A SPECT low-dose sinogram restoration method based on poisson distribution noise diffusion probability model

Abstract

The present application relates to the technical field of low-dose SPECT chord diagram recovery reconstruction, and more particularly to a SPECT low-dose chord diagram recovery method based on a Poisson distribution noise diffusion probability model. A PDPM preliminary framework containing a forward process and a reverse process is first proposed, and the preliminary framework is perfected to contain two parts: the first part is to discard the forward process and use an ideal reverse process based method to generate a data set for training, which solves the mismatch problem between low-dose and normal-dose chord diagrams in the forward process of the preliminary framework; the second part is to introduce a time prediction aggregation module TPAM, which effectively improves the performance of model recovery by aggregating and predicting across multiple time steps.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-dose SPECT sinogram restoration reconstruction, and more particularly, to a SPECT low-dose sinogram restoration method based on a Poisson distribution noise diffusion probability model. BACKGROUND

[0002] Single Photon Emission Computed Tomography (SPECT) is a nuclear medicine imaging modality that utilizes gamma rays for diagnostic imaging. In this process, a radiotracer is injected into a patient's body through an intravenous line and is selectively absorbed by different organs or tissues. When the tracer undergoes radioactive decay, it releases gamma photons, which are then detected for imaging. Subsequently, a gamma camera, as a detection imaging device, rotates around the patient's body to capture gamma rays and generates projection data from multiple angles, also known as sinograms. These sinograms are then processed using tomographic reconstruction algorithms to generate anatomical images representing the internal body structure of the patient. Since the radiation released by the radiotracer used in SPECT studies can alter cell structure or function, even increasing the risk of cancer, it is crucial to minimize the tracer dose in SPECT procedures. However, reducing the tracer dose can exacerbate the photon attenuation effect during the sinogram generation process. As gamma photons traverse tissue, they interact with atomic electrons, causing many photons to deviate from their original path. This deflection results in some photons being deflected away from the detector, a phenomenon known as photon attenuation. In normal-dose SPECT sinograms, the effects of photon attenuation can be negligible, but in low-dose SPECT sinograms, it becomes significant. Therefore, compared to normal-dose sinograms, low-dose SPECT sinograms exhibit higher noise levels, reduced spatial resolution, and can introduce more apparent artifacts. These factors often lead to a decrease in the quality of the reconstructed images, making it more difficult for doctors to diagnose.

[0003] There are many traditional model-driven methods proposed for low-dose SPECT sinogram restoration and reconstruction, such as iterative reconstruction methods, but the model parameters of such methods are often determined by experience, so the generalization ability for projection data under various noise levels is usually poor. With the increasing popularity of deep learning, methods that can automatically learn discriminative features from training data have attracted the attention of researchers. Various deep learning-based low-dose reconstruction methods can be roughly divided into three categories: end-to-end methods, post-processing methods and preprocessing methods based on sinogram domain. Specifically, the end-to-end reconstruction method takes the projection data or k-space data as the input and directly generates a high-quality reconstructed image, but this requires the model to independently learn the underlying physical characteristics, and the model building and training are difficult; The post-processing method uses the low-dose reconstructed image as the input of the neural network to restore the high-quality image, but this usually requires a large amount of data set and complex network architecture to achieve better performance in practical applications, which can easily lead to inaccurate restored images; The preprocessing method based on the sinogram domain limits the denoising and restoration problem to the sinogram domain, first uses a neural network to reduce the noise of the sinogram, and then uses traditional algorithms such as iterative reconstruction to reconstruct the image from the high-quality sinogram. The calculation process is complex and the accuracy is low. Therefore, how to restore the low-dose SPECT sinogram to the normal dose level has become a problem to be solved in SPECT clinical imaging. SUMMARY

[0004] The purpose of the present application is to overcome the difficulty and low accuracy of the prior art in restoring low-dose SPECT sinograms to normal dose levels, and to provide a SPECT low-dose sinogram restoration method based on a Poisson distribution noise diffusion probability model, which realizes low-dose restoration of SPECT sinograms and has high restoration accuracy.

