A CT Metal Artifact Correction Method Based on Physical Dual Regression Learning

Abstract

A CT metal artifact correction method based on physical dual regression learning corrects uncorrected projection data to obtain a corrected image through nine steps. This CT metal artifact correction algorithm and system based on physical dual regression learning enhances metal artifact correction performance and improves the algorithm's stability and reliability. The beneficial effects of this invention are: 1. The metal artifact correction results are more stable and reliable, and the learning process is more stable. 2. By modeling the physical mechanism that generates metal artifacts during the learning process, the mapping learned by the model can be effectively constrained to be consistent with the real imaging system.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and computed tomography (CT) technology, and in particular to a CT metal artifact correction method based on physical dual regression learning. Background Technology

[0002] X-ray computed tomography (CT) provides three-dimensional anatomical information of patients and is an indispensable imaging tool in modern medicine. However, when highly attenuating substances such as metals are present in the patient's body, severe metallic artifacts, such as stripes, bands, or shadows, will appear in the reconstructed images due to physical mechanisms such as beam hardening, scattering, photon starvation, and nonlinear partial volume effects. Metallic artifacts cause signal distortion, create false structures in the image, and mask real lesions, seriously affecting subsequent clinical diagnosis. Metal artifact correction (MAR) is one of the most challenging and long-standing problems in CT imaging.

[0003] Traditional metal artifact correction algorithms either treat the projected data at the metal trajectory as missing data and fill in the missing portion; or correct the physical mechanism that causes metal artifacts; or use iterative reconstruction algorithms for reconstruction. However, traditional metal artifact correction algorithms struggle to handle complex metal artifacts, often producing oversmoothing effects or introducing secondary artifacts in the image. Deep learning-based metal artifact correction algorithms leverage the feature learning capabilities of neural networks, exhibiting significant advantages in prior image learning, image anatomical structure restoration, and artifact suppression. However, existing deep learning-based metal artifact correction algorithms still have some shortcomings. First, current algorithms only learn the mapping from data containing metal projections to data without metal projections, or to images without metal artifacts, based on limited training data. This process learns a one-way mapping from one data domain to another, lacking constraints on the model output, leading to insufficient stability and reliability of the reconstruction results. Second, current deep learning-based metal artifact correction models lack modeling of the corresponding physical mechanisms during the modeling process. When using neural networks to learn the mapping from data with metallic projections to data without metallic projections, or images without artifacts, multiple possible mappings exist due to the limited quantity and quality of training data. Traditional deep learning-based metal artifact correction algorithms struggle to learn the optimal mapping, and the learning process has poor interpretability, resulting in low acceptance by clinicians.

[0004] Therefore, it is essential to provide a CT metal artifact correction method based on physical dual regression learning to address the shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a CT metal artifact correction method based on physical dual regression learning. This CT metal artifact correction algorithm and system based on physical dual regression learning can enhance metal artifact correction performance and improve the stability and reliability of the algorithm.

[0006] The above-mentioned objectives of the present invention are achieved through the following technical measures:

[0007] A method for correcting CT metal artifacts based on physical dual regression learning is provided, comprising the following steps:

[0008] S1. Reconstruct the uncorrected projection data to obtain the uncorrected reconstructed image;

[0009] S2. Use the trained first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in S1 to obtain a pre-restored image.

[0010] S3. Input the pre-recovered image obtained in S2 into the trained second image domain network for prediction to obtain the bone base map, water base map and metal base map;

[0011] S4. Perform front projection on the bone base map, water base map and metal base map obtained in S3 to obtain the corresponding bone base map projection data, water base map projection data and metal base map projection data.

[0012] S5. Segment the metal base map projection data obtained in S4 to obtain the metal projection trajectory map;

[0013] S6. Based on the uncorrected projection data of S1, the bone base map projection data, water base map projection data and metal base map projection data of S4, and the metal projection trajectory map of S5, photon hunger correction is performed through the trained photon hunger correction module to obtain the non-metal region projection data and the photon hunger corrected metal projection data.

[0014] S7. The photon-starved corrected metal projection data obtained in S6 is subjected to beam hardening correction by the trained beam hardening correction module, and combined with the non-metal region projection data obtained in S6 to obtain the beam hardening corrected projection data.

[0015] S8. Reconstruct the beam-hardened and corrected projection data obtained in S7 to obtain the beam-hardened and corrected image.

[0016] S9. The beam-hardened image obtained in S8 is subjected to scattering correction and denoising through the trained scattering correction and denoising module to finally obtain the corrected image.

[0017] This invention relates to a CT metal artifact correction method based on physical dual regression learning. A CT metal artifact correction model is constructed through dual learning, and the CT metal artifact correction model is used to jointly train a first image domain network, a second image domain network, a photon starvation correction module, a beam hardening correction module, a scattering correction and denoising module, a beam hardening module, a scattering signal module, and a photon starvation module.

