A low-count PET image quality enhancement method and system based on a fusion multi-input cyclic consistent generative adversarial network

Abstract

The application provides a low-count PET image quality enhancement method and system based on a fusion multi-input cycle consistent generative adversarial network. The method comprises the following steps: obtaining a PET image set, including a plurality of training sample pairs, each training sample pair including an SCPET image sample and a corresponding LCPET image sample; constructing a fusion multi-input cycle consistent generative adversarial network, including: expanding the input of the generator and the discriminator in the LCPET image domain into a plurality of input channels, expanding the output of the generator in the SCPET image domain into a plurality of output channels, one input channel or output channel corresponding to one PET image, and the count ratios of the plurality of PET images being different; designing a total loss function, including a perception loss, an adversarial loss and a cycle consistency loss; based on the total loss function, training the network using the image set to obtain a trained image quality enhancement model; and inputting a low-count PET image to be enhanced into the trained generator in the LCPET image domain to obtain an enhanced PET image.

Description

Technical Field

[0001] This invention relates to the field of PET image processing technology, and in particular to a method and system for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network. Background Technology

[0002] Positron emission tomography (PET) is an advanced molecular imaging technique with unique clinical applications in oncology, cardiology, and neurology. PET imaging is based on coincidence detection technology, and its image quality is affected by inherent physical factors, detector structure and performance, and the coincidence photon count. In clinical practice, PET image quality primarily depends on the coincidence photon count, which is positively correlated with the radioactive tracer dose and acquisition time. Higher coincidence photon counts in clinical practice mean higher ionizing radiation doses, tracer costs, and longer scan times, thus limiting the scope and efficiency of PET's clinical use. Low-count (LC) PET corresponds to lower radioactive tracer doses and faster PET scans in clinical practice. Lower radioactive tracer doses reduce radiation doses to both patients and staff, lower tracer costs, and faster PET scans reduce motion artifacts caused by patient movement, while also improving PET efficiency. However, both methods suffer from reduced signal-to-noise ratios, image blurring, and loss of detail due to the lower coincidence photon count, making them less suitable for clinical diagnostic needs. Therefore, improving the image quality of low-count PET (LCPET) imaging to obtain diagnostic information comparable to that of standard-count (SC) PET imaging, which is currently used in clinical practice, thereby reducing the radiation dose of PET examinations and improving PET efficiency, is of great significance and practical value in the clinical application of PET.

[0003] Current research can be divided into two directions: The first direction is to integrate improved methods into the PET image reconstruction process, such as direct reconstruction methods that use end-to-end networks to map raw PET data to PET images, and deep learning reconstruction (DLR) that combines deep learning with Bayesian reconstruction methods. This direction is computationally intensive and often requires large datasets. The second direction is to directly apply improved methods to the reconstructed images, which belongs to image post-processing methods. Related research can be mainly divided into two categories: traditional methods and deep learning methods.

[0004] Traditional methods such as nonlinear Gaussian filtering, nonlocal mean (NLM), and wavelet transform suffer from limitations. LCPET images contain more complex spatial variations and variable statistical noise than SCPET images, making it difficult to achieve ideal outputs through hyperparameter settings. This results in the loss of much geometric and texture information in the generated images, thus limiting the effectiveness of traditional methods.

