Image reconstruction method and system based on fast sampling score generation model

Abstract

The application discloses an image reconstruction method and system based on a fast sampling score generation model, and the method comprises the following steps: acquiring a score generation model and a limited-angle CT image; performing noise injection processing on the limited-angle CT image through a forward process of the score generation model to generate Gaussian noise features; and performing image reconstruction on the limited-angle CT image through a backward process of the score generation model according to the Gaussian noise features to generate a CT reconstructed image. The application greatly reduces the sampling steps of image reconstruction, significantly speeds up the sampling process, improves the efficiency and rate of image reconstruction, effectively alleviates the negative influence of directional artifacts generated by limited-angle CT scanning in the image reconstruction process, maintains clear edges and detailed features of the image while realizing fast reconstruction, and improves the quality and effect of image reconstruction. The application is applied to the technical field of image reconstruction.

Description

Technical Field

[0001] This invention relates to the field of image reconstruction technology, and in particular to image reconstruction methods and systems based on a fast sampling score generation model. Background Technology

[0002] Limited-Angle Computed Tomography (LACT) reconstruction is a classic inverse problem, producing reconstructed images exhibiting edge divergence and artifacts. These degraded images pose a significant challenge to clinical diagnosis, making the acquisition of high-quality images from limited-angle data a formidable task. In recent years, score-based generative models (SGMs) have garnered considerable attention due to their excellent modeling capabilities for complex data distributions. A typical SGM employs a two-stage generative scheme, including a forward stage that introduces noise into the original data and an inverse stage that recovers the original data from the noise. SGMs have demonstrated remarkable success in limited-angle CT reconstruction applications.

[0003] In limited-angle CT reconstruction applications, existing SGM (Sampling Generalization) techniques generally suffer from a slow sampling rate. Taking the latest SGM models DOLCE and DPS as examples, their sampling step size is typically set to 1000 or 2000, resulting in a reconstruction time of approximately 40 minutes per image. This severely limits the application of SGM models in clinical diagnosis. To improve sampling efficiency, related techniques employ skip sampling by increasing the sampling interval, thereby accelerating the sampling rate. For instance, Denoising Diffusion Implicit Models (DDIM) improves the sampling process by using skip sampling, which achieves rapid sampling at fixed intervals and continuously increases the speed with a constant coefficient. However, when implementing skip sampling, the noise variance range is too wide, and the number of sampling points is relatively small. The large noise range leads to the loss of some complex details, while the reduced number of sampling points results in higher model instability, thus reducing the image reconstruction effect and leading to unstable and low-quality sampling results. Summary of the Invention

[0004] The purpose of this invention is to at least partially solve one of the technical problems existing in the prior art.

[0005] Therefore, the purpose of this invention is to provide an image reconstruction method and system based on a fast sampling score generation model.

[0006] To achieve the above-mentioned technical objectives, the technical solutions adopted in the embodiments of the present invention include:

[0007] On one hand, embodiments of the present invention provide an image reconstruction method based on a fast sampling score generation model, comprising the following steps:

[0008] Acquire the scoring generation model and finite-angle CT images;

[0009] The finite-angle CT image is denoised using the forward process of the scoring generation model to generate Gaussian noise features.

[0010] Based on the Gaussian noise characteristics, the finite-angle CT image is reconstructed through the backward process of the scoring generation model to generate a CT reconstructed image.

[0011] In addition, the image reconstruction method based on the fast sampling score generation model according to the above embodiments of the present invention may also have the following additional technical features:

[0012] Furthermore, in one embodiment of the present invention, the step of performing noise injection processing on the finite-angle CT image through the forward process of the scoring generation model to generate Gaussian noise features includes:

[0013] The forward process of the score generation model is defined by stochastic differential equations;

[0014] The noise features are injected into the finite-angle CT image through the forward process to generate Gaussian noise features.

[0015] Further, in one embodiment of the present invention, the step of reconstructing the finite-angle CT image based on the Gaussian noise features through the backward process of the score generation model to generate a CT reconstructed image includes:

[0016] The timing information of the backward process is obtained, and the backward process is divided into several timing reconstruction stages based on the timing information, wherein the Gaussian noise feature is used as the input feature of the first timing reconstruction stage.

[0017] Determine the current temporal reconstruction stage, sample the input features of the current temporal reconstruction stage, and generate the image sampling features of the current temporal reconstruction stage;

[0018] Using the image sampling features of the current temporal reconstruction stage as the input features of the next temporal reconstruction stage, and taking the next temporal reconstruction stage as the current temporal reconstruction stage, the process returns to the step of determining the current temporal reconstruction stage.

[0019] When all temporal reconstruction stages have been sampled, the image sampling features from the last temporal reconstruction stage are used as the CT reconstructed image and output.

[0020] Further, in one embodiment of the present invention, sampling the input features of the current temporal reconstruction stage to generate image sampling features for the current temporal reconstruction stage includes:

[0021] Skip sampling is performed on the input features of the current time-series reconstruction stage to obtain the first sampled features;

[0022] The first sampling feature is subjected to time back sampling to obtain the second sampling feature, wherein the time sequence of the second sampling feature is earlier than that of the first sampling feature;

[0023] The second sampling feature is resampled to obtain a third sampling feature, wherein the timing of the third sampling feature is the same as that of the first sampling feature;

[0024] Compressed sensing processing is performed on the third sampling feature to generate the image sampling feature for the current temporal reconstruction stage.

[0025] Furthermore, in one embodiment of the present invention, the step of skipping sampling of the input features in the current time-series reconstruction stage to obtain the first sampled features includes:

[0026] Obtain the noise interval for skip sampling, and calculate the first noise scale based on the noise interval;

[0027] The first sampled feature is obtained by skipping the input features of the current time-series reconstruction stage according to the first noise scale.

[0028] Furthermore, in one embodiment of the present invention, the step of performing time-backtracking sampling on the first sampling feature to obtain the second sampling feature includes:

[0029] Obtain the step size information of time backtracking, and calculate the second noise scale based on the step size information;

[0030] The first sampling feature is backsampled according to the second noise scale, so that the temporal sequence of the first sampling feature gradually approaches the temporal sequence of the input feature in the current temporal reconstruction stage, thereby generating the second sampling feature.

[0031] Furthermore, in one embodiment of the present invention, the resampling of the second sampling feature to obtain the third sampling feature includes:

[0032] Obtain a sampler for resampling, the sampler including a predictor and a corrector;

[0033] The predictor resamples the second sampled feature, and the corrector corrects the resampled feature to obtain the third sampled feature.

