A three-dimensional PET image reconstruction method and system based on a diffusion multi-scale generative adversarial network and a storage medium

Abstract

This invention discloses a method, system, and storage medium for three-dimensional PET image reconstruction based on a diffusion multi-scale generative adversarial network (DMGAN). The method includes: inputting the PET image to be reconstructed into a pre-trained DMGAN to generate a near-realistic full-dose PET image; the DMGAN includes a diffusion generator and a U-Net discriminator; the PET image to be reconstructed, i.e., a low-dose L-PET image, is sliced ​​and then input into the diffusion generator to synthesize a full-dose image, i.e., an F-PET image, which contains a series of corresponding target slices; simultaneously, a noisy F-PET image is generated; the generated image is input into the U-Net discriminator to extract details from global and local views to improve the quality of the generated F-PET image. This invention can effectively capture and restore image details, making the generated image visually and structurally close to a real full-dose PET image.

Description

Technical Field

[0001] This invention relates to the field of medical image engineering technology, and in particular to a method, system and storage medium for three-dimensional PET image reconstruction based on diffusion multiscale generative adversarial networks. Background Technology

[0002] As one of the most widely used medical imaging technologies, PET plays a crucial role in navigation surgery, medical evaluation, and clinical examination. Unlike other imaging technologies such as magnetic resonance imaging (MRI) and computed tomography (CT), PET can detect biochemical and physiological changes. Since biochemical and physiological changes often precede anatomical changes, PET is also widely used for preventative treatment and early disease identification. PET can assess molecular changes in the human body within the body. Despite the significant advantages of PET, there is growing concern about the potential health risks associated with radiation exposure during the scanning process. For example, in clinical practice, injection doses are often limited by radiation exposure, as higher doses may increase the risk of cancer and cause some degree of harm to the body. Therefore, low-dose L-PET, which allows for image acquisition with minimal radiation exposure, has attracted considerable interest from researchers. However, compared to full-dose F-PET images, L-PET images exhibit higher noise levels, reduced image contrast, and more artifacts, making accurate diagnosis difficult for physicians. Therefore, acquiring high-quality images from low-dose images to minimize exposure while maintaining image quality is of significant practical importance.

[0003] To improve the quality of PET images, numerous methods have been proposed. One approach to achieving high-quality PET images is to integrate prior information into the image reconstruction process. This method allows for the direct incorporation of imaging physical information. However, it faces challenges related to computationally intensive processing and requires access to the physical projection model. Many studies have explored voxel estimation methods after image reconstruction. These methods include random forest-based regression methods, mapping-based sparse representation methods, semi-supervised triple dictionary learning methods, and multi-level canonical correlation analysis frameworks. While these existing methods have shown promising results, they tend to produce overly smoothed images.

[0004] In recent years, deep learning methods have been extensively studied in the field of medical imaging. Generative adversarial networks (GANs) and convolutional neural networks (CNNs) have proven successful in denoising low-dose CT images. Due to the widespread application of deep learning across various fields, it has also had a significant impact on L-PET image tasks. Xiang et al. proposed a sophisticated CNN model with automatic context learning to predict high-dose PET images using only a quarter dose of the full-dose PET image and the corresponding MR T1 image. Wang et al. developed a comprehensive framework using 3D conditional generative adversarial networks (GANs) to generate high-quality PET images from corresponding L-PET images. Kaplan and Zhu proposed a model that incorporates specific image features into the loss function to denoise a 1 / 10 dose full-dose PET image slice and estimate its full-dose counterpart. Chen et al. proposed a method to synthesize high-quality and accurate PET images using only PET data or by combining PET and MR information. Ouyang et al. proposed that GANs can achieve similar performance levels even without MR information. Recently, Yu et al. introduced a simplified L-PET image reconstruction framework that can quickly generate F-PET images and improve the overall quality of the final 3D F-PET image by utilizing the spatial details of the generated F-PET slices.

