A neural network-based beam hardening artifact correction method and system

Abstract

The application discloses a kind of based on neural network's beam hardening artifact correction method and system, and its basic idea is by with artifact image as network input, using neural network generates not containing artifact water-based graph, bone-based graph and metal-based graph, again combining energy spectrum information to the water-based graph, bone-based graph and metal-based graph not containing artifact are fused, obtain artifact correction image under multiple energy conditions.It is specific, training data is obtained by artificial construction, and the prediction neural network is trained in combination with multidimensional loss function, and the prediction effect of prediction neural network is relatively accurate;In the synthesis process of artifact correction image, after the predicted water-based graph and the predicted bone-based graph are fused and reconstructed in the corresponding chord diagram information, then directly fused with the predicted metal-based graph, can ensure that the artifact correction effect of the artifact correction image obtained finally is better, and has good accuracy and practicality in actual application.

Description

Technical Field

[0001] This invention relates to the field of medical imaging quality control, and more specifically, to a method and system for beam hardening artifact correction based on neural networks. Background Technology

[0002] Computed tomography (CT) is a commonly used medical imaging technique that provides detailed images of the human body's interior, aiding doctors in diagnosis and treatment. However, when the scanned object contains highly attenuating materials, such as teeth or bones, low-energy rays are easily absorbed, causing the rays to harden. Reconstruction algorithms generally assume that the measurement data is single-energy, so direct reconstruction will introduce errors, resulting in cupping artifacts or bone hardening artifacts. If the scanned object contains metal, such as metal implants or dental implants in the chest or abdomen, the absorption rate of metal to radiation is very high. As the rays pass through, they are almost entirely absorbed, making them undetectable by the detector on the opposite side. This results in abnormal projection data, causing discontinuous jumps in the radiation attenuation coefficient. After filtering, this weak continuity is further amplified, resulting in alternating bright and dark stripe artifacts in the reconstructed image, known as metal artifacts.

[0003] Metal artifact correction in CT images has always been a challenge. The metal artifact correction methods proposed by researchers at home and abroad can be roughly divided into three categories: projection restoration method, iterative reconstruction method, and deep learning method.

[0004] Projection restoration methods refer to correcting the projection data damaged by metal in CT projections using physical models or interpolation methods, and then using the corrected projections for filtered backprojection reconstruction to obtain an image with metal artifacts removed. However, the structural details restored by this method are still limited, especially when the metal shape is large or there are many metal objects. Projection restoration methods for correcting metal artifacts are widely used clinically due to their simplicity and effectiveness, but problems such as excessive smoothing of adjacent tissues and the generation of new artifacts still exist.

[0005] The basic idea of ​​iterative reconstruction is to obtain high-quality images by iteratively solving for the optimal solution of an optimization model. Iterative reconstruction can be divided into Algebraic Reconstruction Technique (ART) and Statistical Algebraic Reconstruction Technique (SART) depending on the method of constructing the optimization model. The metal artifact correction method in iterative reconstruction involves, on the one hand, using projection data of missing metal trajectories to iteratively reconstruct CT images, and on the other hand, adding constraint regularization terms to the CT images during the iterative reconstruction process. This method reconstructs high-quality images, but it requires multiple preprojection and reconstruction operations, resulting in a large computational load and long processing time, making it difficult to use in practical clinical applications.

[0006] In recent years, with the rapid development of deep learning, a large number of deep learning-based methods have been proposed. Supervised networks are trained using paired data, which has stable results but requires a large amount of labeled datasets. Unsupervised methods use autoencoders to learn the features and structure of data, which has a wide range of applications, but the results are random and uncertain. Dual-domain networks use projection and image as network inputs. This method obtains a large amount of information by fully fusing information from the projection domain and the image domain. It has good correction results and can preserve structural information. However, projection data is generally only available from manufacturers, and it requires a large amount of training data and computing resources.

[0007] Therefore, it is necessary to develop a more practical method and system for beam hardening artifact correction. Summary of the Invention

[0008] This invention provides a beam hardening artifact correction method based on a neural network. The basic idea is to use an image containing artifacts as network input, and use the neural network to generate artifact-free water-based images, bone-based images, and metal-based images. Then, the artifact-free water-based images, bone-based images, and metal-based images are fused with energy spectrum information to obtain an artifact-corrected image under multi-energy conditions. It has good artifact correction effect in practical applications.