[0018] Preferably, the joint training of the above-mentioned CT metal artifact correction module is carried out according to the following steps:

[0019] A1. Reconstruct the uncorrected projection data to obtain the uncorrected reconstructed image;

[0020] A2. Remove metal artifacts from the uncorrected reconstructed image obtained in A1 using the first image domain network to obtain the pre-restored image;

[0021] A3. Input the pre-recovered image obtained in A2 into the second image domain network for prediction to obtain the bone base map, water base map and metal base map;

[0022] A4. Perform front projection on the bone base map, water base map, and metal base map obtained in A3 to obtain the corresponding bone base map projection data, water base map projection data, and metal base map projection data;

[0023] A5. Segment the metal base map projection data obtained in A4 to obtain the metal projection trajectory map;

[0024] A6. Based on the uncorrected projection data of A1, the bone base map projection data, water base map projection data, and metal base map projection data of A4, and the metal projection trajectory map of A5, photon hunger correction is performed through the photon hunger correction module to obtain the non-metallic region projection data and the photon hunger corrected metal projection data.

[0025] A7. The photon-starved corrected metal projection data obtained in A6 is subjected to beam hardening correction by the beam hardening correction module and combined with the non-metal region projection data obtained in S6 to obtain the beam hardened corrected projection data.

[0026] A8. Reconstruct the beam-hardened and corrected projection data obtained in A7 to obtain the beam-hardened and corrected image;

[0027] A9. The beam-hardened image obtained in A8 is subjected to scattering correction and denoising through the scattering correction and denoising module to obtain the corrected image;

[0028] A10. Input the corrected image obtained from A9 into the second image domain network for prediction to obtain the bone base map, water base map and metal base map;

[0029] A11. Perform front projection on the bone base map, water base map and metal base map obtained in A10 to obtain the corresponding bone base map projection data, water base map projection data and metal base map projection data;

[0030] A12. The bone base map projection data, water base map projection data and metal base map projection data obtained in A11 are processed by the beam hardening module to obtain beam hardening projection data.

[0031] A13. The beam hardening projection data obtained in A12 is processed through the scattering signal module to obtain projection data containing scattering signals;

[0032] A14. The projection data containing the scattered signal obtained in A13 is processed through the photon starvation module to obtain the estimated projection data;

[0033] A15. Construct a consistency loss by combining the estimated projection data obtained from A14 with the uncorrected projection data from A1, and use this to constrain network training.

[0034] In S5 or A5, segmentation involves binarizing the metal base map projection data and setting values ​​greater than or equal to a specific threshold to 1 and values ​​less than the threshold to 0, where the threshold size th is given in advance.

[0035] The metal projection trajectory in S5 or A5 is obtained by equation (1):

[0036]

[0037] Where d metal This is the metal base map projection data, where x and y are coordinate indices, th is the threshold, and M is the base map projection data. proj This is a projection trajectory diagram of the metal.

[0038] In S6 or A6, the photon hunger correction method of the photon hunger correction module is performed by the following steps:

[0039] B1. Using the metal projection trajectory diagram M proj Extract uncorrected projection data p origin Non-metallic region projection data p non-metal and metal projection data p metal Among them, the projection data p of the non-metallic region non-meta With metal projection data p metal It is obtained from equations (2) and (3);

[0040] p non-metal =p origin ⊙(1-M proj ...Equation (2);

[0041] p metal =p origin ⊙Mproj ...Equation (3);

[0042] Where ⊙ represents the dot product operation;

[0043] B2. Using the calculated projection data p est Replace in metal projection data p metal Greater than a specific threshold th metal The data is finally obtained from equation (4);

[0044]

[0045] Where p metal-ps The metal projection data after photon starvation correction, th metal p is the threshold. est This is the projection data.

[0046] Preferably, the above threshold th metal Based on uncorrected projection data p origin The maximum value is selected, specifically the threshold th. metal It is obtained from equation (5);

[0047] th metal =0.8×max(p) origin ... Equation (5).

[0048] Preferably, the above projection data p est Based on the bone base map projection data of S4 or A4 d b Water base map projection data d w Projection data d with metal base map me The multi-energy projection data, which is synthesized by combining the material decay curves of metal, water, and bone with X-ray spectral data, is specifically obtained by equation (6).

[0049]

[0050] Where φ(E) represents the X-ray spectral data, E represents the X-ray energy, and m b (E) is the bone material decay curve, m w (E) is the decay curve of water, m me (E) is the material decay curve of the metal.

[0051] In S7 or A7, the beam hardening correction method of the beam hardening correction module is performed by the following steps:

[0052] C1, Metal projection data p after photon starvation correction metal-psc Perform tissue correction;

[0053] C2. Use a multilayer perceptron model (MLP) to learn the transformation from multi-energy projection data to single-energy projection data, then perform tissue inverse correction on the MLP output and compare it with the corresponding non-metallic region projection data p. non-meta Combined, the projection data after beam hardening correction is obtained, specifically obtained from equation (7);

[0054] p bb =p non-metal +f inv-tiss (MLP(f tiss (p metal-psc Equation (7)

[0055] Where f tiss For organizational correction, f inv-tiss For reverse correction of the organization.

[0056] Preferably, the above tissue correction is based on the metal projection data p after photon starvation correction. metal-psc Subtracting the projection data p of non-metallic substances contained in the middle tissu Specifically, this is done through equation (8):

[0057] f tiss (p metal-psc ) = p metal -p tissu ⊙M proj ...Equation (8);

[0058] Where M proj This is the metal projection trajectory diagram, and ⊙ represents the dot product operation.

