An artifact migration method for generating paired data based on real metal projection maps

Abstract

The application discloses a kind of based on real metal projection graph to generate the artifact migration method of matching data, including the following steps: from the image affected by artifact with threshold segmentation way extraction metal region, then obtain metal track image by forward projection;Using LI algorithm in the projection graph affected by artifact, in accordance with metal track, linear interpolation is obtained in corresponding area to remove metal track preprocessed sinogram;By subtracting LI sinogram from the sinogram affected by artifact to obtain residual sinogram containing only metal projection;Residual sinogram is added with artifact-free sinogram, and new sinogram affected by artifact is synthesized;Using step S4 synthesized sinogram containing artifact to reconstruct CT image containing metal artifact.The application provides data generation method, overcomes the negative influence brought by network training in traditional metal artifact removal model based on deep learning, unknown X-ray spectrum, unknown metal material and different detector sensitivity and the like.

Description

Technical Field

[0001] This invention relates to the field of image processing technology in computed tomography, and specifically to a method for metal artifact migration based on real metal projection images to generate paired data. Background Technology

[0002] In CT imaging, the presence of metallic objects often leads to severe metal artifacts in CT images, which can significantly degrade image quality and negatively impact clinical diagnosis and dosage calculations in radiotherapy. Over the past few decades, numerous methods for reducing metal artifacts have been proposed. These methods can be broadly categorized into six types: metal implant optimization methods, acquisition process optimization methods, physical model-based preprocessing methods, projection domain interpolation methods, iterative reconstruction methods, and image post-processing methods. Recently, with the rise of deep learning methods, many deep learning-based metal artifact removal models have been proposed, bringing revolutionary breakthroughs to CT metal artifact removal. Among these, supervised models demonstrate particularly outstanding performance in metal artifact removal, exhibiting excellent capabilities in both artifact removal and image detail restoration. However, in practical applications, the lack of effectively paired data severely limits the practicality of supervised models.

[0003] Current practices typically involve simulating artifact images from phantoms or clinical CT data using idealized physical models, assuming that parameters such as the energy spectrum of the incident X-ray source and the attenuation coefficient of the metal material are known. However, in real clinical scenarios, due to various practical limitations, such as unknown X-ray spectra, unknown metal materials, and varying detector sensitivities, this synthetic method struggles to generate artifact images that closely resemble those found in clinical settings. Therefore, to develop a supervised metal artifact removal method suitable for real-world scenarios, a solution is needed to address the lack of paired training datasets. The core task of this solution is to generate synthetic artifact images that closely resemble the features of metal artifact images in real-world scenarios. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies by proposing a paired data generation method based on real artifact data in a deep learning-based metal artifact removal model. This method uses deep learning algorithms to remove metal artifacts from CT images, thereby obtaining a set of paired data that includes both metal artifacts and those without them.

[0005] Technical solution: This invention provides a method for generating paired data artifact migration based on real metal projection images, comprising the following steps:

[0006] S1: Extract the metal region from the image affected by artifacts using threshold segmentation, and then obtain the metal trajectory image through forward projection;

[0007] S2: Using the LI algorithm, linear interpolation is performed in the corresponding region of the projection map affected by artifacts to obtain a preprocessed sine curve with the metal trajectory removed.

[0008] S3: Obtain a residual sine map containing only the metal projection by subtracting the preprocessed sine map from the sine map affected by artifacts;

[0009] S4: Add the residual sine plot to the artifact-free sine plot to synthesize a sine plot affected by artifacts;

[0010] S5: Use the artifact-containing sine wave synthesized in step S4 to reconstruct the CT image containing the metal artifact.

[0011] The present invention further defines the technical solution. The aforementioned artifact migration method based on real metal projection images to generate paired data can also be applied to the synthesis of metal artifacts in cone-beam CT and various other real-world scenarios such as fan-beam CT, and has a wide range of applications.

