CT metal artifact correction method based on physical dual regression learning

Abstract

The invention discloses a CT (Computed Tomography) metal artifact correction method based on physical dual regression learning. Uncorrected projection data is corrected through nine steps to obtain a corrected image. According to the CT metal artifact correction algorithm and system based on physical dual regression learning, the metal artifact correction performance can be enhanced, and the stability and reliability of the algorithm are improved. The method has the beneficial effects that 1, the metal artifact correction result is more stable and reliable, and the learning process is more stable; 2, modeling is carried out on a physical mechanism generating metal artifacts in the learning process, so that the mapping of model learning can be effectively restrained to be consistent with a real imaging system;

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and computer tomography, and in particular to a CT metal artifact correction method based on physical dual regression learning. Background Art

[0002] X-ray computed tomography (CT) can provide information about the patient's three-dimensional anatomical structure and is an indispensable imaging tool in modern medicine. However, when high-attenuation materials such as metals are present in the patient's body, due to physical mechanisms such as beam hardening, scattering, photon starvation, and nonlinear partial volume effects, severe metal artifacts such as strips, bands, or shadows will be generated in the reconstructed image. Metal artifacts can cause signal distortion, produce pseudo-structures in the image, obscure the true lesions, and seriously affect 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 regard the projection data at the metal trajectory as missing data and complete this part of the data; or correct the physical mechanism of metal artifacts; or reconstruct using iterative reconstruction algorithms. However, traditional metal artifact correction algorithms are difficult to handle complex metal artifacts, and often produce over-smoothing effects or introduce secondary artifacts in the image. Metal artifact correction algorithms based on deep learning use the feature learning ability of neural networks and have significant advantages in prior image learning, image anatomical structure restoration, and artifact suppression. However, existing metal artifact correction algorithms based on deep learning still have some defects. First, the current algorithms only learn the mapping from metal-containing projection data to metal-free projection data, or to metal-free images based on limited training data. This process learns a one-way mapping from one data domain to another data domain, lacks constraints on the model output, and results in insufficient stability and reliability of the reconstruction results. Secondly, the current metal artifact correction model based on deep learning lacks the modeling of the corresponding physical mechanism when modeling. When using neural networks to learn the mapping of metal-containing projection data to metal-free projection data or artifact-free images, there are multiple possible mappings due to the limited quantity and quality of training data. Traditional deep learning-based metal artifact correction algorithms are difficult to learn the optimal mapping, and the learning process is poorly interpretable, which is not widely accepted by clinicians.

[0004] Therefore, in view of the shortcomings of the existing technology, it is necessary to provide a CT metal artifact correction method based on physical dual regression learning to solve the shortcomings of the existing technology. Summary of the invention

[0005] The purpose of the present invention is to avoid the shortcomings of the prior art and provide a CT metal artifact correction method based on physical dual regression learning. The CT metal artifact correction algorithm and system based on physical dual regression learning can enhance the metal artifact correction performance and improve the stability and reliability of the algorithm.

[0006] The above-mentioned purpose of the present invention is achieved by the following technical measures:

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

[0008] S1, reconstructing the uncorrected projection data to obtain an uncorrected reconstructed image;

[0009] S2, using the trained first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in S1 to obtain a pre-recovered image;

[0010] S3, inputting the pre-recovered image obtained in S2 into the trained second image domain network for prediction, and obtaining a bone-based image, a water-based image, and a metal-based image;

[0011] S4, forward-projecting the bone-based map, water-based map, and metal-based map obtained in S3 to obtain bone-based map projection data, water-based map projection data, and metal-based map projection data accordingly;

[0012] S5, segmenting the metal base image projection data obtained in S4 to obtain a metal projection trajectory map;

[0013] S6, performing photon starvation correction through a trained photon starvation correction module according to the uncorrected projection data of S1, the bone-based image projection data, the water-based image projection data and the metal-based image projection data of S4, and the metal projection trajectory map of S5, and correspondingly obtaining non-metallic area projection data and metal projection data after photon starvation correction;

[0014] S7, performing beam hardening correction on the metal projection data after photon starvation correction obtained in S6 through the trained beam hardening correction module, and combining it with the non-metal area projection data obtained in S6 to obtain projection data after beam hardening correction;

[0015] S8, reconstructing the beam hardening corrected projection data obtained in S7 to obtain a beam hardening corrected image;

[0016] S9, performing scattering correction and denoising on the beam hardening corrected image obtained in S8 through a trained scattering correction and denoising module, and finally obtaining a corrected image.

