weights, the optimization evaluation weights include the artifact optimization evaluation weights, the convergence speed optimization evaluation weights and the computational complexity optimization evaluation weights.

[0081] The method for obtaining the optimization effect evaluation index is as follows:

[0082] ;

[0083] Where EA represents the optimization effect evaluation index, τ1 represents the artifact optimization evaluation weight, and AR y Indicates that the optimization artifact removal meets the evaluation index, AR c represents the initial artifact removal compliance evaluation index, τ2 represents the convergence speed optimization evaluation weight, ΔV represents the optimized convergence speed change, τ3 represents the computational complexity optimization evaluation weight, FI c Indicates the initial computational complexity, FI y represents the optimization computational complexity, ΔF represents the reference change in complexity, and e represents a natural constant.

[0084] In an embodiment of the present invention, specifically, the initial computational complexity and the optimized computational complexity are obtained through PyTorch Profiler, the initial convergence speed and the optimized convergence speed are obtained through PyTorch Lightning, and the complexity reference change is represented by summing and averaging the collected historical complexity changes; wherein, the initial artifact removal compliance evaluation index refers to the artifact removal compliance evaluation index before the execution model optimization, the initial computational complexity refers to the computational complexity before the execution model optimization, and the initial convergence speed refers to the convergence speed before the execution model optimization.

[0085] The optimized artifact removal conformity evaluation index refers to the artifact removal conformity evaluation index after the execution model optimization, the optimized computational complexity refers to the computational complexity after the execution model optimization, and the optimized convergence speed refers to the convergence speed after the execution model optimization.

[0086] The artifact optimization evaluation weight is a numerical value of the influence of the quantitative artifact removal optimization preset in the preset database on the optimization effect evaluation index. When used, the weight corresponding to the preset artifact optimization evaluation can be directly obtained from the preset database. The corresponding relationship can be a preset mapping relationship. For example, the artifact removal optimization and the weights corresponding to the preset artifact optimization evaluation in the preset database form a mapping set, and the real-time artifact removal optimization is input into the mapping set to obtain the corresponding weight. The mapping relationship can be one-to-one or many-to-one. In this example, the value range is [0, 1].

[0087] The convergence speed optimization evaluation weight is a numerical value of the influence of the quantitative convergence speed optimization situation preset in the preset database on the optimization effect evaluation index. When used, the weight corresponding to the preset convergence speed optimization evaluation can be directly obtained from the preset database, and the corresponding relationship can be a pre-set mapping relationship. For example, the convergence speed optimization situation and the weight corresponding to the convergence speed optimization evaluation preset in the preset database form a mapping set, and the real-time convergence speed optimization situation is input into the mapping set to obtain the corresponding weight. The mapping relationship can be a one-to-one correspondence or a many-to-one relationship. In the embodiment of the present invention, the value range is [0, 1].

[0088] It should be noted that the sum of the artifact optimization evaluation weight, the convergence speed optimization evaluation weight, and the computational complexity optimization evaluation weight is 1, where the computational complexity optimization evaluation weight is used to quantify the influence of computational complexity optimization on the optimization effect evaluation index.

[0089] The optimization effect evaluation index is used to quantitatively evaluate the effect of model optimization. Specifically, the performance optimization evaluation index includes parameters from multiple aspects, and each parameter is related to each other and does not exist independently. For example, as the optimization artifact removal meets the evaluation index, more complex numerical methods and iteration times are usually required, which leads to an increase in the complexity of the optimization calculation, and then leads to an increase in the change in the optimization convergence speed. Therefore, the embodiment of the present invention considers the correlation and mutual influence between various factors in a quantitative manner, and obtains the optimization effect evaluation index through comprehensive analysis, thereby realizing the numerical evaluation of the model optimization effect, and judging the model optimization effect through numerical evaluation, thereby realizing the improvement of the accuracy of the model optimization effect evaluation.

[0090] Furthermore, the specific steps for determining whether to re-acquire the simulation data set based on the image processing quality evaluation index and the artifact removal compliance evaluation index are as follows: M1, determining whether the acquired image processing quality evaluation index is within the preset processing quality evaluation threshold range. If the image processing quality evaluation index is within the preset processing quality evaluation threshold range, M2 is executed; otherwise, the simulation data set is not re-acquired; M2, determining whether the acquired artifact removal compliance evaluation index is within the preset artifact compliance threshold range. If the artifact removal compliance evaluation index exceeds the preset artifact compliance threshold range, M3 is executed; otherwise, the simulation data set is not re-acquired; M3, determining whether the acquired optimization effect evaluation index is within the preset optimization effect threshold range. If the optimization effect evaluation index is within the preset optimization effect threshold range, the simulation data set is not re-acquired; otherwise, the simulation data set is re-acquired.

