A method for generating PET dual tracer concentration maps based on a diffusion antagonism model

Abstract

The present invention discloses a method for generating a PET dual-tracer concentration map based on a diffusion adversarial model. Conditional GAN ​​is used to model each denoising step in the diffusion model, and pairing loss and cycle consistency loss are introduced to generate high-quality PET dual-tracer concentration maps. By combining the stepwise optimization mechanism of the diffusion model and the global characteristics of the adversarial generation network, the problems of slow generation speed and insufficient image quality are effectively solved, the balance between generation speed and quality is achieved, and an effective means is provided for the rapid generation of multi-tracer concentration maps. Through the stepwise optimization mechanism of the diffusion process, the generated concentration map has a high fidelity in structural details and can truly reflect the distribution characteristics of multiple tracers. This mechanism effectively solves the problem of loss of generated image details in the prior art and realizes the accurate generation of multi-tracer distribution characteristics.

Description

Technical Field

[0001] The invention belongs to the technical field of PET imaging, and in particular relates to a method for generating a PET double tracer concentration map based on a diffusion confrontation model. Background Art

[0002] Positron emission tomography (PET), as a non-invasive molecular imaging technology, is of great significance for early diagnosis, efficacy evaluation and prognosis of diseases. Traditional PET imaging usually relies on a single radioactive tracer, however, a single tracer often cannot fully reflect complex biological processes.

[0003] To address this limitation, dual-tracer PET imaging can provide multi-parameter information by using two different radioactive tracers, helping doctors to have a more comprehensive understanding of physiological and pathological processes. With dual-tracer images, the accuracy and sensitivity of diagnosis have been significantly improved. This is because dual tracers can provide complementary biological information, making the detection, localization and assessment of diseases more precise. However, traditional dual-tracer imaging may face the problems of excessive dose and high cost. For this reason, it is of great significance to develop dual-tracer generation technology.

[0004] Diffusion models [J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Proceedings of the 34th International Conference on Neural Information Processing Systems (NeurIPS), vol. 33, Article No. 574, pp. 6840–6851, 2020.] have received widespread attention in recent years. This generative model stands out for its powerful expressiveness and stable training process, especially in the task of high-quality image generation. Compared with traditional generative adversarial networks (GANs), diffusion models generate data in a step-by-step optimization manner, which can better capture complex image distribution characteristics and thus generate highly realistic images. However, the generation process of DDPM requires step-by-step denoising (usually hundreds to thousands of steps), resulting in a significantly slower generation speed than GAN.

[0005] [Z. Xiao, K. Kreis, and A. Vahdat, “Tackling the generative learning trilemma with denoising diffusion GANs,” International Conference on Learning Representations (ICLR), Apr. 2022.] In order to solve the problem of long sampling time of diffusion model, a method of using complex multi-peak distribution to simulate denoising distribution is proposed. By introducing denoising diffusion generative adversarial network and using conditional GAN ​​to model each denoising step, the generation time is significantly reduced.

[0006] The prior art has the following technical problems:

[0007] 1) The contradiction between high-quality sampling and diverse coverage The current generative learning framework has an obvious contradiction between high-quality generation and pattern coverage and sample diversity. Most image synthesis techniques focus on generating high-quality samples, but often perform poorly in covering the complete distribution of data and sample diversity.

[0008] 2) Limitations of generation speed and computational cost Current deep generative models usually require a lot of computing resources and have slow sampling speeds, making it difficult to meet the real-time or interactivity requirements of practical applications.

