A coaxial phase contrast CT reconstruction method and system based on fractional generative model

Abstract

The present invention discloses a coaxial phase contrast CT reconstruction method and system based on a score generation model, which belongs to the field of radiological medical image processing technology and is applied to X-ray phase contrast CT imaging, including: S1, using the coaxial phase contrast CT images reconstructed under full projection angles to establish a training data set; S2, first applying SDE perturbations at different times to the training images in the training data set to obtain perturbed images, and then inputting the perturbed images and the training images into a convolutional neural network for training; S3, learning the scores under SDE perturbations at different times through a score matching network; S4, using the trained convolutional neural network to reconstruct coaxial phase contrast CT. By adopting the above technical solution, the present invention can directly reconstruct high-quality coaxial phase contrast CT images from the intensity projection information obtained from the detector under an incomplete projection angle without the need for an additional phase recovery step.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiological medical image processing, and in particular relates to a coaxial phase contrast CT reconstruction method and system based on a fractional generation model. Background Art

[0002] Compared with traditional X-ray absorption computed tomography (CT) imaging technology, X-ray phase contrast CT imaging technology can provide higher contrast and spatial resolution when imaging weakly absorbing samples, so it shows great application prospects in the fields of biomedical imaging and clinical imaging. However, the main challenges currently faced by this technology include long CT scanning time and high radiation dose. To overcome these difficulties, phase contrast CT reconstruction technology based on incomplete projection data has become a feasible solution.

[0003] Traditional coaxial phase-contrast CT (PB-PCCT) reconstruction algorithms are mainly divided into three categories: filtering methods, model-based iterative reconstruction methods, and end-to-end deep learning-based methods. These methods usually require phase recovery before CT reconstruction. Filtering methods are known for their simplicity of calculation and low demand for computing resources, but they fail to make full use of additional prior information, and the reconstruction effect under incomplete projection conditions is often unsatisfactory. Model-based iterative reconstruction methods, such as total variation (TV) algorithms, although they use additional prior knowledge and suppress artifacts and noise in reconstructed images under incomplete projection conditions to a certain extent, they rely on mathematical models to model prior knowledge, which may lead to over-smoothing problems in the reconstruction results. End-to-end deep learning methods, such as (FBPConvnet, DDNet), effectively eliminate artifacts and noise under incomplete projection conditions by utilizing data priors, but they require paired annotated data for training and do not make full use of the physical priors of coaxial phase-contrast CT imaging, which may lead to pseudo structures in the reconstructed images.

[0004] In view of the limitations of the existing technology, it is particularly urgent to develop a coaxial phase contrast CT reconstruction method and system based on a fractional generation model to address the shortcomings of the existing technology in the reconstruction of incomplete projections of X-ray phase contrast CT images. Summary of the invention

[0005] The present invention aims at common problems in incomplete projection reconstruction of coaxial phase contrast CT, such as the need for phase recovery, the presence of artifacts and noise in the reconstruction results, and provides a coaxial phase contrast CT reconstruction method and system based on a fractional generative model. The method and system can directly reconstruct high-quality coaxial phase contrast CT images from the intensity projection information obtained by the detector at the incomplete projection angle without the need for an additional phase recovery step.

[0006] To achieve the above-mentioned object of the invention, the first object of the present invention is to provide a coaxial phase contrast CT reconstruction method based on a fractional generation model, which is applied to X-ray phase contrast CT imaging, comprising the following steps:

[0007] S1. Establish a training data set using in-axial phase contrast CT images reconstructed at all projection angles;

[0008] S2, first applying SDE perturbations at different times to the training images in the training data set to obtain perturbed images, and then inputting the perturbed images and the training images into a convolutional neural network for training;

[0009] S3. Through the score matching network Learn the scores under SDE perturbations at different times;

[0010] S4. Reconstructing coaxial phase contrast CT using the trained convolutional neural network; specifically including:

[0011] S401, initialize the following variables: Reconstruct the starting noise , incomplete angle projection matrix A and various variables required for iterative algorithms;

[0012] S402, start the sampling process of the score generation model. In each sampling process, first use the score matching network trained in S3 Obtaining noisy phase-contrast CT images via a predictive-corrected sampler ;

[0013] S403, calculating noisy phase contrast CT image by single material assumption Corresponding absorption CT image ;

[0014] S404, using noisy phase contrast CT images and absorption CT images , the intensity information predicted by the forward projection physical model of phase contrast CT is calculated;

[0015] S405, using an iterative algorithm based on a mathematical priori model, the corresponding phase contrast CT image is obtained from the intensity information obtained from the actual detector and the predicted intensity information. ;

[0016] S406: The phase contrast CT image As the starting value for the next fraction generation model sampling, the next round of iteration begins. After the iteration is completed, the final phase CT reconstruction result is output .