[0034] Furthermore, in one embodiment of the present invention, the step of performing compressed sensing processing on the third sampling feature to generate the image sampling feature for the current temporal reconstruction stage includes:

[0035] Obtain the sampling operator;

[0036] The sampling operator performs compressed sensing processing on the third sampling feature to generate the image sampling feature for the current temporal reconstruction stage.

[0037] Furthermore, in one embodiment of the present invention, the sampling operator is a sampling operator based on a diagonal total score regularization term.

[0038] On the other hand, embodiments of the present invention provide an image reconstruction system based on a fast sampling score generation model, comprising:

[0039] The acquisition module is used to acquire the scoring generation model and finite-angle CT images;

[0040] The forward processing module is used to perform noise injection processing on the finite-angle CT image through the forward process of the score generation model to generate Gaussian noise features;

[0041] The image reconstruction module is used to reconstruct the image from the Gaussian noise features through the backward process of the score generation model, and generate a CT reconstructed image.

[0042] The beneficial effects of this invention are: it provides an image reconstruction method and system based on a fast sampling score generation model, which can significantly reduce the sampling steps in image reconstruction, significantly accelerate the sampling process, improve the efficiency and speed of image reconstruction, and achieve rapid reconstruction; at the same time, during the image reconstruction process, it can effectively alleviate the negative impact of directional artifacts generated by limited-angle CT scanning, and can maintain the clear edges and details of the image while rapidly reconstructing, thereby improving the quality and effect of image reconstruction.

[0043] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0045] Figure 1 This is a flowchart of the image reconstruction method based on a fast sampling score generation model provided by the present invention;

[0046] Figure 2 This is a schematic diagram of the image reconstruction method based on a fast sampling score generation model provided by the present invention.

[0047] Figure 3 This is a schematic diagram of the fast sampling principle provided by the present invention;

[0048] Figure 4 This is a flowchart of fast sampling in any time-series reconstruction stage provided by the present invention;

[0049] Figure 5 This is a distortion comparison diagram of a CT image with limited angles provided by the present invention;

[0050] Figure 6 This is a comparison chart of the effects of the TIFA method and other comparative methods on the reconstruction task of 90-degree CT images provided by this invention;

[0051] Figure 7 This is another comparison chart showing the effect of the TIFA method and other comparative methods in the reconstruction task of 90-degree CT images provided by this invention;

[0052] Figure 8 This is a comparison chart of the effects of the TIFA method and other comparative methods on the reconstruction task of 60-degree CT images provided by this invention.

[0053] Figure 9 This is another comparison chart showing the effect of the TIFA method and other comparative methods in the reconstruction task of 60-degree CT images provided by this invention;

[0054] Figure 10 This is another comparison chart showing the effect of the TIFA method and other comparative methods in the reconstruction task of 90-degree CT images provided by this invention;

[0055] Figure 11 This is a comparison chart of the effects of the TIFA method, DPS, and DOLCE methods on the reconstruction of 90-degree CT images provided by this invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0058] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0060] Limited-Angle Computed Tomography (LACT) reconstruction is a classic inverse problem, producing reconstructed images exhibiting edge divergence and artifacts. These degraded images pose a significant challenge to clinical diagnosis, making the acquisition of high-quality images from limited-angle data a formidable task. In recent years, score-based generative models (SGMs) have garnered considerable attention due to their excellent modeling capabilities for complex data distributions. Examples include Denoising Diffusion Probabilistic Models (DDPM), Stochastic Differential Equations (SDE), and Denoising Diffusion Implicit Models (DDIM). SGMs have demonstrated remarkable success in limited-angle CT reconstruction applications. A typical SGM employs a two-stage generative scheme, including a forward stage that introduces noise into the original data and a backward stage that recovers the original data from the noise. This backward process is typically implemented using parameterized deep neural networks; for instance, SGMs can be fine-tuned using training data.

[0061] In limited-angle CT reconstruction applications, existing SGM (Scanning Generalized Measuring) techniques generally suffer from a problem: slow sampling speed. Taking the latest SGM models DOLCE and DPS as examples, their sampling step size is typically set to 1000 or 2000, resulting in a reconstruction time of approximately 40 minutes for a single image. This severely limits the application of SGM models in clinical diagnosis. To improve the efficiency of the sampling process, related techniques employ methods such as increasing the sampling interval to achieve skip sampling, thereby accelerating the sampling rate. For example, DDIM improved the sampling process by using a skip sampling technique. This technique achieves rapid sampling at fixed intervals and continuously increases the speed with a constant coefficient.

[0062] However, achieving this accelerated sampling typically requires a trade-off between reconstruction quality and sampling rate. Specifically, in skip sampling, the noise variance range is too wide, and the number of sampling points is relatively small. While the noise distribution and sampling point distribution of skip sampling can effectively speed up the sampling process, the large noise range will cause the loss of some complex details, and the reduction in the number of sampling points will lead to higher model instability, thereby reducing the image reconstruction effect and resulting in unstable and low-quality sampling results. Since skip sampling cannot achieve a balance between sampling rate and reconstruction quality, simply increasing the sampling interval to pursue accelerated sampling has proven to be inappropriate.

[0063] Therefore, in the application of CT reconstruction with limited angles, how to achieve rapid image reconstruction while maintaining the quality and effect of image reconstruction has become one of the urgent problems to be solved.

[0064] To address the problems and deficiencies of related technologies, this invention provides an image reconstruction method and system based on a fast sampling score generation model. It introduces a derivative-based generative model (SDE), using the forward process of SDE to gradually inject noise into the finite-angle CT image to be reconstructed. Subsequently, the backward process of SDE is used to gradually denoise the noisy data, obtaining the reconstructed CT image, thus achieving the reconstruction of the finite-angle CT image. To accelerate sampling of coarse noise while simultaneously sampling fine noise, this invention introduces large-scale skip sampling, time-backtracking sampling with resampling, and compressed sensing sampling in the backward process. Large-scale skip sampling and time-backtracking sampling with resampling effectively capture coarse structural details while obtaining more refined features under significant noise levels. By introducing a diagonal total variation (DTV) regularization term in compressed sensing sampling to reduce the negative impact of diagonal orientation artifacts in the finite-angle CT image, this invention can achieve rapid reconstruction while maintaining clear edge and detail features of the image.

[0065] First, the following will describe in detail, with reference to the accompanying drawings, one implementation step of the image reconstruction method based on the fast sampling score generation model provided by the present invention.