[0005] In recent years, deep learning methods have demonstrated significant advantages and enormous application potential in medical image generation. Deep learning techniques have been shown to better remove noise while predicting texture details in PET images and preserving lesion information as much as possible. Kaplan et al. divided PET images by body part and trained a residual convolutional neural network to estimate full-dose PET images from 1 / 10-dose PET images, conducting high-quality PET image prediction and generation studies for different body regions (brain, chest, abdomen, and pelvis). Research by Wang, Nie, and others has shown that Generative Adversarial Networks (GANs) have superior performance in medical image denoising and generative modeling compared to standalone Convolutional Neural Networks (CNNs). However, because both the input and output data of the LCPET to SCPET training model contain noise, it is difficult to ensure that the GAN generator learns meaningful mappings that result in a unique output for a given input. The generator may create non-existent features or collapse to a narrow distribution. Therefore, this invention adopts a Cycle-Consistent Generative Adversarial Network (CycleGAN) architecture based on GAN. By introducing inverse transformations in a cyclic manner, more constraints are added to the generator, effectively preventing model collapse and ensuring that the generator finds a unique and meaningful mapping. Experimental results also show that CycleGAN outperforms GAN for LCPET image quality enhancement tasks. Lei et al. used the CycleGAN model to predict full-dose PET images from 1 / 8 dose PET images, and the model was trained and tested using 25 and 10 whole-body PET images, respectively. Amirhossein et al. used data from 100 patients to compare the performance of CycleGAN and Residual Neural Network (ResNET) in generating full-dose PET images from low-dose whole-body PET, and the results showed that CycleGAN has better qualitative and quantitative indicators. Their research also demonstrates the advantages of the CycleGAN network in PET image quality enhancement. In 2020, Lei et al. further developed a CT-assisted CycleGAN network for denoising low-count PET images, improving the accuracy and quality of the reconstructed images. However, the training and use of this model require structural information provided by CT images and cannot be used for PET single-modal images. Furthermore, in 2021, Xue et al. proposed a CycleGAN network incorporating domain transform (LCPR-Net), using PET image data from 30 patients to directly reconstruct high-quality full-count PET images from LC sinusoidal PET data.However, this study simulates sinogram data with different noise levels by forward projection of the system matrix, introducing Poisson noise, and normalization, which differs from real clinical data. Furthermore, the LCPR-Net network uses least squares loss, which makes training more stable than cross-entropy loss, but still faces the problem of gradient vanishing or saturation. For PET image generation, a better network structure and loss function are still needed to optimize the training process and image generation quality.

[0006] Despite the progress made in CycleGAN-based low-count PET image quality enhancement methods, the following problems and limitations remain: First, insufficient model generalization ability: small sample training and site specificity limit the model's versatility and clinical application scope. Previous studies typically used small sample sizes, reducing model robustness and affecting generalization, especially for abnormal cases. Many models are trained separately for different body parts, while clinical practice usually involves scanning the body as a whole; if a model is only effective for one site, it increases the complexity of model use. Some models require anatomical structures provided by CT or MR images as an aid and cannot be used for PET single-modal images, reducing versatility. Second, low detail fidelity; traditional architectures easily lose information about the edges of small lesions or metabolic heterogeneity. Third, previous studies have focused primarily on quantitative indicators of image quality in model evaluation, with insufficient attention to clinical needs such as lesion detectability, lacking multi-dimensional validation of standard uptake value (SUV) quantitative error and lesion detectability. Most studies only evaluated general quantitative indicators of the overall image, such as normalized root mean square error (NRMSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). They did not comprehensively evaluate indicators crucial for clinical practice, such as lesion detectability and the SUV values ​​of lesion regions and typical structural regions like the hepatic and mediastinal blood pools. Fourth, the quantitative accuracy was insufficient, particularly the SUV value, which is of greater clinical concern, resulting in significant differences between the generated SCPET images and the actual SCPET images. Summary of the Invention

[0007] In view of the problems existing in the prior art, this invention proposes a method and system for enhancing the quality of low-count PET images based on MI-CycleWGAN, realizing the predictive generation of SCPET images from LCPET images. By optimizing the network structure and loss function, training with a large sample size of whole-body PET image data, and combining comprehensive qualitative and quantitative clinical evaluation methods, the method improves the poor PET image quality and clinical quantitative indicators after enhancement model processing in the prior art.

[0008] In a first aspect, the present invention provides a method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network, comprising:

[0009] Obtain a PET image set, which includes several training sample pairs. Each training sample pair includes SCPET image samples and three corresponding LCPET image samples with different counting ratios. SCPET is standard count PET, and LCPET is low count PET.

[0010] Constructing a multi-input cyclic consistent generative adversarial network includes: expanding the inputs of the generator and discriminator in the LCPET image domain to three input channels, expanding the output of the generator in the SCPET image domain to three output channels, with one input channel or output channel corresponding to one PET image, and the counting ratios of the three PET images corresponding to the three input channels or output channels being different;

[0011] Design a total loss function, which includes perceptual loss, adversarial loss, and periodic consistency loss; the perceptual loss includes the difference between the first generated image and the SCPET image sample in the feature space and the difference between the second generated image and the LCPET image sample in the feature space; the first generated image is the image reconstructed from the LCPET image sample by the generator in the LCPET image domain; the second generated image is the image reconstructed from the SCPET image sample by the generator in the SCPET image domain.