[0017] Preferably, in S2, the formula of SDE perturbation is as follows:

[0018] (1)

[0019] in, represents the training image, represents the drift function, is the diffusion function; d represents the differential, t represents the continuous time variable, and w represents the forward Brownian motion.

[0020] Preferably, in S3, the formula of the loss function is as follows:

[0021] (2)

[0022] in, represents the score estimation function in the score generation model, which is parameterized by a neural network; ψ represents the time-dependent weighting coefficient; Represents the score of the training data distribution, that is, the gradient of the logarithmic probability of the training data distribution under the conditions at time t; express The square of the L2 norm.

[0023] Preferably, in S404, the calculation process of the intensity information is as follows:

[0024] (3)

[0025] in, , represents the wavelength of the X-ray, D represents the distance the X-ray travels after passing through the sample, represents the absorption projection, through Multiply it by the CT projection matrix A to get, represents the phase projection, Multiply it by the CT projection matrix A to get; Indicates the intensity information obtained on the detector; represents the transport equation of the sample; represents the Fresnel propagation operator; s represents the one-dimensional coordinate variable on the detector; i represents the imaginary unit; θ represents different detector sampling angles.

[0026] Preferably, in S405, the iterative algorithm is a Chambolle-Pock algorithm.

[0027] The second object of the present invention is to provide a coaxial phase contrast CT reconstruction system based on a fractional generation model, which is applied to X-ray phase contrast CT imaging, comprising:

[0028] The data set construction module uses the in-line phase contrast CT images reconstructed at all projection angles to establish a training data set;

[0029] A convolutional neural network training module first applies SDE perturbations at different times to the training images in the training data set to obtain perturbed images, and then inputs the perturbed images and the training images into the convolutional neural network for training;

[0030] Score matching network learning module, through the score matching network Learn the scores under SDE perturbations at different times;

[0031] The reconstruction module uses the trained convolutional neural network to perform in-axial phase contrast CT reconstruction; specifically includes:

[0032] Parameter initialization unit, initialize the following variables: Reconstruct the starting noise , incomplete angle projection matrix A and various variables required for iterative algorithms;

[0033] The sampling unit starts the sampling process of the score generation model. In each sampling process, the score matching network trained by the score matching network learning module is first used. Obtaining noisy phase-contrast CT images via a predictive-corrected sampler ;

[0034] The first calculation unit calculates the noisy phase contrast CT image by single material assumption Corresponding absorption CT image ;

[0035] The second calculation unit uses the noisy phase contrast CT image and absorption CT images , the intensity information predicted by the forward projection physical model of phase contrast CT is calculated;

[0036] The third calculation unit uses an iterative algorithm based on a mathematical priori model to solve the corresponding phase contrast CT image from the intensity information obtained from the actual detector and the predicted intensity information. ;

[0037] Iteration unit, the phase contrast CT image As the starting value for the next fraction generation model sampling, the next round of iteration begins. After the iteration is completed, the final phase CT reconstruction result is output .

[0038] Preferably, in the convolutional neural network training module, the formula for SDE perturbation is as follows:

[0039] (1)

[0040] in, represents the training image, represents the drift function, is the diffusion function; d represents the differential, t represents the continuous time variable, and w represents the forward Brownian motion.

