Three-dimensional PET image reconstruction method and system based on diffusion multi-scale generative adversarial network, and storage medium

Through diffusion multi-scale generation adversarial networks, combined with diffusion generator and U-Net discriminator, the problem of high noise in low-dose PET images is solved, and high-quality full-dose PET images are generated, which improves the detail and clarity of the image and ensures clinical practice. Reliability of the application.

CN118941718AActive Publication Date: 2024-11-12ZHEJIANG UNIV

Patent Information

Application Number
CN202411136520.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-12
Estimated Expiration
2044-08-19

AI Technical Summary

Technical Problem

Due to high noise and low image contrast, low image contrast makes it difficult to accurately diagnose low-dose PET images. The prior art is prone to excessive smoothing problems during image reconstruction.

Method used

Using a three-dimensional PET image reconstruction method based on diffusion multi-scale generation adversarial network, low-dose PET images are input through diffusion generator and U-Net discriminator to generate high-quality full-dose images. The diffusion generator enhances generalization ability by introducing noise ,U-Net discriminator improves image detail and clarity through multi-scale feature extraction.

Benefits of technology

The generated images are visually and structurally close to real full-dose PET images, improving the quality and reliability of the images and ensuring practicality in clinical applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941718A_ABST
    Figure CN118941718A_ABST
Patent Text Reader

Abstract

The invention discloses a three-dimensional PET image reconstruction method and system based on a diffusion multi-scale generative adversarial network and a storage medium, and the method comprises the steps: inputting a PET image needing to be reconstructed into a trained diffusion multi-scale generative adversarial network DMGAN, and generating a full-dose PET image which is close to the reality; the diffusion multi-scale generative adversarial network comprises a diffusion generator and a U-Net discriminator, the PET image needing to be reconstructed, namely a low-dose L-PET image, is sliced and then input into the diffusion generator to synthesize a full-dose image, namely an F-PET image which comprises a series of corresponding target slices, and meanwhile, a noise-added F-PET image is generated; the generated image is input into a U-Net discriminator, and details are extracted from global and local views to improve the quality of the generated F-PET image. According to the method, image details can be effectively captured and restored, so that the generated image is close to a real full-dose PET image visually and structurally.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical image engineering technology, and in particular to a three-dimensional PET image reconstruction method, system and storage medium based on a diffusion multi-scale generative adversarial network. Background Art

[0002] As one of the most widely used medical imaging techniques, PET plays a key role in navigation surgery, medical evaluation, and clinical examinations. Unlike other imaging techniques such as magnetic resonance imaging and computed tomography, PET is able to detect biochemical and physiological changes. Since biochemical and physiological changes often precede anatomical changes, PET is also widely used for preventive treatment and early disease identification. PET can evaluate molecular changes in the human body in vivo. Despite the significant advantages of PET, people are increasingly concerned about the possible health risks of radiation exposure during scanning. For example, in clinical practice, the injection dose is usually limited by the radiation dose, because more radiation dose has the potential to increase the risk of cancer and cause a certain degree of damage to the body. Therefore, low-dose L-PET, which can acquire images with minimal radiation exposure, has attracted great interest from researchers. However, compared with full-dose F-PET images, L-PET images show higher noise levels, reduced image contrast, and more artifacts, making it difficult for doctors to make accurate diagnoses. Therefore, it is of great practical significance to obtain high-quality images from low-dose images to minimize image exposure while maintaining image quality.

[0003] Many methods have been proposed to improve the quality of PET images. One approach to achieve high-quality PET images is to integrate prior information into the image reconstruction process. This approach allows for the direct incorporation of imaging physics information. However, it faces challenges associated with intensive computation and must have access to a physical projection model. Many studies have explored methods for voxel estimation 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. Although these existing methods have shown promising results, they often produce over-smoothed images.

[0004] In recent years, deep learning methods have been widely studied in the field of medical imaging. Generative adversarial networks and convolutional neural networks have been shown to be successful in denoising low-dose CT images. As deep learning is widely used in various fields, it has also had a significant impact in the L-PET image task. Xiang et al. proposed a complex convolutional neural network model with automatic context learning to predict high-dose PET images using only full-dose PET images and 1 / 4 dose of corresponding MR T1 images. 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 1 / 10 dose full-dose PET image slices and estimate their full-dose counterparts. Chen et al. proposed methods to synthesize high-quality and accurate PET images using PET-only data or combining PET and MR information. Ouyang et al. proposed generative adversarial networks to achieve similar performance levels without MR information. Recently, Yu et al. introduced a simplified L-PET image reconstruction framework that can quickly generate F-PET images and utilize the spatial details of the generated F-PET slices to improve the overall quality of the final 3D F-PET images.