[0012] Based on the total loss function, the image set is used to train the fused multi-input cyclic consistent generative adversarial network to obtain a trained image quality enhancement model;

[0013] The low-count PET image to be enhanced is input into the generator of the trained LCPET image domain to obtain the enhanced PET image.

[0014] Furthermore, the generator adopts a U-Net structure and introduces residual blocks in the encoder and decoder, as well as uses an attention mechanism with adaptive instance normalization at the end of the decoder.

[0015] Furthermore, the discriminator includes three sub-discriminators arranged in parallel, each of which adopts a patchGAN structure and uses spectral normalization; correspondingly, for the same generated image, it is downsampled to three different resolutions, and the generated images at the three different resolutions are respectively input into the three sub-discriminators.

[0016] Furthermore, the total loss function L is:

[0017] L=λ cycle L cycle +Ladv +λ pept L pept

[0018]

[0019] Among them, L cycle For periodic consistency loss; L adv To combat the loss, a Wasserstein distance formula with a gradient penalty term is used; L pept To perceive loss, L pept1 L represents the difference between the first generated image and the SCPET image samples in the feature space. pept2 To represent the difference between the second generated image and the LCPET image samples in the feature space, L pept =L pept1 +L pept2 ;λ cycle and λ pept The corresponding weights are N, the batch size is N, φ is the given feature extractor, and G(x) = N(b). i ) and y i These are the generated image and the real image, respectively. A and B are defined as the LCPET image domain and the SCPET image domain, respectively. X A X represents multiple original LCPET images with different counting ratios. B This represents the corresponding original SCPET image, generator G. A Represents the mapping from A to B, generator G B This represents the mapping from B to A.

[0020] Furthermore, the training process also includes: clinical evaluation of the enhanced PET images, quantitative index analysis, and key anatomical region analysis to optimize model parameters; wherein the key anatomical regions include lesion areas and normal tissue areas.

[0021] Furthermore, the clinical evaluation specifically includes: visual assessment of image quality using a five-point Likert scale; and binary evaluation of lesion detectability.

[0022] Furthermore, the quantitative index analysis specifically includes: quantitatively evaluating the image as a whole based on normalized root mean square error, structural similarity index, and peak signal-to-noise ratio; analyzing the joint histogram of the generated image and the real SCPET image; and comparing and evaluating the results by combining linear regression parameters.

[0023] Furthermore, the key anatomical region analysis specifically includes: drawing ROIs in key anatomical regions; calculating the average and maximum standard uptake values ​​for each ROI based on the quantitative accuracy requirements of standard uptake values ​​in clinical diagnosis; and using SCPET images as a reference standard to evaluate the consistency of standard uptake values ​​using Bland-Altman plots.

[0024] Secondly, the present invention provides a low-count PET image quality enhancement system based on a fused multi-input cyclic consistent generative adversarial network, comprising:

[0025] The data acquisition and preprocessing module is used to acquire a PET image set, which includes several training sample pairs. Each training sample pair includes SCPET image samples and three corresponding LCPET image samples with different count ratios. SCPET is standard count PET, and LCPET is low count PET.

[0026] A network model construction module is used to construct a fused multi-input cyclic consistent generative adversarial network (CGI). This includes: expanding the inputs of the generator and discriminator in the LCPET image domain to three input channels, and expanding the output of the generator in the SCPET image domain to three output channels. Each input or output channel corresponds to one PET image, and the counting ratios of the three PET images corresponding to the three input or output channels are different. The module also designs a total loss function, which includes perceptual loss, adversarial loss, and cyclic consistency loss. The perceptual loss refers to the difference in feature space between the first generated image and the SCPET image sample, and the difference in feature space between the second generated image and the LCPET image sample. The first generated image is the image reconstructed from the LCPET image sample by the generator in the LCPET image domain; the second generated image is the image reconstructed from the SCPET image sample by the generator in the SCPET image domain.

[0027] The training optimization module is used to train the fused multi-input cyclic consistent generative adversarial network based on the total loss function and the image set, so as to obtain a trained image quality enhancement model.

[0028] The image quality enhancement module is used to input the low-count PET image to be enhanced into the generator of the trained LCPET image domain to obtain the enhanced PET image.

[0029] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in the first aspect.