[0041] Preferably, the formula of the loss function is as follows:

[0042] (2)

[0043] in, represents the score estimation function in the score generation model, which is parameterized by a neural network; ψ represents the time-dependent weighting coefficient; Represents the score of the training data distribution, that is, the gradient of the logarithmic probability of the training data distribution under the conditions at time t; express The square of the L2 norm;

[0044] The intensity information is calculated as follows:

[0045] (3)

[0046] in, , represents the wavelength of the X-ray, D represents the distance the X-ray travels after passing through the sample, represents the absorption projection, through Multiply it by the CT projection matrix A to get, represents the phase projection, Multiply it by the CT projection matrix A to get; Indicates the intensity information obtained on the detector; represents the transport equation of the sample; represents the Fresnel propagation operator; s represents the one-dimensional coordinate variable on the detector; i represents the imaginary unit; θ represents different detector sampling angles.

[0047] Preferably, the iterative algorithm is a Chambolle-Pock algorithm.

[0048] A third object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned in-axial phase contrast CT reconstruction method based on a fractional generation model.

[0049] A fourth object of the present invention is to provide a computer program product, comprising a computer program, which, when executed by a processor, implements the above-mentioned in-axial phase contrast CT reconstruction method based on a fractional generation model.

[0050] The advantages and positive effects of this application are:

[0051] (1) Compared with model-based iterative algorithms, this technology can provide better reconstructed images. At the same time, compared with traditional end-to-end deep learning methods, the accuracy and quality of its reconstructed images are also superior. In this study, a score generation model was used, which integrated data prior knowledge to achieve better image reconstruction. Compared with the end-to-end deep learning method, the score generation model used in the present invention focuses on learning the probability distribution of high-quality in-axial phase contrast CT images. Compared with traditional methods, the method proposed in the present invention has stronger generalization ability in learning the mapping relationship between high-quality and low-quality images, and the quality of reconstructed images is higher. In addition, this method also uses the physical model of phase contrast CT and the original intensity information of the detector as constraints to achieve more accurate image reconstruction.

[0052] (2) Compared with traditional end-to-end deep learning networks, this technology can be trained without paired data. The score generation model used in this method learns the scores of the training data under the perturbation of stochastic differential equations at different times through a deep learning network. During the training process, only high-quality phase-contrast CT images with the same distribution are required, and their corresponding low-quality images are not required, thus avoiding the large amount of manpower and material resources required to collect paired data.

[0053] (3) Compared with traditional reconstruction methods, this technology does not require an additional phase recovery step. The reconstruction problem of coaxial phase contrast CT is a complex nonlinear inverse problem. Traditional model-based iterative algorithms are difficult to update due to the lack of prior constraints. Usually, they need to use phase recovery methods to obtain phase projections before reconstruction. The end-to-end deep learning method learns the mapping relationship between images. However, due to the incomplete projection of detector intensity information, the reconstructed image has less information and a large number of artifacts, making it difficult for the network to learn the mapping relationship between it and high-quality coaxial phase contrast CT. Therefore, traditional end-to-end networks also require an additional phase recovery step. The image probability distribution learned by the fractional generative model provides strong prior information. This method uses the physical model of coaxial phase contrast CT to link the coaxial phase contrast CT images sampled by the fractional generative model with the detector intensity information. During the image sampling and generation process, the sampling results of the fractional generative model are continuously constrained, and finally high-quality coaxial phase contrast CT images are directly output. This process does not require an additional phase recovery step. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 A first flow chart provided by a preferred embodiment of the present invention is shown;

[0056] Figure 2 A second flow chart provided by a preferred embodiment of the present invention is shown;

[0057] Figure 3 A diagram showing the reconstruction result under limited angle projection in a preferred embodiment of the present invention is shown;

[0058] Figure 4 The figure shows the reconstruction result under sparse angular projection in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Compared with the traditional technology, the present invention does not require an additional phase recovery process, and while ensuring the quality of the reconstructed image, it can also greatly reduce the time, range and radiation dose of phase contrast CT scanning. The specific embodiments are as follows:

[0061] See also Figures 1 to 4 ;

[0062] The first embodiment is a coaxial phase contrast CT reconstruction method based on a fractional generative model, which is applied to X-ray phase contrast CT imaging, comprising the following steps:

[0063] S1. Establish a training data set using coaxial phase contrast CT images reconstructed at all projection angles; specifically:

[0064] First, a large number of reconstructed coaxial phase-contrast CT images at all projection angles are collected;

[0065] Then, high-quality in-line phase contrast CT images reconstructed at all projection angles are selected as training data sets;

[0066] S2, firstly applying SDE perturbations at different times to the training images in the training data set to obtain perturbed images, and then inputting the perturbed images and the training images into a convolutional neural network for training; specifically comprising:

[0067] By applying stochastic differential equation (SDE) perturbations at different times to the training images, perturbed images are obtained;

[0068] The perturbed image and the original training image are input into the convolutional neural network for training. The formula of the stochastic differential equation is as follows:

[0069] (1)

[0070] in, represents the training image, represents the drift function, is the diffusion function, d represents the differential, t represents the continuous time variable, and w represents the forward Brownian motion.