[0009] 3) The trade-off between generation quality, pattern coverage, and sampling efficiency Generative models usually find it difficult to simultaneously meet the three key requirements of high-quality sampling, pattern coverage and sample diversity, and fast generation. This trilemma forces existing models to make a trade-off between the three, often resulting in significant deficiencies in one aspect. Summary of the invention

[0010] In view of the above, the present invention proposes a method for generating a PET dual tracer concentration map based on a diffusion adversarial model. The method combines the advantages of the diffusion model and the adversarial generative network, and can significantly improve the quality and accuracy of the generated image while ensuring the generation speed. The present invention is achieved through the following technical solutions:

[0011] The present invention discloses a method for generating a PET dual tracer concentration map based on a diffusion confrontation model, comprising the following steps:

[0012] 1) Scanning the same biological tissue phantom injected with radioactive tracer A and radioactive tracer B respectively, acquiring radioactivity concentration maps of tracer A and tracer B, and using the obtained radioactivity concentration maps to form radioactivity concentration distribution data sets X and data sets Y, respectively, wherein data set X contains the concentration map of tracer A, and data set Y contains the concentration map of tracer B;

[0013] 2) Pairing the radioactivity concentration maps of tracer A and tracer B obtained above, using the two tracer radioactivity concentration maps of the same biological tissue phantom as a set of paired training data, and dividing all paired data into a training set, a validation set, and a test set;

[0014] 3) Construct a diffusion model, including:

[0015] Forward diffusion: Gaussian noise is gradually added to the original tracer concentration map to form a noise image;

[0016] Reverse denoising: denoising the noisy image and restoring the noisy image to a tracer concentration map;

[0017] 4) Construct a diffusion adversarial model and design the loss function. The diffusion adversarial model includes a generator , Generator , Discriminator , Discriminator , and introduce pairing loss and cycle consistency loss;

[0018] 5) Inputting the paired radioactive concentration maps in the training set into the diffusion confrontation model for training to obtain a trained diffusion confrontation model;

[0019] 6) For the trained diffusion adversarial model, the generator Responsible for converting the images in the tracer A radioactive concentration distribution dataset X Converted into an image that conforms to the distribution of tracer B radioactivity concentration data set Y , through the reverse denoising in step 3), the denoising sampling of each step is realized, and finally the concentration map that conforms to the distribution of the tracer B radioactive concentration distribution data set Y is obtained, and the generator Responsible for converting the images in the tracer B radioactivity concentration distribution dataset Y Converted into an image that conforms to the distribution of tracer A radioactive concentration data set X Then, through the reverse denoising in step 3), the denoising sampling of each step is realized, and finally a concentration map that conforms to the distribution of the radioactive concentration distribution data set X of the tracer domain A is obtained.

[0020] As a further improvement, the forward diffusion in step 3) of the present invention is specifically as follows:

[0021]

[0022] in and Two data sets of radioactivity concentration distribution from tracer A and tracer B respectively and Sample, initial input sample is considered as the initial noise-free image before starting the diffusion process ,sample is considered as the initial noise-free image before starting the diffusion process , through the forward diffusion process, in the predefined variance plan of Step by step, gradually move to data Add Gaussian noise, initial image is gradually transformed into an image close to noise , Indicates that the condition Next Generation The probability distribution of the forward process is used to describe the distribution of each step of the noise addition process. Represents a mean value of , the variance is Gaussian distribution, is the identity matrix, where Used to control the degree of noise accumulation. is the noise intensity during the diffusion process, The same operation is performed to obtain .

[0023] As a further improvement, the reverse denoising process in step 3) of the present invention is specifically as follows:

[0024]

[0025] Symmetrically, The same operation is performed for reverse denoising.

[0026] As a further improvement, the diffusion adversarial model in step 4) of the present invention includes two generators and , and two discriminators and , generator take over , , latent variables and time step , generates an image of the target tracer domain Y , and then use the posterior distribution Calculate the prediction sample , generator take over , , and , generates an image of the target tracer domain X and through Calculate the prediction sample , discriminator Used to determine whether the generated concentration map conforms to the distribution characteristics of the tracer domain X image, the discriminator Used to determine whether the generated concentration map conforms to the distribution characteristics of the tracer domain Y image, the latent variable is a high-dimensional standard normal distribution vector.