[0030] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0031] The beneficial effects of this invention are as follows:

[0032] 1. This invention redesigns and optimizes the CycleGAN network (new generator structure, introduction of attention mechanism, residual blocks, perceptual loss, and multiple sub-discriminators, etc.) to achieve superior network model performance. This allows processed low-count PET images to more closely resemble standard-count images while better preserving detail. Compared to previous studies, low-count PET images enhanced using this invention exhibit better global metrics such as PSNR and SSIM. Simultaneously, it demonstrates comparable detection performance to standard-count images in lesion detectability, and the SUV values ​​for lesion areas and typical anatomical regions show high accuracy. Using this method, the problem of decreased PET image quality affecting clinical diagnosis due to reduced tracer dosage, shortened acquisition time, or equipment limitations in current clinical PET examinations can be effectively addressed. This significantly reduces radiation dose to patients and staff during PET examinations, lowers the environmental and economic costs associated with radioactive tracers, and improves the efficiency of PET examinations.

[0033] 2. Supervised learning method, more efficient

[0034] Although the CycleGAN network was initially designed for unsupervised learning tasks, this invention employs paired supervised learning for training, taking into account the characteristics of PET images and the need for image accuracy in clinical diagnosis. This maintains a fixed number of pixel values, removes large geometric mismatches, allows the network to focus on mapping details, and accelerates training, thus improving both training speed and accuracy. Low-count images are obtained through retrospective reconstruction with shortened acquisition time, eliminating the need for additional scans. The paired supervised learning mode is only involved in the model training phase. After model training is complete, subsequent users only need to provide low-count PET images; paired data is not required, thus not increasing the complexity of model usage.

[0035] 3. Deep learning image enhancement technology is more feasible.

[0036] Unlike the Deep Learning Image Reconstruction (DLR) method, which directly applies deep learning to the PET image reconstruction process, this invention performs post-processing on the reconstructed images from the PET device. It belongs to the Deep Learning Image Enhancement (DLE) technology. Compared with the DLR method, it requires less computing resources and can be trained and deployed directly in the back end of the current clinical workflow, rather than requiring reconstruction algorithm reconstruction by the device manufacturer, making it more clinically feasible.

[0037] 4. Through comparative experiments, the shortest scan time and lowest count ratio per clinical bed were determined. This provides a useful reference for the minimum level of radioactive tracer activity that can be used clinically and the shortest achievable acquisition time per bed. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the low-count PET image quality enhancement method based on a fused multi-input cyclic consistent generative adversarial network according to the present invention.

[0039] Figure 2 This is a schematic diagram of the MI-CycleWGAN network model structure used in this invention;

[0040] Figure 3 This is a schematic diagram of the MI-CycleWGAN network generator structure used in this invention;

[0041] Figure 4 This is a schematic diagram of the MI-CycleWGAN network discriminator structure used in this invention;

[0042] Figure 5 This is a schematic diagram of the clinically-oriented comprehensive evaluation protocol of the present invention;

[0043] Figure 6 (a) is a real standard count PET image (corresponding to an acquisition time of 180s), also known as an SCPET image and corresponding lesion; (b) is a PET image generated after processing with the MI-CycleWGAN network of this invention, also known as an MI-CycleWGAN generated image; (c) is a raw low count PET image (corresponding to an acquisition time of 30s), also known as an LCPET image and corresponding lesion.

[0044] Figure 7 (a) is the joint histogram of LCPET and SCPET images; (b) is the joint histogram of MI-CycleWGAN generated image and SCPET image.

[0045] Figure 8 Bland-Altman plots of SUVmean (first row) for typical anatomical regions and SUVmax (second row) for lesion regions. Calculated using SCPET images as the standard, for LCPET (left) and MI-CycleWGAN generated images (right). Solid red lines and dashed blue lines represent the mean and 95% confidence interval of the SUV difference, respectively.

[0046] Figure 9 This is a schematic diagram of the structure of a low-count PET image quality enhancement system based on a fused multi-input cyclic consistent generative adversarial network according to the present invention;

[0047] Figure 10 A structural block diagram of an electronic device provided for the implementation of this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0049] This invention proposes a method and system for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network (MI-CycleWGAN). The core of this method is to establish a nonlinear end-to-end mapping model through deep learning technology to map low-count PET (LCPET) images to the standard-count PET (SCPET) image domain, thereby achieving a significant improvement in the quality of LCPET images.