[0017] The present invention is 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 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 CT metal artifact correction module is performed according to the following steps:

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

[0020] A2, using the first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in A1 to obtain a pre-recovered image;

[0021] A3, input the pre-recovered image obtained in A2 into the second image domain network for prediction, and obtain a bone-based image, a water-based image, and a metal-based image;

[0022] A4, forward-projecting the bone-based map, water-based map and metal-based map obtained in A3, and correspondingly obtaining bone-based map projection data, water-based map projection data and metal-based map projection data;

[0023] A5, segmenting the metal base image projection data obtained in A4 to obtain a metal projection trajectory map;

[0024] A6, according to the uncorrected projection data of A1, the bone-based image projection data, the water-based image projection data and the metal-based image projection data of A4, and the metal projection trajectory map of A5, photon starvation correction is performed through a photon starvation correction module to obtain non-metallic area projection data and metal projection data after photon starvation correction;

[0025] A7, performing beam hardening correction on the metal projection data after photon starvation correction obtained in A6 through a beam hardening correction module, and combining it with the non-metal area projection data obtained in S6 to obtain beam hardening corrected projection data;

[0026] A8, reconstructing the beam hardening corrected projection data obtained in A7 to obtain a beam hardening corrected image;

[0027] A9, performing scatter correction and denoising on the beam hardening corrected image obtained in A8 through a scatter correction and denoising module to obtain a corrected image;

[0028] A10, inputting the corrected image obtained in A9 into the second image domain network for prediction to obtain a bone-based image, a water-based image, and a metal-based image;

[0029] A11, forward-projecting the bone-based map, water-based map, and metal-based map obtained in A10, and correspondingly obtaining bone-based map projection data, water-based map projection data, and metal-based map projection data;

[0030] A12, processing the bone-based image projection data, the water-based image projection data, and the metal-based image projection data obtained in A11 through a beam hardening module to obtain beam hardening-containing projection data;

[0031] A13, processing the beam hardening projection data obtained in A12 through a scattered signal module to obtain scattered signal projection data;

[0032] A14, processing the projection data containing the scattered signal obtained in A13 through a photon starvation module to obtain estimated projection data;

[0033] A15. Construct a consistency loss by combining the estimated projection data obtained in A14 with the uncorrected projection data of A1 to constrain network training.

[0034] In S5 or A5, segmentation is to perform binarization processing on the metal base image projection data, and set it to 1 when it is greater than or equal to a specific threshold, and set it to 0 when it is less than the threshold, wherein the threshold value th is given in advance.

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

[0036]

[0037] where d metal is the metal base image projection data, x and y are coordinate indexes, th is the threshold, M proj This is the metal projection trajectory diagram.

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

[0039] B1. Using metal projection trajectory diagram M proj , extract the uncorrected projection data p origin Non-metallic area projection data p non-metal and metal projection data p metal ; Among them, the non-metallic area projection data p non-meta With metal projection data p metal From equation (2) and equation (3), we can get:

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

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

[0042] Where ⊙ is the dot product operation;

[0043] B2. Using the calculated projection data p est Replace the metal projection data p metal Greater than a certain threshold th metal The data is finally obtained by formula (4);

[0044]

[0045] where p metal-ps After correcting the metal projection data for photon starvation, th metal is the threshold, p est The projection data.

[0046] Preferably, the threshold th metal According to the uncorrected projection data p origin The maximum value is selected, and the specific threshold value th metal From formula (5), we get:

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

[0048] Preferably, the projection data p est is the bone base projection data according to S4 or A4 b , water base map projection data d w With the metal base map projection data d me , and the multi-energy projection data synthesized by combining the material attenuation curves of metal, water, and bone with the X-ray spectrum data is specifically obtained by formula (6);

[0049]

[0050] Where φ(E) is the X-ray spectrum data, E is the ray energy, m b (E) is the material attenuation curve of bone, m w (E) is the material attenuation curve of water, m me (E) is the material attenuation curve of 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 Conduct tissue correction;

[0053] C2. Use the multi-layer perceptron model MLP to learn the transformation from multi-energy projection data to single-energy projection data, and then perform inverse correction on the output of the multi-layer perceptron model MLP and compare it with the corresponding non-metal area projection data p non-meta Combined, the projection data after beam hardening correction is obtained, which is specifically obtained by formula (7);

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

[0055] ...Formula (7);

[0056] where f tiss For tissue correction, f inv-tiss To reverse the correction of the organization.

[0057] Preferably, the tissue correction is performed from the metal projection data p after photon starvation correction. metal-psc Subtract the projection data of non-metallic materials contained in tissu , specifically through formula (8):

[0058] f tiss (p metal-psc )=p metal -p tissu ⊙M proj ...Formula (8);

[0059] Among them, M proj is the metal projection trajectory diagram, and ⊙ is the dot multiplication operation.

[0060] Preferably, the tissue inverse correction is to add the non-metallic material projection data p to the output result of the multi-layer perceptron model MLP. tissu , specifically through formula (9):

[0061] f inv-tiss (x) = x + p tis ⊙M proj ...Formula (9);

[0062] Where x is the output of MLP.