[0005] However, there are still some limitations, such as how to enhance the understanding of semantic information in L-PET reconstruction from different viewpoints. Summary of the invention

[0006] The present invention provides a three-dimensional PET image reconstruction method, system and storage medium based on a diffusion multi-scale generative adversarial network, aiming to solve the above problems. The method includes a diffusion generator and a U-Net discriminator. Through experimental verification, the method of the present invention shows superiority over other L-PET image reconstruction methods.

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

[0008] A three-dimensional PET image reconstruction method based on a diffusion multi-scale generative adversarial network, comprising:

[0009] The PET image to be reconstructed is input into the trained diffusion multi-scale generative adversarial network DMGAN to generate a full-dose PET image close to the real one;

[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 global and local views to improve the quality of the generated F-PET image.

[0011] In the above technical solution, further, before reconstruction, both the training data sets L-PET and F-PET images need to be preprocessed using statistical parameter mapping for realignment and standardization.

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

[0013] Among them, 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] At the same time, the Charbonnier Loss penalty is introduced to generate the Euclidean difference between the noisy F-PET image and the original F-PET image:

[0019]

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

[0021]

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

[0023] The U-Net discriminator includes a downsampling network and an upsampling network. Through downsampling, the network increases the number of channels of the feature map while reducing the spatial resolution, thereby capturing the high-level abstract features of the image. The goal of downsampling is to gradually reduce the spatial size of the image and enhance the receptive field of the network to capture global information; through upsampling, the network gradually restores the spatial resolution of the image and reconstructs the details of the image as accurately as possible. The goal of upsampling is to gradually expand the spatial size of the feature map to generate an output image of the same size as the input image. The downsampling network and the upsampling network are connected by skip-connection and bottleneck connection.

[0024] Furthermore, the U-Net discriminator performs discrimination on a per-pixel basis, and the encoder loss is:

[0025]

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

[0027]

[0028] [D dec (x)] i,j and [D dec (G(z))] i,j They represent the judgment of the discriminator at pixel (i, j); D dec The per-pixel output of is obtained by skipping connections from the middle layers 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 layers;

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

[0030]

[0031] Encouraging the generator to focus on synthesizing images by effectively capturing global structures and local details, aiming to deceive the discriminator more effectively;

[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 output of the decoder and encoder respectively. The intermediate loss is used to describe the loss of the U-Net discriminator encoder, 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 a diffusion multi-scale generative adversarial network is used to implement any of the methods described above.

[0040] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements any of the methods described above.

[0041] An electronic device, comprising:

[0042] one or more processors;

[0043] A memory for storing one or more programs;

[0044] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the methods described above.

[0045] Compared with the prior art, the present invention has the following specific beneficial effects:

[0046] The present invention realizes high-quality reconstruction of L-PET images by designing an end-to-end neural network architecture, which is a novel diffusion multi-scale generative adversarial network DMGAN, which can effectively capture and restore image details, so that the generated image is close to the real full-dose PET image in both visual and structural aspects. The original low-dose PET image slices are input into the diffusion generator to synthesize the full-dose image. The diffusion generator also contains a series of corresponding target slices. In the diffusion generator, a noisy L-PET image is generated by introducing noise into the input low-dose PET image. Then, the diffusion generator generates a full-dose PET (F-PET) image and a noisy F-PET image. The generated image is input into the U-Net discriminator, which extracts details from the overall and specific perspectives to enhance the quality of the generated F-PET image. Through this process, our method is able to generate high-quality F-PET images, ensuring its practicality and reliability in clinical applications. The diffusion generator makes the generated image more realistic by gradually adding and removing noise, while the U-Net discriminator further improves the details and clarity of the image through multi-scale feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[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 compared visually by color visualization. The color spectrum ranges from blue to red, indicating metabolic values ​​increasing from low to high.

[0050] Figure 4 Comparison of pseudo-color variance maps between images generated by existing methods and the proposed method and the original F-PET image. The colors transition from blue to red, indicating differences ranging from slight to significant.

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

[0052] Figure 6 The peak signal-to-noise ratio for each method is shown.

[0053] Figure 7 Displays the structural similarity index of various methods

[0054] Figure 8 Bland-Altman 2D dot plot of brain SUVR for all reconstructed F-PET images based on different GAN models. In the AAL brain template, each blue dot represents its corresponding brain region. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0056] In an embodiment of the present invention, the experimental data set includes brain PET scan data of 45 children with epilepsy, aiming to improve the diagnostic accuracy through improved image processing methods, wherein low-dose brain PET scans and corresponding full-dose brain PET scans are collected.

[0057] In order to generate high-quality F-PET images from L-PET images, the present invention proposes a novel diffusion multi-scale generative adversarial network (DMGAN). The design of DMGAN aims to effectively capture and restore image details, so that the generated images are close to the real full-dose PET images in both visual and structural aspects. In the embodiments of the present invention, the performance of DMGAN is verified through a series of qualitative and quantitative comparisons.

