A method for metal artifact removal from CT images

Abstract

The application discloses a metal artifact removal method of a CT image, and the method comprises the following steps: constructing an initial adaptive iterative learning model based on wavelet transform, decomposing a CT image by using a target adaptive iterative learning model obtained by optimizing an optimization objective function, calculating the area of a metal artifact, removing the metal artifact in the CT image, and obtaining a CT image with reduced metal artifacts; the initial adaptive iterative learning model based on wavelet transform is used for artifact removal, so that the spatial distribution characteristics of the metal artifact under different domains and resolutions can be fully utilized, the artifact removal is better, and the interpretability is high; in addition, when a first optimization objective function is solved by combining a proximal gradient descent algorithm and a Taylor formula, a proximal gradient operator obtained can be replaced by a simple network module, so that the network can be more easily constructed, and the adaptability of the network is enhanced. The application can be widely applied to the technical field of CT image processing.

Description

Technical Field

[0001] This invention relates to the field of CT image processing technology, and in particular to a method for removing metal artifacts from CT images. Background Technology

[0002] Reducing metal artifacts (MAR) to improve CT image quality without compromising diagnostic value has always been a significant challenge. Metal implants can cause bright and dark bands in reconstructed images, obscuring detailed structures. Large metal implants can even affect the identification of the entire CT image, severely interfering with physician diagnoses. Particularly in cancer treatment, inaccurate tumor location and identification of surrounding tissues can lead to errors in radiation dose calculations, with serious consequences. With the increasing use of metal implants, such as dental fillings, artificial hip joints, and spinal implants, metal artifacts in CT images have become increasingly prevalent. Therefore, reducing metal artifacts and accurately restoring tissue details has direct clinical significance.

[0003] Because metal artifacts are structured and non-local, removing them directly from images is difficult. Therefore, before the application of deep learning to MAR (Metal Augmentation Reduction), most research proposed reducing metal artifacts in the sinusoidal domain. However, severe secondary artifacts are an inherent problem faced by traditional methods. Deep learning seems to offer a good solution for MAR; however, its generalization performance is poor, resulting in disappointing performance on clinical data. Clearly, improving the MAR performance of deep learning methods on clinical data is a challenging task.

[0004] For traditional MAR methods, linear interpolation (LI) removes metal artifacts by replacing the data in the metal trace region with interpolated data in the sine plot. However, interpolation errors lead to severe secondary artifacts. While normalized metal artifacts removal (NMAR) achieves more meaningful results than LI by interpolating on a normalized sine plot, unavoidable secondary artifacts blur the reconstructed image. Another widely used traditional method is iterative reconstruction algorithms, which aim to minimize the error between the generated sine plots. However, they are computationally expensive. Furthermore, due to the limited capabilities of a single traditional MAR method, multiple traditional methods are often combined.

[0005] Deep learning-based MAR methods can be divided into three main categories: sinusoidal graph domain, image domain, and dual domain.

[0006] (1) Sine Map Domain: Unlike the global distribution of metal artifacts in the image domain, the projection data errors in the sine map converge on the metal traces, exhibiting greater regularity and regionality. Therefore, the metal artifact features reflected in the sine map can be easily captured by neural networks. However, on the one hand, the method of obtaining the sine map mainly involves projecting the image, which may introduce new data errors; on the other hand, due to the lack of constraints from the image domain, even slight disturbances in the sine map can cause severe artifacts in the image.

[0007] (2) Image Domain: Image domain-based MAR methods typically utilize residual learning or adversarial learning to construct deep learning networks. For example, Wang et al. encoded metal artifacts as a weighted convolutional dictionary model, enhancing the interpretability of neural networks. Unfortunately, the features learned from the image domain are limited, which restricts the improvement of network performance.

[0008] (3) Two-domain: Recent studies generally employ two sub-networks to recover the metal artifact image and the sine curve, respectively. Li et al. studied the distribution of metal artifacts in the frequency domain and integrated Fourier transform into the network instead of directly reconstructing from the Randon and image domains. However, due to the limitations of Fourier transform in handling non-stationary signals, the improvement is limited. Inspired by iterative reconstruction mechanisms, two-domain data consistency cyclic networks iteratively recover CT images and sine curves; Wang et al. embedded the CT imaging process into a deeply expanded two-domain network. Although two-domain networks have been widely used and achieved significant results in reducing metal artifacts, the difficulty in obtaining the sine curve limits their clinical application. In addition, many two-domain networks lack interpretability, and their MAR process is generally unclear. Summary of the Invention

[0009] In view of this, embodiments of the present invention provide a method for removing metal artifacts from CT images that is highly adaptable, has good processing effect, and is highly interpretable.

[0010] One aspect of this invention provides a method for removing metal artifacts from CT images, comprising: constructing an initial adaptive iterative learning model based on wavelet transform; determining a CT image expression based on clean image regions, binary non-metallic regions, and metal artifact regions of the CT image; constructing a first optimization objective function based on the CT image expression, a regularization term, and the image wavelet transform; solving the first optimization objective function using a proximal gradient descent algorithm and Taylor's formula to obtain a first calculation result; optimizing the initial adaptive iterative learning model based on the first calculation result, using L2 output loss and a ground-based real image to obtain a target adaptive iterative learning model; and performing image decomposition and stitching on the CT image using the target adaptive iterative learning model to remove metal artifacts from the CT image.