[0027] As a further improvement, the loss function of the diffusion adversarial model in step 4) of the present invention is Including pairing loss , Fighting Losses , cycle consistent loss The three parts are as follows:

[0028]

[0029] in, , and are the weighted coefficients of pairing loss, adversarial loss and cycle consistency loss respectively.

[0030] As a further improvement, the pairing loss described in the present invention Specifically:

[0031]

[0032]

[0033] in The concentration of tracer A is shown in The noisy image of the step, The tracer B concentration graph is shown in The noisy image of the step, represents the clean concentration map of tracer A, represents the clean concentration graph of tracer B, is a latent variable, and the time step Encoded via sinusoidal position embeddings to ensure consistency with the temporal conditions in the generator, represents L1 loss.

[0034] As a further improvement, the anti-loss Specifically:

[0035]

[0036]

[0037] in represents the clean concentration map of tracer A, represents the clean concentration graph of tracer B, Concentration graph of tracer A In the The noisy image of the step, The tracer B concentration graph is shown in The noisy image of the step, The concentration of tracer A is shown in The noisy image of the step, Concentration diagram of tracer B In the The noisy image of the step, is through Generated Output , and further obtain the false samples through the posterior distribution sampling, is through Generated Output , and further obtain the false samples through the posterior distribution sampling, is a latent variable, and the time step Encoded via sinusoidal position embedding to ensure consistency of temporal information.

[0038] As a further improvement, the cycle consistency loss described in the present invention The definition is as follows:

[0039]

[0040]

[0041] in,

[0042]

[0043]

[0044] in represents the clean concentration map of tracer A, represents the clean concentration graph of tracer B, Concentration graph of tracer A In the The noisy image of the step, The tracer B concentration graph is shown in The noisy image of the step, is a latent variable, and the time step Encoded by sinusoidal position embedding to ensure the consistency of temporal information, represents L1 loss.

[0045] The beneficial effects of the present invention are as follows:

[0046] 1. This paper proposes a method for generating PET dual tracer concentration maps based on a diffusion adversarial model. This method uses conditional GAN ​​to model each denoising step in the diffusion model, and introduces pairing loss and cycle consistency loss to generate high-quality PET dual tracer concentration maps. By combining the stepwise optimization mechanism of the diffusion model and the global characteristics of the adversarial generative network, the problems of slow generation speed and insufficient image quality are effectively solved, and a balance between generation speed and quality is achieved, providing an effective means for the rapid generation of multi-tracer concentration maps.

[0047] 2. The present invention uses a gradual optimization mechanism of the diffusion process to make the generated concentration map have high fidelity in structural details and can truly reflect the distribution characteristics of multiple tracers. This mechanism effectively solves the problem of loss of image details in the prior art and realizes the accurate generation of the distribution characteristics of multiple tracers.

[0048] 3. The present invention uses the global characteristics of the adversarial generative network to model each denoising step, so that the generated concentration map is consistent with the target distribution in terms of overall distribution, while improving the stability of the generation process. This method effectively solves the problems of inconsistent distribution or unstable training in existing generation models, and significantly improves the reliability and consistency of the generation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A schematic diagram of the steps of the PET image generation method of the present invention;

[0050] Figure 2 It is a structural schematic diagram of the diffusion countermeasure model in the present invention;

[0051] Figure 3 (a) is the concentration distribution diagram of tracer A;

[0052] Figure 3 (b) is the concentration distribution diagram of tracer B;

[0053] Figure 3(c) is a concentration distribution diagram of the corresponding tracer A generated using the present invention;

[0054] Figure 3 (d) is a concentration distribution diagram of the corresponding tracer B generated using the present invention. DETAILED DESCRIPTION

[0055] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0056] Figure 1 The present invention discloses a method for generating a PET double tracer concentration map based on a diffusion confrontation model, comprising the following steps:

[0057] (1) The same biological tissue phantom injected with radioactive tracer A and radioactive tracer B is scanned respectively to acquire radioactivity concentration maps of tracer A and tracer B, and these concentration maps are respectively used to form radioactivity concentration distribution data sets X and Y, where data set X contains the concentration map of tracer A and data set Y contains the concentration map of tracer B.