[0063] Preferably, the non-metallic material projection data p tis Projection data d through the bone base map corresponding to S4 or A4 b , water base map projection data d w , combined with the material attenuation curves of metal, water, and bone and the X-ray spectrum data, is obtained by formula (10):

[0064]

[0065] In A12, the processing method of the beam hardening module is based on the bone base image projection data d obtained in A11 b , water base map projection data d w With the metal base map projection data d me , combining the material attenuation curves of metal, water, and bone with X-ray spectrum data, the process of synthesizing multi-energy projection data through formula (6).

[0066] In A13, the processing method of the scattered signal module is a projection domain machine learning model, and the scattered projection signal is learned by the projection domain machine learning model; and the input is the projection data containing beam hardening, and the output is the scattered projection signal. The modeling process is expressed by formula (11):

[0067]

[0068] in is the projection domain machine learning model, p bbh-model is the projection data containing beam hardening, p scatter-model is the projection data containing scattered signals.

[0069] In A14, the photon starvation module processes the projection data p containing the scattered signal obtained in A13. scatter-mode Perform truncation processing.

[0070] Preferably, the above truncation process is to scatter-model Greater than the threshold th metal The value of th metal , and th metal Take the maximum value among the uncorrected projection data.

[0071] In A15, the consistency loss is calculated by converting the estimated projection data p obtained in A14 into dual The uncorrected projection data p of A1 otigin Keeping the same, the corresponding loss is constructed using the mean square error (MSE), which is expressed by formula (12):

[0072]

[0073] The present invention discloses a CT metal artifact correction method based on physical dual regression learning, comprising the following steps: S1, reconstructing uncorrected projection data to obtain an uncorrected reconstructed image; S2, using a trained first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained by S1 to obtain a pre-recovered image; S3, inputting the pre-recovered image obtained by 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, forward-projecting the bone base map, the water base map and the metal base map obtained by S3 to obtain bone base map projection data, water base map projection data and metal base map projection data respectively; S5, segmenting the metal base map projection data obtained by S4 to obtain a metal projection trajectory map; S6, based on the uncorrected projection data of S1, S4 The bone-based image projection data, water-based image projection data and metal-based image projection data, as well as the metal projection trajectory map of S5 are subjected to photon starvation correction through the trained photon starvation correction module, and the corresponding non-metallic area projection data and the metal projection data after photon starvation correction are obtained; S7, the metal projection data after photon starvation correction obtained by S6 is subjected to beam hardening correction through the trained beam hardening correction module, and combined with the non-metallic area projection data obtained by S6, to obtain the projection data after beam hardening correction; S8, the projection data after beam hardening correction obtained by S7 is reconstructed to obtain the beam hardening corrected image; S9, the beam hardening corrected image obtained by S8 is subjected to scattering correction and denoising through the trained scattering correction and denoising module, and finally the corrected image is obtained. Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention is based on dual regression learning, so the metal artifact correction result is more stable and reliable, and the learning process is more stable. Compared with the traditional metal artifact correction algorithm that only learns the one-way mapping from uncorrected projection data to the corrected image, the present invention adds a consistency constraint between the corrected image and the uncorrected projection data, thereby ensuring that the corrected image is accurate and stable. 2. The present 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 metal artifact correction algorithms based on deep learning are difficult to learn the optimal mapping. The present invention can effectively constrain the mapping learned by the model to be consistent with the real imaging system by modeling the physical mechanism of metal artifact generation during the learning process. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The present invention is further described with reference to the accompanying drawings, but the contents in the accompanying drawings do not constitute any limitation to the present invention.

[0075] Figure 1 This is a flow chart of a CT metal artifact correction method based on physical dual regression learning.

[0076] Figure 2Flowchart of joint training.

[0077] Figure 3 is the data or image in S1 or A1, where Figure 3 (a) is the uncorrected projection data, Figure 3 (b) Uncorrected reconstructed image.

[0078] Figure 4 is the pre-recovered image in S2 or A2.

[0079] Figure 5 are the base graphs of S3 or A3, where Figure 5 (a) Bone base diagram, Figure 5 (b) is the water base map, Figure 5 (c) is the metal base image.

[0080] Figure 6 is the projection data of each base map of S4 or A4, where Figure 6 (a) Bone base projection data, Figure 6 (b) is the water base map projection data, Figure 6 (c) is the metal base map projection data.

[0081] Figure 7 It is the S5 or A5 metal projection trajectory diagram.

[0082] Figure 8 is the projection data of S6 or A6, where Figure 8 (a) is the projection data of non-metallic area, Figure 8 (b) Metal projection data after photon starvation correction.

[0083] Fig. 9 Projection data after beam hardening correction for S7 or A7.

[0084] Fig.10 This is the image after beam hardening correction for S8 or A8.

[0085] Fig.11 The image is corrected on S9 or A9.

