A neural network-based metal artifact reduction method and apparatus

Abstract

The application relates to a metal artifact suppression method and device based on a neural network, and belongs to the technical field of CT imaging, and solves the technical problem that the existing metal artifact removal is not applicable to security CT. The method comprises the following steps: acquiring a CT image comprising a detection target; determining an equivalent metal atomic number of a metal region according to the CT image, wherein the equivalent metal atomic number is used to represent the average atomic number of each pixel in the metal region in the CT image; determining a neural network model to be used according to the equivalent metal atomic number, wherein the neural network model to be used takes an image containing artifacts of the equivalent metal atomic number as input and takes an image not containing artifacts of the equivalent metal atomic number as output; and removing metal artifacts in the CT image through the neural network model to be used. The technical scheme provided in the application can select different processing methods for artifacts formed by different metal materials, so that the precision of metal artifact removal is improved.

Description

Technical Field

[0001] This application relates to the field of CT imaging technology, and in particular to a method and apparatus for suppressing metal artifacts based on neural networks. Background Technology

[0002] When using CT scanners, metallic objects are frequently encountered. The reconstructed CT images show numerous dark bands and bright radial streaks around these objects; these are known as metal artifacts. These artifacts severely degrade image quality, making it extremely difficult for inspectors to interpret the results.

[0003] To address the problem of metal artifacts, technicians typically reduce the pitch to minimize their impact, or eliminate them through image processing.

[0004] However, in the field of security inspection technology, in order to ensure the efficiency of item detection, CT must use a large pitch, which leads to a decrease in the clarity of CT images. Obviously, the above methods are not suitable for the security inspection field. Therefore, there is an urgent need to develop a metal artifact suppression method and device based on neural networks to achieve efficient metal identification. Summary of the Invention

[0005] In view of the above analysis, this application aims to provide a method and apparatus for suppressing metal artifacts based on neural networks, so as to solve at least one of the above problems.

[0006] In a first aspect, embodiments of this application propose a method for suppressing metal artifacts based on a neural network, comprising:

[0007] Acquire a CT image including a target to be detected; wherein the target to be detected is metal, and the CT image contains a metal region and metal artifacts;

[0008] Based on the CT image, the equivalent metal atomic number of the metal region is determined, and the equivalent metal atomic number is used to characterize the average atomic number of each pixel in the metal region of the CT image;

[0009] Based on the equivalent metal atomic number, a neural network model to be used is determined. The neural network model to be used takes an image containing artifacts of the equivalent metal atomic number as input and an image not containing artifacts of the equivalent metal atomic number as output.

[0010] The neural network model to be used is used to remove metal artifacts from the CT image.

[0011] Furthermore, the CT image is a slice image; determining the equivalent metal atomic number based on the CT image includes:

[0012] Calculate the atomic number of each pixel in the CT image;

[0013] The metallic region in the CT image is determined based on the atomic number;

[0014] The average atomic number of each pixel in each of the aforementioned metal regions is determined as the equivalent metal atomic number.

[0015] Furthermore, multiple neural network models are pre-set, including a first neural network model, a second neural network model, and a general neural network model;

[0016] Each of the first neural network models corresponds to a preset model atomic number, each of the second neural network models corresponds to a preset model atomic number range, and the general neural network model corresponds to the atomic numbers of all metals;

[0017] The atomic number of the model is the weighted average of the atomic numbers of each component in the metallic material, with the weight being the percentage content of each component; the range of the atomic number of the model is the set of weighted averages of the atomic numbers of each component in the metallic material.

[0018] Further, determining the neural network model to be used based on the equivalent metal atomic number includes:

[0019] Calculate the difference between the atomic number of each model and the atomic number of the equivalent metal;

[0020] Determine whether the minimum difference is within the preset range;

[0021] When the minimum difference is within the preset value range, the first neural network model corresponding to the minimum difference is determined as the neural network model to be used.