[0011] Optionally, the CT image expression is determined based on the clean image region, the binary non-metallic region, and the metallic artifact region of the CT image, wherein the CT image expression is:

[0012] Y = I⊙X + I⊙A

[0013] Where X represents the clean image region; Y represents the original CT image region; I is the binary non-metallic region of Y; A is the metallic artifact region of the CT image; and ⊙ represents element-wise multiplication.

[0014] Optionally, constructing the first optimization objective function based on the CT image expression and wavelet transform includes: constructing an initial optimization objective function based on the CT image expression; adjusting the initial optimization function through regularization terms and wavelet transform to obtain an intermediate optimization objective function; and optimizing the intermediate optimization objective function through an error feedback term to obtain the first optimization objective function.

[0015] Optionally, the expression for the initial objective optimization function is:

[0016]

[0017] in, ε is the Frobenius norm; X represents the clean image region; Y represents the original CT image region; I is the binary non-metallic region of Y; A is the metallic artifact region of the CT image; ⊙ is element-wise multiplication.

[0018] The expression for the first optimization objective function is:

[0019]

[0020] Where U is the image with metal artifacts removed; W is the adaptive wavelet transform; λ1, λ2, and λ3 are regularization weights; and f1(), f2(), and f3() are regularization terms.

[0021] Optionally, the step of combining the proximal gradient descent algorithm and Taylor's formula to solve the first optimization objective function and obtain the first calculation result includes: determining the expression for the metal artifact region and the expression for the clean image region based on the first optimization objective function; determining a first iterative equation for solving the metal artifact region in the wavelet domain based on the metal artifact region expression using the proximal gradient descent algorithm and Taylor's formula; determining a second iterative equation for solving the clean image region in the wavelet domain based on the clean image region expression using the proximal gradient descent algorithm and Taylor's formula; determining a third iterative equation based on the first iterative equation and the second iterative equation; wherein the third iterative equation is the iterative equation for the intermediate image after removing the metal artifacts; and using the first iterative equation, the second iterative equation, and the third iterative equation in the image domain as the first calculation result.

[0022] Optionally, the first iterative equation is:

[0023]

[0024] Where, is the result of the (k+1)th iteration when solving for the metal artifact region; A (k+0.5) Represents approximation A (k+1) Taylor polynomial; W is adaptive wavelet transform; WT is inverse wavelet transform calculation; WA (k+0.5) This indicates that the metal artifact region is solved iteratively in the wavelet domain; It is the proximal gradient operator for calculating the region of metal artifacts; α and η1 are trainable parameters.

[0025] The second iterative equation is:

[0026]

[0027] Among them, X (k+1) This is the result of the (k+1)th iteration when solving for the clean image region; X (k+0.5) Represents approximation X (k+1) Taylor polynomial; WX (k+0.5) This indicates that the clean image region is solved iteratively in the wavelet domain; It is the proximal gradient operator for calculating clean image regions; α and η2 are trainable parameters.

[0028] The third iterative equation is:

[0029]

[0030] Among them, U (k+1) It is the (k+1)th iteration result of solving the image with metal artifacts removed; U (k+0.5 ) represents approximate U (k +1)Taylor polynomials; It is the proximal gradient operator for calculating the image with metal artifacts removed; α and η3 are trainable parameters.

[0031] Optionally, the step of optimizing the initial adaptive iterative learning model based on the first calculation result, combined with L2 output loss and ground-based real images, to obtain a target adaptive iterative learning model includes: obtaining wavelet components of the metal artifact region, clean image region, and first image region of the CT image; solving the first optimization objective function based on the wavelet components and the first calculation result to obtain an intermediate adaptive iterative learning model; evaluating the intermediate adaptive iterative learning model during the iteration process using L2 output loss and ground-based real images; iteratively executing the step of solving the first optimization objective function based on the first calculation result to obtain the intermediate adaptive iterative learning model until the evaluation result meets a predetermined threshold to obtain the target adaptive iterative learning model.

[0032] Embodiments of the present invention also provide a metal artifact removal system for CT images, comprising: a first module for constructing an initial adaptive iterative learning model based on wavelet transform; a second module for determining a CT image expression based on clean image regions, binary non-metallic regions, and metal artifact regions of the CT image; a third module for constructing a first optimization objective function based on the CT image expression, a regularization term, and an image wavelet transform; a fourth module for solving the first optimization objective function by combining a proximal gradient descent algorithm and a Taylor formula to obtain a first calculation result; a fifth module for optimizing the initial adaptive iterative learning model based on the first calculation result, combined with L2 output loss and a ground-based real image to obtain a target adaptive iterative learning model; and a sixth module for performing image decomposition and stitching on the CT image using the target adaptive iterative learning model to remove metal artifacts from the CT image.

[0033] Embodiments of the present invention also provide an electronic device including a processor and a memory; the memory is used to store a program; the processor executes the program to implement the method described above.

[0034] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described above.

[0035] The embodiments of the present invention have the following beneficial effects: By constructing an initial adaptive iterative learning model based on wavelet transform, and optimizing the initial adaptive iterative learning model according to the first optimization objective function, the resulting target adaptive iterative learning model can decompose CT images, calculate the regions of metal artifacts, and remove metal artifacts from CT images, resulting in CT images with reduced metal artifacts; the embodiments of the present invention construct an initial adaptive iterative learning model based on wavelet transform for artifact removal, which can make full use of the spatial distribution characteristics of metal artifacts under different domains and resolutions, and better remove artifacts, with high interpretability; in addition, when combining the proximal gradient descent algorithm and Taylor formula to solve the first optimization objective function, the obtained proximal gradient operator can be replaced by a simple network module, which not only makes it easier to construct the network, but also enhances the adaptability of the network. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0037] Figure 1 This is a flowchart of the method steps according to an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the metal artifact features after adaptive wavelet transform according to an embodiment of the present invention;

[0039] Figure 3 This is a diagram of the overall architecture of the T-times iterative adaptive transformation network according to an embodiment of the present invention.