[0005] To solve the above technical problems, the technical scheme adopted by the present application is:

[0006] A SPECT low-dose sinogram restoration method based on a Poisson distribution noise diffusion probability model is provided, comprising the following steps:

[0007] S1. Training: for a data set containing n pairs of normal dose / low dose SPECT sinogram data, use the ideal back-projection Markov chain to generate the intermediate images of the ideal back-projection, and use these intermediate images together with the low-dose sinogram as the input data, and the corresponding normal dose sinogram as the label, to form the training data set of the U-Net model, and train the U-Net model using the training data set;

[0008] S2. Inference: input the image x t and the time t into the neural network U-Net to obtain the prediction of the normal dose sinogram The prediction results of all time steps are input into a temporal prediction aggregation module TPAM together, in which the prediction results from time step T to time step t are respectively input into the recovery operator R(·) corresponding to time t to obtain image distribution, and the image generated after inverse transform sampling is taken as output; the outputs of the recovery operators corresponding to the prediction results from time step T to time step t are averaged to obtain the recovered generated image x t The prediction results of all time steps are input into a temporal prediction aggregation module TPAM together, in which the prediction results from time step T to time step t are respectively input into the recovery operator R(·) corresponding to time t to obtain image distribution, and the image generated after inverse transform sampling is taken as output; the outputs of the recovery operators corresponding to the prediction results from time step T to time step t are averaged to obtain the recovered generated image x t-1 ;

[0009] S3. Iteration: in the inference of the PDPM, the process of step S2 is iterated continuously starting from the low-dose SPECT sinogram x T T-1 times, and finally the U-Net prediction result obtained by taking x1 as input is taken as the final recovered image x0.

[0010] The SPECT low-dose sinogram recovery method based on a Poisson distribution noise diffusion probabilistic model provided by the application has a core part, which is a Poisson noise diffusion-based probabilistic model, referred to as a Poisson diffusion probabilistic model (PDPM). Considering the physical principle of the formation of a low-dose SPECT sinogram, the PDPM replaces the Gaussian noise used in a traditional diffusion model with Poisson noise, and uses a low-dose and normal-dose SPECT sinogram as the starting point and the end point of the recovery process. Specifically, the application is based on a preliminary framework of the PDPM containing a forward process and a reverse process, and the preliminary framework is perfected to contain two parts: the first part is to discard the forward process and use an ideal reverse process-based method to generate a data set for training, which solves the mismatch problem between the low-dose and normal-dose sinograms in the forward process of the preliminary framework; the second part is to introduce a temporal prediction aggregation module (TPAM), which effectively improves the recovery performance of the model by aggregating predictions across multiple time steps. The PDPM provided by the application is based on the noise properties generated in the formation process of a low-dose SPECT sinogram, uses Poisson noise to disturb the image, and conforms to the actual noise generation and propagation mechanism observed in SPECT imaging, so the model of the application is more consistent with the physical principle of the degradation of a low-dose SPECT sinogram than a traditional diffusion model. On the other hand, the TPAM introduced in the reverse process aggregates the prediction results in multiple time steps to reduce the error between the generated image in the reverse process of the model and the actual image in the ideal reverse process, so that the model performs better in image recovery performance than a traditional diffusion model. In summary, compared with a traditional diffusion model, the PDPM provided by the application is more suitable for a low-dose SPECT sinogram recovery task.

[0011] As preferably, the ideal inverse process comprises: the inverse process image distribution p θ (x t-1 |x t ) is replaced by the neural network prediction result The normal dose SPECT chord diagram x0.

[0012] As preferably, the Markov chain is expressed as:

[0013]

[0014] Wherein,

[0015]

[0016] In the formula, x t-1 represents the recovery image of the t-1 step, x t represents the recovery image of the t step, x0 represents the original image, that is, the normal dose SPECT chord diagram, and t represents the time step.