[0059] Preferably, the above-mentioned tissue inverse correction involves adding non-metallic material projection data p to the output of the multilayer perceptron model (MLP). tissu Specifically, it is done through equation (9):

[0060] f inv-tiss (x)=x+p tis ⊙M proj ...Equation (9);

[0061] Where x is the output of the MLP.

[0062] Preferably, the above-mentioned non-metallic material projection data p tis Based on the bone base map projection data corresponding to S4 or A4 d b Water base map projection data d w The results, obtained by combining the decay curves of metals, water, and bone with X-ray spectral data, are specifically derived from equation (10):

[0063]

[0064] In A12, the beam hardening module processes the data based on the bone base map projection data d obtained in A11. b Water base map projection data d w Projection data d with metal base map me The process of synthesizing multi-energy projection data by combining the material decay curves of metal, water, and bone with X-ray spectral data through equation (6).

[0065] In A13, the scattering signal module is processed using a projection domain machine learning model, which learns the scattering projection signal. The input is beam-hardened projection data, and the output is the scattering projection signal. The modeling process is represented by equation (11):

[0066]

[0067] in For the projection domain machine learning model, p bbh-model For beam-hardened projection data, p scatter-model This is projection data containing scattered signals.

[0068] In A14, the processing method of the photon-starved module is to process the projection data p containing the scattered signal obtained in A13. scatter-mode Perform truncation.

[0069] Preferably, the above truncation process is to cut p scatter-model Greater than the threshold th metal The value is set to th metal , and th metal Take the maximum value from the uncorrected projection data.

[0070] In A15, the consistency loss is the estimated projection data p obtained from A14. dual The uncorrected projection data p of A1 otigin To maintain consistency, the corresponding loss is constructed using the mean squared error (MSE), as expressed by equation (12):

[0071]

[0072] This invention discloses a CT metal artifact correction method based on physical duality regression learning, comprising the following steps: S1, reconstructing uncorrected projection data to obtain an uncorrected reconstructed image; S2, removing metal artifacts from the uncorrected reconstructed image obtained in S1 using a trained first image domain network to obtain a pre-reconstructed image; S3, inputting the pre-reconstructed image obtained in S2 into a trained second image domain network for prediction to obtain a bone base map, a water base map, and a metal base map; S4, performing pre-projection on the bone base map, water base map, and metal base map obtained in S3 to obtain corresponding bone base map projection data, water base map projection data, and metal base map projection data; S5, segmenting the metal base map projection data obtained in S4 to obtain a metal projection trajectory map; S6, based on the uncorrected projection data in S1, S4... The bone base map projection data, water base map projection data, and metal base map projection data, as well as the metal projection trajectory map from S5, are subjected to photon starvation correction by a trained photon starvation correction module to obtain non-metallic region projection data and photon starvation corrected metal projection data. In S7, the photon starvation corrected metal projection data obtained in S6 is subjected to beam hardening correction by a trained beam hardening correction module and combined with the non-metallic region projection data obtained in S6 to obtain beam hardened corrected projection data. In S8, the beam hardened corrected projection data obtained in S7 is reconstructed to obtain a beam hardened corrected image. In S9, the beam hardened corrected image obtained in S8 is subjected to scattering correction and denoising by a trained scattering correction and denoising module to finally obtain the corrected image. Compared with the prior art, the beneficial effects of this invention are: 1. This invention is based on dual regression learning, therefore the metal artifact correction results are more stable and reliable, and the learning process is more stable. Compared to traditional metal artifact correction algorithms that only learn a one-way mapping from uncorrected projection data to the corrected image, this invention adds a consistency constraint between the corrected image and the uncorrected projection data, thereby ensuring the accuracy and stability of the corrected image. 2. This invention models the physical mechanism of metal artifact generation, thus constraining the model to learn the solution space of the mapping from uncorrected projection data to the corrected image. Due to the limited quantity and quality of training data, there are multiple possible mappings from uncorrected projection data to the corrected image. Traditional deep learning-based metal artifact correction algorithms struggle to learn the optimal mapping. This invention, by modeling the physical mechanism of metal artifact generation during the learning process, effectively constrains the model-learned mapping to be consistent with the real imaging system. Attached Figure Description

[0073] The invention will be further described with reference to the accompanying drawings, but the contents of the drawings do not constitute any limitation on the invention.

[0074] Figure 1 This is a flowchart of a CT metal artifact correction method based on physical dual regression learning.

[0075] Figure 2This is a flowchart for joint training.

[0076] Figure 3 For data or images in S1 or A1, where Figure 3 (a) is uncorrected projection data. Figure 3 (b) Uncorrected reconstructed image.

[0077] Figure 4 The pre-recovered image in S2 or A2.

[0078] Figure 5 For each base graph of S3 or A3, where Figure 5 (a) Bone base map, Figure 5 (b) is a water-based map. Figure 5 (c) is the metal base diagram.

[0079] Figure 6 For each base map projection data in S4 or A4, where Figure 6 (a) Bone base projection data, Figure 6 (b) Projection data for the water-based map Figure 6 (c) is the metal base map projection data.

[0080] Figure 7 This is an S5 or A5 metal projection trajectory diagram.