[0040] Figure 4 This is a subnetwork structure diagram at the k-th iteration of this embodiment of the invention;

[0041] Figure 5 This is a structural diagram of the adaptive wavelet module transform according to an embodiment of the present invention;

[0042] Figure 6 These are detailed structural diagrams of different proximal operators in embodiments of the present invention;

[0043] Figure 7 This is a comparison diagram showing the results of comparing the method of this invention, the traditional method LI, the traditional method NMAR, and the deep learning methods ACDNet, DuDoNet++, DuDoNet, and CNNMAR on simulated images.

[0044] Figure 8This is a box plot of the SSIM evaluation metric values ​​of the method of this invention, the conventional method LI, the conventional method NMAR, and the deep learning methods ACDNet, DuDoNet++, DuDoNet, and CNNMAR.

[0045] Figure 9 This is a box plot of the PSNR values ​​of the method of the present invention, the conventional method LI, the conventional method NMAR, and the deep learning methods ACDNet, DuDoNet++, DuDoNet and CNNMAR.

[0046] Figure 10 This is a box plot of the SSIM values, an evaluation index of the method of this invention, for different metal sizes;

[0047] Figure 11 This is a box plot of the PSNR values, an evaluation index of the method of this invention, for different metal sizes;

[0048] Figure 12 This is a comparison diagram showing the results of comparing the method of this invention, the conventional method LI, the conventional method NMAR, the deep learning methods ACDNet, DuDoNet++, DuDoNet, and CNNMAR on clinical metal artifact images;

[0049] Figure 13 This is a comparison chart of the image processing results provided by the embodiments of the present invention, including real ground images, metal damage images, the methods of the embodiments of the present invention, image domain only methods, wavelet domain only methods, and ordinary wavelet domain methods. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] To address the issues of lack of interpretability and limited performance improvement in many existing dual-domain networks, this invention proposes a method for removing metal artifacts from CT images, referring to... Figure 1 , Figure 1This is a flowchart of the method steps provided in this embodiment of the invention, including: constructing an initial adaptive iterative learning model based on wavelet transform; determining a CT image expression based on the clean image region, binary non-metallic region, and metallic artifact region of the CT image; constructing a first optimization objective function based on the CT image expression, regularization term, and image wavelet transform; solving the first optimization objective function by combining the proximal gradient descent algorithm and Taylor formula to obtain a first calculation result; optimizing the initial adaptive iterative learning model based on the first calculation result, combining L2 output loss and ground real image to obtain a target adaptive iterative learning model; and performing image decomposition and stitching on the CT image through the target adaptive iterative learning model to remove metallic artifacts from the CT image.

[0052] Specifically, the embodiments of the present invention employ a multi-level MAR method based on wavelet transform, which can better capture the spatial features of different frequency components of an image, such as... Figure 2 As shown, Figure 2 This is a schematic diagram of metal artifact features after adaptive wavelet transform according to an embodiment of the present invention. Metal artifacts exhibit different feature distributions in the wavelet components and the image. Different wavelet components also exhibit different characteristics. Before proceeding with the introduction, the definitions of some of the relevant variables are explained as follows:

[0053] X represents the clean image region; Y represents the original CT image region; I is the binary non-metallic region of Y; A is the metallic artifact region of the CT image; ⊙ is element-wise multiplication; W is adaptive wavelet transform; λ1, λ2, and λ3 are regularization weights; f1(), f2(), and f3() are regularization terms.

[0054] The following describes in detail, with reference to the accompanying drawings, a method for removing metal artifacts from CT images according to an embodiment of the present invention:

[0055] The method includes the following steps S100 to S600:

[0056] S100. Construct an initial adaptive iterative learning model based on wavelet transform.

[0057] Specifically, refer to Figure 3 , Figure 3 This is an overall architecture diagram of the T-iteration adaptive transform network of an embodiment of the present invention. The network of this model includes T modules, corresponding to the T iterations of the optimization function. Each module contains three key components, which are used to solve for the iterative results of the metal artifact region, the clean image region, and the image with metal artifacts removed.

[0058] S200. Determine the CT image expression based on the clean image region, binary non-metallic region, and metallic artifact region of the CT image.

[0059] Specifically, in the image domain, metallic regions with higher CT values ​​are easily segmented, while non-metallic regions can be viewed as a combination of metallic artifacts A and the undamaged, clean CT image X. Therefore, the image Y ∈ RH × W with metallic artifacts can be decomposed as follows:

[0060] Y = I⊙X + I⊙A + M (1)

[0061] Where M represents the metal implant in the image, I is the binary nonmetallic region of Y, and ⊙ represents element-wise multiplication.

[0062] This invention does not focus on the metal regions of human tissues, and equation (1) can be derived as follows:

[0063] I⊙Y=I⊙X+I⊙A (2)

[0064] S300. Construct the first optimization objective function based on the CT image expression, regularization term, and image wavelet transform.

[0065] Specifically, step S300 includes the following steps S310 to S320:

[0066] S310. Construct an initial optimization objective function based on the CT image expression.

[0067] Specifically, the goal in the derivation process is to solve for X and A so that their sum completely replaces Y. Because there are two variables to solve, it is difficult to find an analytical solution to equation (2). As an alternative, an initial optimization function is constructed as shown in equation (3):

[0068]

[0069] in, ε is the Frobenius norm; ε is the target threshold for optimization.