[0017] As preferably, in step S1, the U-Net model is trained, and the training target is expressed as:

[0018]

[0019] In the formula, is the prediction result of the neural network U-Net, and x0 is the normal dose SPECT chord diagram.

[0020] As preferably, in step S2, the recovery operator R(·) is expressed as:

[0021]

[0022] In the formula, represents the prediction result of the neural network U-Net at the t time step, x t represents the recovery image of the t time step, t represents the time step, represents the recovery image of the t-1 time step generated by the recovery operator.

[0023] As preferably, in step S2, the recovery operator R(·) uses the formula The image distribution p θ (x t-1 |x t ) is calculated.

[0024] As preferably, in step S2, the recovery operator output corresponding to the prediction result of the time step T to t is averaged, and is processed by the following formula:

[0025]

[0026] The result calculated by the above formula is taken as the recovered generated image x t-1 .

[0027] The application also provides a SPECT low-dose sinogram recovery system based on a Poisson distribution noise diffusion probability model, comprising:

[0028] The training module is configured to use an ideal inverse process Markov chain to generate intermediate images of the ideal inverse process for a set of data sets containing n normal dose / low-dose SPECT sinogram data pairs, use the intermediate images and the low-dose sinogram as input data, and use the corresponding normal dose sinogram as a label to form a training data set of a U-Net model, and train the U-Net model using the training data set.

[0029] The inference module is configured to input the image x t and the time t as input of the neural network U-Net to obtain the prediction The prediction results of all time steps are input into a time prediction aggregation module TPAM together, and in the TPAM, the prediction results from the time step T to the time step t are respectively input into the recovery operator R(·) with the image x t and the time t to calculate the image distribution, and the image generated after inverse transform sampling is taken as output; the outputs of the recovery operators corresponding to the prediction results from the time step T to the time step t are averaged to obtain the recovered generated image x t-1 .

[0030] The iteration module is configured to start from the low-dose SPECT sinogram x T in the inference of the PDPM, and iteratively perform the process of step S2 for a total of T-1 steps, and finally take the U-Net prediction result obtained by taking the image x1 as input as the final recovered image x0.

[0031] Preferably, in the training module, the Markov chain is represented as:

[0032]

[0033] wherein,

[0034]

[0035] In the formula, x t-1 represents the recovered image of the t-1th step, x t represents the recovered image of the tth step, x0 represents the original image, i.e., the normal dose SPECT sinogram, and t represents the time step.

[0036] The U-Net model is trained, and the training target is represented as:

[0037]

[0038] wherein, is the prediction result of the neural network U-Net, and x0 is a normal dose SPECT sinogram.

[0039] As preferred, in the inference module, the restoration operator R(·) is represented as:

[0040]

[0041] The restoration operator R(·) uses the formula to calculate the image distribution p θ (x t-1 |x t ), and the image generated after inverse transform sampling is taken as the output; the restoration operator outputs corresponding to the prediction results from time step T to time t are averaged, and processed by the following formula:

[0042]

[0043] The result calculated by the above formula is taken as the restored generated image x t-1 .

[0044] Compared with the prior art, the beneficial effects of the present application are:

[0045] The SPECT low dose sinogram restoration method based on a Poisson distribution noise diffusion probability model of the present application is based on the noise properties generated in the low dose SPECT sinogram formation process, uses Poisson noise to disturb the image, conforms to the actual noise generation and propagation mechanism observed in SPECT imaging, and therefore the model of the present application is more in line with the physical principle of low dose SPECT sinogram degradation than the traditional diffusion model. On the other hand, the TPAM introduced in the reverse process aggregates the prediction results in multiple time steps, reduces the error between the image generated by the model in the reverse process and the actual image in the ideal reverse process, and makes the model perform better than the traditional diffusion model in image restoration performance. In summary, compared with the traditional diffusion model, the PDPM proposed in the present application is more suitable for the low dose SPECT sinogram restoration task. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 FIG. 1 is a model framework schematic diagram of a SPECT low dose sinogram restoration method based on a Poisson distribution noise diffusion probability model. DETAILED DESCRIPTION

[0047] The application will be further described below in conjunction with the specific embodiments. Among them, the drawings are only used for exemplary description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation of the patent; in order to better illustrate the embodiments of the application, some components of the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings can be omitted.