[0081] Figure 8 The projection data is for S6 or A6, where Figure 8 (a) shows the projection data for the non-metallic region. Figure 8 (b) shows the metal projection data after photon starvation correction.

[0082] Figure 9 The projection data is after beam hardening correction for S7 or A7.

[0083] Figure 10 Image after beam hardening correction for S8 or A8.

[0084] Figure 11 The corrected image in S9 or A9.

[0085] Figure 12 The image in A10, where Figure 12 (a) is a bone base map, Figure 12 (b) is a water-based map. Figure 12 (c) is the metal base diagram.

[0086] Figure 13 The projection data in A11, where Figure 13 (a) is the bone base map projection data, Figure 13 (b) Projection data for the water-based map Figure 13 (c) Metal base map projection data.

[0087] Figure 14 This contains beam-hardened projection data in A12.

[0088] Figure 15 This is the projection data containing the scattered signal in A13.

[0089] Figure 16 This is the estimated projection data in A14.

[0090] Figure 17 These are the images before and after correction in Example 2. Detailed Implementation

[0091] The technical solution of the present invention will be further described in conjunction with the following embodiments.

[0092] Example 1

[0093] A CT metal artifact correction method based on physical dual regression learning, such as Figure 1 As shown, it includes the following steps:

[0094] S1. Reconstruct the uncorrected projection data to obtain an uncorrected reconstructed image, such as... Figure 3 ;

[0095] S2. Using the trained first image domain network, remove metal artifacts from the uncorrected reconstructed image obtained in S1 to obtain a pre-restored image, such as... Figure 4 ;

[0096] S3. Input the pre-restored image obtained in S2 into the trained second image domain network for prediction, and obtain the bone base map, water base map, and metal base map, such as Figure 5 ;

[0097] S4. Perform front projection on the bone base map, water base map, and metal base map obtained in S3 to obtain the corresponding bone base map projection data, water base map projection data, and metal base map projection data, such as... Figure 6 ;

[0098] S5. Segment the metal base map projection data obtained in S4 to obtain the metal projection trajectory map, such as... Figure 7 ;

[0099] S6. Based on the uncorrected projection data of S1, the bone base map projection data, water base map projection data, and metal base map projection data of S4, and the metal projection trajectory map of S5, photon hunger correction is performed using the trained photon hunger correction module to obtain the corresponding non-metallic region projection data and the photon hunger-corrected metal projection data, such as... Figure 8 ;

[0100] S7. The photon-starved corrected metal projection data obtained in S6 is subjected to beam hardening correction by the trained beam hardening correction module, and then combined with the non-metallic region projection data obtained in S6 to obtain the beam-hardened corrected projection data, such as... Figure 9 ;

[0101] S8. Reconstruct the beam-hardened and corrected projection data obtained in S7 to obtain the beam-hardened and corrected image, such as... Figure 10 ;

[0102] S9. The beam-hardened corrected image obtained in S8 is then subjected to scattering correction and denoising through the trained scattering correction and denoising module to finally obtain the corrected image, such as... Figure 11 .

[0103] This invention relates to a CT metal artifact correction method based on physical dual regression learning. A CT metal artifact correction model is constructed through dual learning. The CT metal artifact correction model is then used to jointly train a first image domain network, a second image domain network, a photon starvation correction module, a beam hardening correction module, a scattering correction and denoising module, a beam hardening module, a scattering signal module, and a photon starvation module.

[0104] It should be noted that the characteristic of dual learning in this invention is the creation of a dual task for the original task, and the simultaneous learning of both tasks. In this invention, the original task is to learn the transformation or mapping from uncorrected projection data to the corrected image; the dual task learns the transformation or mapping from the corrected image to the uncorrected projection data. Because this invention is a CT metal artifact correction method based on physical dual regression learning, the "physical" aspect refers to modeling the physical mechanism of metal artifact generation in both tasks and constructing corresponding correction methods.

[0105] like Figure 2 As shown, the joint training of the dual models is performed according to the following steps:

[0106] A1. Reconstruct the uncorrected projection data to obtain an uncorrected reconstructed image, such as... Figure 3 ;

[0107] A2. Using the first image domain network, remove metal artifacts from the uncorrected reconstructed image obtained in A1 to obtain a pre-restored image, such as... Figure 4 ;

[0108] A3. Input the pre-restored image obtained in A2 into the second image domain network for prediction to obtain the bone base map, water base map, and metal base map, as follows: Figure 5 ;

[0109] A4. Perform front projection on the bone base map, water base map, and metal base map obtained in A3 to obtain the corresponding bone base map projection data, water base map projection data, and metal base map projection data, such as... Figure 6 ;

[0110] A5. The metal base map projection data obtained in A4 is segmented by a threshold to obtain the metal projection trajectory map, as shown below. Figure 7 ;

[0111] A6. Based on the uncorrected projection data of A1, the bone base map projection data, water base map projection data, and metal base map projection data of A4, and the metal projection trajectory map of A5, photon starvation correction is performed using the photon starvation correction module to obtain the corresponding non-metallic region projection data and the photon starvation corrected metallic projection data, such as... Figure 8 ;

[0112] A7. The photon-starved corrected metal projection data obtained in A6 is subjected to beam hardening correction by the beam hardening correction module, and then combined with the non-metallic region projection data obtained in S6 to obtain the beam-hardened corrected projection data, such as... Figure 9 ;