[0071] S3. Through the score matching network Learn the scores under SDE perturbations at different times; specifically:

[0072] Using the input image, through the score matching network Learn the corresponding scores under SDE perturbations at different times. The specific loss function is as follows:

[0073] (2)

[0074] in, represents the score estimation function in the score generation model, which is parameterized by a neural network; ψ represents the time-dependent weighting coefficient; Represents the score of the training data distribution, that is, the gradient of the logarithmic probability of the training data distribution under the conditions at time t; express The square of the L2 norm.

[0075] S4. Reconstructing coaxial phase contrast CT using the trained convolutional neural network; specifically including:

[0076] S401, initialize the following variables: Reconstruct the starting noise , incomplete angle projection matrix A and various variables required for iterative algorithms;

[0077] S402, start the sampling process of the score generation model. In each sampling process, first use the score matching network trained in S3 Obtaining noisy phase-contrast CT images via a predictive-corrected sampler ;

[0078] S403, calculating noisy phase contrast CT image by single material assumption Corresponding absorption CT image ;

[0079] S404, using noisy phase contrast CT images and absorption CT images , the intensity information predicted by the forward projection physical model of phase contrast CT is calculated;

[0080] The intensity information is calculated as follows:

[0081] (3)

[0082] in, , represents the wavelength of the X-ray, D represents the distance the X-ray travels after passing through the sample, represents the absorption projection, through Multiply it by the CT projection matrix A to get, represents the phase projection, Multiply it by the CT projection matrix A to get; Indicates the intensity information obtained on the detector; represents the transport equation of the sample; represents the Fresnel propagation operator; s represents the one-dimensional coordinate variable on the detector; i represents the imaginary unit; θ represents different detector sampling angles.

[0083] S405, using an iterative algorithm based on a mathematical priori model, the corresponding phase contrast CT image is obtained from the intensity information obtained from the actual detector and the predicted intensity information. ;

[0084] S406: The phase contrast CT image As the starting value for the next fraction generation model sampling, the next round of iteration begins. After the iteration is completed, the final phase CT reconstruction result is output .

[0085] The iterative algorithm may be a Chambolle-Pock algorithm.

[0086] A second embodiment, a coaxial phase contrast CT reconstruction system based on a fractional generation model, is used to implement the method of the first embodiment, including:

[0087] The dataset construction module uses the coaxial phase contrast CT images reconstructed at all projection angles to establish a training dataset; specifically:

[0088] First, a large number of reconstructed coaxial phase-contrast CT images at all projection angles are collected;

[0089] Then, high-quality in-line phase contrast CT images reconstructed at all projection angles are selected as training data sets;

[0090] The convolutional neural network training module first applies SDE perturbations at different times to the training images in the training data set to obtain perturbed images, and then inputs the perturbed images and the training images into the convolutional neural network for training; specifically includes:

[0091] By applying stochastic differential equation (SDE) perturbations at different times to the training images, perturbed images are obtained;

[0092] The perturbed image and the original training image are input into the convolutional neural network for training. The formula of the stochastic differential equation is as follows:

[0093] (1)

[0094] in, represents the training image, represents the drift function, is the diffusion function, d represents the differential, t represents the continuous time variable, and w represents the forward Brownian motion.

[0095] Score matching network learning module, through the score matching network Learn the scores under SDE perturbations at different times; specifically:

[0096] Using the input image, through the score matching network Learn the corresponding scores under SDE perturbations at different times. The specific loss function is as follows:

[0097] (2)

[0098] in, represents the score estimation function in the score generation model, which is parameterized by a neural network; ψ represents the time-dependent weighting coefficient; Represents the score of the training data distribution, that is, the gradient of the logarithmic probability of the training data distribution under the conditions at time t; express The square of the L2 norm.