[0009] This invention provides a neural network-based method for correcting beam hardening artifacts in actual images containing artifact information, including a data simulation process, a training process, and an application process.

[0010] The data simulation process is used to simulate the corresponding metal implant and the corresponding artifacts of the metal implant based on dual-energy CT images without metal artifacts.

[0011] The training process is used to train a predictive neural network based on several simulated images to obtain a trained predictive neural network. The input of the predictive neural network is an input image containing metal artifacts, and the output of the predictive neural network is a predicted water base map, a predicted bone base map, and a predicted metal base map corresponding to the input image.

[0012] The application process includes inputting the actual image into the trained predictive neural network, deriving a predicted water base map, a predicted bone base map, and a predicted metal base map corresponding to the actual image through the predictive neural network, and obtaining an artifact-corrected image based on the predicted water base map, the predicted bone base map, and the predicted metal base map.

[0013] In an optional implementation, the data simulation process includes:

[0014] S1: Acquire a dual-energy CT image free of metal artifacts;

[0015] S2: The dual-energy CT image is decomposed into a water-based image and a bone-based image, and a metal-based image is obtained based on the material of the metal implant and the metal mask corresponding to the metal implant.

[0016] S3: Take the non-metallic region image of the water-based image and perform forward projection to obtain the water-based chord diagram; take the non-metallic region image of the bone-based image and perform forward projection to obtain the bone-based chord diagram; and perform forward projection of the metal-based image to obtain the metal-based chord diagram.

[0017] S4: The water-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the first projection data, the bone-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the second projection data, the metal-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the third projection data, and the first projection data, the second projection data and the third projection data are accumulated to obtain the processing projection data under the corresponding energy.

[0018] S5: Combining the energy spectrum information, multi-energy projection data under multi-energy conditions is obtained based on the processed projection data, and a simulated image is reconstructed based on the multi-energy projection data. The simulated image has simulated metal artifacts.

[0019] In an optional implementation, the training process includes:

[0020] S7: Construct a predictive neural network, wherein the input of the predictive neural network is an input image containing metal artifacts, and the output of the predictive neural network is a predicted water base map, a predicted bone base map, and a predicted metal base map corresponding to the input image;

[0021] S8: Construct the loss function;

[0022] S9: Train the prediction neural network.

[0023] In an optional implementation, in step S7, the prediction neural network is a Unet structure with skip links, and the activation function is SimpleGate.

[0024] In an optional implementation, the loss function includes a first loss function based on the predicted water base map, a second loss function based on the predicted bone base map, a third loss function based on the predicted metal base map, a fourth loss function based on the first fused image obtained by fusing the predicted water base map and the predicted bone base map, and a fifth loss function based on the multi-energy projection image obtained by fusing the predicted water base map, the predicted bone base map, and the predicted metal base map.

[0025] In an optional implementation, the application process includes:

[0026] The actual image containing artifact information is input into the prediction neural network, and the corresponding predicted water-based image, predicted bone-based image, and predicted metal-based image are derived based on the prediction neural network. Combining the energy spectrum information, the artifact-corrected image is derived based on the predicted water-based image, predicted bone-based image, and predicted metal-based image.

[0027] In an optional implementation, deriving the artifact-corrected image based on the predicted water-based image, predicted bone-based image, and predicted metal-based image includes:

[0028] The predicted water-based image and the predicted bone-based image are converted into corresponding predicted water-based chord maps and predicted bone-based chord maps respectively based on methods such as Radon transform. The predicted water-based chord map and the predicted bone-based chord map are multiplied by the corresponding quality attenuation coefficient and then summed to obtain a fused chord map containing information of the predicted water-based image and the predicted bone-based image.

[0029] The intermediate predicted fused image is reconstructed based on the fused chord diagram;

[0030] The predicted metal-based image is multiplied by the corresponding mass attenuation coefficient to obtain an intermediate metal image.

[0031] The intermediate predicted fusion image and the intermediate metal image are merged to obtain the artifact-corrected image.

[0032] In an optional implementation, the X-ray tube voltages selected for the dual-energy CT images are 70kVp and 140kVp, respectively.

[0033] Accordingly, the present invention also provides a flowchart of a neural network-based beam hardening artifact correction method for use in the aforementioned neural network-based beam hardening artifact correction system.