[0005] However, some limitations still exist, such as how to enhance the understanding of semantic information in L-PET reconstruction from different perspectives. Summary of the Invention

[0006] This invention provides a method, system, and storage medium for three-dimensional PET image reconstruction based on diffusion multi-scale generative adversarial networks (GANs), aiming to solve the aforementioned problems. It includes a diffusion generator and a U-Net discriminator. Experimental verification shows that the method of this invention exhibits superiority over other L-PET image reconstruction methods.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for 3D PET image reconstruction based on diffusion multi-scale generative adversarial networks, comprising:

[0009] The PET image to be reconstructed is input into the pre-trained diffusion multiscale generative adversarial network DMGAN to generate a near-realistic full-dose PET image.

[0010] The diffusion multi-scale generative adversarial network includes a diffusion generator and a U-Net discriminator. The PET image to be reconstructed, i.e., a low-dose L-PET image or a slice thereof, is input into the diffusion generator to synthesize a full-dose image, i.e., an F-PET image. At the same time, a noisy F-PET image is generated. The generated image is input into the U-Net discriminator to extract details from the global and local views to improve the quality of the generated F-PET image.

[0011] In the above technical solution, furthermore, before reconstruction, the L-PET and F-PET images used for training need to be preprocessed using statistical parameter mapping in order to re-align and standardize.

[0012] In this invention, according to a specific embodiment, raw low-dose PET image slices are input into a diffusion generator to synthesize a full-dose image. The diffusion generator also contains a series of corresponding target slices. In the diffusion generator, noisy L-PET images are generated by introducing noise into the input low-dose PET images. Then, the diffusion generator generates full-dose PET (F-PET) images and noisy F-PET images. The purpose of this step is to enhance the generator's generalization ability to images and improve training stability. The generated images are input into a U-Net discriminator, which extracts details from both holistic and specific perspectives to enhance the quality of the generated F-PET images. Through this process, the method of this invention can generate high-quality F-PET images, ensuring their practicality and reliability in clinical applications. The diffusion generator makes the generated images more realistic by progressively adding and removing noise, while the U-Net discriminator further improves the detail and clarity of the images through multi-scale feature extraction.

[0013] The loss function in the diffusion generator is:

[0014] L GAN (D,G)=-E x,y [(D(x,y)-1) 2 ]-E x [D(x,G(x)) 2 ]

[0015] Where x represents the L-PET image, G(x) represents the F-PET image generated by the generator, and y is the corresponding original F-PET image;

[0016] Charbonnier Loss is introduced to penalize the Euclidean difference between the generated F-PET image and the original F-PET image:

[0017]

[0018] Simultaneously, Charbonnier Loss penalty is introduced to generate Euclidean discretization between noisy F-PET images and the original F-PET images:

[0019]

[0020] Furthermore, considering the perceptual differences between the generated F-PET image, the generated noisy F-PET image, and the original F-PET image, VGG16-Net trained on ImageNet is used to extract feature representations of G(x) and y, with the perceptual loss being:

[0021]

[0022] Where V represents the feature map extracted by VGG16-Net.

[0023] The U-Net discriminator comprises a downsampling network and an upsampling network. Downsampling reduces spatial resolution while increasing the number of channels in the feature map, thereby capturing high-level abstract features of the image. The goal of downsampling is to progressively reduce the spatial size of the image and enhance the network's receptive field to capture global information. Upsampling progressively restores the spatial resolution of the image and reconstructs image details as accurately as possible. The goal of upsampling is to progressively expand the spatial size of the feature map to generate an output image of the same size as the input image. The downsampling and upsampling networks are connected via skip connections and bottleneck connections.

[0024] Furthermore, the U-Net discriminator performs discrimination based on each pixel z, and the encoder loss is:

[0025]

[0026] The decoder loss for the average of all pixel values ​​is:

[0027]

[0028] [D dec (x)] i,j and [D dec (G(z))] i,j These represent the discriminator's judgment at pixel (i,j); D dec The output of each pixel is obtained by skip connections from the middle layer of the encoder network, combining the specific details of the low-level features with the global information of the high-level features obtained by upsampling from the upsampling layer.