[0086] Fig.12 is the image in A10, where Fig.12 (a) is the bone base diagram, Fig.12 (b) is the water base map, Fig.12 (c) is the metal base image.

[0087] Fig.13 is the projection data in A11, where Fig.13 (a) is the bone base projection data, Fig.13 (b) is the water base map projection data, Fig.13 (c) Metal base map projection data.

[0088] Fig.14 A12 contains beam hardening projection data.

[0089] Fig.15 This is the projection data of the scattered signal in A13.

[0090] Fig.16 is the estimated projection data in A14.

[0091] Fig.17 These are the images before and after correction of Example 2. DETAILED DESCRIPTION

[0092] The technical solution of the present invention is further described in conjunction with the following embodiments.

[0093] Example 1

[0094] A CT metal artifact correction method based on physical dual regression learning, such as Figure 1 As shown, the following steps are included:

[0095] S1. Reconstruct the uncorrected projection data to obtain an uncorrected reconstructed image, such as Figure 3 ;

[0096] S2, using the trained first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in S1, to obtain a pre-recovered image, such as Figure 4 ;

[0097] S3, input the pre-recovered 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 ;

[0098] S4, forward-project the bone-based map, water-based map and metal-based map obtained in S3, and obtain bone-based map projection data, water-based map projection data and metal-based map projection data, such as Figure 6 ;

[0099] S5, segment the metal base image projection data obtained in S4 to obtain a metal projection trajectory map, such as Figure 7 ;

[0100] S6, based on the uncorrected projection data of S1, the bone-based image projection data, water-based image projection data and metal-based image projection data of S4, and the metal projection trajectory map of S5, photon starvation correction is performed through the trained photon starvation correction module, and the non-metallic area projection data and the metal projection data after photon starvation correction are correspondingly obtained, such as Figure 8 ;

[0101] S7, performing beam hardening correction on the metal projection data after photon starvation correction obtained in S6 through the trained beam hardening correction module, and combining it with the non-metal area projection data obtained in S6 to obtain beam hardening corrected projection data, such as Fig. 9 ;

[0102] S8, reconstructing the beam hardening corrected projection data obtained in S7 to obtain a beam hardening corrected image, such as Fig.10 ;

[0103] S9, the beam hardening corrected image obtained in S8 is subjected to scattering correction and denoising through the trained scattering correction and denoising module, and finally a corrected image is obtained, such as Fig.11 .

[0104] The present invention is a CT metal artifact correction method based on physical dual regression learning. 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.

[0105] It should be noted that the characteristic of the dual learning of the present invention is to create a dual task of the original task and learn two tasks at the same time. In the present invention, the original task is to learn the transformation or mapping from the 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 the present invention is based on the CT metal artifact correction method of physical dual regression learning, the physics in the present invention refers to modeling the physical mechanism of metal artifact generation in the two tasks and constructing the corresponding correction method.

[0106] like Figure 2 As shown in Figure 2, the joint training of the dual model is carried out in the following steps:

[0107] A1. Reconstruct the uncorrected projection data to obtain an uncorrected reconstructed image, such as Figure 3 ;

[0108] A2, using the first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in A1, to obtain a pre-recovered image, such as Figure 4 ;

[0109] A3, input the pre-recovered image obtained in A2 into the second image domain network for prediction, and obtain the bone base map, water base map and metal base map, such as Figure 5 ;

[0110] A4, forward-project the bone-based image, water-based image and metal-based image obtained in A3, and obtain bone-based image projection data, water-based image projection data and metal-based image projection data, such as Figure 6 ;

[0111] A5, segment the metal base image projection data obtained in A4 by threshold, and obtain the metal projection trajectory map, such as Figure 7 ;

[0112] A6, based on the uncorrected projection data of A1, the bone-based projection data, water-based projection data and metal-based projection data of A4, and the metal projection trajectory map of A5, photon starvation correction is performed through a photon starvation correction module, and corresponding non-metallic area projection data and metal projection data after photon starvation correction are obtained, such as Figure 8 ;

[0113] A7, the metal projection data after photon starvation correction obtained in A6 is subjected to beam hardening correction through a beam hardening correction module, and combined with the non-metal area projection data obtained in S6 to obtain projection data after beam hardening correction, such as Fig. 9 ;

[0114] A8, reconstructing the beam hardening corrected projection data obtained in A7 to obtain a beam hardening corrected image, such as Fig.10 ;

[0115] A9, the beam hardening corrected image obtained in A8 is subjected to scatter correction and denoising by a scatter correction and denoising module to obtain a corrected image, such as Fig.11 ;

[0116] A10, input the corrected image obtained in A9 into the second image domain network for prediction, and obtain the bone base map, water base map and metal base map, such as Fig.12 ;

[0117] A11, forward-project the bone-based image, water-based image, and metal-based image obtained in A10, and obtain bone-based image projection data, water-based image projection data, and metal-based image projection data, such as Fig.13 ;

[0118] A12, the bone-based image projection data, the water-based image projection data and the metal-based image projection data obtained in A11 are processed by a beam hardening module to obtain beam hardening projection data, such as Fig.14 ;

[0119] A13, processing the beam hardening projection data obtained in A12 through a scattered signal module to obtain scattered signal projection data, such as Fig.15 ;

[0120] A14, perform photon starvation modeling on the projection data containing scattered signals obtained in A13 to obtain estimated projection data, such as Fig.16 ;

[0121] A15. Construct a consistency loss by combining the estimated projection data obtained in A14 with the uncorrected projection data of A1 to constrain network training.