[0058] In the quantitative evaluation, the present invention uses two main indicators: structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR). SSIM is used to measure the similarity of two images in structural information, while PSNR is used to evaluate the reconstruction quality and noise level of the image. Through these indicators, the present invention can objectively compare the differences in reconstructed image quality between DMGAN and other methods, ensuring that the method of the present invention has obvious advantages in generating high-quality F-PET images.

[0059] It can be seen that the research of the present invention not only proposed a new image reconstruction method, but also verified its effectiveness through a strict evaluation system, providing important technical support for the clinical diagnosis of brain diseases such as epilepsy.

[0060] Example

[0061] A three-dimensional PET image reconstruction method based on a diffusion multi-scale generative adversarial network, comprising:

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

[0063] Step 2: Before reconstruction, both L-PET and F-PET images need to be preprocessed using statistical parameter mapping for realignment and standardization. After preprocessing, the voxel size of these PET images is standardized to 1×1×1mm. By processing the three-dimensional brain images in the PET data set, a standardized data set is obtained as the basis for the experimental analysis of the present invention.

[0064] Step 3: Design an end-to-end neural network architecture, which mainly includes two core modules: 3D Fusion Generator module and 3D U-net Discriminator module, which are used to reconstruct low-dose 3D PET images.

[0065] Step 4. Among these two types of modules, the role of the 3D Diffusion Generator module is as follows: the original low-dose PET image is input into the 3D Diffusion Generator module to synthesize the full-dose image, namely the F-PET image. At the same time, the 3D Diffusion Generator module can simultaneously generate F-PET images and noisy F-PET images to improve the discrimination ability of the discriminator.

[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, which can extract details from the overall and specific angles and discriminate the images obtained by the generator to improve the quality of the generated F-PET images.

[0067] Step 6: In order to evaluate the effectiveness of the proposed method, the evaluation performed by the present invention includes qualitative evaluation and quantitative evaluation. In terms of quantitative comparison, the present invention uses two indicators, structural similarity index measurement (SSIM) and peak signal-to-noise ratio (PSNR), to evaluate the performance of different methods.

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

[0069] For PET images in clinical situations, the goal of the present invention is to generate images that are very similar to the original F-PET images from L-PET images. The advantage of the method proposed in the present invention is to generate output samples that use the identified distribution of the actual image and fully exploit its variability. First, the original low-dose PET image slices are input into the diffusion generator to synthesize the full-dose image, which also contains a series of corresponding target slices. Then, the diffusion generator generates F-PET images and noisy F-PET images. The generated images are input into the U-Net discriminator to extract details from the overall and specific angles to improve the quality of the generated F-PET images.

[0070] In step 4, the original low-dose PET image is input into the 3D Diffusion Generator module to synthesize the full-dose image. At the same time, the 3D Diffusion Generator module can simultaneously generate F-PET images and noisy F-PET images to improve the discrimination ability of the discriminator.

[0071] Specifically, a new diffusion generator is designed in step 4. The biggest advantage of this model is that it can obtain different information from different levels, thereby improving the generator's ability to obtain information from the original image. The diffusion generator provides a new perspective on learning the distribution of original images 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. The present invention uses Charbonnier Loss to penalize the Euclidean difference between the generated F-PET image and the original F-PET image:

[0073]

[0074] At the same time, the Euclidean difference between the generated noisy F-PET image and the original F-PET image is penalized:

[0075]

[0076] Specifically, considering the perceptual differences between the generated F-PET images, the generated noisy F-PET images, and the original F-PET images, the present 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. According to the above loss function, 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. The discriminator module can extract information from both the overall and specific details to identify the images generated by the generator, thereby improving the quality of the generated F-PET images.

[0082] Specifically, the present invention introduces a u-net discriminator in the model to obtain the original image information from the global and local views, which can better cooperate with the diffusion generator. The u-net discriminator consists of the original downsampling network and the new upsampling network, which are connected by skip-connection and bottleneck. Compared with the original network, the u-net discriminator performs discrimination on a per-pixel basis. The loss of the encoder is: L Denc =-E x [logD enc (x)]-E z [log(1-D enc (G(z)))]

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

[0084]

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

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

[0087]

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

[0089] The loss of the base 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 the present invention, the u-net diffusion discriminator returns two values, representing the output of the decoder and the encoder respectively. The intermediate loss is used to describe the loss of the u-net discriminator encoder, 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 A structural diagram of the DMGAN neural network architecture designed in the step 3, including a diffusion generator module and a u-net discriminator module. In order to verify the performance of the proposed method, the present invention compares it with cGAN, CycleGAN and transGAN. And in order to eliminate the influence of different initialization parameters on the experiment, the present invention sets the same random seed for all experimental methods. The present invention uses the Pytorch library on the NVIDIA RTX 4090 GPU. The batch size is 1 and the epoch is set to 300. Finally, the present invention evaluates it through qualitative evaluation and quantitative measurement. In order to make a quantitative comparison, the present invention selects two indicators to evaluate the performance of various methods: SSIM and PSNR.