[0029] Considering the loss function above, the goal of the generator is:

[0030]

[0031] The generator is encouraged to focus on synthesized images by effectively capturing global structure and local details, aiming to more effectively deceive the discriminator;

[0032] The basic loss of the discriminator is:

[0033] L adv =E x,y [logD dec [(x,y)]+E x [log(1-D dec (G(x),x))]

[0034] The U-Net discriminator returns two values, representing the outputs of the decoder and encoder, respectively. The intermediate loss describes the encoder loss of the U-Net discriminator, expressed as:

[0035] L middle = -log(1-D encfake )-log(D encreal )

[0036] The total loss of the discriminator is:

[0037] L d =(L adv +L middle )*0.5ν

[0038] Where v is a hyperparameter.

[0039] A three-dimensional PET image reconstruction system based on diffusion multi-scale generative adversarial networks, for implementing the method described in any of the preceding methods.

[0040] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any of the preceding claims.

[0041] An electronic device, the device comprising:

[0042] One or more processors;

[0043] Memory, used to store one or more programs;

[0044] When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in any of the preceding methods.

[0045] Compared with the prior art, the specific beneficial effects of the present invention are as follows:

[0046] This invention achieves high-quality reconstruction of L-PET images through an end-to-end neural network architecture. This architecture is a novel diffusion multi-scale generative adversarial network (DMGAN), which effectively captures and restores image details, resulting in generated images that are visually and structurally close to real full-dose PET images. Original low-dose PET image slices are input into a diffusion generator to synthesize a full-dose image. The diffusion generator also contains a series of corresponding target slices. In the diffusion generator, noisy L-PET images are generated by introducing noise into the input low-dose PET images. The diffusion generator then generates both full-dose PET (F-PET) images and noisy F-PET images. The generated images are input into a U-Net discriminator, which extracts details from both holistic and specific perspectives to enhance the quality of the generated F-PET images. Through this process, our method can generate high-quality F-PET images, ensuring its practicality and reliability in clinical applications. The diffusion generator makes the generated images more realistic by gradually adding and removing noise, while the U-Net discriminator further improves the details and clarity of the images through multi-scale feature extraction. Attached Figure Description

[0047] Figure 1 The overall architecture of the DMGAN network includes a 3D diffusion generator module and a 3D u-net discriminator module.

[0048] Figure 2 The main structure of the 3D diffusion generator module.

[0049] Figure 3 Different models are visually compared using color images. The chromatographic range from blue to red represents an increase in metabolic value from low to high.

[0050] Figure 4 A comparison of pseudo-color variance plots between images generated by existing methods and the method of this invention, and the original F-PET images. Colors transition from blue to red, indicating differences ranging from slight to significant.

[0051] Figure 5 Displays L-PET images of epileptic foci, F-PET images of synthetic epileptic foci generated by various methods, and corresponding F-PET images of real epileptic foci. The color gradient ranges from blue to red, representing the metabolic values ​​of the tracer FDG from low to high.

[0052] Figure 6 The peak signal-to-noise ratios for various methods are displayed.

[0053] Figure 7 Displaying structural similarity indices for various methods.

[0054] Figure 8 Bland-Altman provides a two-dimensional dot map of the brain's SUVR based on all reconstructed F-PET images from different GAN models. In the AAL brain template, each blue dot represents its corresponding brain region. Detailed Implementation

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] In an embodiment of the invention, the experimental dataset includes brain PET scans of 45 children with epilepsy, aiming to improve diagnostic accuracy through improved image processing methods. Low-dose brain PET scans and corresponding full-dose brain PET scans were collected.

[0057] To achieve the generation of high-quality F-PET images from L-PET images, this invention proposes a novel diffusion multi-scale generative adversarial network (DMGAN). DMGAN is designed to effectively capture and restore image details, resulting in generated images that visually and structurally approximate real full-dose PET images. In embodiments of this invention, the performance of DMGAN is validated through a series of qualitative and quantitative comparisons.

[0058] In the quantitative evaluation, this invention uses two main metrics: the Structural Similarity Index (SSIM) and the Peak Signal-to-Noise Ratio (PSNR). SSIM measures the structural similarity between two images, while PSNR assesses the reconstruction quality and noise level. These metrics allow for an objective comparison of the differences in image quality between DMGAN and other methods, ensuring that the method presented in this invention has a significant advantage in generating high-quality F-PET images.