[0099] The reconstruction module uses the trained convolutional neural network to perform in-axial phase contrast CT reconstruction; specifically includes:

[0100] Parameter initialization unit, initialize the following variables: Reconstruct the starting noise , incomplete angle projection matrix A and various variables required for iterative algorithms;

[0101] The sampling unit starts the sampling process of the score generation model. During each sampling process, the score matching network trained by S3 is first used. Obtaining noisy phase-contrast CT images via a predictive-corrected sampler ;

[0102] The first calculation unit calculates the noisy phase contrast CT image by single material assumption Corresponding absorption CT image ;

[0103] The second calculation unit uses the noisy phase contrast CT image and absorption CT images , the predicted intensity information is calculated by the forward projection physical model of phase contrast CT; the calculation process of the intensity information is as follows:

[0104] (3)

[0105] in, , represents the wavelength of the X-ray, D represents the distance the X-ray travels after passing through the sample, represents the absorption projection, through Multiply it by the CT projection matrix A to get, represents the phase projection, Multiply it by the CT projection matrix A to get; Indicates the intensity information obtained on the detector; represents the transport equation of the sample; represents the Fresnel propagation operator; s represents the one-dimensional coordinate variable on the detector; i represents the imaginary unit; θ represents different detector sampling angles.

[0106] The third calculation unit uses an iterative algorithm based on a mathematical priori model to solve the corresponding phase contrast CT image from the intensity information obtained from the actual detector and the predicted intensity information. ;

[0107] Iteration unit, the phase contrast CT image As the starting value for the next fraction generation model sampling, the next round of iteration begins. After the iteration is completed, the final phase CT reconstruction result is output .

[0108] The iterative algorithm may be a Chambolle-Pock algorithm.

[0109] Table 1 shows the quantitative results of sparse angular projection reconstruction.

[0110]

[0111] Table 2 is the quantitative results of limited angle projection reconstruction

[0112]

[0113] As shown in Tables 1 and 2, quantitative indicators such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of reconstruction results of different types of methods and the present invention at sparse angles and limited angles are quantitatively compared. It can be seen from the table that the reconstruction result obtained by the present invention is closer to the ground truth, indicating that the system method of the present invention has good reconstruction results.

[0114] like Figure 3 As shown, the present invention compares the reconstruction results of the proposed method in limited angles among the standard reconstruction method FBP, the model-based iterative TV method, the end-to-end deep learning network-based FBPConvNet and DDNet algorithms, and the deep learning algorithm EGARNet for coaxial phase contrast CT. Visually, it can be seen that the reconstructed image of the FBP algorithm is seriously affected by limited angle artifacts, while the model-based iterative method cannot completely remove the limited angle artifacts, while the FBPConvNet and DDNet methods can remove most of the limited angle artifacts, but the reconstructed image will lose details, and although EGARNet can remove most of the limited angle artifacts, it will remove residual pseudo structures and deformation. The reconstruction result of the system method of the present invention removes limited angle artifacts and has a high similarity with the reference reconstruction result.

[0115] like Figure 4 As shown, the reconstruction results of the system method of the present invention, the standard reconstruction method FBP, the model-based iterative TV method, the end-to-end deep learning network-based FBPConvNet and the DDNet algorithm, and the deep learning algorithm EGARNet for coaxial phase contrast CT on sparse angle projection are compared. Visually, the reconstructed image of the CT image of the FBP algorithm has serious sparse angle artifacts and noise, while the model-based iterative method cannot remove the sparse angle artifacts cleanly and has over-smoothing. The FBPConvNet and DDNet methods can remove most of the sparse angle artifacts, but the reconstructed image will lose details. Although EGARNet can remove most of the sparse angle artifacts, it removes residual pseudo structures and deformations. The reconstruction result of the system method of the present invention removes the sparse angle artifacts and has a high similarity with the reference reconstruction result.

[0116] The present invention has the following characteristics:

[0117] (1) Compared with the traditional coaxial phase contrast CT reconstruction method, the present invention does not require an additional phase recovery step. The present method integrates the physical model of coaxial phase contrast from phase contrast CT image to intensity information into the sampling framework of the fractional generation model, and directly uses the detector intensity information to constrain the generation process of phase contrast CT.