[0122] It should be noted that the non-metallic area projection data in S6 and A6 of the present invention refers to the part of the projection data that does not contain metal. The calculation process is to perform a forward projection on the metal mask to obtain its projection trajectory. The data at the trajectory is the metal area projection data, and the other areas are the non-metallic area projection data. The metal area in the metal area projection data represents the spatial position. In addition to the projection information of the metal, this part of the data also contains the projection information of soft tissues such as bones and fat. The projection data other than the metal is collectively referred to as the non-metallic material projection data.

[0123] The first image domain network and the second image domain network of the present invention are both U-Net models, and the output channel of the first image domain network is 1, and the output channel of the second image domain network is 3. The present invention sets the number of training times of joint training according to actual needs, and the number of training times is comprehensively judged based on the amount of data and the size of the network. When the set number of times is reached, the training is considered to be completed. The number of training times of joint training in this embodiment is specifically 100 times. Each joint training is as follows.

[0124] In S5 or A5, segmentation is to perform binarization processing on the metal base image projection data, and set the binarization value to 1 when it is greater than or equal to a specific threshold, and set the binarization value to 0 when it is less than the threshold, wherein the threshold value th is given in advance;

[0125] The metal projection trajectory in S5 or A5 is obtained by formula (1):

[0126]

[0127] where d metal is the metal base image projection data, x and y are coordinate indexes, th is the threshold, M proj This is the metal projection trajectory diagram.

[0128] The photon starvation correction method in S6 or A6 is performed as follows:

[0129] B1. Using metal projection trajectory diagram M proj , extract the uncorrected projection data p origin Non-metallic area projection data p non-metal and metal projection data p metal ; Among them, the non-metallic area projection data p non-metal With metal projection data p metal From equation (2) and equation (3), we can get:

[0130] p non-metal =p origin ⊙(1-Mproj )……Formula (2);

[0131] p metal =p origin ⊙M proj ...Formula (3);

[0132] Where ⊙ is the dot product operation;

[0133] B2. Using the calculated projection data p est Replace the metal projection data p metal Greater than a certain threshold th metal The data is finally obtained by formula (4);

[0134]

[0135] where p metal-psc After correcting the metal projection data for photon starvation, th metal is the threshold, p est is the projection data;

[0136] Threshold metal According to the uncorrected projection data p origin The maximum value is selected, and the specific threshold value th metal From formula (5), we get:

[0137] th metal =0.8×max(p origin )……Formula (5);

[0138] Projection data p est is the bone base projection data according to S4 or A4 b , water base map projection data d w With the metal base map projection data d me , and the multi-energy projection data synthesized by combining the material attenuation curves of metal, water, and bone with the X-ray spectrum data is specifically obtained by formula (6);

[0139]

[0140] Where φ(E) is the X-ray spectrum data, E is the ray energy, m b (E) is the material attenuation curve of bone, m w (E) is the material attenuation curve of water, m me (E) is the material attenuation curve of metal.

[0141] It should be noted that the material attenuation curves of bone, water and metal in the present invention are obtained by simulation software XCIST.

[0142] The beam hardening correction method in S7 or A7 is performed as follows:

[0143] C1, metal projection data p after photon starvation correction metal-psc Conduct tissue correction;

[0144] C2. Use the multi-layer perceptron model MLP to learn the transformation from multi-energy projection data to single-energy projection data, and then perform inverse correction on the output of the multi-layer perceptron model MLP and compare it with the corresponding non-metal area projection data p non-metal Combined, the projection data after beam hardening correction is obtained, which is specifically obtained by formula (7);

[0145] p bbh =p non-metal +f inv-tiss (MLP(f tiss (p metal-psc )))

[0146] ...Formula (7);

[0147] where f tiss For tissue correction, f inv-tiss To reverse the correction of the organization.

[0148] Tissue correction is obtained from the metal projection data after photon starvation correction p metal-ps Subtract the projection data of non-metallic materials contained in tissue , specifically through formula (8):

[0149] f tiss (p metal-psc )=p metal-ps -p tissue ⊙M proj ...Formula (8);

[0150] Among them, M proj is the metal projection trajectory diagram, and ⊙ is the dot multiplication operation.