[0048] The same or similar reference numerals in the drawings of the embodiments of the application correspond to the same or similar components; in the description of the application, it is understood that if the terms "upper", "lower", "left", "right" and the like indicate the orientation or positional relationship shown in the drawings, only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, therefore the positional relationship described in the drawings is only used for exemplary description, and cannot be understood as a limitation of the patent, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific situation.

[0049] Embodiment one

[0050] The first embodiment of the SPECT low-dose chord diagram recovery method based on Poisson distribution noise diffusion probability model is provided, which solves the problem of poor image quality caused by photon attenuation in the imaging process in the low-dose SPECT chord diagram, and realizes the low-dose recovery of the SPECT chord diagram.

[0051] The embodiment proposes a SPECT low-dose sinogram recovery method based on a Poisson distribution noise diffusion probability model. The core part of the method is a Poisson noise diffusion-based probability model, referred to as a Poisson diffusion probability model (PDPM). Considering the physical principle of low-dose SPECT sinogram formation, PDPM replaces the Gaussian noise used in the traditional diffusion model with Poisson noise, and uses low-dose and normal-dose SPECT sinograms as the starting point and endpoint of the recovery process. Specifically, the embodiment first proposes a PDPM preliminary framework containing a forward process and a reverse process. This preliminary framework strictly derives the changes of image probability distribution in the forward and reverse processes by mathematical formula, laying a foundation for the physical principle of the method. Then, the preliminary framework is improved, including two parts: the first part is to discard the forward process and use an ideal reverse process-based method to generate the data set used for training, which solves the mismatch problem between low-dose and normal-dose sinograms in the forward process of the preliminary framework; the second part is to introduce a temporal prediction aggregation module (TPAM), which effectively improves the performance of the model recovery by aggregating and predicting across multiple time steps. The PDPM preliminary framework and the improvement of the preliminary framework are described in detail below.

[0052] (1) PDPM preliminary framework:

[0053] The preliminary framework of PDPM includes two parts: forward process and reverse process. The forward process simulates the sinogram degradation process from normal dose to low dose by gradually subtracting Poisson random noise from the data; the reverse process restores the low-dose sinogram to the corresponding normal-dose sinogram by learning the transformation of the forward process. Given a set of normal-dose SPECT sinograms x0~q(x0), in the forward process, Poisson random noise is gradually subtracted from the image by the following formula:

[0054] x t :=x t-1 -∈ t ,t=1,2,…,T,

[0055] where x0and x T represent the starting and ending images in the forward process, i.e., normal-dose SPECT sinograms and low-dose SPECT sinograms, respectively, and x t is the image after t steps of the forward process. ∈ t is a random Poisson noise with mean λ, denoted as ∈ t ~P(λ). The Markov chain of the forward process is defined as:

[0056]

[0057] where,

[0058]

[0059] The inverse process of PDPM is defined as a Markov chain with learnable parameters θ:

[0060]

[0061] where,

[0062]

[0063] Here, An approximation of x0is generated by a neural network with learnable parameters θ. In PDPM, a U-Net is used as this trainable neural network, and since x t-1 is a 2D image, the input of the U-Net is x θ and the output is x t-1 . t The image distribution p t-1 (x N |x L ) does not follow any known distribution, so Inverse Transform Sampling is used to generate images x N from the image distribution.

[0064] (2) Improvement of the preliminary framework: the generation method of the PDPM training dataset.