[0058] (2) Pairing the radioactivity concentration maps of tracer A and tracer B obtained above, using the two tracer radioactivity concentration maps of the same biological tissue phantom as a set of paired training data, and dividing all paired data into a training set, a validation set, and a test set.

[0059] (3) Construct a diffusion model, including:

[0060] Forward diffusion: Gaussian noise is gradually added to the original tracer concentration map to form a noise image;

[0061] Reverse denoising: denoising the noise image to restore the noise image to a tracer concentration map. The specific process is as follows:

[0062] 3.1 Forward diffusion is defined as:

[0063] and Two data sets of radioactivity concentration distribution from tracer A and tracer B respectively and The initial input sample is considered as the initial noise-free image before starting the diffusion process Then, through the forward diffusion process, in the predefined variance plan of Step by step, gradually move to data Add Gaussian noise, initial image is gradually transformed into an image close to noise , the forward diffusion process can be expressed as:

[0064]

[0065] in Indicates that the condition Next Generation The probability distribution of the forward process is used to describe the distribution of each step of the noise addition process. Represents a mean value of , the variance is The Gaussian distribution of is used to define the noise-added distribution. is the unit matrix, indicating that the noise of each pixel in the noise adding process is independent and identically distributed, and the noise has the same variance in each dimension. Used to control the degree of noise accumulation. is the noise intensity during the diffusion process, The constant , , .definition .right The same operation is performed to obtain .

[0066] 3.2 Reverse denoising is defined as:

[0067]

[0068] Symmetrically, The same operation is performed for reverse denoising.

[0069] (4) Construct a diffusion adversarial model and design the loss function. The diffusion adversarial model includes the generator , Generator , Discriminator , Discriminator , and introduce pairing loss and cycle consistency loss, Figure 2 This is a schematic diagram of the structure of the diffusion confrontation model in the present invention. The specific process is as follows:

[0070] 4.1 The diffusion adversarial model includes two generators and And two discriminators and . Generator Enter , , latent variables , time step Converted into an image of the target tracer domain B distribution . In getting Then, using the posterior distribution To calculate the prediction sample , realize the denoising sampling at each step. Generator Responsible for , , latent variables , time step The conditions are used as input to generate an image that conforms to the distribution characteristics of tracer A. . In getting Then, using the posterior distribution To calculate the prediction sample , to achieve denoising sampling at each step. Discriminator To evaluate the authenticity of the generated image, it receives the generated sample pair And the real sample , used to distinguish between real images and generated images. Discriminator To evaluate the authenticity of the generated image, it receives the generated sample pair And the real sample , used to distinguish between real images and generated images. Latent variables is a 100-dimensional vector that helps enhance the expressiveness of the generator, enabling it to generate dual-tracing images that are more consistent with the target characteristics and accelerate the convergence of the network. Each dimension of So the sampling satisfy ,in is the unit covariance matrix.