[0113] A8. Reconstruct the beam-hardened and corrected projection data obtained in A7 to obtain the beam-hardened and corrected image, such as... Figure 10 ;

[0114] A9. The beam-hardened image obtained from A8 is then subjected to scattering correction and denoising through a scattering correction and denoising module to obtain the corrected image, as shown below. Figure 11 ;

[0115] A10. Input the corrected image obtained from A9 into the second image domain network for prediction to obtain the bone base map, water base map, and metal base map, as follows: Figure 12 ;

[0116] A11. Perform front projection on the bone base map, water base map, and metal base map obtained from A10 to obtain the corresponding bone base map projection data, water base map projection data, and metal base map projection data, such as... Figure 13 ;

[0117] A12. The bone base map projection data, water base map projection data, and metal base map projection data obtained in A11 are processed by the beam hardening module to obtain beam hardening projection data, such as... Figure 14 ;

[0118] A13. The beam-hardening projection data obtained in A12 is processed through the scattering signal module to obtain projection data containing scattering signals, such as... Figure 15 ;

[0119] A14. Perform photon starvation modeling on the projection data containing the scattered signal obtained in A13 to obtain estimated projection data, such as... Figure 16 ;

[0120] A15. Construct a consistency loss by combining the estimated projection data obtained from A14 with the uncorrected projection data from A1, and use this to constrain network training.

[0121] It should be noted that the non-metallic region projection data in S6 and A6 of this invention refers to the portion of the projection data that does not contain metal. The calculation process involves forward projection of the metal mask to obtain its projection trajectory. The data at this trajectory is the metallic region projection data, while other regions are the non-metallic region projection data. In the metallic region projection data, the metallic region represents a spatial location. This data includes not only the projection information of the metal but also the projection information of soft tissues such as bone and fat. All projection data excluding metals are collectively referred to as non-metallic material projection data.

[0122] In this invention, both the first and second image domain networks are U-Net models, with the first image domain network having 1 output channel and the second image domain network having 3 output channels. The number of training iterations for joint training is set according to actual needs, and the number of training iterations is determined comprehensively based on factors such as the amount of data and the network size. Training is considered complete when the set number of iterations is reached. In this embodiment, the number of joint training iterations is specifically 100. Each joint training iteration is detailed as described above.

[0123] In S5 or A5, segmentation involves binarizing the metal base map projection data and setting values ​​greater than or equal to a specific threshold to 1 and values ​​less than the threshold to 0, where the threshold value th is given in advance.

[0124] The metal projection trajectory in S5 or A5 is obtained by equation (1):

[0125]

[0126] Where d metal This is the metal base map projection data, where x and y are coordinate indices, th is the threshold, and M is the base map projection data. proj This is a projection trajectory diagram of the metal.

[0127] The photon starvation correction method in S6 or A6 is performed by the following steps:

[0128] B1. Using the metal projection trajectory diagram M proj Extract uncorrected projection data p origin Non-metallic region projection data p non-metal and metal projection data p metal Among them, the projection data p of the non-metallic region non-metal With metal projection data p metal It is obtained from equations (2) and (3);

[0129] p non-metal =p origin ⊙(1-Mproj ...Equation (2);

[0130] p metal =p origin ⊙M proj ...Equation (3);

[0131] Where ⊙ represents the dot product operation;

[0132] B2. Using the calculated projection data p est Replace in metal projection data p metal Greater than a specific threshold th metal The data is finally obtained from equation (4);

[0133]

[0134] Where p metal-psc The metal projection data after photon starvation correction, th metal p is the threshold. est For projection data;

[0135] Threshold th metal Based on uncorrected projection data p origin The maximum value is selected, specifically the threshold th. metal It is obtained from equation (5);

[0136] th metal =0.8×max(p origin ...Equation (5);

[0137] Projection data p est Based on the bone base map projection data of S4 or A4 d b Water base map projection data d w Projection data d with metal base map me The multi-energy projection data, which is synthesized by combining the material decay curves of metal, water, and bone with X-ray spectral data, is specifically obtained by equation (6).

[0138]

[0139] Where φ(E) represents the X-ray spectral data, E represents the X-ray energy, and m b (E) is the bone material decay curve, m w (E) is the decay curve of water, m me (E) is the material decay curve of the metal.

[0140] It should be noted that the material decay curves of bone, water, and metal in this invention were obtained using the simulation software XCIST.

[0141] The beam hardening correction method in S7 or A7 is performed by the following steps:

[0142] C1, Metal projection data p after photon starvation correction metal-psc Perform tissue correction;

[0143] C2. Use a multilayer perceptron model (MLP) to learn the transformation from multi-energy projection data to single-energy projection data, then perform tissue inverse correction on the MLP output and compare it with the corresponding non-metallic region projection data p. non-metal Combined, the projection data after beam hardening correction is obtained, specifically obtained from equation (7);

[0144] p bbh =p non-metal +f inv-tiss (MLP(f tiss (p metal-psc Equation (7)

[0145] Where f tiss For organizational correction, f inv-tiss For reverse correction of the organization.