[0070] S320. The initial optimization function is adjusted by regularization terms and wavelet transform to obtain the intermediate optimization objective function.

[0071] Specifically, to obtain a better solution, a regularization term representing the prior constraints and an adaptive wavelet transform are introduced, resulting in:

[0072]

[0073] Where W is the wavelet transform; λ1, λ2, and λ3 are regularization weights; f1(), f2(), and f3() are regularization terms. f1() introduces prior knowledge of X in the image domain into the model, and f2() and f3() represent the wavelet domain constraints of X and A, respectively. Furthermore, equation (4) can be derived as:

[0074]

[0075] Equation (5) is the intermediate optimization objective function.

[0076] S330. The intermediate objective function is optimized through the error feedback term to obtain the first objective function.

[0077] Specifically, when wavelet components are transformed into an image through inverse wavelet transform, the image becomes blurred due to data loss. To recover a more accurate image, let U = I⊙X, and introduce an error feedback term into equation (5). The intermediate optimization objective function is optimized through the error feedback term, resulting in the first optimization objective function as shown in equation (6):

[0078]

[0079] Where U is the image with metal artifacts removed; W is the adaptive wavelet transform; λ1, λ2, and λ3 are regularization weights; and f1(), f2(), and f3() are regularization terms.

[0080] S400. The first optimization objective function is solved by combining the proximal gradient descent algorithm and Taylor formula to obtain the first calculation result.

[0081] Specifically, step S400 includes the following steps S410 to S450:

[0082] S410. Determine the expressions for the metal artifact region and the clean image region based on the first optimization objective function.

[0083] Specifically, taking A as an example, the metallic artifact A is constrained only in the wavelet domain, represented by f1() in equation (6). When solving for A, X and U can be regarded as constants, so in the (k+1)th iteration, A (k+1) It can be exported as:

[0084]

[0085] Update A (k+1) To minimize the sum of differentiable and non-differentiable convex functions, this embodiment of the invention employs the Proximal Gradient Method (PGD) to solve this problem. Because the proximal gradient operator obtained by the algorithm can be replaced by a simple module, this not only facilitates network construction but also enhances the network's adaptability through trainable operational modules. Similarly, X represents the clean image region, and U is the image with metal artifacts removed. (k+1) and U (k+1) The following can be derived:

[0086]

[0087]

[0088] Although U is constrained only in the wavelet domain, X is related to U, so the wavelet components of U also participate in the calculation in Equation (8). In Equation (9), the error feedback term is used to recover the loss caused by the domain transformation. Since U is an optimization based on X and is only related to X, U is solved by transforming to X in the image domain using the inverse wavelet transform, and then A is solved separately. (k+1) X (k+1) and U (k+1) .

[0089] S420. Based on the expression for the metal artifact region, the first iterative equation for solving the metal artifact region in the wavelet domain is determined using the near-end gradient descent algorithm and Taylor formula.

[0090] Specifically, the first iterative equation for solving A is determined, and equation (7) can be rewritten as:

[0091]

[0092] Here, α is a non-zero parameter. According to the near-end gradient descent algorithm, directly solving the equation requires a large amount of computation; therefore, Taylor's formula is used to approximate A. (k+1) , when A=A (k+1) Then, the second-order approximation of equation (10) can be derived as:

[0093]

[0094] in, constant Is g1(A) in A (k) The second derivative of η1. In the code implementation, a trainable parameter can be used to represent η1, and the application of adaptive parameters can better help the network fit A. Furthermore, g1(A) (k) Replace with another constant This forms a perfect square trinomial. Therefore, equation (11) yields:

[0095]

[0096] To more clearly illustrate the iterative process, this embodiment of the invention introduces an intermediate variable A. (k+0.5) ,in

[0097]

[0098] Based on this, the first iterative equation is:

[0099]

[0100] Among them, A (k+1)This is the result of the (k+1)th iteration when solving for the metal artifact region; A (k+0.5) Represents approximation A (k+1) The Taylor polynomial; W is the adaptive wavelet transform; W T It is a wavelet inverse transform calculation; WA (k+0.5) This indicates that the metal artifact region is solved iteratively in the wavelet domain; It is the proximal gradient operator for calculating the region of metal artifacts; α and η1 are trainable parameters.

[0101] prox αη1 It is the proximal gradient operator related to f1(). Since f1() represents the constraint on metal artifacts in the wavelet domain, A is iteratively solved in the wavelet domain, where WA is used. (k+0.5) Indicated. (Through W) T The inverse wavelet transform was implemented to obtain the desired metal artifacts in the image domain.

[0102] S430. Based on the expression for the clean image region, determine the second iterative equation for solving the clean image region in the wavelet domain using the proximal gradient descent algorithm and Taylor formula.

[0103] Specifically, determine the second iterative equation for solving X, let the trainable parameter (α+β) be non-zero, and use A obtained from equation (14) (k +1) Substituting into equation (8), we get X. (k+1) for:

[0104]

[0105] X (k+0.5) It can be approximated as a Taylor polynomial of equation (15), then the iterative process of X is:

[0106]

[0107] Here, η2 is a trainable parameter.

[0108] Based on this, the second iterative equation is:

[0109]

[0110] Among them, X (k+1) This is the result of the (k+1)th iteration when solving for the clean image region; X (k+0.5) Represents approximation X (k+1) Taylor polynomial; WX (k+0.5) This indicates that the clean image region is solved iteratively in the wavelet domain; It is the proximal gradient operator for calculating clean image regions; α and η2 are trainable parameters.