[0022] Furthermore, determining the neural network model to be used based on the equivalent metal atomic number further includes:

[0023] When it is determined that the minimum difference is not within the preset value range, it is determined whether the equivalent metal atomic number falls within the range of the model atomic number.

[0024] Based on the range of model atomic numbers where the equivalent metal atomic number is located, the corresponding second neural network model is determined as the neural network model to be used.

[0025] Furthermore, when the equivalent metal atomic number does not fall within the range of any of the model atomic numbers, the general neural network model is determined as the neural network model to be used.

[0026] Furthermore, the training method for the first neural network model includes:

[0027] Based on the preset model atomic number, obtain an artifact-free CT slice image, wherein the CT slice image includes at least one metallic region corresponding to the preset model atomic number:

[0028] Metal artifacts are added to the CT slice image based on the preset model atomic number;

[0029] The CT slice images without artifacts and the CT slice images with added metal artifacts are used as training samples for the corresponding neural network models.

[0030] Furthermore, the training method for the second neural network model includes:

[0031] Based on the preset model atomic number range, obtain an artifact-free CT slice image, wherein the CT slice image includes at least one metallic region corresponding to the preset model atomic number range:

[0032] Metal artifacts are added to the CT slice images according to the preset model atomic number range;

[0033] The CT slice images without artifacts and the CT slice images with added metal artifacts are used as training samples for the corresponding neural network models.

[0034] Furthermore, the training method for the general neural network model includes:

[0035] Take a CT slice image without artifacts. The CT slice image includes multiple overlapping or superimposed metal regions corresponding to the detection targets. The atomic numbers of the metals of the detection targets do not have a fixed correlation.

[0036] Metal artifacts are added to the CT slice images based on the atomic number of the metals of each of the detected targets.

[0037] The CT slice images without artifacts and the CT slice images with added metal artifacts are used as training samples for the corresponding neural network models.

[0038] Furthermore, the CT image includes multiple metallic regions; correspondingly, based on the equivalent metal atomic number, the neural network model to be used is determined, including:

[0039] Based on the multiple metal regions, the CT image is divided into multiple sub-images, and each sub-image corresponds one-to-one with a metal region;

[0040] Based on the equivalent metal atomic number of each metal region, the neural network model to be used for each sub-image is determined.

[0041] Furthermore, after removing the metal artifacts from the CT image using the neural network model to be used, the method further includes:

[0042] The sub-images with metal artifacts removed are stitched together to obtain a CT image with metal artifacts removed.

[0043] In a second aspect, embodiments of this application provide a metal artifact suppression device based on a neural network, comprising: an acquisition module, an image processing module, a data processing module, and an artifact removal module;

[0044] The acquisition module is used to acquire a CT image including the target to be detected, wherein the target to be detected is metal and the CT image contains a metal region;

[0045] The image processing module is used to determine the equivalent metal atomic number of the metal region based on the CT image. The equivalent metal atomic number is used to characterize the average atomic number of each pixel in the metal region of the CT image.

[0046] The data processing module is used to determine the neural network model to be used based on the equivalent metal atomic number. The neural network model to be used takes an image containing the artifact of the equivalent metal atomic number as input and an image not containing the artifact of the equivalent metal atomic number as output.

[0047] The artifact removal module is used to remove metal artifacts from the CT image using the neural network model to be used.

[0048] Based on the above technical solution, the beneficial effects of this application are as follows:

[0049] 1. For cases where the detection targets are mostly alloys or elemental metals, the equivalent atomic number of the detection target is used to represent the metal type of the detection target, so as to quantify the complex chemical composition; according to the metal type of the detection target, the corresponding neural network model is selected to select different processing methods for artifacts formed by different metal materials, thereby improving the accuracy of metal artifact removal.

[0050] 2. Based on the components in the alloy, generate model atomic numbers or model atomic number ranges to distinguish each alloy. Then, set up more neural network models based on the model atomic numbers or model atomic number ranges to more comprehensively consider the influence of each component in the alloy on metal artifacts, thereby improving the accuracy of metal artifact removal.