[0059] As can be seen, the research of this invention not only proposes a new image reconstruction method, but also verifies its effectiveness through a rigorous evaluation system, providing important technical support for the clinical diagnosis of brain diseases such as epilepsy.

[0060] Example

[0061] A method for 3D PET image reconstruction based on diffusion multi-scale generative adversarial networks, comprising:

[0062] Step 1: The dataset used included 45 pediatric subjects diagnosed with epilepsy, comprising low-dose brain PET scans collected in 2020 and corresponding full-dose brain PET scans. FDG PET images of the brains of all subjects were obtained using a PET / MR system designed specifically for whole-body imaging.

[0063] Step 2: Before reconstruction, both L-PET and F-PET images require preprocessing using statistical parametric mapping for realignment and standardization. After preprocessing, the voxel size of these PET images is standardized to 1×1×1 mm. By processing the three-dimensional brain images in the PET dataset, a standardized dataset is obtained, which serves as the basis for the experimental analysis of this invention.

[0064] Step 3: An end-to-end neural network architecture was designed, which mainly includes two core modules: the 3D Diffusion Generator module and the 3D U-net Discriminator module, for the reconstruction of low-dose 3D PET images.

[0065] Step 4: In these two types of modules, the 3D Diffusion Generator module functions as follows: The original low-dose PET image is input into the 3D Diffusion Generator module to synthesize a full-dose image, i.e., an F-PET image. Simultaneously, the 3D Diffusion Generator module can generate both F-PET images and noisy F-PET images to improve the discriminator's discrimination capability.

[0066] Step 5: Input the F-PET image and the noisy F-PET image obtained by the 3D Diffusion Generator module into the 3D U-net Discriminator module. This module can extract details from the whole and from specific angles and identify the image obtained by the generator to improve the quality of the generated F-PET image.

[0067] Step 6: To evaluate the effectiveness of the proposed method, the evaluation conducted by this invention includes both qualitative and quantitative assessments. For quantitative comparison, this invention uses two metrics: the Structural Similarity Index (SSIM) and the Peak Signal-to-Noise Ratio (PSNR), to evaluate the performance of different methods.

[0068] Specifically, the experimental dataset used in step 1 included 45 pediatric subjects diagnosed with epilepsy, comprising low-dose brain PET scans collected in 2020 and corresponding full-dose brain PET scans. FDG-PET images of the brains of all subjects were obtained using a PET / MR system designed specifically for whole-body imaging. In clinical practice, this invention does not exclude images of relatively lower quality. L-PET images were generated by reconstructing list-pattern F-PET data that underwent a 5% undersampling process.

[0069] For PET images in clinical settings, the goal of this invention is to generate images from L-PET images that are highly similar to the original F-PET images. The advantage of the proposed method is that it generates output samples using the identified distribution of the actual images and fully utilizes their variability. First, slices of the original low-dose PET image are input into a diffusion generator to synthesize a full-dose image, which also includes a series of corresponding target slices. Then, the diffusion generator generates both an F-PET image and a noisy F-PET image. The generated images are input into a U-Net discriminator to extract details from both the overall picture and specific angles, thereby improving the quality of the generated F-PET image.

[0070] In step 4, the original low-dose PET image is input into the 3D Diffusion Generator module to synthesize a full-dose image. Simultaneously, the 3D Diffusion Generator module can generate both F-PET images and noisy F-PET images to improve the discriminator's ability to distinguish between different doses.