[0034] In summary, this invention provides a neural network-based method for beam hardening artifact correction. The basic idea is to use an image containing artifacts as network input, and then utilize the neural network to generate artifact-free water-based, bone-based, and metal-based images. These artifact-free water-based, bone-based, and metal-based images are then fused with energy spectrum information to obtain an artifact-corrected image under multi-energy conditions. Specifically, training data is obtained through manual construction, and a multi-dimensional loss function is used to train the prediction neural network, resulting in relatively accurate predictions. During the synthesis of the artifact-corrected image, the predicted water-based and bone-based images are fused and reconstructed using their corresponding chordal information before being directly fused with the predicted metal-based image. This ensures a better artifact correction effect in the final image, demonstrating good accuracy and practicality in real-world applications. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the beam hardening artifact correction method based on neural networks according to an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of the energy spectrum information of an embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of a dual-energy CT image without metal artifacts according to an embodiment of the present invention.

[0039] Figure 4 This is a schematic diagram of a simulated image according to an embodiment of the present invention.

[0040] Figure 5 This is a schematic diagram of the structure of a neural network block according to an embodiment of the present invention.

[0041] Figure 6 This is a schematic diagram of the SimpleGate activation function structure according to an embodiment of the present invention.

[0042] Figure 7 This is a schematic diagram of artifact correction images according to an embodiment of the present invention.

[0043] Figure 8 This is a flowchart of the data processing in the neural network-based beam hardening artifact correction method according to an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Figure 1 This is a flowchart of the beam hardening artifact correction method based on neural networks according to an embodiment of the present invention.

[0046] This invention provides a neural network-based beam hardening artifact correction method, mainly used to correct beam hardening artifacts in the original image, including a data simulation process, a training process, and an application process.

[0047] Specifically, during a CT scan, when a stable DC high voltage is applied to the X-ray tube, it emits rays of different energies. For example, at 120 kVp, the energy range of the X-ray tube's output is 0-120 keV. Figure 2 The diagram shows energy spectrum information, where the horizontal axis represents the energy range and the vertical axis represents the number of X-ray photons.

[0048] Specifically, the simulation process mainly functions to generate a simulated image containing metal implants and metal artifacts based on a common dual-energy CT image without metal artifacts. The data simulation process includes:

[0049] S1: Acquire a dual-energy CT image free of metal artifacts;

[0050] For details, please refer to Figure 3 The dual-energy CT image shown is free of metal artifacts. In practice, dual-energy CT images free of metal artifacts can be obtained through relevant databases. However, it is difficult to obtain training data with metal artifacts that can be well applied to the embodiments of this invention in practical applications. Therefore, in order to ensure that the subsequent neural network training has a sufficient number of training samples and sufficiently accurate training samples, it is necessary to artificially construct simulated images containing metal implants and metal artifacts.

[0051] S2: The dual-energy CT image is decomposed into a water-based image and a bone-based image, and a metal-based image is obtained based on the material of the metal implant and the metal mask corresponding to the metal implant.

[0052] Specifically, dual-energy CT images are images reconstructed from dual-energy CT data obtained by X-ray tubes under two different tube voltages. In practical applications, different materials such as soft tissue, bone, and metal in human tissues can be identified through spectral information.

[0053] Specifically, in this embodiment of the invention, the tube voltages are 70kVp and 140kVp, respectively. For dual-energy CT images, the linear attenuation coefficient of the dual-energy CT image can be approximated as the weighted sum of the two matrix images. In this embodiment of the invention, the two matrix materials are soft tissue and compact bone, respectively, and the corresponding matrix images are called water-based image and bone-based image, respectively.

[0054] Specifically, the weighting formula for dual-energy CT images can be expressed as follows: In this diagram, subscripts H and L represent high-energy X-rays and low-energy X-rays, respectively; subscripts 1 and 2 represent two base materials, corresponding to the water-based diagram and the bone-based diagram, respectively; and μ H，i，j and μ L，i，j Let μ represent the linear decay coefficient of the material at image position (i, j) under high and low energies, respectively. kt Let μ represent the mass decay coefficient of the k-th base substance at energy t∈{1,2}. Specifically, μ kt The numerical value can be found in relevant literature (such as the website of the National Institute of Standards and Technology, USA).

[0055] Specifically, water-based images and bone-based images correspond to two types of objects in solid structures: soft tissue and dense bone, respectively. Correspondingly, for practical applications, metal-based images of metal implant components are also required.