[0051] 3. Based on the atomic number of the model, and using the artifact addition algorithm, a sufficient number of training samples are generated to provide technical support for the removal of metal artifacts by multiple neural networks.

[0052] 4. Sequentially determine whether the equivalent metal atomic number corresponds to a fixed value (model atomic number) and a fixed value range (model atomic number range) to quickly find the neural network model corresponding to the equivalent metal atomic number.

[0053] 5. The technical solution provided in this application can improve the accuracy of metal artifact removal without changing the pitch, so it is more suitable for security CT scenarios compared with existing technologies.

[0054] Other features and advantages of this application will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained from the description and the accompanying drawings. Attached Figure Description

[0055] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Throughout the drawings, the same reference numerals denote the same parts.

[0056] Figure 1 A schematic diagram of a method for suppressing metal artifacts based on a neural network, provided in an embodiment of this application;

[0057] Figure 2 This is a schematic diagram of a neural network-based metal artifact suppression device provided in an embodiment of this application. Detailed Implementation

[0058] The preferred embodiments of this application are described in detail below with reference to the accompanying drawings, which constitute a part of this application and are used together with the embodiments of this application to illustrate the principles of this application, but are not intended to limit the scope of this application.

[0059] CT scans can identify the atomic number and electron cloud density of an object, both of which can uniquely identify the type of matter. Therefore, CT scans have a significant advantage in dealing with the diverse material forms of explosives and cutting tools.

[0060] However, CT scans also have their limitations. Specifically, to ensure security screening efficiency, large-pitch CT scanners must be used, resulting in low-resolution CT images that hinder subsequent removal of metal artifacts. Although some researchers have used image processing and neural network technologies to further process security CT images, existing technologies and neural networks do not consider the impact of metal type on artifact area, applying a uniform algorithm to different metals, resulting in a generic neural network model that ultimately leads to unsatisfactory image processing results. Its PSNR (Peak Signal-to-Noise Ratio) typically does not exceed 30dB.

[0061] Specifically, firstly, different metals have different artifact areas. For image processing, the difficulty of accurately determining the artifact boundaries varies depending on the artifact area; generally, the larger the artifact area, the easier it is to determine its boundaries. Secondly, when the metal region of the target object attenuates X-rays more than other regions, metal artifacts appear at the boundaries between the metal region and other regions. Different metals attenuate X-rays to varying degrees, resulting in slightly different image characteristics of the artifacts, such as artifact brightness and the smoothness of artifact boundaries. These factors all have varying degrees of impact on subsequent image processing results. Finally, the materials of metal parts in bags and knives are usually alloys or treated with gold plating. Therefore, in the above cases, metal artifacts are determined by multiple elements. However, the general model only removes artifacts from an image processing perspective and does not distinguish the influence of different substances on metal artifacts from a chemical perspective.

[0062] To address the aforementioned technical problems, this application provides a method for suppressing metal artifacts based on neural networks, comprising the following steps:

[0063] Step 1: Acquire a CT image including the target to be detected.

[0064] In this embodiment, the CT image is a CT slice image, which includes metallic regions and metallic artifacts. Specifically, acquiring the CT sinusoidal image of the target to be detected includes reconstructing the CT sinusoidal image to obtain the CT slice image. The target to be detected as a metallic substance includes: elemental metals and alloys.

[0065] Step 2: Determine the equivalent metal atomic number of the metal region based on the CT image.

[0066] In this embodiment, the metal equivalent atomic number is used to characterize the average atomic number of each pixel in the metal region of a CT image. Specifically, the atomic number of each pixel in the CT image is calculated, and the metal region in the CT image is determined based on the atomic number. Then, the average atomic number of each pixel in each metal region is determined as the equivalent metal atomic number.

[0067] During security checks, the metallic areas in CT images typically fall into one of the following three categories:

[0068] In the first scenario, the metal regions correspond one-to-one with the detection target. In this case, the metal artifacts are single artifacts, and they do not overlap or superimpose. The equivalent metal atomic number is a fixed value or exists within a fixed range of variation in this situation.