[0091] In this embodiment, the preset optimization effect threshold range is obtained from a preset database. Specifically, the preset optimization effect threshold range is set by professionals according to standards in the field. For example, the preset optimization effect threshold range is set to 0.5 to 1. By making a judgment based on the preset optimization effect threshold range, a more accurate judgment of the model optimization effect evaluation is achieved.

[0092] To summarize, the embodiment of the present invention evaluates the effect of data set preprocessing to determine whether to execute the multi-level dual-domain progressive denoising network model construction, then evaluates the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to determine whether to execute model optimization, and finally determines whether to re-acquire the simulation data set, thereby realizing the evaluation and optimization of the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts, thereby achieving more accurate removal of metal artifacts in CT images, and effectively solving the problem of low accuracy of metal artifact removal in CT images based on dual-domain progressive enhancement network in the prior art.

[0093] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0095] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0097] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0098] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network, characterized in that: The following steps are involved: S1, performing data set preprocessing on the acquired simulation data set to obtain a CT image modeling data set; the simulation data set includes a CT image data set and a metal mask data set; S2, evaluating the CT image modeling data set to obtain an image processing quality evaluation index, and judging whether to execute the multi-level dual-domain progressive denoising network model construction based on the image processing quality evaluation index; the image processing quality evaluation index is used to quantify the effect of data set preprocessing; S3, after executing the multi-level dual-domain progressive denoising network model construction, evaluating the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to obtain an artifact removal conformity evaluation index, and judging whether to perform model optimization based on the artifact removal conformity evaluation index, wherein the artifact removal conformity evaluation index is used to quantitatively evaluate the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts; S4, obtaining an optimization effect evaluation index after the execution model optimization, and judging whether to execute re-acquisition of the simulation data set according to the image processing quality evaluation index and the artifact removal compliance evaluation index, wherein the optimization effect evaluation index is used to quantify the effect of the evaluation model optimization; The specific steps for evaluating the CT image modeling dataset to obtain the image processing quality assessment index are as follows: Acquire initial evaluation related data before performing data set preprocessing, wherein the initial evaluation related data includes the number of missing values ​​of the initial data, the mean of the initial data set, the signal mean of the initial data set, and the noise variance of the initial data set; Obtaining processing evaluation related data after executing data set preprocessing, wherein the processing evaluation related data includes the number of missing values ​​of the processing data, the data set related evaluation value, the data set Euclidean distance, the processing data set signal mean, the processing data set noise variance, and the data set processing time; Obtaining a data set quality assessment value and a preprocessing efficiency assessment value according to the initial assessment related data, the processing assessment related data, and the reference processing data obtained from a preset database, wherein the reference processing data includes a processing missing reduction rate reference value, a processing noise reference value, and a data set processing reference time; The image processing quality evaluation index is obtained by combining the data set quality evaluation value, the preprocessing efficiency evaluation value, the preprocessing influence factor, the data set quality evaluation weight, and the preprocessing efficiency evaluation weight obtained from the preset database; The method to obtain the image processing quality evaluation index is as follows: ； ； ； In the formula, IQ represents the image processing quality evaluation index, τ pf represents the preprocessing impact factor, ε1 represents the data set quality assessment weight, DQ represents the data set quality assessment value, ε2 represents the preprocessing efficiency assessment weight, PQ represents the preprocessing efficiency assessment value, Q c represents the number of missing values ​​in the initial data, Q y represents the number of missing values ​​in the processed data, ΔQ represents the reference value of the missing value reduction rate, μ′ represents the relevant evaluation value of the data set, and μ c represents the mean of the initial data set, D(c,y) represents the Euclidean distance of the data set, μ x,y Represents the mean value of the processed data set signal, μ x,c represents the initial data set signal mean, σ n,y represents the noise variance of the processed data set, σ n,c represents the noise variance of the initial data set, ΔS represents the reference value of the processed noise, and T y represents the dataset processing time, T′ represents the dataset processing reference time; The specific steps of constructing the multi-level dual-domain progressive denoising network model are as follows: A1, constructing an attention embedding sinusoidal domain restoration network, wherein the attention embedding sinusoidal domain restoration network includes an embedding input layer, an embedding hidden layer and an embedding output layer; A2, constructing an image domain restoration network, wherein the image domain restoration network includes a restoration input layer, a restoration hidden layer, and a restoration output layer; A3, constructing an image detail enhancement network, wherein the image detail enhancement network includes a parallel expansion fusion block and a pixel-level attention block; The specific steps of evaluating the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts to obtain the artifact removal conformity evaluation index are as follows: Obtain artifact removal related evaluation values ​​after training the multi-level dual-domain progressive denoising network model, wherein the artifact removal related evaluation values ​​include artifact peak signal-to-noise ratio, initial peak signal-to-noise ratio, removed edge similarity, removed texture similarity, and conformed structure similarity; The artifact removal related evaluation value is combined with the removal reference data obtained from a preset database to obtain an artifact removal compliance evaluation index, wherein the removal reference data includes a similarity reference value and a removal evaluation weight, wherein the similarity reference value includes a reference edge similarity, a reference texture similarity, and a structure similarity reference value, and the removal evaluation weight includes an artifact removal compliance evaluation weight, a detail compliance evaluation weight, and a structure similarity evaluation weight.