[0071] Specifically, step 4 involves designing a novel diffusion generator. The greatest advantage of this model lies in its ability to acquire different information from different levels, thereby improving the generator's ability to extract information from the original image. The diffusion generator provides a completely new perspective on learning the distribution of the original image compared to other common generators: L GAN (D,G)=-E x,y [(D(x,y)-1) 2 ]-E x [D(x,G(x)) 2 ]

[0072] Specifically, where x represents the L-PET image, G(x) represents the F-PET image generated by the generator, and y is the corresponding F-PET image. This invention uses Charbonnier Loss to penalize the Euclidean difference between the generated F-PET image and the original F-PET image:

[0073]

[0074] Simultaneously penalize the Euclidean difference between the generated noisy F-PET image and the original F-PET image:

[0075]

[0076] Specifically, considering the perceptual differences between the generated F-PET image, the generated noisy F-PET image, and the original F-PET image, this invention uses VGG16-Net trained on ImageNet to extract feature representations of G(x) and y. The perceptual loss is:

[0077]

[0078] Where V represents the feature map extracted by VGG16-Net. Based on the loss function described above, the total loss function of the diffusion generator is expressed as:

[0079] L dmgan (G,V,D)=L GAN (D,G)+α(L(G)+L(G noised ))+βL perc (G,V).

[0080] Both α and β are set to 100

[0081] In step 5, after the 3D diffusion generator module generates the F-PET image and the noisy F-PET image, these images are input into the 3D U-net discriminator module. This discriminator module can extract information from both the overall and specific detail levels to distinguish the images generated by the generator, thereby improving the quality of the generated F-PET images.

[0082] Specifically, this invention introduces a u-net discriminator into the model to obtain raw image information from both global and local views, which better complements the diffusion generator. The u-net discriminator consists of an original downsampling network and a novel upsampling network, connected via skip-connections and bottlenecks. Compared to the original network, the u-net discriminator performs discrimination on a per-pixel basis. The encoder loss is: L Denc =-E x [logD enc (x)]-E z [log(1-D enc (G(z)))]

[0083] Specifically, the decoder loss for the average of all pixel values ​​can be expressed as:

[0084]

[0085] [D dec (x)] i,j and [D dec (G(z))] i,j These represent the discriminator's judgment at pixel (i,j). D dec The output of each pixel is obtained by skip connections from the middle layer of the encoder network, combining the specific details of the low-level features with the global information of the high-level features obtained by upsampling from the upsampling layer.

[0086] Specifically, considering the loss function mentioned above, the goal of the generator is:

[0087]

[0088] The generator is encouraged to focus on synthesizing images by effectively capturing global structure and local details, aiming to more effectively deceive the discriminator.

[0089] The loss of the basic discriminator is:

[0090] L adv =E x,y [logD dec [(x,y)]+E x [log(1-D dec (G(x),x))]

[0091] Specifically, in the method proposed in this invention, the u-net diffusion discriminator returns two values, representing the outputs of the decoder and encoder, respectively. The intermediate loss is used to describe the loss of the u-net discriminator encoder, and is expressed as:

[0092] L middle = -log(1-D encfake )-log(D encreal )

[0093] Therefore, the total loss of the discriminator is:

[0094] L d =(L adv +L middle )*0.5ν

[0095] Here, the value of ν is set to 1.

[0096] like Figure 1 , Figure 2 This is a schematic diagram of the DMGAN neural network architecture designed in step 3, including a diffusion generator module and a u-net discriminator module. To verify the performance of the proposed method, it is compared with cGAN, CycleGAN, and transGAN. Furthermore, to eliminate the influence of different initialization parameters on the experiments, the same random seed is set for all experimental methods. This invention uses the PyTorch library on an NVIDIA RTX 4090 GPU. The batch size is 1, and the epoch is set to 300. Finally, this invention is evaluated through qualitative assessment and quantitative measurement. For quantitative comparison, two metrics are selected to evaluate the performance of various methods: SSIM and PSNR.

[0097] like Figure 3 As shown, the present invention can be achieved through... Figure 3Metabolic diagrams from various methods are used to observe the metabolic processes more clearly. The method of this invention provides more metabolic detail than other methods.

[0098] like Figure 4 As shown, a more intuitive comparison is as follows: Figure 4 As shown, the generated F-PET image and the original F-PET image are calculated in the pseudochromatic difference map, indicating that the proposed method exhibits the smallest voxel-scale difference compared to other methods.