[0069] In the second scenario, multiple items are placed close together, such as two mobile phones with metal casings stacked on top of each other. In this case, the two metal casings become the detection targets, and their artifacts overlap or merge. Since the atomic number of the detected targets is random during security checks, the equivalent metal atomic number has no fixed value or range of variation in this situation.

[0070] The third scenario involves metal products with metallic coatings, where the coating and the product itself are made of different metals. In this case, the artifacts of the coating and the product overlap. The equivalent metal atomic number is either a fixed value or exists within a fixed range of variation. This is because, from a manufacturing perspective, the materials of the metal product and its coating are usually a fixed combination, thus the equivalent metal atomic number is either fixed or exists within a fixed range of variation. Therefore, in this case, for a specific item, the equivalent metal atomic number is either a fixed value or exists within a fixed range of variation.

[0071] Therefore, there are three cases for the equivalent metal atomic number: it is a fixed value, within a certain fixed range of variation, or it is neither a fixed value nor has a fixed range of values.

[0072] Step 3: Determine the neural network model to be used based on the equivalent metal atomic number.

[0073] In this embodiment, the neural network model to be used takes an image containing artifacts of equivalent metal atomic numbers as input and an image without artifacts of equivalent metal atomic numbers as output. X-rays attenuate after passing through an object, and the degree of attenuation is determined by the atomic number of the material. Therefore, the neural network model is classified and trained according to the atomic number, and then, based on the equivalent atomic number of the target metal region, a neural network model with a matching atomic number is selected for image processing.

[0074] Specifically, for the three scenarios in step 2, multiple neural network models are pre-set, including a first neural network model, a second neural network model, and a general neural network model; each first neural network model corresponds to a preset model atomic number, each second neural network model corresponds to a preset model atomic number range, and the general neural network model corresponds to the atomic numbers of all metals.

[0075] The model atomic number is the weighted average of the atomic numbers of all components in a metallic material, with the weights representing the percentage content of each component. The range of model atomic numbers is the set of weighted averages of the atomic numbers of all components in a metallic material. For example, in the field of alloys, the components of an alloy typically vary within a range. For instance, a high-strength ferritic steel might have the following partial chemical composition: Ni 2–15%, Mn 2–10%, Al 1–6%, Cu 1.5–4%, Cr 8–12%. In this case, the equivalent metallic atomic number has a fixed range. Alternatively, if the partial chemical composition is: Ni 2%, Mn 10%, Al 6%, Cu 4%, Cr 12%, the equivalent metallic atomic number has a fixed value.

[0076] The specific training methods for each neural network model are as follows:

[0077] First neural network model training method:

[0078] Based on the preset model atomic number, obtain an artifact-free CT slice image. The CT slice image includes at least one metal region corresponding to the preset model atomic number.

[0079] Add metal artifacts to CT slice images based on the preset atomic number of the model;

[0080] CT slice images without artifacts and CT slice images with added metal artifacts were used as training samples for the corresponding neural network models.

[0081] Training method for the second neural network model:

[0082] Based on the preset atomic number range of the model, obtain CT slice images without artifacts. The CT slice images include at least one metal region corresponding to the preset atomic number range of the model.

[0083] Metal artifacts are added to CT slice images based on the preset atomic number range;

[0084] CT slice images without artifacts and CT slice images with added metal artifacts were used as training samples for the corresponding neural network models.

[0085] Training methods for general neural network models include:

[0086] Take CT slice images without artifacts. These CT slice images include multiple overlapping or superimposed metal regions corresponding to the detection targets. The atomic numbers of the metals in each detection target do not have a fixed correlation.

[0087] Metal artifacts are added to CT slice images based on the atomic number of the metals of each detected target.

[0088] CT slice images without artifacts and CT slice images with added metal artifacts were used as training samples for the corresponding neural network models.