2. The method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network according to claim 1, characterized in that: The specific steps to obtain the CT image modeling data set are as follows: Performing data set cleaning on the simulation data set to obtain a first CT image modeling data set; Performing data set standardization and normalization processing on the first CT image modeling data set to obtain a second CT image modeling data set; The second CT image modeling data set is enhanced and segmented to obtain a CT image modeling data set.

3. The method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network as claimed in claim 1, characterized in that: The specific steps for determining whether to perform the construction of a multi-level dual-domain progressive denoising network model based on the image processing quality evaluation index are as follows: comparing the image processing quality assessment index with a preset processing quality assessment threshold range obtained from a preset database; If the image processing quality assessment index is within the preset processing quality assessment threshold range, the multi-level dual-domain progressive denoising network model is constructed, and the conformity of the multi-level dual-domain progressive denoising network model in removing artifacts is evaluated to obtain an artifact removal conformity assessment index; If the image processing quality assessment index exceeds the preset processing quality assessment threshold range, the multi-level dual-domain progressive denoising network model construction will not be executed, and the preset personnel will be reminded to re-preprocess the data set.

4. The method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network as claimed in claim 1, characterized in that: The specific steps for determining whether to perform model optimization based on the artifact removal compliance evaluation index are as follows: comparing the artifact removal compliance evaluation index with a preset artifact compliance threshold range obtained from a preset database; If the artifact removal compliance evaluation index is within the preset artifact compliance threshold range, model optimization is not performed; If the artifact removal compliance evaluation index exceeds the preset performance compliance threshold range, the model optimization is performed, and the optimization effect evaluation index after the execution of the model optimization is obtained.

5. The method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network as claimed in claim 4, characterized in that: The specific steps to obtain the optimization effect evaluation index after executing model optimization are as follows: Obtaining initial optimization effect evaluation related values ​​before executing model optimization, wherein the initial optimization effect evaluation related values ​​include an initial artifact removal compliance evaluation index, an initial computational complexity, and an initial convergence speed; Obtaining optimization evaluation related values ​​after the execution model optimization, wherein the optimization evaluation related values ​​include optimization artifact removal compliance evaluation index, optimization calculation complexity, and optimization convergence speed; The initial convergence speed and the optimized convergence speed are subtracted to obtain the optimized convergence speed change. The optimization effect evaluation index is obtained according to the initial optimization effect evaluation related value, the optimization evaluation related value, the optimized convergence speed change, the complexity reference change obtained from the preset database and the optimization evaluation weight. The optimization evaluation weight includes the artifact optimization evaluation weight, the convergence speed optimization evaluation weight and the computational complexity optimization evaluation weight.

6. The method for removing metal artifacts from CT images based on a dual-domain progressive enhancement network as claimed in claim 1, characterized in that: The specific steps for determining whether to reacquire the simulation data set based on the image processing quality evaluation index and the artifact removal compliance evaluation index are as follows: M1, judging whether the acquired image processing quality assessment index is within the preset processing quality assessment threshold range, if the image processing quality assessment index is within the preset processing quality assessment threshold range, executing M2, otherwise, not executing, re-acquiring the simulation data set; M2, judging whether the acquired artifact removal compliance evaluation index is within a preset artifact compliance threshold range, if the artifact removal compliance evaluation index exceeds the preset artifact compliance threshold range, executing M3, otherwise, not executing and re-obtaining the simulation data set; M3, judging whether the obtained optimization effect evaluation index is within the preset optimization effect threshold range, if the optimization effect evaluation index is within the preset optimization effect threshold range, then the simulation data set is not re-acquired, otherwise, the simulation data set is re-acquired.

Images