[0066] The method described in this invention can be applied to a terminal, a server, or software running on either a terminal or a server. The terminal can be a tablet computer, a laptop computer, a desktop computer, etc., but is not limited to these. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0067] Reference Figure 1 and Figure 2 , Figure 1 This is a flowchart of the image reconstruction method based on a fast sampling score generation model provided by the present invention. Figure 2 This is a schematic diagram of the image reconstruction method based on a fast sampling score generation model provided by the present invention. Figure 2 The training process is a forward process, which follows the training of the stochastic differential equation (SDE). The testing process is a backward process, which introduces large-scale jump sampling, time-backtracking sampling with resampling, and compressed sensing sampling. The method may include, but is not limited to, the following steps:

[0068] S100, acquires the scoring generation model and finite-angle CT images.

[0069] It should be noted that the score generation model is an improved SDE model, which is a typical example in the field of generative models and is widely used in the field of medical imaging to solve inverse problems.

[0070] Specifically, based on the typical SDE model, the embodiments of the present invention maintain the original forward process of SDE, and introduce large-scale jump sampling, time backtracking sampling with resampling, and compressed sensing sampling based on DTV regularization in the backward process of SDE, thereby constructing an improved SDE model, and using the improved SDE model to realize image reconstruction.

[0071] S200 performs noise injection on CT images at limited angles through the forward process of the scoring generation model to generate Gaussian noise features.

[0072] In this step, the forward process is the training process. During the forward process of the score generation model, noise is gradually injected into clean CT images with limited angles to shape the training process of the neural network and generate Gaussian noise features.

[0073] S300, based on Gaussian noise characteristics, performs image reconstruction on CT images at limited angles through a backward process of a scoring generation model, generating CT reconstructed images.

[0074] In this step, the backward process of the score generation model, also known as the iterative denoising process or testing process, represents the image reconstruction stage in image reconstruction. It depicts a perturbation process that reverses the initial forward process. In the backward process of the score generation model, the backward process is divided into multiple ordered reconstruction stages. In each stage, the noisy data is sequentially subjected to large-scale jump sampling, time-backtracking sampling with resampling, and compressed sensing sampling. Through multiple stages of sampling operations, the noisy image is gradually denoised, thereby generating a CT reconstructed image.

[0075] In some embodiments of the present invention, reference is made to... Figure 2 In step S200, the step of generating Gaussian noise features by annotating the CT image at a limited angle through the forward process of the scoring generation model includes:

[0076] S210 defines the forward process of the score generation model through stochastic differential equations.

[0077] In this step, during the entire forward process, the input data x is defined as x(t) = xt, and the time t∈[0,1]. Then, the forward process of the score generation model can be represented by the following formula (1):

[0078] dx=f(x,t)dt+g(t)dw,(1)

[0079] Where f(·, t) represents the linear drift function, g(t) represents the scalar diffusion coefficient, the vector w follows standard Brownian motion, and dt represents an infinitesimal time step. Intuitively, given a small time step dt, the data x is supplemented with normally distributed random value noise with mean f(x, t) and variance g(t)dw.

[0080] Optionally, different combinations of linear drift functions and scalar diffusion coefficients can generate different types of SDEs. This embodiment of the invention employs a typical explosion mutation SDE model, where the linear drift function and scalar diffusion coefficient can be set to f = 0. Where σ(t) represents a proportional function that gradually increases the time variation of the noise. However, it should be noted that in other embodiments of the present invention, the linear drift function and the scalar diffusion coefficient can also be set to other parameters, and the present invention does not specifically limit them.

[0081] S220 injects noise features into finite-angle CT images through a forward process to generate Gaussian noise features.

[0082] In this step, through the forward process defined by the previous formula (1), noise features are gradually injected into the finite-angle CT image x0 to generate Gaussian noise features x. T ,like Figure 2 As shown. Here, x0 represents the distribution of the original data, i.e., finite-angle CT images, x... T It approximates a spherical Gaussian distribution, where T represents the number of sampling steps.

[0083] In some embodiments of the present invention, given a forward process as defined in formula (1) above, a corresponding reverse process can be constructed. The reverse process is represented by the inverse of formula (1), as shown in formula (2) below:

[0084]

[0085] in, Let be a time-dependent scoring function. Optionally, an estimator s for this scoring function. θ The score is obtained by training using a denoising score matching method, as shown in formula (3) below:

[0086]

[0087] Where p(x) t |x0) corresponds to a perturbation kernel, which is used to perturb the data x0 to the noisy sample x. T Fraction estimator θ (xt ,t) can replace the scoring function in formula (2).

[0088] Reference Figure 3 , Figure 3 This is a schematic diagram of the fast sampling principle provided by the present invention. Based on the reverse process defined by formula (2), the embodiments of the present invention introduce large-scale jump sampling, time backtracking sampling with resampling and compressed sensing sampling in the reverse process, aiming to perform fast sampling through coarse noise scale and fine noise scale.

[0089] Specifically, in step S300, based on the Gaussian noise characteristics, the CT image at a limited angle is reconstructed through the backward process of the scoring generation model. The process of generating the CT reconstructed image may include, but is not limited to, the following steps:

[0090] S310: Obtain the timing information of the backward process, and divide the backward process into several timing reconstruction stages based on the timing information.

[0091] It should be noted that Gaussian noise features are used as the input features for the first time-series reconstruction stage.

[0092] In this step, based on the temporal sequence of the backward process, the entire image reconstruction stage is divided into several temporal reconstruction stages with sequential and temporal relationships. Specifically, the output of the forward process, i.e., the Gaussian noise feature, serves as the input to the first temporal reconstruction stage, and the output of the last temporal reconstruction stage serves as the CT reconstructed image. For other temporal reconstruction stages besides the first one, the output of the previous temporal reconstruction stage serves as the input to the next temporal reconstruction stage.

[0093] S320, determine the current temporal reconstruction stage, sample the input features of the current temporal reconstruction stage, and generate image sampling features for the current temporal reconstruction stage.

[0094] In this step, for the i-th temporal reconstruction stage, the input features of the i-th temporal reconstruction stage are subjected to large-scale skip sampling, time-backtracking sampling with resampling, and compressed sensing sampling to generate the image sampling features of the i-th temporal reconstruction stage.

[0095] S330, using the image sampling features of the current temporal reconstruction stage as the input features of the next temporal reconstruction stage, and taking the next temporal reconstruction stage as the current temporal reconstruction stage, return to step S320.

[0096] In this step, when the sampling of the i-th temporal reconstruction stage is completed, the image sampling features of the i-th temporal reconstruction stage are used as the input of the (i+1)-th temporal reconstruction stage, and the (i+1)-th temporal reconstruction stage is used as the i-th temporal reconstruction stage, that is, let i = i+1, and then return to step S320.