[0146] Tissue correction is derived from metal projection data p after photon starvation correction. metal-ps Subtracting the projection data p of non-metallic substances contained in the middle tissue Specifically, this is done through equation (8):

[0147] f tiss (p metal-psc ) = p metal-ps -p tissue ⊙M proj ...Equation (8);

[0148] Where M proj This is the metal projection trajectory diagram, and ⊙ represents the dot product operation.

[0149] Tissue inverse correction involves adding non-metallic material projection data p to the output of a multilayer perceptron (MLP) model. tissue Specifically, it is done through equation (9):

[0150] f inv-tiss (x)=x+p tissue ⊙M proj ...Equation (9);

[0151] Where x is the output of the MLP.

[0152] Non-metallic material projection data p tissue Based on the bone base map projection data corresponding to S4 or A4 d b Water base map projection data d w The results, obtained by combining the decay curves of metals, water, and bone with X-ray spectral data, are specifically derived from equation (10):

[0153]

[0154] In S9 and A9 of this invention, scattering correction and denoising refer to the use of arbitrary image domain machine learning models, including but not limited to convolutional neural networks and graph neural networks, to remove residual noise and scattering artifacts in the image after beam hardening correction. The input is the image after beam hardening correction, and the output is the corrected image.

[0155] In A12, beam hardening is modeled as the bone base map projection data d obtained from A11. b Water base map projection data d w Projection data d with metal base map me The process of synthesizing multi-energy projection data by combining the material decay curves of metal, water, and bone with X-ray spectral data through equation (6).

[0156] In A13, the scattered signal is modeled as a projection domain machine learning model, which learns the scattered projection signal; the input is beam-hardened projection data, and the output is the scattered projection signal. The modeling process is represented by equation (11):

[0157]

[0158] in For the projection domain machine learning model, p bbh-model For beam-hardened projection data, p scatter-mode This is projection data containing scattering signals.

[0159] It should be noted that the projection domain machine learning model of the present invention can be a convolutional neural network or a graph neural network, and the projection domain machine learning model in this embodiment is specifically a convolutional neural network.

[0160] In A14, photon hunger is modeled as a projection of the scattered signal data p obtained in A13. scatter-mod Perform truncation. Truncation involves setting p... scatter-mode Greater than the threshold th metal The value is set to th metal , and th metal Take the maximum value from the uncorrected projection data.

[0161] In A15, the consistency loss is the estimated projection data p obtained from A14. dual The uncorrected projection data p of A1 origin To maintain consistency, the corresponding loss is constructed using the mean squared error (MSE), as expressed by equation (12):

[0162]

[0163] It should be noted that, before joint training, the first image domain network and the second image domain network can be pre-trained separately, with each image domain network undergoing 50 epochs of pre-training and a learning rate starting from 1×10⁻⁶. -4 linearly decreases to 1×10 -6 The Adam optimizer is used, with parameters (β1, β2) = (0.9, 0.999). In the pre-training of the first image domain network, the input is an image with artifacts. In the pre-training of the second image domain network, the input is an image with and without artifacts, and the output is a bone base map, a water base map, and a metal base map. Alternatively, this invention can also directly train the first and second image domain networks together with other modules without pre-training.

[0164] The joint training of this invention can also introduce consistency loss, and further add loss constraints to the outputs of the first image domain network and the second image domain network, summing all loss functions to constrain network training. Specifically, the loss function used is MSE loss, expressed by equation (13):

[0165]

[0166] Where x is the output of the first image domain network or the second image domain network, and y is the corresponding target image. The target image of the first image domain network is an artifact-free image, and the target image of the second image domain network is a target bone base map, a target water base map, and a target metal base map.

[0167] It should be noted that the corrected projection data and estimated projection data in this invention refer to the projection data after logarithmic transformation. The bone base map, water base map, and metal base map refer to the bone density image, water density image, and metal density image obtained by decomposing the CT attenuation image.

[0168] In this invention, the reconstruction methods for S1, S8, A1, and A8 can be the FBP algorithm, the 3D spiral reconstruction algorithm, or the Feldkamp-Davis-Kress (FDK) algorithm; the front projections for S4, A4, and A1 can be either 2D or 3D front projections; the metal artifact correction in this invention can be either 2D fan-beam CT metal artifact correction or 3D cone-beam CT metal artifact correction, depending on the specific circumstances.

[0169] This CT metal artifact correction algorithm and system based on physical dual regression learning enhances metal artifact correction performance and improves algorithm stability and reliability. Compared with existing technologies, the advantages of this invention are: 1. This invention is based on dual regression learning, thus the metal artifact correction results are more stable and reliable, and the learning process is more stable. Compared with traditional metal artifact correction algorithms that only learn the unidirectional mapping from uncorrected projection data to the corrected image, this invention adds consistency constraints between the corrected image and the uncorrected projection data, thereby ensuring the accuracy and stability of the corrected image. 2. This invention models the physical mechanism of metal artifact generation, thereby constraining the model to learn the solution space of the mapping from uncorrected projection data to the corrected image. Due to the limited quantity and quality of training data, there are multiple possible mappings from uncorrected projection data to the corrected image. Traditional deep learning-based metal artifact correction algorithms struggle to learn the optimal mapping. This invention, by modeling the physical mechanism of metal artifact generation during the learning process, can effectively constrain the model-learned mapping to be consistent with the real imaging system.