[0111] prox αη2 It is the proximal gradient operator associated with f2(). The second iterative equation for X (k+1) Corrections were made in the wavelet domain. For X (k+1) Perform inverse wavelet transform on the wavelet components.

[0112] S440. Determine the third iterative equation based on the first and second iterative equations; wherein, the third iterative equation is the iterative equation for the intermediate image after removing metal artifacts.

[0113] Specifically, the third iterative equation for solving U is determined, and A is obtained in the wavelet domain. (k+1) and X (k+1) Then, U is calculated in the image domain. (k+1) From equation (9), we can obtain:

[0114]

[0115] Following a similar calculation process as step S430 or S440, we obtain:

[0116] U (k+0.5) =(1-2η3)U (k) +2η1X (k+1) (19)

[0117] Here, η3 is a trainable parameter.

[0118] The third iterative equation is:

[0119] Among them, U (k+1) It is the (k+1)th iteration result of solving the image with metal artifacts removed; U (k+0.5) Represents approximation U (k+1) Taylor polynomials; It is the proximal gradient operator for calculating images with metal artifacts removed; prox αη3 This relates to the regularization function f3(U). Unlike A and X, U is enhanced in the image domain.

[0120] S450. The first, second, and third iterative equations of the image domain are used as the first calculation result.

[0121] Specifically, the first, second, and third iterative equations in the image domain are used as the first calculation result, which is then used to optimize and train the initial adaptive iterative learning model.

[0122] S500. Based on the first calculation result, the initial adaptive iterative learning model is optimized by combining the L2 output loss and the real ground image to obtain the target adaptive iterative learning model.

[0123] Specifically, the process of optimizing and training the initial adaptive iterative learning model is as follows:

[0124] At the (k+1)th block (the (k+1)th iteration) of the initial adaptive iterative learning model, the adaptive wavelet transform module is first used to process A. (k) X (k) and U (k) Decompose the components. In the wavelet domain, calculate A according to formula (13). k+0.5 Then input into proxNet A proxNetA is built based on DnCNN to mimic the solution process of equation (14). Similarly, proxNetX with five convolutional blocks was also built. (k+0.5) and X (k+1) Calculate according to equations (17) and (16). Then apply the inverse wavelet transform to transform A. (k+1) and X (k+1) Transform to the image domain. Based on equations (20) and (19), calculate U in the image domain sequentially. (k+0.5) and U (k +1) Meanwhile, the lightweight Unet is used as the proxy proxNetU.

[0125] Furthermore, embodiments of the present invention can also construct an initialization module for initializing A, X, and U. In this initialization module, an image X of metal destruction is provided. ma and linear interpolated image X LI The stitched image is sent as input to the network. Although X LI It provides an artifact-free prior image, but it also introduces interpolation errors, so X... ma and X LI Together, they are sent as input to the network of the initial adaptive transformation learning model.

[0126] The network first decomposes the input image using an adaptive wavelet transform module. The four decomposed wavelet components are then concatenated and input into UNet, which mainly consists of a feature extraction network (encoder) and an upsampling network (decoder). The feature extraction network uses five convolutional layers and max pooling layers, while the upsampling network mainly consists of an upsampling module and convolutional layers. The outputs of each layer are concatenated along the channel dimension using residual learning. The output of UNet then passes through DnCNN, which consists of convolutional layers, batch normalization, and activation layers. The output of DnCNN is then subjected to inverse wavelet transform to obtain the image domain initialization X0. Finally, X0 is compared with X... ma and X LI The Unet input to the image domain is concatenated again to obtain U0, and finally A0 is calculated, where A0 = X. ma -U0.

[0127] During the training and optimization process, this embodiment of the invention uses the L2 output loss function and ground truth images for optimization, ultimately obtaining a target adaptive iterative learning model. The L2 output loss function of this embodiment of the invention is:

[0128]

[0129] Among them, Y out It is the network output image; X gt It is a ground truth image.

[0130] S600: The target adaptive iterative learning model is used to decompose and stitch CT images to remove metal artifacts from CT images.

[0131] Specifically, the CT image with metal artifacts is processed by a target adaptive iterative learning model to perform image decomposition and stitching similar to step S500, which can remove the metal artifacts in the CT image. Furthermore, the target adaptive iterative learning model of this embodiment is more interpretable, and each module has its own mathematical meaning.

[0132] Embodiments of the present invention also provide a metal artifact removal system for CT images, comprising: a first module for constructing an initial adaptive iterative learning model based on wavelet transform; a second module for determining a CT image expression based on clean image regions, binary non-metallic regions, and metal artifact regions of the CT image; a third module for constructing a first optimization objective function based on the CT image expression, a regularization term, and image wavelet transform; a fourth module for solving the first optimization objective function by combining a proximal gradient descent algorithm and Taylor's formula to obtain a first calculation result; a fifth module for optimizing the initial adaptive iterative learning model based on the first calculation result, combined with L2 output loss and a ground-based real image to obtain a target adaptive iterative learning model; and a sixth module for performing image decomposition and stitching on the CT image using the target adaptive iterative learning model to remove metal artifacts from the CT image.

[0133] Embodiments of the present invention also provide an electronic device including a processor and a memory; the memory is used to store a program; the processor executes the program to implement the method described above.

[0134] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described above.

[0135] The embodiments of the present invention have the following beneficial effects:

[0136] 1. It can fully utilize the spatial distribution characteristics of metal artifacts in different domains and resolutions through an adaptive iterative learning model, thereby better removing metal artifacts from CT images. The processing is clear and more interpretable.