[0065] In the training phase of the model, an ideal inverse process-based method is adopted to generate the dataset for training the U-Net. Specifically, for a pair of normal dose / low dose SPECT sinograms (x N , x L ), set x0= x N and x T = x L . Based on the formula of the inverse process, the following Markov chain is used to generate images

[0066]

[0067] where,

[0068]

[0069] In fact, it is the neural network prediction result θ (x t-1 |x t ) in the inverse process image distribution p This is obtained by replacing the normal dose chord graph x0, which is why the image generation process represented by this Markov chain is called the ideal inverse process. Similar to the inverse process in PDPM, an inverse transform is used to generate the image from the image distribution. These images were then used as the dataset for training U-Net in PDPM. Compared to methods that use a forward process to generate training data, this improved method avoids the previously mentioned mismatch between the low-dose and normal-dose chordograms.

[0070] (3) Improvement of the initial framework II: Introducing TPAM into the reverse process of PDPM.

[0071] In the model's inverse process, i.e., the inference phase, a Time Prediction Aggregation (TPAM) module was designed to mitigate the generation of problems during model inverse. In the process of reversing the ideal, the actual x t-1 The error between them. Specifically, in the reverse process before improvement, using Calculated The image distribution is obtained, and inverse transform sampling is used to generate... The entire process is represented by the recovery operator R(·), defined as follows:

[0072]

[0073] After introducing TPAM to improve the model, when TPAM uses the recovery operator to generate... At that time, not only using The calculations also incorporate previous U-Net predictions, i.e. The final The calculation is performed using the following formula:

[0074]

[0075] In this way, TPAM can effectively mitigate the negative impact of accumulated errors during the inversion process, thereby further improving the model's recovery performance.

[0076] like Figure 1 The diagram shows the model framework proposed in this invention. The overall architecture of the model includes a recovery process with a low-dose SPECT chordogram as the starting point and a normal-dose SPECT chordogram as the ending point, with a time step of T. In the single-step recovery process, the U-Net model is used to predict and estimate the high-dose chordogram, and then the Time Prediction Aggregation Module (TPAM) is used to generate the recovery result for that step.

[0077] Specifically, the following steps are included:

[0078] Step 1: Training.

[0079] For a dataset containing n pairs of normal / low dose SPECT sinogram data Using idealized reverse process Markov chain Generate intermediate images of idealized reverse process, use these intermediate images together with low dose sinogram as input data, corresponding normal dose sinogram as label, compose training dataset of U-Net model, denoted as Train U-Net model with this training dataset, training objective is as follows:

[0080]

[0081] where, is the prediction (i.e. output) of neural network U-Net; x0 is normal dose SPECT sinogram.

[0082] Step 2: Inference.

[0083] Take x0 as input image from image x t , and generate image x t-1 . t Take x t and time t as input of neural network U-Net, get prediction for normal dose sinogram. t Take image x t and prediction of this step together with prediction of all previous time steps, i.e. , as input of TPAM; in TPAM, prediction from time step T to time step t respectively and x t and t input recovery operator R(·), which uses formula to calculate image distribution p θ (x t-1 | x t ), and the output of recovery operator is the image generated after inverse transform sampling. Take average of the output of recovery operator corresponding to prediction from time step T to t, i.e. process with the following formula:

[0084]

[0085] Take the result calculated by the above formula as recovered generated image x t-1 .

[0086] Step 3: Iteration.

[0087] In inference of PDPM, start from low dose SPECT sinogram x T , constantly iterate the above process, totally iterate T-1 steps, and finally take U-Net prediction with x1 as input, i.e. , as final recovered image x0.

[0088] Compared with other methods for low-dose SPECT sinogram restoration, the diffusion model works in a step-by-step manner to process the image, which makes it advantageous in solving the key challenge of processing the image while not losing important information contained therein in medical imaging. The step-by-step processing method of the diffusion model makes the model only process a small amount of noisy data at each step while minimizing the loss of important information. Therefore, using the diffusion model for low-dose SPECT sinogram restoration not only improves the accuracy of image restoration, but also effectively preserves key features such as chord edge and tissue characteristics that are crucial for medical diagnosis.