[0056] Specifically, in clinical applications, metal implants include pacemakers, spinal implants, and other similar devices. Correspondingly, the actual materials used in metal implants vary depending on the type. Specifically, the metal matrix is ​​obtained by multiplying the metal density of the implant by the corresponding metal mask. A common material for metal implants is titanium, with a density of 4.3 g / cm³. 3 Metal masking refers to an image of the morphological structure of a metal implant, where the pixel value of the area where the metal implant is located is 1, and the pixel value of the remaining areas is 0.

[0057] S3: Take the non-metallic region image of the water-based image and perform forward projection to obtain the water-based chord diagram; take the non-metallic region image of the bone-based image and perform forward projection to obtain the bone-based chord diagram; and perform forward projection of the metal-based image to obtain the metal-based chord diagram.

[0058] Specifically, the metal mask reflects the pixels involved in the metal implant in the image. Therefore, the water-based image and the bone-based image can be obtained by logically inverting the relevant pixels according to the metal mask, resulting in the non-metallic region images of the water-based image and the bone-based image, respectively. By performing front projection on the two non-metallic region images, the corresponding water-based chord map and bone-based chord map can be obtained. For the metal-based image, the corresponding metal-based chord map can be obtained directly by performing front projection.

[0059] S4: The water-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the first projection data, the bone-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the second projection data, the metal-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the third projection data, and the first projection data, the second projection data and the third projection data are accumulated to obtain the processing projection data under the corresponding energy.

[0060] In the corresponding chord graph data, different basis chord graphs are multiplied by their corresponding mass decay coefficients and then accumulated. This allows multiple basis graphs to be integrated into one graph based on their different mass decay coefficients (which are essentially differences in material composition). Specifically, in this embodiment of the invention, the processed projection data obtained by accumulating the first projection data, the second projection data, and the third projection data includes water-based data, bone-based data, and metal-based data.

[0061] S5: Combining the energy spectrum information, multi-energy projection data under multi-energy conditions is obtained based on the processed projection data, and a simulated image is reconstructed based on the multi-energy projection data, wherein the simulated image has simulated metal artifacts;

[0062] Specifically, in this embodiment of the invention, for a given ray condition, the processing projection data is processed in conjunction with energy spectrum information to obtain corresponding multi-energy projection data.

[0063] Specifically, in this invention, under the radiation condition of 140 kVp, the number of incident photons at different energies from 0 to 140 keV can be calculated using the following statistical formula:

[0064] in, This represents chord diagram data at different energies (kev).

[0065] Specifically, multi-energy projection data can be converted into corresponding multi-energy projection images after passing through filtering back projection and other methods. In the process of generating multi-energy projection, the water-based map and bone-based map data include the most basic CT images of the target area. In the process of generating multi-energy projection through water-based map projection, a metal-based map is added. Since the quality attenuation coefficient of the metal-based map is much larger than that of the water-based map and bone-based map, this weak continuity is further amplified after the filtering operation in the filtering back projection. This results in the appearance of alternating bright and dark stripe artifacts in the reconstructed image, i.e., metal artifacts are produced.

[0066] refer to Figure 4 The schematic diagram of the simulated image shown can be used to simulate a simulated image containing a metal implant and metal artifacts based on a dual-energy CT image without metal artifacts, after processing steps S1 to S5.

[0067] S6: Repeat steps S1 to S5 to obtain a specified number of training data;

[0068] Specifically, according to subsequent requirements, each set of training data includes a simulated image containing metal artifacts and the corresponding simulated image chord graph, a dual-energy CT image without metal artifacts and the corresponding dual-energy CT chord graph, a water-based graph corresponding to the simulated image, a bone-based graph corresponding to the simulated image, and a metal-based graph corresponding to the simulated image. In subsequent applications of the neural network, the simulated image serves as the input image to the network, while the water-based graph, bone-based graph, and metal-based graph corresponding to the simulated image are all outputs of the neural network. The dual-energy CT image and the corresponding dual-energy CT chord graph can be used to calculate the loss function.