[0071] 4.2 Generator and The network is the same as that of the previous work, based on the NCSN++ architecture, but some adjustments have been made to better suit the task requirements. The core structure of the generator still uses the U-Net architecture, combined with residual blocks and attention blocks to enhance feature capture and representation capabilities. Some hyperparameters have been modified: the initial number of channels is set to 64, and different channel magnifications are used at each scale, starting from the initial number of channels in order of 1, 2, 2, and 4, to better capture multi-scale features step by step. At each scale, 2 ResNet blocks are used to improve the depth and refinement of feature extraction. Latent variables is a 100-dimensional vector, each dimension is independently derived from a standard normal distribution So the sampling satisfy ,in is the unit covariance matrix. It is regulated by the mapping network during the generation process. Setting the potential embedding dimension to 256 increases the representation capacity of the latent space to more flexibly express different features. In the generator, the adaptive group normalization (AdaGN) layer is used to replace the standard group normalization layer to allow the input of latent variables. Specifically, the latent variable First, it is transformed through a fully connected network (called a mapping network) to obtain an embedding vector , and pass the embedding vector to each AdaGN layer. Each AdaGN layer contains an independent fully connected layer, which is As input, it outputs the scaling and offset parameters of each channel, which are used to perform affine transformation on the feature map. This design enables the generator to flexibly control the characteristics of the generated image by adjusting the parameters in the latent space. In addition, sinusoidal position embedding is used to embed the integer time step Provides conditional embeddings whose embedding dimension is 4 times the initial number of channels.

[0072] Discriminator and The architecture of the generator is the same, which first contains a convolutional layer for preliminary feature extraction, followed by four ResNet blocks with residual connections to further extract features. The design of these ResNet blocks is similar to that of the generator. A LeakyReLU activation function with a negative slope of 0.2 is used after each ResNet block. After all ResNet blocks, a small batch standard deviation layer is added to increase the model's sensitivity to input diversity. Finally, after global pooling, the discriminant result is output through a fully connected layer.

[0073] 4.3 Diffusion Adversarial Model Loss Function Including pairing loss , Fighting Losses , cycle consistent loss There are three parts. The details are as follows:

[0074]

[0075] in, , and are the weighted coefficients of pairing loss, adversarial loss, and cycle consistency loss, respectively. , , . Pairing loss Specifically:

[0076]

[0077]

[0078] in The concentration of tracer A is shown in The noisy image of the step, The tracer B concentration graph is shown in The noisy image of the step, represents the clean concentration map of tracer A, represents the clean concentration graph of tracer B, is a latent variable, and the time step Encoded via sinusoidal position embeddings to ensure consistency with the temporal conditions in the generator, represents L1 loss;

[0079] Fighting Losses Specifically:

[0080]

[0081]

[0082] in represents the clean concentration map of tracer A, represents the clean concentration graph of tracer B, Concentration graph of tracer A In the The noisy image of the step, The tracer B concentration graph is shown in The noisy image of the step, The concentration of tracer A is shown in The noisy image of the step, Concentration diagram of tracer B In the The noisy image of the step, is through Generated Output , and further obtain the false samples through the posterior distribution sampling, is through Generated Output , and further obtain the false samples through the posterior distribution sampling, is a latent variable, and the time step Encoded via sinusoidal position embedding to ensure consistency of temporal information.

[0083] Cycle consistency loss Specifically:

[0084]

[0085]

[0086] in,

[0087]

[0088]

[0089] in represents the clean concentration map of tracer A, represents the clean concentration graph of tracer B, Concentration graph of tracer A In the The noisy image of the step, The tracer B concentration graph is shown in The noisy image of the step, is a latent variable, and the time step Encoded by sinusoidal position embedding to ensure the consistency of temporal information, represents L1 loss.

[0090] (5) The paired radioactive concentration maps in the training set are input into the diffusion adversarial model for training, and the loss function of the diffusion adversarial model is maximized. To update the generator and generator , minimize the diffusion adversarial model loss function To update the discriminator and the discriminator , obtain the trained diffusion adversarial model.