[0089] In practical application, first determine whether the equivalent metal atomic number corresponds to the first neural network model, and then determine whether the equivalent metal atomic number corresponds to the second neural network model. If the equivalent metal atomic number corresponds to neither the first nor the second neural network model, then the equivalent metal atomic number corresponds to the general neural network model. Specifically, calculate the difference between the atomic number of each model and the equivalent metal atomic number; determine whether the minimum difference is within a preset range; if the minimum difference is within the preset range, then the first neural network model corresponding to the minimum difference is determined as the neural network model to be used.

[0090] When the minimum difference is determined to be outside the preset range, it is determined whether the equivalent metal atomic number falls within the range of the model atomic number. Based on the range of the model atomic number in which the equivalent metal atomic number is located, the corresponding second neural network model is determined as the neural network model to be used.

[0091] When the equivalent metal atomic number does not fall within the range of atomic numbers of any model, a general neural network model is determined as the neural network model to be used.

[0092] Metallic materials are diverse, and it is easy to obtain a large number of training samples using a general model. However, it is difficult to obtain sufficient training samples for a single metallic material. Furthermore, this application requires setting up multiple neural network models, which significantly increases the training sample shortage and hinders the training of different neural network models based on the material type. To address these technical problems, in this embodiment, CT slice images without artifacts are obtained based on a preset model atomic number or a preset model atomic number range. These CT slice images include at least one metallic region corresponding to a preset model atomic number range or a preset model atomic number. Metal artifacts are added to the CT slice images based on the preset model atomic number. The CT slice images without artifacts and those with added artifacts are used as training samples for the corresponding neural network models. For example, after obtaining CT slice images without artifacts, a metal artifact addition algorithm in MATLAB is used to add metal artifacts to different regions of the CT slice images, thereby obtaining sufficient training samples. In this way, a large number of training samples for different metal types can be generated.

[0093] Specifically, this application employs an autoencoder, taking a CT slice image with added metal artifacts as input and a CT slice image with removed metal artifacts as output, and using the difference between the CT slice image with removed metal artifacts and the CT slice image without artifacts to adjust the parameters of the autoencoder.

[0094] Step 4: Remove metal artifacts from CT images using the neural network model to be used.

[0095] In this embodiment, when multiple metal regions exist in a CT image, the CT image is first divided into multiple sub-images, with each sub-image corresponding to a metal region. Metal artifacts are removed from each sub-image using a corresponding neural network model. After removing the metal artifacts from each sub-image, the sub-images output by each neural network model are then stitched together to obtain a CT image with metal artifacts removed.

[0096] To illustrate the feasibility of the above embodiments of this application, three neural network models T1, T2, and T3 trained according to the atomic number classification method of this application were used as the experimental group, and a neural network model trained according to the general mode was used as the control group. Artifact removal processing was performed on the CT images of copper-iron-nickel-chromium and alloys (T1), titanium and titanium alloys (T2), and platinum (T3), respectively.

[0097] Specifically, 10,000 slice images without metal artifacts were used, and then metal artifacts were added to them to obtain the training samples for the experimental group. These included 30,000 images for class T1, 30,000 images for class T2, and 30,000 images for class T3. The classification training method described above was used to train the models A1, A2, and A3. Simultaneously, 90,000 images with generated metal artifacts were used as input, and images without metal artifacts were used as output for training, resulting in the general model B.

[0098] It should be noted that the training algorithms for both models are the same:

[0099] The autoencoder is trained by using the portion of the training set containing artifact-laden images as input and the portion containing the data without artifacts as output.

[0100] An autoencoder consists of an encoder and a decoder, both of which are trained using a convolutional neural network algorithm.

[0101] A. The encoder consists of an input layer, a convolutional layer, a pooling layer, and an output layer:

[0102] The input data for the input layer is two-dimensional projection data containing artifacts of a specific metal type;

[0103] The convolutional layer consists of 4 layers with a kernel size of 3×3 and a stride of 2, which are used to extract features from the input data from the input layer.