[0118] (2) For the model-based incomplete projection iterative reconstruction algorithm of coaxial phase contrast CT, the present invention eliminates the need to manually set the model prior, thereby avoiding the over-smoothing phenomenon and residual artifact problems that may occur in the reconstructed image. This method learns the high-quality image distribution through the score generation model, and combines the physical model of coaxial phase contrast CT as a constraint, so that the reconstructed image is subject to the dual constraints of model prior and data prior, thereby achieving higher quality and more accurate coaxial phase contrast CT image reconstruction.

[0119] (3) Traditional end-to-end deep learning networks achieve CT reconstruction under incomplete projection angles by learning the mapping relationship in paired data. However, due to the scarcity of coaxial phase contrast CT data, it is difficult to obtain a large amount of paired data for training. In addition, the network fails to fully utilize the measured projection information and the physical model of coaxial phase contrast imaging, which may lead to the generation of pseudo structures. In view of this, the present invention adopts a fractional generation model, whose training process does not require paired data. It only needs to use the coaxial phase contrast CT images reconstructed under the complete projection angle to learn the probability distribution of high-quality coaxial phase contrast CT images. In the image sampling and generation process, the present invention combines the physical model of coaxial phase contrast CT and the intensity information measured on the detector to constrain image generation, ensure that the generated image is subject to physical constraints, prevent the generation of pseudo structures, and can directly reconstruct high-quality phase contrast CT from the intensity information without the need for an additional phase recovery step.

[0120] Figure 3 Explanation of the English abbreviations in:

[0121] FBP: Filtered Back Projection TV iterative algorithm based on total variation regularization term;

[0122] FBPConvnet: Filtered Back-Projection Convolutional Neural Network;

[0123] DDNet: Dense Atrous Convolutional Neural Network;

[0124] EGARNet: Edge-guided Artifact Removal Network;

[0125] Proposed Method: Proposed method.

[0126] A third embodiment is a computer-readable storage medium storing a computer program, which implements the above-mentioned in-axial phase contrast CT reconstruction method based on the fractional generation model when executed by a processor.

[0127] A fourth embodiment is a computer program product, comprising a computer program, which implements the above-mentioned in-axial phase contrast CT reconstruction method based on the fractional generation model when executed by a processor.

[0128] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0129] The above is only a preferred embodiment of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A coaxial phase contrast CT reconstruction method based on a fractional generative model, characterized in that: Applications in X-ray phase contrast CT imaging, including: S1. Establish a training data set using in-axial phase contrast CT images reconstructed at all projection angles; S2, first applying SDE perturbations at different times to the training images in the training data set to obtain perturbed images, and then inputting the perturbed images and the training images into a convolutional neural network for training; S3. Through the score matching network Learn the scores under SDE perturbations at different times; S4. Reconstructing coaxial phase contrast CT using the trained convolutional neural network; specifically including: S401, initialize the following variables: Reconstruct the starting noise , incomplete angle projection matrix A and various variables required for iterative algorithms; S402, start the sampling process of the score generation model. In each sampling process, first use the score matching network trained in S3 Obtaining noisy phase-contrast CT images via a predictive-corrected sampler ; S403, calculating noisy phase contrast CT image by single material assumption Corresponding absorption CT image ; S404, using noisy phase contrast CT images and absorption CT images , the intensity information predicted by the forward projection physical model of phase contrast CT is calculated; S405, using an iterative algorithm based on a mathematical priori model, the corresponding phase contrast CT image is obtained from the intensity information obtained from the actual detector and the predicted intensity information. ; S406: The phase contrast CT image As the starting value for the next fraction generation model sampling, the next round of iteration begins. After the iteration is completed, the final phase CT reconstruction result is output .

2. The in-axial phase contrast CT reconstruction method based on the fractional generative model according to claim 1, characterized in that: In S2, the formula for the SDE perturbation is as follows: (1) in, represents the training image, represents the drift function, is the diffusion function; d represents the differential, t represents the continuous time variable, and w represents the forward Brownian motion.