[0091] (6) For the trained diffusion adversarial model, the generator Responsible for converting the images in the tracer A radioactive concentration distribution dataset X Converted into an image that conforms to the distribution of tracer B radioactivity concentration data set Y , through the reverse denoising in step (3), the denoising sampling of each step is realized, and finally the concentration map that conforms to the distribution of the tracer B radioactive concentration distribution data set Y is obtained, and the generator Responsible for converting the images in the tracer B radioactivity concentration distribution dataset Y Converted into an image that conforms to the distribution of tracer A radioactive concentration data set X , and then through the reverse denoising in step (3), the denoising sampling of each step is realized, and finally the concentration map that conforms to the distribution of the radioactive concentration distribution data set X of the tracer domain A is obtained;

[0092] The following experiments are conducted based on simulation data to verify the effectiveness of the present invention: For tracer A and tracer B, respectively, 11 C-FMZ and 11 C-acetate, the phantom is a 3D brain phantom, and the equipment used in the simulation is Siemens' Inveon micro-PET scanner, equipped with 16 detector modules, each module contains 4 detectors, and a total of 80 rings are defined. The diameter of the detector ring is 161 mm, the cross-sectional field of view is 100 mm, the axial field of view is 127 mm, the spatial resolution is 1.4 mm, and the time resolution is 1.5 nanoseconds. The data set used for this experiment includes 18,000 pairs of training data and 3,600 pairs of test data.

[0093] Next, we use pytorch 1.10 to implement the training on an Ubuntu system with an NVIDIA GTX TITAN X. We set different learning rates for the generator and discriminator: the learning rate for the generator is 0.00016, and the learning rate for the discriminator is 0.0001. In the process of optimizing the model, we choose the Adam optimizer with a momentum parameter of The number of training iterations of the model is set to 450,000, the batch size is 8, and the total diffusion steps are Set to 4. In addition, in order to stabilize the training of the generator, only the exponential moving average (EMA) strategy is implemented on the generator, and the decay coefficient of EMA is set to 0.999, so as to smooth the weight update of the generator and improve the stability of the generation effect.

[0094] like Figure 3 (a)~ Figure 3 As shown in (d), the method of the present invention can realize the generation of PET dual tracer concentration map, and the generated image has good contrast, especially in some small structural areas, it shows good consistency and continuity, and better preserves the details of the original image.