[0151] Tissue inverse correction is to add non-metallic material projection data p to the output of the multi-layer perceptron model MLP tissue , specifically through formula (9):

[0152] f inv-tiss (x) = x + p tissue ⊙M proj ...Formula (9);

[0153] Where x is the output of MLP.

[0154] Non-metallic material projection data tissue Projection data d through the bone base map corresponding to S4 or A4 b , water base map projection data d w, combined with the material attenuation curves of metal, water, and bone and the X-ray spectrum data, is obtained by formula (10):

[0155]

[0156] In S9 and A9 of the present invention, scatter correction and denoising refers to using any image domain machine learning model, including but not limited to convolutional neural networks and graph neural networks, to remove residual noise and scatter artifacts in the image after beam hardening correction, wherein the input is the image after beam hardening correction and the output is the corrected image.

[0157] In A12, beam hardening is modeled as the bone base map projection data d obtained in A11 b , water base map projection data d w With the metal base map projection data d me , combining the material attenuation curves of metal, water, and bone with X-ray spectrum data, the process of synthesizing multi-energy projection data through formula (6).

[0158] In A13, the scattered signal is modeled as a projection domain machine learning model, and the scattered projection signal is learned by the projection domain machine learning model; and the input is the projection data containing beam hardening, and the output is the scattered projection signal. The modeling process is expressed by formula (11):

[0159]

[0160] in is the projection domain machine learning model, p bbh-model is the projection data containing beam hardening, p scatter-mode is the projection data containing scattered signals.

[0161] 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. The projection domain machine learning model of this embodiment is specifically a convolutional neural network.

[0162] In A14, photon starvation is modeled as the projection data p containing the scattered signal obtained in A13. scatter-mod The truncation process is to convert p scatter-mode Greater than the threshold th metal The value of th metal , and th metal Take the maximum value among the uncorrected projection data.

[0163] In A15, the consistency loss is calculated by converting the estimated projection data p obtained in A14 into dual The uncorrected projection data p of A1 origin Keeping the same, the corresponding loss is constructed using the mean square error (MSE), which is expressed by formula (12):

[0164]

[0165] It should be noted that, before the joint training, the first image domain network and the second image domain network can be pre-trained separately, and the pre-training of each image domain network is 50 epochs, and the learning rate is increased from 1×10 -4 Linearly decreases to 1×10 -6 , using the Adam optimizer, where the parameters (β 1 ,β 2 )=(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 or without artifacts, and the output is a bone-based image, a water-based image, and a metal-based image. Of course, the present invention can also directly train the first image domain network and the second image domain network with other modules without pre-training.

[0166] In addition to introducing consistency loss, the joint training of the present invention can also add loss constraints to the outputs of the first image domain network and the second image domain network, add all loss functions, and constrain network training. The specific loss function used is MSE loss, which is expressed by formula (13):

[0167]

[0168] Wherein x is the output of the first image domain network or the second image domain network, and y is the corresponding target image, wherein 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-based image, a target water-based image, and a target metal-based image.

[0169] It should be noted that the corrected projection data and estimated projection data of the present invention refer to the projection data after logarithmic transformation. The bone-based image, water-based image and metal-based image refer to the bone density image, water density image and metal density image obtained by material decomposition of the CT attenuation image.

[0170] In the present invention, the reconstruction method of S1, S8, A1 and A8 can be FBP algorithm, 3D spiral reconstruction algorithm or Feldkamp-Davis-Kress (FDK) algorithm; the front projection of S4, A4 and A1 can be 2D front projection or 3D front projection; the metal artifact correction of the present invention can be 2D fan beam CT metal artifact correction or 3D cone beam CT metal artifact correction, which can be determined according to actual situation.

[0171] The CT metal artifact correction algorithm and system based on physical dual regression learning can enhance the metal artifact correction performance and improve the stability and reliability of the algorithm. Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention is based on dual regression learning, so the metal artifact correction result is more stable and reliable, and the learning process is more stable. Compared with the traditional metal artifact correction algorithm that only learns the one-way mapping from uncorrected projection data to the corrected image, the present invention adds consistency constraints between the corrected image and the uncorrected projection data, thereby ensuring that the corrected image is accurate and stable. 2. The present 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. It is difficult for traditional metal artifact correction algorithms based on deep learning to learn the optimal mapping. The present invention can effectively constrain the mapping learned by the model to be consistent with the real imaging system by modeling the physical mechanism of metal artifact generation during the learning process.

[0172] Example 2

[0173] A CT metal artifact correction method based on physical dual regression learning, the other features are the same as those of Example 1, except that. In this embodiment, the training data of the joint training includes 27948 images, of which 15000 are head data and the rest are chest and abdomen data. The test data includes 1000 images, of which 136 are head images and the rest are chest and abdomen images. The metal materials in the training data include stainless steel, titanium, amalgam, copper, and cobalt. The simulation scanning voltage is 120kVp, the current is 500mA, and Poisson noise and Gaussian noise are added during the simulation process.