[0097] S340. When the sampling of all the time-series reconstruction stages is completed, the image sampling features of the last time-series reconstruction stage are used as the CT reconstruction image and output.

[0098] In this step, the Gaussian noise feature x is sampled and denoised successively through multiple time-series reconstruction stages, T such that the Gaussian noise feature x T gradually transitions to x0, and then the CT reconstruction image x0 is obtained.

[0099] In some embodiments of the present invention, referring to Figure 3 and Figure 4 , Figure 4 is the flowchart of the fast sampling of any time-series reconstruction stage provided by the present invention. In step S320, when sampling the input features of the current time-series reconstruction stage to generate the image sampling features of the current time-series reconstruction stage, the implementation process may include but is not limited to the following steps:

[0100] S321. Perform skip sampling on the input features of the current time-series reconstruction stage to obtain the first sampling feature.

[0101] In this step, accelerate the sampling on the coarse noise β t scale through skip sampling, that is, perform one-step sampling on the large-scale noise, which is equivalent to performing m-step sampling on the small-scale noise, to obtain the first sampling feature x'. t .

[0102] S322. Perform time-backtracking sampling on the first sampling feature x' t to obtain the second sampling feature x' t+L .

[0103] It should be noted that the time sequence of the second sampling feature is earlier than that of the first sampling feature.

[0104] S323. Resample the second sampling feature x' t+L to obtain the third sampling feature

[0105] It should be noted that the time sequence of the third sampling feature is the same as that of the first sampling feature.

[0106] In the above steps, after performing one-step sampling on a larger scale, perform time backtracking on a smaller scale to regress the reconstructed image from x' t to x' t+L , where L represents the time-backtracking steps that are satisfied, and L < m. Subsequently, perform image resampling and reconstruction under small-scale noise, transitioning from x' t+L to where, and x′ t The timing sequence is the same. Through the above process, time backtracking provides refinement correction for large-scale reconstruction, improving the speed and stability of the image reconstruction process.

[0107] S324, for the third sampling feature Compressed sensing processing is performed to generate image sampling features x for the current temporal reconstruction stage. t .

[0108] In this step, to further improve the stability of image reconstruction and achieve faster sampling efficiency and more stable image reconstruction, a diagonal total variational regularization term is introduced into compressed sensing, and the third sampled features are processed using compressed sensing. The image sampling features x at the current temporal reconstruction stage are obtained through processing. t This enables CT reconstruction at limited angles.

[0109] Reference Figure 3 Traditional score-based reconstruction methods, such as VE-SDE, use sequential sampling to make the Gaussian noise features x T The process gradually transitions to CT-reconstructed image x0. In practical applications, the choice of T is crucial; T is typically set to values ​​such as 2000 or 4000, which leads to significant computational resource overhead when solving the inverse problem. Therefore, to reduce computational resource overhead and improve sampling efficiency, methods such as DDIM implement skip sampling by increasing the sampling interval. However, the noise scale of skip sampling spans a wide range and is not corrected under large strides, which results in skip sampling failing to achieve high-quality image reconstruction.

[0110] In the traditional T-step stepwise VE-SDE, the choice of noise variance satisfies and as well as in, This represents the distribution of perturbation data. In image reconstruction and sampling, the span of noise is defined as α. t =σ t+1 2 -σ t 2 In the skip sampling scoring model, the noise span is relatively large. Taking m-step skip sampling as an example, the noise span can be expressed as β. t =σ t+m 2 -σ t 2 Clearly, regardless of the value of t, β t Always greater than α tThis means that the noise is coarse when implementing skip sampling, and coarse noise usually leads to poor reconstruction results. In order to obtain better image reconstruction results, the embodiments of the present invention introduce large-scale skip sampling, time-backtracking sampling with resampling, and compressed sensing sampling in the backward process.

[0111] The following will refer to Figures 2 to 5 The implementation process of large-scale skip sampling, time-backtracking sampling with resampling, and compressed sensing sampling introduced in the backward process of the embodiments of the present invention is described in detail.

[0112] 1. The implementation process of large-scale skip sampling is as follows:

[0113] In this embodiment of the invention, the basis for fast sampling lies in performing large-scale skip sampling. Specifically, large-scale skip sampling is achieved by adding a noise interval during the sampling process, that is, sampling is performed at a coarse noise level. Step S321, the process of skip sampling the input features of the current time-series reconstruction stage to obtain the first sampled features mainly includes the following steps:

[0114] The first step is to obtain the noise interval of the skip sampling, and then calculate the first noise scale based on the noise interval.

[0115] In common scoring models, T is typically set to 2000 or 4000, corresponding to a sampling noise scale of α. t =σ t+1 2 -σ t 2 In the large-scale jump of this embodiment of the invention, the noise interval m is obtained, and the first noise scale β is calculated based on the interval m. t =σ t+m 2 -σ t 2 This increases the sampling rate by m times compared to the traditional scoring model.

[0116] The second step is to analyze the input features x of the current time-series reconstruction stage based on the first noise scale. t+m Perform skip sampling to obtain the first sampled feature x′ t .

[0117] Optionally, in the input features x of the current time series reconstruction stage t+m The steps for performing skip sampling may include:

[0118] Obtain the predictor and corrector for skip sampling, and apply the predictor to the input features x of the current time series reconstruction stage. t+m To make a prediction, the output of the predictor is corrected by a corrector to obtain the first sampled feature x'. t .

[0119] In the process of skip sampling, a data consistency method is used to iteratively constrain large-scale skip sampling, thereby realizing forward iteration in the backward process. Optionally, the Simultaneous Iterative Reconstruction Technique (SIRT) is used as the data consistency method.

[0120] In this step, the sampler is pre-set with a certain number of iterations, and at least one iteration is performed based on this number. In a single iteration, the predictor processes the input features x of the current time-series reconstruction stage. t+m Prediction is performed through skip sampling. Then, the predictor's output is constrained for data consistency using SIRT. The updated predictor output is fed into the corrector for correction. SIRT is then used to constrain the corrector's output for data consistency. The updated corrector output is returned to the predictor for iteration. The final iteration outputs the first sampled feature x'. t .

[0121] 2. The implementation process of time-backtracking sampling with resampling is as follows:

[0122] In this embodiment of the invention, the time backtracking sampling process is related to forward noise injection. During the time backtracking process, from x' t to x' t+1 The image can be represented by the following formula (4):

[0123]

[0124] Based on this, we continue the recursive derivation to x' t+L We can obtain the following formula (5):

[0125]

[0126] By reparameterizing the above formula (5), we can obtain the following formula (6):

[0127]

[0128] This means that it is possible to directly obtain the mean x' t The sum and covariance matrix are Sampling is performed in the distribution to achieve time backtracking sampling.