[0104] The pooling layer uses average pooling with a step size of 2;

[0105] The output layer is the result of extracting features from the convolutional and pooling layers, and provides input for the next step of the decoder;

[0106] B. The decoder consists of an input layer, an unpooling layer, a deconvolution layer, and an output layer:

[0107] The input data for the input layer is the output of the encoder described above;

[0108] The anti-pooling layer is an anti-average pooling layer with a step size of 2;

[0109] The deconvolutional layer consists of 4 layers of deconvolutional layers with a kernel size of 3×3 and a stride of 2.

[0110] The output layer is the result G'1(a11,a12,a13...)...G'N(a11,a12,a13...) after restoring the features of the unpooling layer and the deconvolution layer;

[0111] Loss function used in training:

[0112] Where G'i is the image vector reconstructed after the encoder and de-encoder, Gi is the artifact-free image sample corresponding to the image data with artifacts, N is the number of samples, i represents the sample index, and loss is used to characterize the difference between the model output and the observation.

[0113] The results of the above processing are shown in Table 1:

[0114] Table 1 Comparison of artifact removal effects

[0115] Restore using a classification model 36.78125 35.26854 34.23817 Restore using a general model 23.12689 23.15799 22.95864

[0116] PSNR is the most common and widely used objective image quality evaluation metric. It is based on the error between corresponding pixels, i.e., it is an error-sensitive image quality evaluation. Generally speaking, a PSNR above 40dB indicates excellent image quality (i.e., very close to the original image), 30-40dB usually indicates good image quality (i.e., distortion is perceptible but acceptable), 20-30dB indicates poor image quality, and finally, a PSNR below 20dB indicates unacceptable image quality. As shown in Table 1, the image processing performance of this application is better than that of the general neural network model.

[0117] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0118] This application embodiment also provides a metal artifact suppression device based on a neural network, including: an acquisition module 101, an image processing module 102, a data processing module 103, and an artifact removal module 104;

[0119] The acquisition module 101 is used to acquire a CT image including the target to be detected, wherein the target to be detected is metal and the CT image contains a metal region.

[0120] The image processing module 102 is used to determine the equivalent metal atomic number of the metal region based on the CT image. The equivalent metal atomic number is used to characterize the average atomic number of each pixel in the metal region of the CT image.

[0121] The data processing module 103 is used to determine the neural network model to be used based on the equivalent metal atomic number. The neural network model to be used takes the image containing the artifact of the equivalent metal atomic number as input and the image without the artifact of the equivalent metal atomic number as output.

[0122] The artifact removal module 104 is used to remove metal artifacts from CT images using the neural network model to be used.

[0123] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for suppressing metal artifacts based on neural networks, characterized in that, include: Acquire a CT image including a target to be detected; wherein the target to be detected is metal, and the CT image contains a metal region and metal artifacts; Based on the CT image, the equivalent metal atomic number of the metal region is determined, and the equivalent metal atomic number is used to characterize the average atomic number of each pixel in the metal region of the CT image; Based on the equivalent metal atomic number, a neural network model to be used is determined. The neural network model to be used takes an image containing artifacts of the equivalent metal atomic number as input and an image not containing artifacts of the equivalent metal atomic number as output. The neural network model to be used is used to remove metal artifacts from the CT images; Multiple neural network models are pre-set, including a first neural network model, a second neural network model, and a general neural network model; Each of the first neural network models corresponds to a preset model atomic number, each of the second neural network models corresponds to a preset model atomic number range, and the general neural network model corresponds to the atomic numbers of all metals; Based on the equivalent metal atomic number, the neural network model to be used is determined, including: Calculate the difference between the atomic number of each model and the atomic number of the equivalent metal; Determine whether the minimum difference is within the preset range; When the minimum difference is within the preset value range, the first neural network model corresponding to the minimum difference is determined as the neural network model to be used; When it is determined that the minimum difference is not within the preset value range, it is determined whether the equivalent metal atomic number falls within the range of the model atomic number. Based on the range of model atomic numbers in which the equivalent metal atomic number is located, the corresponding second neural network model is determined as the neural network model to be used; When the equivalent metal atomic number does not fall within the range of atomic numbers of any of the models, the general neural network model is determined as the neural network model to be used.