[0174] In S5 or A5, the threshold value is 0.001. The number of training times in this embodiment is 100 times.

[0175] in Fig.17 The head and chest and abdomen images before correction, and the head and chest and abdomen images after correction using the CT metal artifact correction method based on physical dual regression learning of the present invention. Fig.17 It can be seen that before metal artifact correction, the artifacts in the image cause signal distortion and produce pseudo structures. These artifacts or pseudo structures overlap, block or change the shape of the real tissue structure, making it difficult for doctors to correctly interpret the image. After the CT metal artifact correction method based on physical dual regression learning of the present invention is applied, Fig.17The images in the image have changed significantly. In the corrected image, the artifacts originally caused by metal are effectively suppressed or eliminated, the overall uniformity and contrast of the image are significantly improved, and the detailed structure of the soft tissue becomes more clearly discernible.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A CT metal artifact correction method based on physical dual regression learning, characterized in that: The steps include: S1, reconstructing the uncorrected projection data to obtain an uncorrected reconstructed image; S2, using the trained first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in S1 to obtain a pre-recovered image; S3, inputting the pre-recovered image obtained in S2 into the trained second image domain network for prediction, and obtaining a bone-based image, a water-based image, and a metal-based image; S4, forward-projecting the bone-based map, water-based map, and metal-based map obtained in S3 to obtain bone-based map projection data, water-based map projection data, and metal-based map projection data accordingly; S5, segmenting the metal base image projection data obtained in S4 to obtain a metal projection trajectory map; S6, performing photon starvation correction through a trained photon starvation correction module according to the uncorrected projection data of S1, the bone-based image projection data, the water-based image projection data and the metal-based image projection data of S4, and the metal projection trajectory map of S5, and correspondingly obtaining non-metallic area projection data and metal projection data after photon starvation correction; S7, performing beam hardening correction on the metal projection data after photon starvation correction obtained in S6 through the trained beam hardening correction module, and combining it with the non-metal area projection data obtained in S6 to obtain projection data after beam hardening correction; S8, reconstructing the beam hardening corrected projection data obtained in S7 to obtain a beam hardening corrected image; S9, performing scattering correction and denoising on the beam hardening corrected image obtained in S8 through a trained scattering correction and denoising module, and finally obtaining a corrected image.

2. The CT metal artifact correction method based on physical dual regression learning according to claim 1 is characterized in that: A CT metal artifact correction model is constructed through dual learning, and 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 are jointly trained through the CT metal artifact correction model; The joint training of the CT metal artifact correction module is performed according to the following steps: A1. Reconstruct the uncorrected projection data to obtain an uncorrected reconstructed image; A2, using the first image domain network to remove metal artifacts from the uncorrected reconstructed image obtained in A1 to obtain a pre-recovered image; A3, input the pre-recovered image obtained in A2 into the second image domain network for prediction, and obtain a bone-based image, a water-based image, and a metal-based image; A4, forward-projecting the bone-based map, water-based map and metal-based map obtained in A3, and correspondingly obtaining bone-based map projection data, water-based map projection data and metal-based map projection data; A5, segmenting the metal base image projection data obtained in A4 to obtain a metal projection trajectory map; A6, according to the uncorrected projection data of A1, the bone-based image projection data, the water-based image projection data and the metal-based image projection data of A4, and the metal projection trajectory map of A5, photon starvation correction is performed through a photon starvation correction module to obtain non-metallic area projection data and metal projection data after photon starvation correction; A7, performing beam hardening correction on the metal projection data after photon starvation correction obtained in A6 through a beam hardening correction module, and combining it with the non-metal area projection data obtained in S6 to obtain beam hardening corrected projection data; A8, reconstructing the beam hardening corrected projection data obtained in A7 to obtain a beam hardening corrected image; A9, performing scatter correction and denoising on the beam hardening corrected image obtained in A8 through a scatter correction and denoising module to obtain a corrected image; A10, inputting the corrected image obtained in A9 into the second image domain network for prediction to obtain a bone-based image, a water-based image, and a metal-based image; A11, forward-projecting the bone-based map, water-based map, and metal-based map obtained in A10, and correspondingly obtaining bone-based map projection data, water-based map projection data, and metal-based map projection data; A12, processing the bone-based image projection data, the water-based image projection data, and the metal-based image projection data obtained in A11 through a beam hardening module to obtain beam hardening-containing projection data; A13, processing the beam hardening projection data obtained in A12 through a scattered signal module to obtain scattered signal projection data; A14, processing the projection data containing the scattered signal obtained in A13 through a photon starvation module to obtain estimated projection data; A15. Construct a consistency loss by combining the estimated projection data obtained in A14 with the uncorrected projection data of A1 to constrain network training.