[0129] Based on this, the step of performing time-backtracking sampling on the first sampling feature to obtain the second sampling feature in step S322 mainly includes:

[0130] The first step is to obtain the step size information of time backtracking, and then calculate the second noise scale based on the step size information.

[0131] In this step, the step size information l of the time backtracking is obtained, and the second noise scale σ is calculated based on the step size information l. t+l 2 -σ t+l-1 2 .

[0132] The second step is to perform time-backtracking sampling on the first sampling feature according to the second noise scale, so that the temporal sequence of the first sampling feature gradually approaches the temporal sequence of the input feature in the current temporal reconstruction stage, and generate the second sampling feature.

[0133] In this step, based on the second noise scale σ t+l 2 -σ t+l-1 2 From the previous formula (6), the mean value shown by formula (6) is x' t The sum and covariance matrix are The distribution of the first sampled feature x′ t Perform time-backtracking sampling so that the first sampled feature x′ t Reverse transition to the second sampling feature x′ t+L .

[0134] The second sampling feature x′ obtained after time backtracking t+L It requires a reverse process of resampling to return to x′. t In this embodiment of the invention, a PC sampler is used to sample the second feature x′. t+L Resampling and reconstruction are performed. The PC sampler consists of two components: a predictor for numerically solving the SDE and a corrector for refining the results using a score-based method. Specifically, by discretizing the sampling process based on the contrast-driven scoring model, i.e., VE-SDE, the formula (7) characterizing the prediction process of the predictor can be obtained:

[0135]

[0136] The initial value l is equal to L.

[0137] The steps for updating the corrector follow formula (8):

[0138]

[0139] Based on this, in step S323, the second sampling feature is resampled to obtain the third sampling feature, including:

[0140] The first step is to obtain the sampler used for resampling, which includes a predictor and a corrector;

[0141] The second step involves using the predictor to process the second sampled feature x. t+L Resampling is performed, and the resampled features are corrected by a corrector to obtain the third sampled feature x. t .

[0142] Optionally, during the resampling process, a data consistency method is used to iteratively constrain the resampling process, thereby achieving forward iteration in the backward process.

[0143] Optionally, SIRT can be used as the data consistency method.

[0144] In this step, the sampler is pre-set with a certain number of iterations, and at least one iteration is performed based on this number. In a single iteration, the predictor samples the second feature x. t+L Prediction is performed, i.e., resampling. Then, the predictor's output is constrained for data consistency using SIRT. The updated predictor's output is fed into the corrector for correction. The corrector's output is then constrained for data consistency using SIRT. The updated corrector's output is returned to the predictor for iteration. The final iteration outputs the third sampled feature x. t .

[0145] For example, the output of the corrector is constrained for data consistency using SIRT, as shown in the following formula (9):

[0146]

[0147] Where D = diag{1 / ||b1||, 1 / ||b2||, ..., 1 / ||b q ||} represents a diagonal matrix, b q Representation matrix The q-th row vector.

[0148] 3. The implementation process of compressed sensing sampling is as follows:

[0149] Reference Figure 5 , Figure 5 This is a distortion comparison image of a finite-angle CT image provided by the present invention, which is composed of... Figure 5It is evident that the deformation in finite-angle CT images primarily originates from diagonal stretching. Obvious artifacts are visible at the diagonal locations indicated by the circles, with significantly more artifacts along the diagonal than in other areas. In compressed sensing, traditional TV regularization only uses gradient operators in the horizontal and vertical directions to sparsely represent the image, while the shape distortion caused by finite angles is mainly manifested along the diagonal direction. Therefore, calculating the TV operator along the diagonal direction of the image is more suitable for the reconstruction task of finite-angle CT images. Furthermore, during the large-scale jump of m steps, the huge noise span will lead to unstable and low-quality sampling results. Therefore, considering both the noise span problem caused by large-scale jumps and the diagonal artifact problem in finite-angle CT images, this embodiment of the invention introduces a diagonal total variational regularization term, namely the DTV regularization term, into the compressed sensing sampling.

[0150] Specifically, in step S324, the process of performing compressed sensing processing on the third sampling features to generate image sampling features for the current temporal reconstruction stage may include, but is not limited to, the following steps:

[0151] The first step is to obtain the sampling operator.

[0152] It should be noted that the sampling operator is a sampling operator based on the DTV regularization term.

[0153] In this step, as t decreases, the magnitude of image change also decreases accordingly. Therefore, the weight μ(t) of the DTV regularization term is adjusted as t changes. The operation of the sampling operator based on the DTV regularization term is shown in the following formula (10):

[0154]

[0155] in, Denotes the sampling operator based on the DTV regularization term, x i,j This represents the pixel value x.

[0156] The second step involves performing compressed sensing processing on the third sampling feature using a sampling operator to generate the image sampling feature for the current temporal reconstruction stage.

[0157] In this step, DTV transformation is used to correct artifacts in the diagonal region, thereby achieving image reconstruction in the current temporal reconstruction stage.

[0158] The following section will elaborate on another implementation step of the image reconstruction method based on the fast sampling score generation model provided in this embodiment of the invention.

[0159] The finite-angle reconstruction problem in finite-angle CT imaging can be expressed as: in, Let p represent an underdetermined matrix, y represent the measured finite-angle projection, and x represent the image to be reconstructed. In the context of the reconstruction task, due to the observability of y as prior information, unlike the sampling process shown in Equation (2), p... t (x) is affected by the observed prior information y, which causes the sampling process shown in formula (2) to change, as shown in formula (11) below:

[0160]

[0161] Where, p t (x|y) represents the posterior probability term affected by y. Using Bayes' theorem expansion formula, p... t (x|y), we can obtain the result shown in formula (12) below:

[0162]

[0163] The hyperparameter λ balances the prior information of the model with the posterior distribution term of the measurement data. To achieve the sampling process outlined in formula (12), this embodiment of the invention constructs the following initial objective function, as shown in formula (13):

[0164]

[0165]

[0166] The initial objective function represents the optimization equation for solving the finite-angle inverse problem using a scoring model, where u represents the prior information of the data, solved using the SDE model, and λ1 represents the weight of the prior information. Typically, the sampling process for u requires 2000 or more iterations, consuming significant time and computational resources.

[0167] In order to reduce sampling time while improving the effect and quality of image reconstruction, this embodiment of the invention introduces large-scale skip sampling, time-backtracking sampling with resampling and compressed sensing sampling in the backward process, and optimizes the initial objective function shown in the above formula (13) based on this.