[0012] Furthermore, in the aforementioned artifact migration method based on real metal projection images to generate paired data, the reconstruction in step S5 uses the FBP method or FDK method to reconstruct the image, thereby generating a paired dataset with extremely high similarity to real metal artifacts.

[0013] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows: The paired data generation method based on real artifact data in the deep learning metal artifact removal model provided by the present invention overcomes the negative impact of unknown X-ray spectra, unknown metal materials and different detector sensitivities on network training in traditional deep learning-based metal artifact removal models; The metal artifact transmission method of the present application generates metal artifact images similar to those in real application scenarios, generating synthetic artifact CT images, making the synthetic artifact CT images very similar to the artifacts in real images. Attached Figure Description

[0014] Figure 1 This is a flowchart of artifact migration in a real-world scenario according to the present invention.

[0015] Figure 2 This is an example of synthesized artifact images in a real-world scenario according to the present invention.

[0016] Figure 3 In this embodiment, conventional methods are used to synthesize metal artifacts on Micro CT images.

[0017] Figure 4 This is a qualitative comparison of two different methods of synthesizing artifacts in the embodiments on real cone-beam MicroCT data; the image display window is [-330 700]HU.

[0018] Figure 2 (a) A true CT image with metal artifacts, (b) A true CT image without artifacts, and (c) A new synthetic CT image with artifacts obtained by transferring the artifacts from (a) to (b).

[0019] Figure 3 In the image, (a) is an image containing metal artifacts, and (b) is the ground truth image corresponding to (a); the image display window is [-330 700] HU.

[0020] Figure 4 The first column is the real artifact image, the third column is the training result of the paired data generated based on the traditional numerical simulation method, and the fourth column is the training result of the paired data generated based on the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the embodiments described.

[0022] I. Traditional Artifact Removal Methods

[0023] The paper CNNMAR provides a relatively classic synthesis scheme, which synthesizes metal artifacts by simulating beam hardening and Poisson noise in CT projection, and then generates paired data in simulation experiments. The specific process is as follows: First, prepare an artifact-free image x and an image x containing only simulated metal. m Then, a soft thresholding method is used to decompose the artifact-free image x into two independent parts, water x. w and skeleton x b If we assume the energy distribution of the X-rays is I(E), then for a given X-ray path L, the images of water, bone, and metal will show the energy distribution. i The linear integral is and Path L i The analog projection on is written as:

[0024]

[0025] Where m w (E), m b (E) and m m (E) represent the mass decay coefficients of water, bone, and metal at energy E, respectively. E0 is the equivalent monoenergetic energy of the X-ray. By removing the metal-related terms from the exponential function, the metal-free projection p can be obtained. i Finally, by adding Poisson noise to the projected image and using the Filter Back Projection (FBP) method, a set of paired data containing both metal artifacts and those without metal artifacts is obtained.

[0026] As described above regarding the generation process of the simulated paired data, it's clear that the process is based on several assumptions. For example, it requires knowledge of the X-ray energy distribution and equivalent monochromatic energy, as well as the mass attenuation coefficients of bone and metal. Appropriate thresholds also need to be set to separate water and bone. However, in actual clinical applications, it's difficult to accurately obtain these parameters, and the actual parameters often differ significantly from the simulated parameters.

[0027] II. Method of this embodiment

[0028] This embodiment proposes an artifact migration method for generating paired data based on real metal projection images. The artifact migration process is as follows: Figure 1 As shown, the first step extracts the metal region from the artifact-affected image using thresholding segmentation, and then obtains a metal trace image through forward projection. The second step uses the LI algorithm to linearly interpolate the corresponding regions in the artifact-affected projection image according to the metal traces to obtain a preprocessed sine map with the metal traces removed. The third step obtains a residual sine map containing only the metal projection by subtracting the LI sine map from the artifact-affected sine map. The fourth step adds the residual sine map to the artifact-free sine map to synthesize the artifact-affected sine map. Finally, these synthesized artifact-containing sine maps are used to reconstruct the CT image containing metal artifacts.