[0099] like Figure 5 As shown, given the clinical importance of accurate low-dose reconstruction of epileptic foci in pediatric patients, this invention presents, in section 5, L-PET images of epileptic foci, synthetic F-PET images of epileptic foci, and F-PET images of ground-based real epileptic foci. Clearly, compared to other methods, the synthetic F-PET images of epileptic foci generated by DMGAN exhibit a more accurate distribution of voxel metabolic intensity, thus providing a clearer depiction of the epileptic foci.

[0100] like Figure 6 As shown, peak signal-to-noise ratio is used to evaluate the reconstruction quality and noise level of the image, quantitatively demonstrating the significant advantages of the method of the present invention in generating high-quality F-PET images.

[0101] like Figure 7 As shown, the structural similarity index is used to evaluate the similarity of two images in terms of structural information, quantitatively demonstrating that the method of the present invention has significant advantages in generating high-quality F-PET images.

[0102] like Figure 8 As shown, the whole-brain SUVR of the present invention has a smaller consistency limit compared with other methods.

[0103] It can be seen that the method of the present invention has a superior reconstruction effect compared with other existing methods.

[0104] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0106] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0107] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention are included within the protection scope of the present invention.

Claims

1. A method for reconstructing three-dimensional PET images based on diffusion-multiscale generative adversarial networks, characterized in that, include: The PET images to be reconstructed are input into the pre-trained diffusion multiscale generative adversarial network DMGAN to generate near-realistic full-dose PET images. The diffusion multi-scale generative adversarial network includes a diffusion generator and a U-Net discriminator. The PET image to be reconstructed, i.e., a low-dose L-PET image or a slice thereof, is input into the diffusion generator to synthesize a full-dose image, i.e., an F-PET image. At the same time, a noisy F-PET image is generated. The generated image is input into the U-Net discriminator to extract details from global and local views to improve the quality of the generated F-PET image. The loss function in the diffusion generator is: ， Where x represents the L-PET image, G(x) represents the F-PET image generated by the generator, and y is the corresponding original F-PET image; Charbonnier Loss is introduced to penalize the Euclidean difference between the generated F-PET image and the original F-PET image: ， Simultaneously, Charbonnier Loss penalty is introduced to generate Euclidean discretization between noisy F-PET images and the original F-PET images: ， Furthermore, considering the perceptual differences between the generated F-PET image, the generated noisy F-PET image, and the original F-PET image, VGG16-Net trained on ImageNet is used to extract feature representations of G(x) and y, with the perceptual loss being: ， Where V represents the feature map extracted by VGG16-Net.

2. The three-dimensional PET image reconstruction method based on diffusion-multiscale generative adversarial networks according to claim 1, characterized in that, Before reconstruction, the L-PET and F-PET images used for training need to be preprocessed using statistical parameter mapping for realignment and standardization.

3. The three-dimensional PET image reconstruction method based on diffusion-multiscale generative adversarial networks according to claim 1, characterized in that, The U-Net discriminator comprises a downsampling network and an upsampling network, which are connected via skip-connections and bottleneck connections.

4. The three-dimensional PET image reconstruction method based on diffusion-multiscale generative adversarial networks according to claim 3, characterized in that, The U-Net discriminator performs discrimination based on each pixel z, and the encoder loss is: ， The decoder loss for the average of all pixel values ​​is: ， and These represent the discriminant's judgment at pixel (i, j); The output of each pixel is obtained by skip connections from the middle layer of the encoder network, combining the specific details of the low-level features with the global information of the high-level features obtained by upsampling from the upsampling layer. Considering the loss function above, the goal of the generator is: ， The generator is encouraged to focus on synthesized images by effectively capturing global structure and local details, aiming to more effectively deceive the discriminator; The basic loss of the discriminator is: ， The U-Net discriminator returns two values, representing the outputs of the decoder and encoder, respectively. The intermediate loss describes the encoder loss of the U-Net discriminator, expressed as: ， The total loss of the discriminator is: , Where v is a hyperparameter.

5. A three-dimensional PET image reconstruction system based on diffusion-multiscale generative adversarial networks, characterized in that, Used to implement the method as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-4.

7. An electronic device, characterized in that, The device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-4.

Images