[0137] 2. When solving the first optimization objective function by combining the proximal gradient descent algorithm and Taylor's formula, the obtained proximal gradient operator can be replaced by a simple network module, which not only makes it easier to build the network, but also enhances the network's adaptability.

[0138] The technical effects of the embodiments of the present invention will be further explained below with reference to experimental data:

[0139] (1) Synthetic dataset

[0140] First, 1200 images and 100 metal masks were collected for image synthesis. A training set was created using 1000 CT images and 90 metal masks. The remaining 200 CT images were paired with another 10 metal masks to synthesize test data. The size of the 10 masks was [2054, 879, 878, 448, 242, 115, 115, 111, 53, 32] pixels. The metal implant was inserted into a clean image by setting parameters, and a fan beam projection was used. 640 projection views were uniformly sampled between 0 and 360 degrees. The final generated CT image with metal artifacts had a value range limited to 0 to 120 and a size of 416×416 pixels.

[0141] (2) Clinical Dataset

[0142] The model's effectiveness was validated on the publicly available clinical dataset CLINIC-metal, which primarily collects postoperative images of pelvic fractures and consists of 75 3D metal fragments. A metal mask with a threshold of 2000 HU was used for segmentation, and the same method was used to compute the linear interpolated image X for the synthesized data. LI Image X that ultimately destroys the metal. ma and linear interpolated image X LI Send to network

[0143] (3) Benchmarks and evaluation indicators

[0144] During the experiments, the embodiments of the present invention were compared with traditional linear interpolation methods (LI), normalized metal artifact removal (NMAR), and deep learning methods such as Adaptive Convolutional Dictionary Network (ACDNet), DuDoNet++ (metal mask projection dual-domain network), DuDoNet (dual-domain metal artifact removal network), and Convolutional Neural Network (CNNMAR). In the experiments, the test results of the synthetic data were quantitatively compared using root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Furthermore, in clinical practice, metal artifact removal relies solely on visual comparison due to the lack of reference images.

[0145] (4) Experimental results of the synthetic dataset

[0146] Reference Figure 7 , Figure 7 This is a comparison diagram showing the results of comparing the embodiments of the present invention with traditional method L1, and deep learning methods ACDNet, DuDoNet++, DuDoNet, and CNNMAR on simulated images; Figure 7 In the diagram, (a) is a ground truth image, (b) is a metal damage image, (c) is the simulation image processing result of the conventional method LI, (d) is the simulation image processing result of the conventional method NMAR, (e) is the simulation image processing result of the deep learning method CNNMAR, (f) is the simulation image processing result of the deep learning method DuDoNet, (g) is the simulation image processing result of the deep learning method DuDoNet++, (h) is the simulation image processing result of the deep learning method ACDNet, and (i) is the simulation image processing result of the method of this embodiment.

[0147] from Figure 7 It can be concluded that interpolation errors cause severe secondary artifacts in both LI and NMAR results, and a large amount of detailed structure is blurred. Compared with traditional methods, although CNNMAR achieves better results, the detailed structure around the metal is still destroyed. DuDoNet, DuDoNet++, and ACDNet achieve better performance in reducing metal artifacts; however, from... Figure 3 It can be seen that the reconstructed spine has slight deformation.

[0148] Compared with the methods described above, referring to Table 1, which is a quantitative evaluation table of the embodiments of the present invention with the traditional method LI and the deep learning methods ACDNet, DuDoNet++, DuDoNet and CNNMAR, the network of the embodiments of the present invention has the best performance in reducing metal artifacts and restoring details. The three quantitative indicators, PANR, SSIM and RMSE, reached the best values ​​of 46.31, 0.984 and 0.628, respectively.

[0149] Table 1

[0150] method PSNR SSIM RMSE LI 35.5±3.03 0.898±0.032 2.21±0.833 NMAR 36.08±2.93 0.91±0.028 2.671±0.823 CNNMAR 41.76±3.31 0.962±0.016 1.077±0.437 DuDoNet 42.23±3.24 0.974±0.007 1.032±0.41 DuDoNet++ 45.15±3.1 0.974±0.006 0.716±0.296 ACDNet 45.47±2.59 0.98±0.005 0.68±0.209 Embodiments of the present invention 46.31±3.2 0.984±0.005 0.628±0.263

[0151] Reference Figure 8 and Figure 9 , Figure 8 This is a box plot of the SSIM evaluation metric values ​​for the method of this invention, the conventional method LI, the conventional method NMAR, and the deep learning methods ACDNet, DuDoNet++, DuDoNet, and CNNMAR. Figure 9 This is a box plot on the PSNR value, which is the evaluation metric. Figure 8 and Figure 9 The statistical results characterizing the performance of different methods on the test set show that the RMSE box plot of the method in this embodiment is very short, indicating that more data is distributed within a small range. Furthermore, the median of the box plot is close to the bottom, indicating that most data values ​​are relatively small. Similarly, similar conclusions can be drawn from the box plots of SSIM and PSNR.

[0152] Reference Figure 10 and Figure 11 , Figure 10 and Figure 11 Box plots showing the SSIM and PSNR values ​​of the method according to embodiments of the present invention on different metal sizes represent the statistical results of the method on different metal masks, with the metal mask size gradually increasing from 1 to 10. It can be observed that smaller metal implants have better MAR performance, but this is not positively correlated with the size of the metal.

[0153] (5) Experimental results of the clinical dataset

[0154] The embodiments of the present invention not only achieved good results in synthetic datasets, but also performed well in clinical datasets.