[0168] Considering the DTV regularization term, large-scale jump sampling, and time-backtracking sampling with resampling, the initial objective function shown in formula (13) above can be transformed into the final objective function shown in formula (15) below:

[0169]

[0170]

[0171] Among them, u tThis represents the term optimized using large-scale jump sampling and time-backtracking with resampling based on the scoring model, μ. t It is the weight of the DTV regularization term. It is a regularization term.

[0172] Based on this, the image reconstruction method based on the fast sampling score generation model proposed in this embodiment of the invention can also include the following steps:

[0173] A100, obtains scoring generation models and finite-angle CT images.

[0174] Step A100 is executed in the same way as step S100 in the previous embodiment.

[0175] A200 uses a scoring generation model to perform noise injection on CT images at limited angles, generating Gaussian noise features.

[0176] Step A200 is executed in the same way as step S200 in the previous embodiment.

[0177] A300 constructs an objective function to characterize the backward process of the score generation model, solves the objective function to reconstruct images from finite-angle CT images, and generates CT reconstructed images.

[0178] Unlike the execution process of step S300 in the previous embodiment, this step further integrates the entire process of image reconstruction of finite angle CT images into an objective function, as shown in the above formula (15). Then, by solving the objective function, the reconstruction of finite angle CT images can be realized.

[0179] For the objective function shown in formula (15) above, this embodiment of the invention uses an initial Gaussian noise feature x. T The solution is obtained iteratively. Starting with large-scale jump sampling, the objective function shown in equation (15) can be transformed into the following equation (16):

[0180]

[0181] Equation (16) represents the predictor in the PC sampler used in large-scale jump sampling, and its numerical solution scoring model.

[0182] Subsequently, we need to use a corrector to correct the solution, and the correction process is shown in the following formula (17):

[0183]

[0184] To impose constraints on image generation, this embodiment of the invention introduces the SIRT iterative method, which is applied to the correction of prior model reconstruction using finite angle sinusoidal image maps, as shown in the following formula (18):

[0185]

[0186] Finally, since large-scale skip sampling is unstable, the previously constructed DTV regularization term is used to correct the result, as shown in the following formula (19):

[0187]

[0188] Through the processes shown in formulas (16) to (19) above, this embodiment of the invention has achieved one-step reconstruction under large-scale noise δ. Using large-scale noise δ for reconstruction allows us to reduce the number of parameters T from the traditional 2000 or 4000 to 200 or 100. Compared with the traditional SDE model, this embodiment of the invention significantly reduces the number of image reconstruction steps. However, because the jump reconstruction of large-scale noise δ bypasses small-scale noise, it leads to unstable and inaccurate image reconstruction results. To solve this problem, this embodiment of the invention implements a time-backtracking and resampling process. The small-scale noise σ is reconstructed. As described in the previous embodiments, this embodiment of the invention reconstructs x′. t Perform a time backtracking to x′ t+L Then, the time backtracking result x' t+L Perform resampling to return to x' t The time sequence in which this process takes place is entirely under small-scale noise σ. By iteratively executing the optimization process proposed in this embodiment of the invention, the pure Gaussian noise features x can be extracted. T Downsampling produces a clear reconstructed image x0.

[0189] In summary, in the reconstruction of finite-angle CT images, this invention introduces a fast sampling mechanism in SDE, which can maintain the clear edges and details of the image while achieving rapid reconstruction. The reconstruction process is as follows: noise is gradually injected into the finite-angle CT image to be reconstructed using the forward process of SDE, and then the noisy data is gradually denoised using the backward process of SDE to obtain the CT reconstructed image, thereby realizing the reconstruction of the finite-angle CT image.

[0190] The fast sampling mechanism is concentrated in the backward process of SDE. The entire sampling process follows the principles of robust optimization theory and can be built on a complete data model (such as the objective function shown in formula (15) above). The fast sampling mechanism includes large-scale skip sampling, time-backtracking sampling with resampling, and compressed sensing sampling implemented in each stage. First, in the initial large-scale skip stage, multiple sampling steps are skipped to accelerate the acquisition of the initial result. Then, in the time-backtracking sampling step, small-scale noise is introduced to controllably destroy the initial result. After that, in the resampling step, a special resampling technique is used to carefully adjust the initially damaged result. Finally, the fine result is fine-tuned by introducing a DTV regularization term in the compressed sensing sampling to complete the sampling of the current stage. Image reconstruction can be achieved through multiple stages of sampling iteration.

[0191] This invention can significantly reduce the sampling steps in image reconstruction, greatly accelerate the sampling process, improve the efficiency and speed of image reconstruction, and achieve rapid reconstruction. At the same time, during the image reconstruction process, it can effectively alleviate the negative impact of orientation artifacts generated by limited-angle CT scanning, and maintain the clear edges and details of the image while reconstructing rapidly, thereby improving the quality and effect of image reconstruction.

[0192] The image reconstruction method based on a fast sampling score generation model proposed in this invention will be verified through the following embodiments. The method proposed in this invention is referred to as the TIFA method in the following embodiments.

[0193] 1. Selection of comparison models: In this embodiment, the Filtered Back Projection (FBP) algorithm, the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA), the Deep Convolutional Neural Network (FBPConvNet), the DPS model, and the DOLCE model are selected as comparison methods.

[0194] 2. Dataset Construction: A simulated dataset from the AAPM2016 CT Low-Dose Challenge was selected as the first dataset for the image reconstruction task. To generate finite-angle CT images at 90 and 60 degrees, this embodiment employed an isoangular fan-beam projection geometry algorithm to process the image data of the first dataset, generating image data from both 60 and 90-degree angles. This generated image data was then added to the first dataset. Subsequently, the first dataset was divided into training and test sets at a 10:1 ratio. Furthermore, a real-world cardiac clinical dataset was selected as the second dataset for the image reconstruction task.

[0195] 3. Evaluation metrics for image reconstruction tasks: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) are used. Higher PSNR and SSIM correspond to higher levels of reconstruction quality.

[0196] 4. Experimental Environment: The PyTorch framework was used for experimental research, and all experiments were conducted on a high-performance computing system equipped with an NVIDIA RTX A6000 48GB graphics processing unit. The Adam optimization algorithm was used during training, with a learning rate set to 2×10⁻⁶. -4 When configuring the noise variance, a fixed value σ is used. min =0.01 and σ max =378.

[0197] 5. Description of experimental data:

[0198] 1) For image reconstruction tasks on the AAPM dataset:

[0199] This embodiment first compares the performance of the method proposed in this invention with other comparative methods on the image reconstruction task on the AAPM dataset. The image reconstruction task includes two aspects: reconstruction of 60-degree CT images and reconstruction of 90-degree CT images.