[0069] The training process includes:

[0070] S7: Construct a predictive neural network;

[0071] Figure 5 This is a schematic diagram of the predictive neural network structure according to an embodiment of the present invention. In this embodiment, referring to SimpleBaselines for Image Restoration, the framework of the predictive neural network is a Unet structure with skip connections (basic structure as follows). Figure 5 As shown, the specific operating mechanism is as follows: Figure 6 (Illustrative image) Specifically, the Unet structure with skip links enables the network to fuse the feature maps of the corresponding encoder positions on the channel during upsampling at each level. Through the fusion of low-level and high-level features, the network can retain more high-resolution detail information contained in the high-level feature maps, thereby improving image segmentation accuracy. Neighbor-layer feature reconstruction and cross-layer feature reconstruction enhance the transfer of feature information between layers while further utilizing the rich detail information in high-level convolutional feature layers, thus maximizing the utilization rate of feature information in each layer of the network and ensuring the accuracy of the image information content after segmentation. It should be noted that the attached diagram... Figure 5 In this context, "Block" refers to a neural network block within the predictive neural network.

[0072] Furthermore, SimpleGate was used to replace the activation function in the prediction neural network. SimpleGate splits the feature map into two parts in the channel direction and then multiplies them, as shown in the attached figure. Figure 6As shown, a simplified channel attention mechanism is used to aggregate global information and achieve channel information interaction. Among them, LayerNorm is a method of normalization processing on the feature dimension, Conv 1x1 is a convolution operation with a 1x1 kernel, DConv 3x3 is a deformable convolution operation with a 3x3 kernel, and SCA refers to Spatial Attention and Channel-wise Attention (an attention mechanism).

[0073] Specifically, the predictive neural network imports an actual image containing metal artifacts, derives a water-based image, a bone-based image, and a metal-based image corresponding to the actual image, first synthesizes an intermediate predictive fusion image without metal implants using the water-based image and the bone-based image, and then fuses the intermediate predictive fusion image with the metal-based image to obtain the final predictive image after artifact correction.

[0074] Correspondingly, the same logic is used to import data for training during the training phase. Specifically, in each set of provided training data, the simulated image serves as the network input, while the water-based map, bone-based map, and metal-based map corresponding to the simulated image serve as the neural network output. The dual-energy CT image and the corresponding dual-energy CT chord map are used for calculating the loss function.

[0075] S8: Construct the loss function;

[0076] Specifically, the direct objective of this predictive neural network is to generate three base maps free of artifacts from images containing artifacts. Correspondingly, the loss functions include:

[0077]

[0078]

[0079]

[0080] Where I w I b and I m These are the water-based map, bone-based map, and metal-based map predicted by the predictive neural network. and These are the reference images (corresponding data in the training data) for the three types of base images.

[0081] Furthermore, an indirect objective of this predictive neural network is to synthesize the final predicted image using three types of base maps. Correspondingly, the training data also includes dual-energy CT images without metal implants. Therefore, by combining spectral information to synthesize multi-energy conditions, the water-based and bone-based base maps from the training data can be used to synthesize the first fused image I without metal implants. com Correspondingly, I can be used w I b Synthesize the first predictive fusion image without metal implants. Through the first fused image I com Fusion image with the first prediction L2 loss can be formed:

[0082]

[0083] Based on a similar approach, and combining energy spectrum information with multi-energy synthesis conditions, multi-energy projection p can be synthesized using metal-based maps, water-based maps, and bone-based maps. train Compared with the simulated projection p containing artifacts poly Perform self-supervised loss:

[0084] loss self =||p poly -p train ||2

[0085] Correspondingly, the total loss function is equal to the sum of the five loss functions mentioned above multiplied by their respective weights:

[0086] loss = λ w loss w +λ b loss b +λ m loss m +λ im gloss img +λ self loss self ;

[0087] In this embodiment of the invention, λ w =λ b =λ m =1,λ img =λ self =100.

[0088] In this embodiment of the invention, specifically, under the condition of 140 kVp, the energy spectrum data represents the number of incident photons I0(E) corresponding to 0 keV to 140 keV at 140 kVp; the corresponding number of emitted photons I(x) can be calculated according to the following statistical formula: in, The string diagram data representing different energies (kev) can be decomposed into the following in this embodiment of the invention: Where, μ E，Ti μ E，water and μ E，bone These represent the mass decay coefficients of metal, water, and bone at energy E, respectively.

[0089] Correspondingly, the multi-energy projection formula Where I0 is the number of incident photons,

[0090] Specifically, the number of emitted photons I(x) and the number of incident photons I0 can be understood as the intensity of X-rays. In the multi-energy projection formula, I0(E) = w i *I0 means that the number of incident photons at a certain energy is equal to the corresponding weight multiplied by the total number of incident photons. Therefore, if energy spectrum information is required, the corresponding weights need to be learned.