[0155] Reference Figure 12 , Figure 12 This is a comparison chart showing the results of the method of this invention, the conventional method L1, the conventional method NMAR, and the deep learning methods ACDNet, DuDoNet++, DuDoNet, and CNNMAR on clinical metal artifact images; Figure 12In the images, (a) is a metal destruction image, (b) is the clinical metal artifact image processing result of the conventional method LI, (c) is the clinical metal artifact image processing result of the conventional method NMAR, (d) is the clinical metal artifact image processing result of the deep learning method CNNMAR, (e) is the clinical metal artifact image processing result of the deep learning method DuDoNet, (f) is the clinical metal artifact image processing result of the deep learning method DuDoNet++, (g) is the clinical metal artifact image processing result of the deep learning method ACDNet, and (h) is the clinical metal artifact image processing result of the method of the embodiment of the present invention.

[0156] Figure 12 This paper presents a visual comparison of different methods on clinical CT images with metal artifacts. It can be seen that secondary artifacts caused by LI are more severe than metal artifacts. While NMAR significantly alleviates secondary artifacts, it still struggles to identify detailed tissue structures. Conversely, Figure 12 The results obtained by CNN, DuDoNet, and DuDoNet++ are acceptable. However, the performance of these methods gradually decreases as the metal volume increases. ACDNet performs best in clinical images, but it over-smooths information in metal artifact regions, causing the tissue near the metal to become blurred. Overall, the MAIL network of this embodiment achieves the best results in reducing metal artifacts and restoring tissue details.

[0157] (6) Ablation test

[0158] In the network framework of this invention embodiment, a scheme combining image and wavelet domain information is adopted to reduce metal artifacts. It is expected that the network can better fit artifacts by learning multi-scale knowledge. To verify the effectiveness of this scheme, this invention embodiment reconstructs the network and trains the model in both the wavelet domain and the image domain. The quantitative results on the synthetic dataset are shown in Table 2. Table 2 is a table of quantitative results of the method of this invention embodiment on the synthetic dataset:

[0159] Table 2

[0160] Method PSNR SSIM RMSE Wavelet domain only 44.04±3.26 0.978±0.009 0.823±0.345 Image domain only 45.80±3.07 <![CDATA[ 0.983±0.005 ]]> 0.664±0.271 Ordinary wavelet domain <![CDATA[ 45.97±3.47 ]]> 0.983±0.006 <![CDATA[ 0.663±0.307 ]]> Embodiments of the present invention 46.31±3.20 0.984±0.005 0.628±0.263

[0161] As shown in Table 2, the method of this embodiment achieves better performance than the "wavelet domain only" MAR scheme. Wavelet transform decomposes the original image into different frequency bands, and the image details and artifacts are mainly concentrated in the high-frequency components. The "wavelet domain only" method effectively reduces metallic artifacts in the image, but also blurs the details of the tissue.

[0162] Reference Figure 13 , Figure 13This is a comparison chart of image processing results provided by embodiments of the present invention, including real ground images, metal damage images, and the methods of the present invention, image domain-only methods, wavelet domain-only methods, and ordinary wavelet domain methods. Figure 13 In the image, (a) is a real ground image, (b) is a metal damage image, (c) is the result of metal artifact processing by the method of the present invention, (d) is the result of metal artifact processing by the image domain only method, (e) is the result of metal artifact processing by the wavelet domain only method, and (f) is the result of metal artifact processing by the ordinary wavelet domain method.

[0163] from Figure 13 As can be seen, the "wavelet domain only" method effectively reduces metallic artifacts in the image, but it also blurs tissue details. Wavelet domain processing causes some image details to disappear, and the image domain is used to recover these details. On the other hand, due to the lack of wavelet transform support, the "image domain only" method is poor in detail representation. It can be observed that the model in the embodiment of the present invention achieves better quantitative indicators than the "image domain only" method, such as PSNR, SSIM, and RMSE in Table 2.

[0164] The adaptive wavelet module enhances the network's ability to depict metal artifacts in this embodiment of the invention, and experiments were conducted to verify the effectiveness of this module. By using the Daubechies wavelet (db3, filter length 3) to reconstruct the network and comparing its trained model with the adaptive wavelet, Table 2 shows that the adaptive wavelet module is more stable when facing various metal artifacts. Furthermore, the network achieves better generalization performance after applying the adaptive module. Figure 11 The visual comparison shows that the adaptive wavelet is superior to the ordinary wavelet in terms of detail representation.

[0165] The adaptive iterative learning model established based on the wavelet transform iterative optimization algorithm in this embodiment of the invention has clearer interpretability. The effect of iteration in the algorithm is obvious and can be verified on the generated dataset through corresponding ablation experiments. The verification results are shown in Table 3, which illustrates the impact of the iteration number T on network performance in this embodiment of the invention. It should be noted that network performance refers to PSNR(dB) / SSIM / RMSE.

[0166] Table 3

[0167]

[0168] As shown in Table 2, the quantitative indicators improved with increasing iteration number T, although the improvement was not significant. The effect of the iteration number was more pronounced when T was small.