[0200] Reference Figure 6 and Figure 7 , Figure 6 This is a comparison chart showing the effectiveness of the TIFA method compared with other methods in the reconstruction of 90-degree CT images provided by this invention. Figure 7 This is another comparison chart showing the effectiveness of the TIFA method with other comparative methods in the reconstruction task of 90-degree CT images provided by this invention. Figure 6 and Figure 7 The “Label” in this context refers to the labeled image, which is the ground truth commonly used in model evaluation. Figure 6 and Figure 7 The first row in the image is the reconstructed CT image, and the second row is the region of interest (ROI) with an angle range of [0°, 90°]. (This is achieved through...) Figure 6 and Figure 7 visible:

[0201] The image reconstruction results of FBP and FISTA exhibit significant image distortion, and both struggle to extract information beyond the skeleton structure. FBPConvNet, a typical example of supervised deep learning methods, has achieved significant improvements in image reconstruction quality, focusing on larger structures and edges. However, due to the inherent limitations of finite angular measurements, FBPConvNet has a significant deficiency in accurately predicting reconstructed details. Compared to other methods, score-based generative models DPS and DOLCE are widely considered benchmark methods, achieving commendable reconstruction results on finite-angle CT reconstruction tasks. However, when detecting ROIs, it is evident that the areas indicated by the arrows are not accurately restored, indicating that DPS and DOLCE fail to accurately capture image details. The difference in image reconstruction results between different methods further highlights the shortcomings of DPS and DOLCE in restoring boundary structures. In contrast, the TIFA method proposed in this invention can still preserve important structural details even with accelerated sampling. Figure 6 and Figure 7 The comparison of results from different methods highlights the superiority of the TIFA method over the current mainstream baseline methods.

[0202] Referring to Table 1 below, "Method" represents the image reconstruction method, "90 Limited-angle" represents the 90-degree angle CT image reconstruction task, and "60 Limited-angle" represents the 60-degree angle image reconstruction task. As can be seen from Table 1, the TIFA method proposed in this invention outperforms other methods in terms of both PSNR and SSIM.

[0203] Table 1. Performance of TIFA and the contrastive method on the AAPM dataset.

[0204]

[0205] Similarly, this embodiment continues to evaluate the effectiveness of the image reconstruction method on a 60-degree CT image task. (Refer to...) Figure 8 , Figure 8 This is a comparison chart showing the effectiveness of the TIFA method compared with other methods in the reconstruction of 60-degree CT images provided by this invention. Figure 8 In this context, "Label" refers to the label image. Figure 8 The first row in the image is the reconstructed image, and the second row is the ROI with an angle range of [0°, 60°]. (This is achieved through...) Figure 8 visible:

[0206] Under these highly sparse view conditions, the image reconstruction results of DPS exhibit significant noise, revealing its limited ability to recover subtle structural details. Similarly, the image reconstruction results of DOLCE demonstrate its difficulty in accurately preserving structural fidelity, a fact further confirmed by examining the extracted ROIs. This indicates that both benchmark methods based on score generation models—DPS and DOLCE—fail to ensure accuracy at each sampling step during iterative sampling. To address this challenge, the proposed TIFA method employs time backtracking and resampling to optimize the sampling steps. This not only accelerates the sampling process but also captures the textural complexity of the image. Compared to other methods, TIFA minimizes the gap between the reconstructed image and the ground truth.

[0207] 2) Image reconstruction task for real clinical cardiac datasets:

[0208] To verify the stable generalization performance of the TIFA method proposed in this invention, this embodiment uses a model trained on the AAPM simulation dataset to reconstruct clinical cardiac data. (Refer to...) Figure 9 and Figure 10 , Figure 9 This is another comparison chart showing the effectiveness of the TIFA method with other comparative methods in the reconstruction task of 60-degree CT images provided by this invention. Figure 10 This is another comparison chart showing the effectiveness of the TIFA method with other comparative methods in the reconstruction task of 90-degree CT images provided by this invention. Figure 9 and Figure 10 In this context, "Label" refers to the label image, and "Ours" indicates the TIFA method. Figure 9 and Figure 10 The first row in the image is the reconstructed image. Figure 9 The second row in the table represents the ROIs with an angle range of [0°, 60°]. Figure 10 The second row in the table represents the ROIs with an angle range of [0°, 90°]. (This is achieved through...) Figure 9 and Figure 10 visible:

[0209] When there are limited angles, the image reconstruction results of other contrast methods all show severe distortion, and they cannot accurately reconstruct the structure of the image. However, the TIFA method proposed in this invention has a greater advantage in terms of structural fidelity.

[0210] Referring to Table 2, it can be seen that the TIFA method proposed in this embodiment of the invention is superior to other methods in terms of PSNR and SSIM, and the TIFA method has good image reconstruction performance.

[0211] Table 2. Performance of TIFA and the comparison method on real-world clinical cardiac datasets.

[0212]

[0213] Reference Figure 11 , Figure 11 This is a comparison chart of the performance of the TIFA method, DPS, and DOLCE methods in the reconstruction task of 90-degree CT images provided by this invention. Figure 11 It is evident that DPS and DOLCE typically involve 2000 or 4000 sampling steps, resulting in approximately 40 minutes of time required to reconstruct an image using either method, leading to low image reconstruction efficiency. The method proposed in this invention, however, employs only 200 sampling steps, achieving rapid image reconstruction while maintaining image reconstruction quality. In typical scenarios, compared to the 2000-step methods DPS (27.67 minutes) and DOLCE, our proposed TIFA method achieves significantly better image reconstruction results in just 50 steps (1.35 minutes).

[0214] The above verification data demonstrates that the image reconstruction method based on the fast sampling score model proposed in this invention achieves significantly better 90° and 60° reconstruction performance than other methods using 2000 or 4000 sampling steps, even with only 200 sampling steps. Furthermore, experimental results show that even with only 100 steps, the image reconstruction method based on the fast sampling score model proposed in this invention can still provide high-quality image reconstruction.

[0215] Therefore, compared with existing methods, this invention can achieve rapid image reconstruction while ensuring clear edges and detailed features of the image, improving the quality and effect of image reconstruction, and has high generalization ability and high availability.

[0216] In addition, embodiments of the present invention provide an image reconstruction system based on a fast sampling score generation model, comprising:

[0217] The acquisition module is used to acquire the scoring generation model and finite-angle CT images;

[0218] The forward processing module is used to perform noise injection on CT images at limited angles through the forward process of the scoring generation model, generating Gaussian noise features;

[0219] The image reconstruction module is used to reconstruct images from Gaussian noise features through the backward process of the scoring generation model, generating CT reconstructed images.