[0091] The specific method involves using a learnable variable of size [1, 140] to predict weights at different energies. Each weight is equal to the number of incident photons at a given energy divided by the total number of incident photons. This implementation allows for the generation of corresponding multi-energy ray images by combining chordal diagram information under different ray conditions.

[0092] S9: Training the predictive neural network;

[0093] Using the loss function of step S7 as the target, and as described above, the training data is imported into the prediction neural network for training, and finally a trained prediction neural network can be obtained.

[0094] The basic function of this predictive neural network is to input an image containing metal artifacts and then output the corresponding predicted water base map, predicted bone base map, and predicted metal base map. Since the loss function is calculated not only based on the three basic base maps, but also supplemented by the L2 loss function and the self-supervised loss function, the final generated predicted water base map, predicted bone base map, and predicted metal base map, as well as the corresponding fused image, have high accuracy.

[0095] After steps S7 to S9, a trained prediction neural network can be obtained, and the application process needs to be executed accordingly.

[0096] Specifically, the application process includes:

[0097] S10: Input the actual image containing artifact information into the prediction neural network, and derive the corresponding predicted water-based image, predicted bone-based image and predicted metal-based image based on the prediction neural network. Combine the energy spectrum information, derive the artifact correction image based on the predicted water-based image, predicted bone-based image and predicted metal-based image.

[0098] Specifically, depending on the purpose of the predictive neural network, after inputting the actual image containing artifact information into the neural network, predictive water-based images, predictive bone-based images, and predictive metal-based images can be obtained respectively. Among them, according to the data used during the training of the predictive neural network, the predictive metal-based image does not have artifacts. Therefore, in this step, the predictive water-based image, predictive bone-based image, and predictive metal-based image are fused together with energy spectrum information to obtain an artifact-corrected image.

[0099] Figure 7 This is a schematic diagram of the artifact correction image 104 according to an embodiment of the present invention. For ease of comparison, the actual image 101, the first image 102 processed based on the DuDo-MAR method, and the second image 103 processed based on the GroundTruth method are included in the comparison. Specifically, through comparison, it is found that the beam hardening artifact correction method based on neural networks provided by the embodiment of the present invention can effectively remove metal artifacts while maintaining the morphological structure well. Moreover, the test results are very close to the industry-leading GroundTruth method, and it can remove metal artifacts better than the DuDo-MAR method based on a dual-domain artifact correction network.

[0100] During the fusion process, the predicted water-based image and the predicted bone-based image are first converted into corresponding chord graph data using methods such as Radon transform. Then, they are multiplied by the corresponding quality attenuation coefficients and summed to obtain a fused chord graph containing information from the predicted water-based image and the predicted bone-based image. An intermediate predicted fused image is reconstructed based on the fused chord graph. The predicted metal-based image is directly multiplied by the corresponding quality attenuation coefficient to obtain an intermediate metal image. The intermediate predicted fused image and the intermediate metal image are directly merged to obtain the final artifact-corrected image.

[0101] Specifically, in combination Figure 8The schematic diagram illustrates the data processing flow of a neural network-based beam hardening artifact correction method, which can be mainly divided into two stages. In the first stage, multiple sets of training data are manually generated using existing data. In the second stage, a predictive neural network is constructed using the generated training data. Ultimately, the predictive neural network can be used to correct artifacts on real CT images. When generating training data, metal artifacts are manually added to artifact-free dual-energy CT images to obtain multiple sets of images containing metal artifacts. Then, in the predictive neural network construction stage, the images containing metal artifacts are used as input, and the artifact-free dual-energy CT images are used as output for training. The objects in the dual-energy CT images are separated into water-based, bone-based, and metal-based categories for processing, which better reflects real-world application scenarios.

[0102] This invention provides a neural network-based method for beam hardening artifact correction. The basic idea is to use a polygon containing artifacts as network input to generate artifact-free water, bone, and metal maps using a neural network. Then, these artifact-free water, bone, and metal maps are fused with spectrum information to obtain an artifact-corrected image under multi-energy conditions. Specifically, training data is obtained through manual construction, and a multi-dimensional loss function is used to train the prediction neural network, resulting in relatively accurate predictions. During the synthesis of the artifact-corrected image, the predicted water and bone maps are fused and reconstructed using the corresponding polygon information before being directly fused with the predicted metal map. This ensures a better artifact correction effect in the final image, demonstrating good accuracy and practicality in real-world applications.