[0170] Example 2

[0171] A CT metal artifact correction method based on physical dual regression learning is presented, with other features identical to Example 1, except for the following: In this example, the training data for joint training contains 27,948 images, of which 15,000 are head images and the remainder are chest and abdominal images. The test data includes 1,000 images, of which 136 are head images and the remainder are chest and abdominal images. The metal materials in the training data include stainless steel, titanium, amalgam, copper, and cobalt. The simulated scanning voltage is 120 kVp, the current is 500 mA, and Poisson noise and Gaussian noise are added during the simulation.

[0172] In S5 or A5, the threshold value is 0.001. The training iterations in this embodiment are 100.

[0173] in Figure 17 Images of the head and chest / abdomen before correction, and images of the head and chest / abdomen after correction using the CT metal artifact correction method based on physical dual regression learning of this invention. Figure 17 As can be seen, before metal artifact correction, artifacts in the image cause signal distortion and produce pseudo-structures. These artifacts or pseudo-structures overlap, obscure, or alter the shape of real tissue structures, making it difficult for doctors to correctly interpret the images. After processing with the CT metal artifact correction method based on physical duality regression learning of this invention, Figure 17The image underwent significant changes. In the corrected image, artifacts caused by metal were effectively suppressed or eliminated, the overall uniformity and contrast of the image were significantly improved, and the detailed structure of soft tissue became clearer.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A CT metal artifact correction method based on physical duality regression learning, characterized in that, Includes the following steps: S1. Reconstruct the uncorrected projection data to obtain the uncorrected reconstructed image; S2. Use the trained first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in S1 to obtain a pre-restored image. S3. Input the pre-recovered image obtained in S2 into the trained second image domain network for prediction to obtain the bone base map, water base map and metal base map; S4. Perform front projection on the bone base map, water base map and metal base map obtained in S3 to obtain the corresponding bone base map projection data, water base map projection data and metal base map projection data. S5. Segment the metal base map projection data obtained in S4 to obtain the metal projection trajectory map; S6. Based on the uncorrected projection data of S1, the bone base map projection data, water base map projection data and metal base map projection data of S4, and the metal projection trajectory map of S5, photon hunger correction is performed through the trained photon hunger correction module to obtain the non-metal region projection data and the photon hunger corrected metal projection data. S7. The photon-starved corrected metal projection data obtained in S6 is subjected to beam hardening correction by the trained beam hardening correction module, and combined with the non-metal region projection data obtained in S6 to obtain the beam hardening corrected projection data. S8. Reconstruct the beam-hardened and corrected projection data obtained in S7 to obtain the beam-hardened and corrected image. S9. The beam-hardened image obtained in S8 is subjected to scattering correction and denoising through the trained scattering correction and denoising module to finally obtain the corrected image.

2. The CT metal artifact correction method based on physical duality regression learning according to claim 1, characterized in that: A CT metal artifact correction model is constructed through dual learning. The first image domain network, the second image domain network, the photon starvation correction module, the beam hardening correction module, the scattering correction and denoising module, the beam hardening module, the scattering signal module, and the photon starvation module are jointly trained through the CT metal artifact correction model. The joint training of the CT metal artifact correction module is carried out according to the following steps: A1. Reconstruct the uncorrected projection data to obtain the uncorrected reconstructed image; A2. Remove metal artifacts from the uncorrected reconstructed image obtained in A1 using the first image domain network to obtain the pre-restored image; A3. Input the pre-recovered image obtained in A2 into the second image domain network for prediction to obtain the bone base map, water base map and metal base map; A4. Perform front projection on the bone base map, water base map, and metal base map obtained in A3 to obtain the corresponding bone base map projection data, water base map projection data, and metal base map projection data; A5. Segment the metal base map projection data obtained in A4 to obtain the metal projection trajectory map; A6. Based on the uncorrected projection data of A1, the bone base map projection data, water base map projection data, and metal base map projection data of A4, and the metal projection trajectory map of A5, photon hunger correction is performed through the photon hunger correction module to obtain the non-metallic region projection data and the photon hunger corrected metal projection data. A7. The photon-starved corrected metal projection data obtained in A6 is subjected to beam hardening correction by the beam hardening correction module and combined with the non-metal region projection data obtained in S6 to obtain the beam hardened corrected projection data. A8. Reconstruct the beam-hardened and corrected projection data obtained in A7 to obtain the beam-hardened and corrected image; A9. The beam-hardened image obtained in A8 is subjected to scattering correction and denoising through the scattering correction and denoising module to obtain the corrected image; A10. Input the corrected image obtained from A9 into the second image domain network for prediction to obtain the bone base map, water base map and metal base map; A11. Perform front projection on the bone base map, water base map and metal base map obtained in A10 to obtain the corresponding bone base map projection data, water base map projection data and metal base map projection data; A12. The bone base map projection data, water base map projection data and metal base map projection data obtained in A11 are processed by the beam hardening module to obtain beam hardening projection data. A13. The beam hardening projection data obtained in A12 is processed through the scattering signal module to obtain projection data containing scattering signals; A14. The projection data containing the scattered signal obtained in A13 is processed through the photon starvation module to obtain the estimated projection data; A15. Construct a consistency loss by combining the estimated projection data obtained from A14 with the uncorrected projection data from A1, and use this to constrain network training.