2. The method according to claim 1, characterized in that, The CT image is a slice image; determining the equivalent metal atomic number based on the CT image includes: Calculate the atomic number of each pixel in the CT image; The metallic region in the CT image is determined based on the atomic number; The average atomic number of each pixel in each of the aforementioned metal regions is determined as the equivalent metal atomic number.

3. The method according to claim 1, characterized in that, The method further includes: The atomic number of the model is the weighted average of the atomic numbers of each component in the metallic material, with the weight being the percentage content of each component; the range of the atomic number of the model is the set of weighted averages of the atomic numbers of each component in the metallic material.

4. The method according to claim 1, characterized in that, The training method for the first neural network model includes: Based on the preset model atomic number, obtain an artifact-free CT slice image, wherein the CT slice image includes at least one metallic region corresponding to the preset model atomic number: Metal artifacts are added to the CT slice image based on the preset model atomic number; The CT slice images without artifacts and the CT slice images with added metal artifacts are used as training samples for the corresponding neural network models.

5. The method according to claim 1, characterized in that, The training methods for the second neural network model include: Based on the preset model atomic number range, obtain an artifact-free CT slice image, wherein the CT slice image includes at least one metallic region corresponding to the preset model atomic number range: Metal artifacts are added to the CT slice images according to the preset model atomic number range; The CT slice images without artifacts and the CT slice images with added metal artifacts are used as training samples for the corresponding neural network models.

6. The method according to claim 1, characterized in that, The training method for the general neural network model includes: Take a CT slice image without artifacts. The CT slice image includes multiple overlapping or superimposed metal regions corresponding to the detection targets. The atomic numbers of the metals of each detection target do not have a fixed correlation. Metal artifacts are added to the CT slice images based on the atomic number of the metals of each of the detected targets. The CT slice images without artifacts and the CT slice images with added metal artifacts are used as training samples for the corresponding neural network models.

7. A device for suppressing metal artifacts based on a neural network, characterized in that, include: The module includes an acquisition module, an image processing module, a data processing module, and an artifact removal module. The acquisition module is used to acquire a CT image including the target to be detected, wherein the target to be detected is metal and the CT image contains a metal region; The image processing module is used to determine the equivalent metal atomic number of the metal region based on the CT image. The equivalent metal atomic number is used to characterize the average atomic number of each pixel in the metal region of the CT image. The data processing module is used to determine the neural network model to be used based on the equivalent metal atomic number. The neural network model to be used takes an image containing the artifact of the equivalent metal atomic number as input and an image not containing the artifact of the equivalent metal atomic number as output. The artifact removal module is used to remove metal artifacts from the CT image using the neural network model to be used; Multiple neural network models are pre-set, including a first neural network model, a second neural network model, and a general neural network model; Each of the first neural network models corresponds to a preset model atomic number, each of the second neural network models corresponds to a preset model atomic number range, and the general neural network model corresponds to the atomic numbers of all metals; Based on the equivalent metal atomic number, the neural network model to be used is determined, including: Calculate the difference between the atomic number of each model and the atomic number of the equivalent metal; Determine whether the minimum difference is within the preset range; When the minimum difference is within the preset value range, the first neural network model corresponding to the minimum difference is determined as the neural network model to be used; When it is determined that the minimum difference is not within the preset value range, it is determined whether the equivalent metal atomic number falls within the range of the model atomic number. Based on the range of model atomic numbers in which the equivalent metal atomic number is located, the corresponding second neural network model is determined as the neural network model to be used; When the equivalent metal atomic number does not fall within the range of atomic numbers of any of the models, the general neural network model is determined as the neural network model to be used.

Images