[0103] Accordingly, this invention also provides a neural network-based beam hardening artifact removal system. Specifically, the aforementioned neural network-based beam hardening artifact correction method is mainly implemented through data calculation, which relies on hardware systems such as computers. Furthermore, the direct source of dual-energy CT images is related storage devices, while the actual source is related X-ray scanning equipment. Therefore, in this invention, the metal artifact correction system should at least include a hardware system carrier. Based on this, the beam hardening artifact removal system of this invention also includes related X-ray scanning equipment and / or data storage devices.

[0104] The foregoing has provided a detailed description of a neural network-based beam hardening artifact correction method and system provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as limiting the present invention.

Claims

1. A method for beam hardening artifact correction based on neural networks, characterized in that, A process for correcting beam hardening artifacts in real images containing artifact information, including data simulation, training, and application procedures. The data simulation process is used to simulate the corresponding metal implant and the corresponding artifacts of the metal implant based on dual-energy CT images without metal artifacts. The training process is used to train a predictive neural network based on several simulated images to obtain a trained predictive neural network. The input of the predictive neural network is an input image containing metal artifacts, and the output of the predictive neural network is a predicted water base map, a predicted bone base map, and a predicted metal base map corresponding to the input image. The application process includes inputting an actual image containing artifact information into the trained predictive neural network, deriving a predicted water-based image, a predicted bone-based image, and a predicted metal-based image corresponding to the actual image through the predictive neural network; obtaining an artifact-corrected image based on the predicted water-based image, the predicted bone-based image, and the predicted metal-based image, including: converting the predicted water-based image and the predicted bone-based image into corresponding predicted water-based chord images and predicted bone-based chord images respectively using the Radon transform method; multiplying the predicted water-based chord images and the predicted bone-based chord images by corresponding quality attenuation coefficients and summing them to obtain a fused chord image containing information from the predicted water-based image and the predicted bone-based image; reconstructing an intermediate predicted fused image based on the fused chord image; multiplying the predicted metal-based image by the corresponding quality attenuation coefficient to obtain an intermediate metal image; and merging the intermediate predicted fused image and the intermediate metal image to obtain the artifact-corrected image. The training process includes: constructing a predictive neural network, wherein the predictive neural network is a Unet structure with skip connections, the activation function is SimpleGate, the input of the predictive neural network is an input image containing metal artifacts, and the output of the predictive neural network is a predicted water base map, a predicted bone base map, and a predicted metal base map corresponding to the input image; constructing a loss function, wherein the loss function includes a first loss function based on the predicted water base map, a second loss function based on the predicted bone base map, a third loss function based on the predicted metal base map, a fourth loss function based on a first fused image obtained by fusing the predicted water base map and the predicted bone base map, and a fifth loss function based on a multi-energy projection image obtained by fusing the predicted water base map, the predicted bone base map, and the predicted metal base map; and training the predictive neural network.

2. The beam hardening artifact correction method based on neural networks as described in claim 1, characterized in that, The data simulation process includes: S1: Acquire a dual-energy CT image free of metal artifacts; S2: The dual-energy CT image is decomposed into a water-based image and a bone-based image, and a metal-based image is obtained based on the material of the metal implant and the metal mask corresponding to the metal implant. S3: Take the non-metallic region image of the water-based image and perform forward projection to obtain the water-based chord diagram; take the non-metallic region image of the bone-based image and perform forward projection to obtain the bone-based chord diagram; and perform forward projection of the metal-based image to obtain the metal-based chord diagram. S4: The water-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the first projection data, the bone-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the second projection data, the metal-based chord diagram is multiplied by the corresponding mass attenuation coefficient to obtain the third projection data, and the first projection data, the second projection data and the third projection data are accumulated to obtain the processing projection data under the corresponding energy. S5: Combining the energy spectrum information, multi-energy projection data under multi-energy conditions is obtained based on the processed projection data, and a simulated image is reconstructed based on the multi-energy projection data. The simulated image has simulated metal artifacts.

3. The beam hardening artifact correction method based on neural networks as described in claim 1, characterized in that, The X-ray tube voltages selected for the dual-energy CT images were 70kVp and 140kVp, respectively.

4. A beam hardening artifact correction system based on a neural network, characterized in that, The beam hardening artifact correction method based on neural networks as described in any one of claims 1 to 3 is adopted.

Images