[0029] The theoretical basis for using this method for artifact migration is as follows:

[0030] Assuming formula (1) It is s(E), It is m(E), in the sine curve affected by metal artifacts. In the image, pixels representing metallic traces can be written as...

[0031]

[0032] The corresponding pixel in the artifact-free sine wave P0 can be written as

[0033]

[0034] Assuming another artifact projection with the same metal implant is Then the pixel x in this image can be represented as

[0035]

[0036] Its corresponding artifact-free projection in the corresponding artifact-free sine curve P1 is:

[0037]

[0038] Subtract p from p(x) a(x), we can obtain

[0039]

[0040] Where I s (E)=I(E)exp(-s(E)) can be regarded as the energy distribution of X-rays passing through an artifact-free image.

[0041] Similarly, we can obtain

[0042]

[0043] When two samples, P0 and P1, produce similar projection maps, it can be assumed that the energy spectrum distribution of X-rays passing through these two samples is almost identical, but the magnitudes may differ. Assume I... s ′(E)≈βI s (E), from which we can obtain equation (7).

[0044]

[0045] Formula (7) is the complete process of generating paired data described above. In real-world applications, p0(x) cannot be obtained, so the LI sine curve above is used instead. Figure 2 This paper demonstrates how the proposed artifact transfer method can be used to transfer metal artifacts from a real artifact-containing CT image to an artifact-free image, generating a synthetic artifact-laden CT image. It can be seen that the synthetic artifact-laden CT image is very similar to the artifacts in the real image.

[0046] III. Performance Validation of Paired Datasets Obtained Based on Different Synthesis Methods

[0047] To verify the performance of different synthesis methods in creating paired training datasets for training networks, this embodiment conducted experiments using two different datasets. Dataset 1 simulates metal artifacts on MicroCT data using the traditional method proposed in the paper CNNMAR. Dataset 2 uses the artifact transfer method proposed in this embodiment to construct a paired training dataset. Five simulated artifact examples from Dataset 1 are shown below. Figure 3 As shown, there is still a gap between the model and real artifacts. The same model was trained on the two different datasets mentioned above and validated on the Micro CT dataset containing real artifacts to test its practicality in real-world applications. A qualitative comparison method was used to evaluate its performance. Figure 4As shown, models trained on datasets constructed using traditional methods cannot completely reduce artifacts; they remain partially affected by artifacts and are more prone to generating false content information. Models trained on datasets constructed using artifact transfer techniques significantly reduce artifacts, resulting in visually more realistic images. Therefore, it can be demonstrated that constructing paired training datasets using artifact transfer can better assist network training in eliminating metallic artifacts and preserving image details.

[0048] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for generating artifact migration of paired data based on real metal projection images, characterized in that... Includes the following steps: S1: Extract the metal region from the image affected by artifacts using threshold segmentation, and then obtain the metal trajectory image through forward projection; S2: Using the LI algorithm, linear interpolation is performed in the corresponding region of the projection map affected by artifacts to obtain a preprocessed sine curve with the metal trajectory removed. S3: Obtain a residual sine map containing only the metal projection by subtracting the preprocessed sine map from the sine map affected by artifacts; S4: Add the residual sine plot to the artifact-free sine plot to synthesize a new sine plot affected by artifacts; S5: Use the artifact-containing sine wave synthesized in step S4 to reconstruct the CT image containing the metal artifact.

2. The artifact migration method for generating paired data based on real metal projection images according to claim 1, characterized in that: This method can be used to synthesize metal artifacts in cone-beam CT.

3. The artifact migration method for generating paired data based on real metal projection images according to claim 1, characterized in that: The reconstruction in step S5 uses the FBP or FDK method to reconstruct the image, thereby generating a paired dataset with extremely high similarity to real metal artifacts.

Images