[0097] like Figure 3 As shown, the present invention can be achieved by Figure 3The metabolic diagrams of various methods in the present invention can be observed more clearly. The method of the present invention has more metabolic details than other methods.

[0098] like Figure 4 As shown, a more intuitive comparison is Figure 4 The generated F-PET image and the original F-PET image are calculated in the pseudo-color difference map, indicating that the proposed method exhibits minimal voxel-scale differences compared to other methods.

[0099] like Figure 5 As shown, in view of the clinical importance of accurate low-dose reconstruction of epileptic lesions in pediatric patients, the present invention shows epileptic lesion L-PET images, synthetic epileptic lesion F-PET images and ground truth epileptic lesion F-PET images in Figure 5. Obviously, compared with other methods, the synthetic epileptic lesion F-PET images generated by DMGAN show more accurate voxel metabolic intensity distribution, thereby depicting the epileptic lesions more clearly.

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

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

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

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

[0104] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented 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] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0106] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions 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 in the protection scope of the present invention.

Claims

1. A three-dimensional PET image reconstruction method based on a diffusion multi-scale generative adversarial network, characterized in that: include: The PET image to be reconstructed is input into the trained diffusion multi-scale generative adversarial network DMGAN to generate a full-dose PET image close to the real one; 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.

2. The three-dimensional PET image reconstruction method based on a diffusion multi-scale generative adversarial network according to claim 1, characterized in that: Before reconstruction, both the training dataset L-PET and F-PET images were preprocessed using statistical parametric mapping for realignment and normalization.

3. The three-dimensional PET image reconstruction method based on a diffusion multi-scale generative adversarial network according to claim 1, characterized in that: The loss function in the diffusion generator is: L GAN (D,G)=-E x,y [(D(x,y)-1) 2 ]-E x [D(x,G(x)) 2 ] 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: At the same time, the Charbonnier Loss penalty is introduced to generate the Euclidean difference between the noisy F-PET image and the original F-PET image: In addition, considering the perceptual differences between the generated F-PET images, the generated noisy F-PET images, and the original F-PET images, VGG16-Net trained on ImageNet is used to extract the feature representations of G(x) and y, and the perceptual loss is: Where V represents the feature map extracted by VGG16-Net.

4. The three-dimensional PET image reconstruction method based on a diffusion multi-scale generative adversarial network according to claim 1, characterized in that: The U-Net discriminator includes a downsampling network and an upsampling network, which are connected by a skip connection and a bottleneck connection.

5. The three-dimensional PET image reconstruction method based on a diffusion multi-scale generative adversarial network according to claim 4, characterized in that: The U-Net discriminator performs discrimination on a per-pixel basis z, and the encoder loss is: The decoder loss of the average of all pixel values ​​is: [D dec (x)] i,j and [D dec (G(z))] i,j They represent the judgment of the discriminator at pixel (i, j); D dec The per-pixel output of is obtained by skipping connections from the middle layers 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 layers; Considering the above loss function, the goal of the generator is: Encouraging the generator to focus on synthesizing images by effectively capturing global structures and local details, aiming to deceive the discriminator more effectively; The basic loss of the discriminator is: L adv =E x,y [logD dec (x,y)]+E x [log(1-D dec (G(x),x))] The U-Net discriminator returns two values, representing the output of the decoder and encoder respectively. The intermediate loss is used to describe the loss of the U-Net discriminator encoder, expressed as: L middle =-log(1-D encfake )-log(D encreal ) The total loss of the discriminator is: L d =(L adv +L middle )*0.5ν Where v is a hyperparameter.

6. A three-dimensional PET image reconstruction system based on a diffusion multi-scale generative adversarial network, characterized in that: Used to implement the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. An electronic device, characterized in that: The device comprises: one or more processors; A memory for storing 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 according to any one of claims 1 to 5.

Citation Information

Patent Citations

  • CGANs-based PET image direct reconstruction method

    CN110717951A

  • Medical image noise reduction method and system, terminal and storage medium

    CN112488953A

  • PET image generation method fusing system response matrix and neural network

    CN116739933A

  • Method for reconstructing CT (Computed Tomography) image in double-view cross-modal manner

    CN117392259A

  • PET (positron emission tomography) iterative reconstruction method and system based on diffusion model

    CN117437152A

Cited By

  • PET double-tracer concentration diagram generation method based on diffusion adversarial model

    CN119417930A

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

    CN119417930B

  • Low-count PET image quality enhancement method and system based on fusion multi-input cyclic consistent generative adversarial network

    CN120807320A