3. The CT metal artifact correction method based on physical dual regression learning according to claim 2 is characterized in that: In S5 or A5, segmentation is to perform binarization processing on the metal base image projection data, and set the binarization value to 1 when it is greater than or equal to a specific threshold, and set the binarization value to 0 when it is less than the threshold, wherein the threshold value th is given in advance; The metal projection trajectory in S5 or A5 is obtained by formula (1): where d metal is the metal base image projection data, x and y are coordinate indexes, th is the threshold, M proj This is the metal projection trajectory diagram.

4. The CT metal artifact correction method based on physical dual regression learning according to claim 3 is characterized in that: In S6 or A6, the photon starvation correction method of the photon starvation correction module is performed by the following steps: B1. Using metal projection trajectory diagram M proj , extract the uncorrected projection data p origin Non-metallic area projection data p non-meta and metal projection data p metal ; Among them, the non-metal area projection data p non-meta With metal projection data p metal From equation (2) and equation (3), we can get: p non-metal =p origin ⊙(1-M proj ) ……Formula (2); p metal =p origin ⊙M proj ...Formula (3); Where ⊙ is the dot product operation; B2. Using the calculated projection data p est Replace the metal projection data p metal Greater than a certain threshold th metal The data is finally obtained by formula (4); where p metal-psc After correcting the metal projection data for photon starvation, th metal is the threshold, p est is the projection data; The threshold th metal According to the uncorrected projection data p origin The maximum value is selected, and the specific threshold value th metal From formula (5), we get: th metal =0.8×max(p origin ) ……Formula (5); The projection data p ese is the bone base projection data d according to S4 or A4 b , water base map projection data d w With the metal base map projection data d me , and the multi-energy projection data synthesized by combining the material attenuation curves of metal, water, and bone with the X-ray spectrum data is specifically obtained by formula (6); Where φ(E) is the X-ray spectrum data, E is the ray energy, m b (E) is the material attenuation curve of bone, m w (E) is the material attenuation curve of water, m me (E) is the material attenuation curve of metal.

5. The CT metal artifact correction method based on physical dual regression learning according to claim 4 is 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 Conduct tissue correction; C2. Use the multi-layer perceptron model MLP to learn the transformation from multi-energy projection data to single-energy projection data, and then perform inverse correction on the output of the multi-layer perceptron model MLP and compare it with the corresponding non-metal area projection data p non-metal Combined, the projection data after beam hardening correction is obtained, which is specifically obtained by formula (7); p bbh =p non-metal +f inv-tiss (MLP(f tiss (p metal-psc ))) ...Formula (7); where f tiss For tissue correction, f inv-tiss To reverse the correction of the tissue.

6. The CT metal artifact correction method based on physical dual regression learning according to claim 5, characterized in that: The tissue correction is obtained from the metal projection data after photon starvation correction. metal-psc Subtract the projection data of non-metallic materials contained in tissue , specifically through formula (8): f tiss (p metal-psc )=p metal-psc -p tissue ⊙M proj ...Formula (8); Among them, M proj is the metal projection trajectory diagram, ⊙ is the dot multiplication operation; The tissue inverse correction is to add the non-metallic material projection data p to the output result of the multi-layer perceptron model MLP. tissu , specifically through formula (9): f inv-tiss (x) = x + p tissue ⊙M proj ...Formula (9); Where x is the output of MLP.

7. The CT metal artifact correction method based on physical dual regression learning according to claim 6, characterized in that: The non-metallic material projection data p tissue Projection data d through the bone base map corresponding to S4 or A4 b , water base map projection data d w , combined with the material attenuation curves of metal, water, and bone and the X-ray spectrum data, is obtained by formula (10):

8. The CT metal artifact correction method based on physical dual regression learning according to claim 7, characterized in that: In A12, the processing method of the beam hardening module is based on the bone base image projection data d obtained in A11 b , water base map projection data d w With the metal base map projection data d me , combining the material attenuation curves of metal, water, and bone with X-ray spectrum data, the process of synthesizing multi-energy projection data through equation (6); In A13, the processing method of the scattered signal module is a projection domain machine learning model, and the scattered projection signal is learned by the projection domain machine learning model; and the input is the projection data containing beam hardening, and the output is the scattered projection signal. The modeling process is expressed by formula (11): in is the projection domain machine learning model, p bbh-model is the projection data containing beam hardening, p scatter-mode is the projection data containing scattered signals.

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

10. The CT metal artifact correction method based on physical dual regression learning according to claim 9, characterized in that: In A15, the consistency loss is calculated by converting the estimated projection data p obtained in A14 into dual The uncorrected projection data p of A1 origin Keeping the same, the corresponding loss is constructed using the mean square error (MSE), which is expressed by formula (12):

Images