[0095] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art to the present invention based on the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A method for generating a PET dual tracer concentration map based on a diffusion confrontation model, characterized in that: The steps include: 1) Scanning the same biological tissue phantom injected with radioactive tracer A and radioactive tracer B respectively, acquiring radioactivity concentration maps of tracer A and tracer B, and using the obtained radioactivity concentration maps to form radioactivity concentration distribution data sets X and data sets Y, respectively, wherein data set X contains the concentration map of tracer A, and data set Y contains the concentration map of tracer B; 2) Pairing the radioactivity concentration maps of tracer A and tracer B obtained above, using the two tracer radioactivity concentration maps of the same biological tissue phantom as a set of paired training data, and dividing all paired data into a training set, a validation set, and a test set; 3) Construct a diffusion model, including: Forward diffusion: Gaussian noise is gradually added to the original tracer concentration map to form a noise image; Reverse denoising: performing a denoising operation on the noise image to restore the noise image into a tracer concentration map; 4) Construct a diffusion adversarial model and design the loss function. The diffusion adversarial model includes the generator G x→y , generator G y→x , Discriminator D x , Discriminator D y , and introduce pairing loss and cycle consistency loss; The generator G x→y Receive y t , x0, latent variable z and time step t, generate the image y′0 of the target tracer domain Y, and then use the posterior distribution q(y t-1 |y t , y′0) calculate the predicted sample y′ t-1 , generator G y→x Receive x t , y0, z and t, generate the image x′0 of the target tracer domain X and pass it through q(x t-1 |x t ,x′0) calculate the predicted sample x′ t-1 , discriminator D x The discriminator D is used to determine whether the generated concentration map conforms to the distribution characteristics of the tracer domain X image. y It is used to determine whether the generated concentration map conforms to the distribution characteristics of the tracer domain Y image. The latent variable z is a high-dimensional standard normal distribution vector; Loss function L of the diffusion adversarial model total Including pairing loss L rec , against the loss L adv , cycle-consistent loss L cyc The three parts are as follows: L total =a rec L rec +a adv L adv +a cyc L cyc ; Among them, a rec , α adv and α cyc are the weighted coefficients of pairing loss, adversarial loss, and cycle consistency loss, respectively; The pairing loss L rec Specifically: where x t represents the noisy image of the tracer A concentration map at step t, y t represents the noisy image of the concentration map of tracer B at step t, x0 represents the clean concentration map of tracer A, y0 represents the clean concentration map of tracer B, z is the latent variable, time step t is encoded by sinusoidal position embedding to ensure consistency with the temporal conditions in the generator, and || ||1 represents the L1 loss; The adversarial loss L adv Specifically: Where x0 represents the clean concentration map of tracer A, y0 represents the clean concentration map of tracer B, and x t represents the noisy image of the tracer A concentration map x0 at step t, y t represents the noisy image of the tracer B concentration map at step t, x t-1 represents the noisy image of the tracer A concentration map at step t-1, y t-1 represents the noise image of the tracer B concentration map y0 at step t-1, x′ t-1 It is through G y→x (x t , y0; z, t) generates the output x′0, and further obtains the false sample y′ by sampling from the posterior distribution t-1 It is through G x→y (y t , x0; z, t) generates the output y′0, and further samples the pseudo samples obtained by the posterior distribution, z is the latent variable, and the time step t is encoded by the sinusoidal position embedding to ensure the consistency of the temporal information; The cycle consistency loss L cyc The definition is as follows: in, Where x0 represents the clean concentration map of tracer A, y0 represents the clean concentration map of tracer B, and x t represents the noisy image of the tracer A concentration map x0 at step t, y t represents the noisy image of the tracer B concentration map at step t, z is the latent variable, time step t is encoded by sinusoidal position embedding to ensure the consistency of temporal information, and || ||1 represents the L1 loss; 5) Inputting the paired radioactive concentration maps in the training set into the diffusion confrontation model for training to obtain a trained diffusion confrontation model; 6) For the trained diffusion adversarial model, the generator G x→y It is responsible for converting the image x0 in the tracer A radioactive concentration distribution dataset X into the image y′0 that conforms to the distribution of the tracer B radioactive concentration distribution dataset Y. Through the reverse denoising in step 3), the denoising sampling of each step is realized, and finally the concentration map that conforms to the distribution of the tracer B radioactive concentration distribution dataset Y is obtained. The generator G y→x It is responsible for converting the image y0 in the radioactive concentration distribution dataset Y of tracer B into the image x′0 that conforms to the distribution of the radioactive concentration distribution dataset X of tracer A, and then implementing the denoising sampling of each step through the reverse denoising in step 3), and finally obtaining the concentration map that conforms to the distribution of the radioactive concentration distribution dataset X of tracer domain A.

2. The method for generating a PET dual tracer concentration map based on a diffusion confrontation model according to claim 1, characterized in that: The forward diffusion in step 3) is specifically as follows: Where x and y are samples from two datasets X and Y of the radioactivity concentration distribution of tracer A and tracer B, respectively. The initial input sample x is regarded as the initial noise-free image x0 before the diffusion process begins, and the sample y is regarded as the initial noise-free image y0 before the diffusion process begins. Through the forward diffusion process, the predefined variance plan β t Gaussian noise is gradually added to the data x0 in T steps, and the initial image x0 is gradually transformed into an image x close to the noise. t ,q(x t |x0) means generating x under condition x0 t The probability distribution of the forward process is used to describe the distribution of each step of the noise addition process. Represents a mean value of The variance is Gaussian distribution, I is the identity matrix, where Used to control the degree of noise accumulation, β t is the noise intensity during the diffusion process. The same operation is performed on y0. y is obtained through the gradual noise addition process. t .

3. The method for generating a PET dual tracer concentration map based on a diffusion confrontation model according to claim 1, characterized in that: The reverse denoising process in step 3) is specifically as follows: Symmetrically, for x t The same operation is performed for reverse denoising.

Images