[0089] Compared with the traditional diffusion model based on Gaussian noise, the PDPM proposed in the embodiment is based on the noise properties generated in the low-dose SPECT sinogram formation process, uses Poisson noise to perturb the image, conforms to the actual noise generation and propagation mechanism observed in SPECT imaging, and therefore the model of the embodiment is more consistent with the physical principle of low-dose SPECT sinogram degradation than the traditional diffusion model. On the other hand, the TPAM introduced in the reverse process reduces the error between the image generated by the model in the reverse process and the actual image in the ideal reverse process by aggregating the prediction results in multiple time steps, so that the model performs better than the traditional diffusion model in terms of image restoration performance. In summary, compared with the traditional diffusion model, the PDPM is more suitable for the low-dose SPECT sinogram restoration task.

[0090] Embodiment two

[0091] The first embodiment of the SPECT low-dose sinogram restoration system based on the Poisson distribution noise diffusion probability model comprises:

[0092] The training module comprises:

[0093] a data set comprising n normal dose / low dose SPECT sinogram data pairs using the ideal reverse process Markov chain generate intermediate images of the ideal reverse process, and use these intermediate images together with the low-dose sinogram as input data, and the corresponding normal dose sinogram as a label, to form a training data set for the U-Net model, denoted as train the U-Net model using the training data set, and the training target is as follows:

[0094]

[0095] wherein, is the prediction result (i.e. output) of the neural network U-Net; x0 is the normal dose SPECT sinogram.

[0096] The inference module comprises:

[0097] for predicting x t and time t as input of the neural network U-Net, obtaining the prediction of the normal dose sinogram inputting the image x t and the prediction result of this step together with the prediction results of all previous time steps, i.e. to the TPAM; in the TPAM, the prediction results from time step T to time step t are respectively and x t and t to the recovery operator R(·), which uses the formula to calculate the image distribution p θ (x t-1 | x t ), and the image generated after inverse transform sampling is taken as the output. The outputs of the recovery operators corresponding to the prediction results from time step T to time step t are averaged, i.e. processed by the following formula:

[0098]

[0099] The result calculated by the above formula is taken as the recovered generated image x t-1 .

[0100] Iteration module:

[0101] for starting from the low-dose SPECT sinogram x T in the inference of the PDPM, constantly iterating the above process, and finally taking the U-Net prediction result obtained by taking x1 as input, i.e. , as the final recovered image x0.

[0102] Embodiment three

[0103] The embodiment provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method in the embodiment one when executing the computer program.

[0104] In the specific content of the above specific embodiments, each technical feature can be combined arbitrarily without contradiction. In order to make the description simple, all possible combinations of the above technical features are not described, but as long as the combination of these technical features does not exist contradiction, it should be considered as the scope of the description.

[0105] Obviously, the above embodiments of the present application are merely exemplary but not intended to limit the embodiments of the present application. Based on the above description, any other variations or changes can be made by those skilled in the art without departing from the spirit and principles of the present application. It is not necessary to list all the embodiments here. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall fall within the scope of the claims of the present application.

Claims

1. A SPECT low-dose sinogram restoration method based on a Poisson-distributed noise diffusion probability model, characterized in that, The method comprises the following steps: S1. Training: for a set of data sets containing n pairs of normal dose / low dose SPECT sinogram data, using the ideal reverse process Markov chain to generate the intermediate images of the ideal reverse process, taking these intermediate images together with the low dose sinogram as input data, and the corresponding normal dose sinogram as a label, to form a training data set of the U-Net model, and training the U-Net model with the training data set; S2. Inference: input the image x t and time t into the neural network U-Net to obtain the prediction Along with the prediction results of all time steps, input into the time prediction aggregation module TPAM, in which the prediction results from time step T to time step t are respectively and the image x t and time t into the recovery operator R(·) to calculate the image distribution, and the image generated after inverse transformation sampling is taken as the output; the outputs of the recovery operators corresponding to the prediction results from time step T to time step t are averaged to obtain the recovered image S3. Iteration: In the inference of PDPM, from the low-dose SPECT sinogram x T Start, iterate the process of step S2 constantly, a total of T-1 steps, and finally take the U-Net prediction result obtained by taking x1 as input as the final recovered image 2. The SPECT low-dose sinogram restoration method based on Poisson distribution noise diffusion probability model according to claim 1, characterized in that, The ideal inverse procedure includes replacing the inverse procedure image distribution p θ (x t-1 | x t ) with the neural network prediction result with the normal dose SPECT sinogram x0.