3. The in-axial phase contrast CT reconstruction method based on the fractional generative model according to claim 1, characterized in that: In S3, the formula of the loss function is as follows: (2) in, represents the score estimation function in the score generation model, which is implemented by neural network parameterization; ψ represents the network parameter; represents the time-dependent weighting coefficient; Represents the score of the training data distribution, that is, the gradient of the logarithmic probability of the training data distribution under the conditions at time t; express The square of the L2 norm.

4. The in-axial phase contrast CT reconstruction method based on the fractional generative model according to claim 1, characterized in that: In S404, the calculation process of the intensity information is as follows: (3) in, , represents the wavelength of the X-ray, D represents the distance the X-ray travels after passing through the sample, represents the absorption projection, through Multiply it by the CT projection matrix A to get, represents the phase projection, Multiply it by the CT projection matrix A to get; Indicates the intensity information obtained on the detector; represents the transport equation of the sample; represents the Fresnel propagation operator; s represents the one-dimensional coordinate variable on the detector; i represents the imaginary unit; θ represents different detector sampling angles.

5. The in-axis phase contrast CT reconstruction method based on the fractional generative model according to claim 1, characterized in that: In S405, the iterative algorithm is the Chambolle-Pock algorithm.

6. A coaxial phase contrast CT reconstruction system based on a fractional generation model, characterized in that: Applications in X-ray phase contrast CT imaging, including: The data set construction module uses the in-line phase contrast CT images reconstructed at all projection angles to establish a training data set; A convolutional neural network training module first applies SDE perturbations at different times to the training images in the training data set to obtain perturbed images, and then inputs the perturbed images and the training images into the convolutional neural network for training; Score matching network learning module, through the score matching network Learn the scores under SDE perturbations at different times; The reconstruction module uses the trained convolutional neural network to perform in-axial phase contrast CT reconstruction; specifically includes: Parameter initialization unit, initialize the following variables: Reconstruct the starting noise , incomplete angle projection matrix A and various variables required for iterative algorithms; The sampling unit starts the sampling process of the score generation model. In each sampling process, the score matching network trained by the score matching network learning module is first used. Obtaining noisy phase-contrast CT images via a predictive-corrected sampler ; The first calculation unit calculates the noisy phase contrast CT image by single material assumption Corresponding absorption CT image ; The second calculation unit uses the noisy phase contrast CT image and absorption CT images , the intensity information predicted by the forward projection physical model of phase contrast CT is calculated; The third calculation unit uses an iterative algorithm based on a mathematical priori model to solve the corresponding phase contrast CT image from the intensity information obtained from the actual detector and the predicted intensity information. ; Iteration unit, the phase contrast CT image As the starting value for the next fraction generation model sampling, the next round of iteration begins. After the iteration is completed, the final phase CT reconstruction result is output .

7. The in-axis phase contrast CT reconstruction system based on the fractional generative model according to claim 6, characterized in that: In the convolutional neural network training module, the formula for SDE perturbation is as follows: (1) in, represents the training image, represents the drift function, is the diffusion function; d represents the differential, t represents the continuous time variable, and w represents the forward Brownian motion.

8. The in-axis phase contrast CT reconstruction system based on the fractional generative model according to claim 6, characterized in that: The formula of the loss function is as follows: (2) in, represents the score estimation function in the score generation model, which is parameterized by a neural network; ψ represents the time-dependent weighting coefficient; Represents the score of the training data distribution, that is, the gradient of the logarithmic probability of the training data distribution under the conditions at time t; express The square of the L2 norm; The intensity information is calculated as follows: (3) in, , represents the wavelength of the X-ray, D represents the distance the X-ray travels after passing through the sample, represents the absorption projection, through Multiply it by the CT projection matrix A to get, represents the phase projection, Multiply it by the CT projection matrix A to get; Indicates the intensity information obtained on the detector; represents the transport equation of the sample; represents the Fresnel propagation operator; s represents the one-dimensional coordinate variable on the detector; i represents the imaginary unit; θ represents different detector sampling angles.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the in-axial phase contrast CT reconstruction method based on the fractional generation model described in any one of claims 1 to 5 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the in-axial phase contrast CT reconstruction method based on the fractional generation model as described in any one of claims 1 to 5 is implemented.

Images