[0220] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0221] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0222] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. An image reconstruction method based on a fast sampling score generation model, characterized in that, Includes the following steps: Acquire the scoring generation model and finite-angle CT images; The finite-angle CT image is denoised using the forward process of the scoring generation model to generate Gaussian noise features. Based on the Gaussian noise characteristics, the finite-angle CT image is reconstructed through the backward process of the scoring generation model to generate a CT reconstructed image. The step of reconstructing the finite-angle CT image based on the Gaussian noise features through the backward process of the score generation model to generate a CT reconstructed image includes: The timing information of the backward process is obtained, and the backward process is divided into several timing reconstruction stages based on the timing information, wherein the Gaussian noise feature is used as the input feature of the first timing reconstruction stage. Determine the current temporal reconstruction stage, sample the input features of the current temporal reconstruction stage, and generate the image sampling features of the current temporal reconstruction stage; Using the image sampling features of the current temporal reconstruction stage as the input features of the next temporal reconstruction stage, and taking the next temporal reconstruction stage as the current temporal reconstruction stage, the process returns to the step of determining the current temporal reconstruction stage. When all temporal reconstruction stages have been sampled, the image sampling features from the last temporal reconstruction stage are used as the CT reconstructed image and output. The step of sampling the input features of the current temporal reconstruction stage to generate the image sampling features of the current temporal reconstruction stage includes: Skip sampling is performed on the input features of the current time-series reconstruction stage to obtain the first sampled features; wherein, the noise variance satisfies , This indicates the number of sampling steps. The skip sampling is used to accelerate sampling on a coarse noise scale, where the coarse noise satisfies... , Indicates noise interval; The first sampling feature is subjected to time-backtracking sampling to obtain the second sampling feature; wherein the time sequence of the second sampling feature is earlier than that of the first sampling feature. The second sampling feature is resampled to obtain a third sampling feature; wherein the timing of the third sampling feature is the same as that of the first sampling feature. A diagonal total variational regularization term is introduced into compressed sensing, and the third sampling feature is processed by compressed sensing to generate the image sampling features for the current temporal reconstruction stage; wherein, the diagonal total variational regularization term satisfies: ； in, This represents a sampling operator based on the DTV regularization term. Represents pixel value.

2. The image reconstruction method based on a fast sampling score generation model according to claim 1, characterized in that, The step of annotating the finite-angle CT image with Gaussian noise features through the forward process of the scoring generation model includes: The forward process of the score generation model is defined by stochastic differential equations; The noise features are injected into the finite-angle CT image through the forward process to generate Gaussian noise features.

3. The image reconstruction method based on a fast sampling score generation model according to claim 1, characterized in that, The step of skipping sampling the input features of the current time-series reconstruction stage to obtain the first sampled features includes: Obtain the noise interval for skip sampling, and calculate the first noise scale based on the noise interval; The first sampled feature is obtained by skipping the input features of the current time-series reconstruction stage according to the first noise scale.

4. The image reconstruction method based on a fast sampling score generation model according to claim 1, characterized in that, The step of performing time-backtracking sampling on the first sampling feature to obtain the second sampling feature includes: Obtain the step size information of time backtracking, and calculate the second noise scale based on the step size information; The first sampling feature is backsampled according to the second noise scale, so that the temporal sequence of the first sampling feature gradually approaches the temporal sequence of the input feature in the current temporal reconstruction stage, thereby generating the second sampling feature.

5. The image reconstruction method based on a fast sampling score generation model according to claim 1, characterized in that, The step of resampling the second sampling feature to obtain the third sampling feature includes: Obtain a sampler for resampling, the sampler including a predictor and a corrector; The predictor resamples the second sampled feature, and the corrector corrects the resampled feature to obtain the third sampled feature.

6. The image reconstruction method based on a fast sampling score generation model according to claim 1, characterized in that, The step of performing compressed sensing processing on the third sampling feature to generate the image sampling features for the current temporal reconstruction stage includes: Obtain the sampling operator; The sampling operator performs compressed sensing processing on the third sampling feature to generate the image sampling feature for the current temporal reconstruction stage.

7. The image reconstruction method based on a fast sampling score generation model according to claim 6, characterized in that, The sampling operator is a sampling operator based on the diagonal total score regularization term.

8. An image reconstruction system based on a fast sampling score generation model, characterized in that, include: The acquisition module is used to acquire the scoring generation model and finite-angle CT images; The forward processing module is used to perform noise injection processing on the finite-angle CT image through the forward process of the score generation model to generate Gaussian noise features; The image reconstruction module is used to reconstruct the image from the Gaussian noise features through the backward process of the score generation model, and generate a CT reconstructed image. The step of reconstructing the image from the Gaussian noise features using the backward process of the score generation model to generate a CT reconstructed image includes: The timing information of the backward process is obtained, and the backward process is divided into several timing reconstruction stages based on the timing information, wherein the Gaussian noise feature is used as the input feature of the first timing reconstruction stage. Determine the current temporal reconstruction stage, sample the input features of the current temporal reconstruction stage, and generate the image sampling features of the current temporal reconstruction stage; Using the image sampling features of the current temporal reconstruction stage as the input features of the next temporal reconstruction stage, and taking the next temporal reconstruction stage as the current temporal reconstruction stage, the process returns to the step of determining the current temporal reconstruction stage. When all temporal reconstruction stages have been sampled, the image sampling features from the last temporal reconstruction stage are used as the CT reconstructed image and output. The step of sampling the input features of the current temporal reconstruction stage to generate the image sampling features of the current temporal reconstruction stage includes: Skip sampling is performed on the input features of the current time-series reconstruction stage to obtain the first sampled features; wherein, the noise variance satisfies , The number of sampling steps is defined as the skip sampling method used to accelerate sampling on a coarse noise scale, where the coarse noise satisfies the following: , Indicates noise interval; The first sampling feature is subjected to time-backtracking sampling to obtain the second sampling feature; wherein the time sequence of the second sampling feature is earlier than that of the first sampling feature. The second sampling feature is resampled to obtain a third sampling feature; wherein the timing of the third sampling feature is the same as that of the first sampling feature. A diagonal total variational regularization term is introduced into compressed sensing. The third sampling feature is processed using compressed sensing to generate the image sampling features for the current temporal reconstruction stage. The diagonal total variational regularization term satisfies the following: ； in, This represents a sampling operator based on the DTV regularization term. Represents pixel value.

Images