[0169] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0170] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0171] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0172] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0173] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0174] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0175] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0176] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0177] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for removing metal artifacts from CT images, characterized in that, include: Construct an initial adaptive iterative learning model based on wavelet transform; The CT image expression is determined based on the clean image region, binary non-metallic region, and metallic artifact region of the CT image; The first optimization objective function is constructed based on the CT image expression, regularization term, and image wavelet transform, including: An initial optimization objective function is constructed based on the CT image expression; the initial optimization objective function is adjusted by regularization terms and wavelet transform to obtain an intermediate optimization objective function; the intermediate optimization objective function is optimized by error feedback terms to obtain a first optimization objective function. The expression for the initial optimization objective function is: ； in, It is the Frobenius norm; It is to optimize the target threshold; Indicates a clean image region; Represents the region of the original CT image; yes The binary nonmetallic region; This refers to the area of ​​metal artifacts in CT images; It is element-wise multiplication; The expression for the first optimization objective function is: ； in, It is an image with metal artifacts removed; For adaptive wavelet transform; 、 、 All are regularized weights; 、 and All are regularization terms; The first optimization objective function is solved by combining the proximal gradient descent algorithm and Taylor's formula, and the first calculation result is obtained; Based on the first calculation result, the initial adaptive iterative learning model is optimized by combining the L2 output loss and the ground real image to obtain the target adaptive iterative learning model. The target adaptive iterative learning model is used to decompose and stitch the CT images to remove metal artifacts.

2. The method for removing metal artifacts from CT images according to claim 1, characterized in that, The CT image expression is determined based on the clean image region, the binary non-metallic region, and the metallic artifact region of the CT image. The CT image expression is as follows: ； in, Indicates a clean image region; Represents the region of the original CT image; yes The binary nonmetallic region; This refers to the area of ​​metal artifacts in CT images; It is element-wise multiplication.

3. The method for removing metal artifacts from CT images according to claim 1, characterized in that, The step of solving the first optimization objective function by combining the proximal gradient descent algorithm and Taylor formula to obtain the first calculation result includes: The expressions for the metal artifact region and the clean image region are determined based on the first optimization objective function. Based on the expression for the metal artifact region, the first iterative equation for solving the metal artifact region in the wavelet domain is determined by the near-end gradient descent algorithm and Taylor formula. Based on the expression for the clean image region, the second iterative equation for solving the clean image region in the wavelet domain is determined by the proximal gradient descent algorithm and the Taylor formula. A third iterative equation is determined based on the first and second iterative equations; wherein, the third iterative equation is an iterative equation for the intermediate image after removing metal artifacts; The first iterative equation, the second iterative equation, and the third iterative equation of the metal artifact region are used as the first calculation result.

4. The method for removing metal artifacts from CT images according to claim 3, characterized in that, The first iterative equation is: ； in, This is the first step in solving for the metal artifact region. The result of the next iteration; Approximation Taylor polynomials; It is an adaptive wavelet transform; It is calculated using the inverse wavelet transform; This indicates that the metal artifact region is solved iteratively in the wavelet domain; It is the near-end gradient operator for calculating the region of metal artifacts; and These are trainable parameters; The second iterative equation is: ； in, It is the first step in solving for a clean image region. The result of the next iteration; Approximation Taylor polynomials; This indicates that the clean image region is solved iteratively in the wavelet domain; It is a proximal gradient operator for calculating clean image regions; and These are trainable parameters; The third iterative equation is: ； in, This is the first step in solving the image with metal artifacts removed. The result of the next iteration; Approximation Taylor polynomials; It is the proximal gradient operator for calculating an image with metal artifacts removed; and These are trainable parameters.

5. The method for removing metal artifacts from CT images according to claim 1, characterized in that, The step of optimizing the initial adaptive iterative learning model based on the first calculation result, combined with the L2 output loss and the ground real image, to obtain the target adaptive iterative learning model includes: Obtain the wavelet components of the metal artifact region, the clean image region, and the first image region in the CT image; The first optimization objective function is solved based on the wavelet components and the first calculation result to obtain the intermediate adaptive iterative learning model. The intermediate adaptive iterative learning model in the iterative process is evaluated using L2 output loss and ground truth images; The process of iteratively executing the step of solving the first optimization objective function based on the wavelet components and the first calculation result to obtain an intermediate adaptive iterative learning model continues until the evaluation result meets a predetermined threshold, thereby obtaining the target adaptive iterative learning model.

6. A system for removing metal artifacts from CT images, characterized in that, include: The first module is used to construct an initial adaptive iterative learning model based on wavelet transform; The second module is used to determine the CT image expression based on the clean image region, the binary non-metal region, and the metal artifact region of the CT image. The third module is used to construct a first optimization objective function based on the CT image expression, a regularization term, and an image wavelet transform. Specifically, the third module is used to: construct an initial optimization objective function based on the CT image expression; adjust the initial optimization objective function through a regularization term and a wavelet transform to obtain an intermediate optimization objective function; and optimize the intermediate optimization objective function through an error feedback term to obtain the first optimization objective function. The expression for the initial optimization objective function is: ； in, It is the Frobenius norm; It is to optimize the target threshold; Indicates a clean image region; Represents the region of the original CT image; yes The binary nonmetallic region; This refers to the area of ​​metal artifacts in CT images; It is element-wise multiplication; The expression for the first optimization objective function is: ； in, It is an image with metal artifacts removed; For adaptive wavelet transform; 、 、 All are regularized weights; 、 and All are regularization terms; The fourth module is used to solve the first optimization objective function by combining the proximal gradient descent algorithm and Taylor formula to obtain the first calculation result; The fifth module is used to optimize the initial adaptive iterative learning model based on the first calculation result, combined with the L2 output loss and the real ground image, to obtain the target adaptive iterative learning model; The sixth module is used to decompose and stitch the CT image using the target adaptive iterative learning model to remove metal artifacts from the CT image.

7. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 5.

Images