3. The CT metal artifact correction method based on physical duality regression learning according to claim 2, characterized in that: In S5 or A5, segmentation involves binarizing the metal base map projection data and setting values ​​greater than or equal to a specific threshold to 1 and values ​​less than the threshold to 0, where the threshold value th is given in advance. The metal projection trajectory in S5 or A5 is obtained by equation (1): Where d metal This is the metal base map projection data, where x and y are coordinate indices, th is the threshold, and M is the base map projection data. proj This is a projection trajectory diagram of the metal.

4. The CT metal artifact correction method based on physical duality regression learning according to claim 3, characterized in that, In S6 or A6, the photon hunger correction method of the photon hunger correction module is performed by the following steps: B1. Using the metal projection trajectory diagram M proj Extract uncorrected projection data p origin Non-metallic region projection data p non-meta and metal projection data p metal ; Among them, the non-metallic region projection data p non-meta With metal projection data p metal It is obtained from equations (2) and (3); p non-metal =p origin ⊙(1-M proj Equation (2) ... p metal =p origin ⊙M proj ...Equation (3); Where ⊙ represents the dot product operation; B2. Using the calculated projection data p est Replace in metal projection data p metal Greater than a specific threshold th metal The data is finally obtained from equation (4); Where p metal-psc The metal projection data after photon starvation correction, th metal p is the threshold. est For projection data; The threshold th metal Based on uncorrected projection data p origin The maximum value is selected, specifically the threshold th. metal It is obtained from equation (5); th metal =0.8×max(p origin ...Equation (5); The projection data p est Based on the bone base map projection data of S4 or A4 d b Water base map projection data d w Projection data d with metal base map me The multi-energy projection data, which is synthesized by combining the material decay curves of metal, water, and bone with X-ray spectral data, is specifically obtained by equation (6). Where φ(E) represents the X-ray spectral data, E represents the X-ray energy, and m b (E) is the bone material decay curve, m w (E) is the decay curve of water, m me (E) is the decay curve of the metal.

5. The CT metal artifact correction method based on physical duality regression learning according to claim 4, characterized in that, In S7 or A7, the beam hardening correction method of the beam hardening correction module is performed by the following steps: C1, Metal projection data p after photon starvation correction metal-psc Perform tissue correction; C2. Use a multilayer perceptron model (MLP) to learn the transformation from multi-energy projection data to single-energy projection data, then perform tissue inverse correction on the MLP output and compare it with the corresponding non-metallic region projection data p. non-metal Combined, the projection data after beam hardening correction is obtained, specifically obtained from equation (7); p bbh =p non-metal +f inv-tiss (MLP(f tiss (p metal-psc Equation (7) Where f tiss For organizational correction, f inv-tiss For reverse correction of the organization.

6. The CT metal artifact correction method based on physical duality regression learning according to claim 5, characterized in that: The tissue correction is derived from the metal projection data p after photon starvation correction. metal-psc Subtracting the projection data p of non-metallic substances contained in the middle tissue Specifically, this is done through equation (8): f tiss (p metal-psc ) = p metal-psc -p tissue ⊙M proj ...Equation (8); Where M proj This is a metal projection trajectory diagram, and ⊙ represents the dot product operation; The tissue inverse correction involves adding non-metallic material projection data p to the output of the multilayer perceptron model (MLP). tissu Specifically, it is done through equation (9): f inv-tiss (x)=x+p tissue ⊙M proj ...Equation (9); Where x is the output of the MLP.

7. The CT metal artifact correction method based on physical duality regression learning according to claim 6, characterized in that: The non-metallic material projection data p tissue Based on the bone base map projection data corresponding to S4 or A4 d b Water base map projection data d w The results, obtained by combining the decay curves of metals, water, and bone with X-ray spectral data, are specifically derived from equation (10):

8. The CT metal artifact correction method based on physical duality regression learning according to claim 7, characterized in that: In A12, the beam hardening module processes the data based on the bone base map projection data d obtained in A11. b Water base map projection data d w Projection data d with metal base map me The process of synthesizing multi-energy projection data by combining the material decay curves of metal, water, and bone with X-ray spectral data through equation (6); In A13, the scattering signal module is processed using a projection domain machine learning model, which learns the scattering projection signal. The input is beam-hardened projection data, and the output is the scattering projection signal. The modeling process is represented by equation (11): in For the projection domain machine learning model, p bbh-model For beam-hardened projection data, p scatter-model This is projection data containing scattered signals.

9. The CT metal artifact correction method based on physical duality regression learning according to claim 8, characterized in that: In A14, the processing method of the photon-starved module is to process the projection data p containing the scattered signal obtained in A13. scatter-model Perform truncation; The truncation process is to p scatter-model Greater than the threshold th metal The value is set to th metal , and th metal Take the maximum value from the uncorrected projection data.

10. The CT metal artifact correction method based on physical duality regression learning according to claim 9, characterized in that: In A15, the consistency loss is the estimated projection data p obtained from A14. dual The uncorrected projection data p of A1 origin To maintain consistency, the corresponding loss is constructed using the mean squared error (MSE), as expressed by equation (12):

Images