3. The SPECT low-dose sinogram restoration method based on Poisson distribution noise diffusion probability model according to claim 2, characterized in that, The Markov chain is expressed as: Wherein, where x t-1 denotes the recovered image of the t-1 step, x t denotes the recovered image of the t step, x0denotes the normal dose SPECT sinogram, and t denotes the time step.

4. The SPECT low-dose reconstruction method based on the Poisson distribution noise diffusion probability model according to claim 1 A chord diagram restoration method, characterized in that, In step S1, the U-Net model is trained, and the training target is expressed as: In the formula, is the prediction result of the neural network U-Net, and x0is the normal dose SPECT sinogram.

5. The SPECT low-dose sinogram restoration method based on Poisson distribution noise diffusion probability model according to claim 1, characterized in that, In step S2, the recovery operator R(·) is expressed as: In the formula, denotes the prediction result of the neural network U-Net at the t-th time step, x t denotes the restored image at the t-th time step, t denotes the time step, denotes the restored image at the t-1-th time step generated by the restoration operator.

6. The SPECT low-dose sinogram restoration method based on Poisson distribution noise diffusion probability model according to claim 5, characterized in that, In step S2, the recovery operator R(·) uses the formula The image distribution p θ (x t-1 | t ) is computed. denotes the prediction result.

7. The SPECT low-dose sinogram restoration method based on Poisson distribution noise diffusion probability model according to claim 6, characterized in that, In step S2, the recovery operator output corresponding to the prediction result of the time step T to t is averaged, and is processed by the following formula: The result calculated from the above equation is used as the restored image 8. A SPECT low-dose sinogram restoration system based on Poisson distribution noise diffusion probability model, characterized in that, Comprise: The training module is used for training a set of data sets containing n pairs of normal dose / low dose SPECT sinogram data, using the ideal reverse process Markov chain to generate the intermediate images of the ideal reverse process, taking these intermediate images together with the low dose sinogram as input data, and the corresponding normal dose sinogram as a label, to form a training data set of the U-Net model, and training the U-Net model with the training data set; Inference module: for taking an image x t and time t as input of a neural network U-Net, obtaining a prediction Along with the prediction results of all time steps, the prediction results are input into a time prediction aggregation module TPAM, in which the prediction results from time step T to time step t are respectively and the image x t and time t are input into a recovery operator R(·), and the image distribution is calculated, and the image generated after sampling of the inverse transform is taken as output; the outputs of the recovery operators corresponding to the prediction results from time step T to time step t are averaged, and the recovered image is taken as output iteration module: for inferring in the PDPM from the low-dose SPECT sinogram x T Initially, the process of step S2 is iterated T-1 times, and the U-Net prediction result obtained by taking x1 as input at the last iteration is taken as the final recovered image 9. The SPECT low-dose sinogram restoration system based on Poisson distribution noise diffusion probability model of claim 8, wherein, In the training module, the Markov chain is expressed as: Wherein, where x t-1 represents the recovered image at the t-1 step, x t represents the recovered image at the t step, x0represents the normal dose SPECT sinogram, and t represents the time step. The U-Net model is trained, and the training target is expressed as: wherein is the prediction result of the neural network U-Net, and xo is the normal dose SPECT sinogram.

10. The SPECT low-dose sinogram restoration system based on Poisson distribution noise diffusion probability model of claim 8, wherein, In the inference module, the recovery operator R(·) is expressed as: The recovery operator R(·) uses the formula The image distribution p is computed θ (x t-1 | x t ), and the inverse-transformed and sampled image is output; the recovery operator outputs corresponding to the prediction results from time step T to time t are averaged, and are processed using the following formula: The result calculated from the above equation is used as the restored image

Images