[0050] This invention significantly enhances the feature representation and image reconstruction performance of the CycleGAN model through multiple improvements and a redesign of the generator network. The technical solution of this invention comprises six main parts: data acquisition and preprocessing, network architecture design, loss function optimization, model training strategy, image enhancement generation, and comprehensive evaluation with clinical guidance.

[0051] Combination Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network, specifically including the following steps:

[0052] S101: Data acquisition and preprocessing. This includes: acquiring a PET image set, which includes several training sample pairs, each training sample pair including SCPET image samples and corresponding LCPET image samples; SCPET is standard count PET, and LCPET is low count PET.

[0053] S102: Construct a fusion multi-input cyclic consistent generative adversarial network, abbreviated as MI-CycleWGAN.

[0054] Specifically, in the MI-CycleWGAN constructed in this embodiment of the invention, the inputs of the generator and discriminator in the LCPET image domain are expanded to three input channels, and the output of the generator in the SCPET image domain is expanded to three output channels. One input channel or one output channel corresponds to one PET image, and the counting ratios of the three PET images corresponding to the three input channels or one output channel are different.

[0055] Specifically, the original CycleGAN architecture typically uses a single channel as input, but this input format has limitations in terms of information dimensionality when processing PET medical images. To capture the changing process of images with different count ratios, obtain more intermediate state information, and thus more fully utilize the detailed features in the images, this embodiment changes the network's input format from a single channel to a three-channel input. The multiple channels represent PET images with different count ratios.

[0056] The raw data of PET images is stored in list mode as "consensus photon counts." Clinically, the concordance photon count is positively correlated with tracer activity and acquisition time. This invention uses a retrospective reconstruction method with shortened acquisition time to obtain low-count PET images. Since the acquisition time for a standard count image is 180 seconds, the corresponding count ratio for a PET image with an acquisition time shortened to 15 seconds is 1 / 12, for a PET image with an acquisition time shortened to 20 seconds is 1 / 9, for a PET image with an acquisition time shortened to 30 seconds is 1 / 6, and so on. In short, PET images with different count ratios represent images obtained by reducing the concordance photon count value of the raw PET data to different ratios of the standard count image.

[0057] This multi-channel fusion strategy not only enriches the input feature dimensions of the model, but also helps to capture more complex spatial structures and texture details, thereby improving the model's ability to identify and recover lesion areas in low-count PET images, and providing a more solid data foundation for the subsequent generation of more biologically consistent images.

[0058] S103: Design the total loss function, which includes perceptual loss, adversarial loss, and periodic consistency loss; the perceptual loss includes the difference between the first generated image and the SCPET image sample in the feature space and the difference between the second generated image and the LCPET image sample in the feature space; the first generated image is the image reconstructed from the LCPET image sample by the generator in the LCPET image domain; the second generated image is the image reconstructed from the SCPET image sample by the generator in the SCPET image domain;

[0059] Specifically, due to the changes in network structure, an adaptive instance normalization (AdaIN) is further introduced into the generator to replace the ontology mapping loss.

[0060] S104: Based on the total loss function, the image set is used to train the fused multi-input cyclic consistent generative adversarial network to obtain a trained image quality enhancement model;

[0061] Specifically, LCPET and SCPET image samples are input into the MI-CycleWGAN network for supervised training. Through a series of ablation experiments, the optimal number of residual blocks and downsampling times are obtained.

[0062] S105: Input the low-count PET image to be enhanced into the generator of the trained LCPET image domain to obtain the enhanced PET image.

[0063] Furthermore, in one embodiment, this embodiment also provides a clinically-oriented comprehensive evaluation method. Correspondingly, the training process also includes: optimizing model parameters in conjunction with this comprehensive evaluation method.

[0064] Specifically, the comprehensive evaluation method includes clinical evaluation, quantitative indicator analysis, and key anatomical region analysis; the key anatomical regions include lesion areas and normal tissue areas.

[0065] In one embodiment, the data acquisition and preprocessing process is as follows: A large volume of whole-body images of the patient, acquired by a fully digital PET / CT system, is collected. 18 F-FDG PET images, excluding those without obvious lesions, were used to form a PET image set encompassing various diseases. Low-count PET images, denoted as low-count PET sample images (LCPET), were obtained using a retrospective reconstruction method with shortened acquisition times. The acquisition times were 15 seconds (1 / 12 count), 20 seconds (1 / 9 count), 30 seconds (1 / 6 count), 40 seconds (2 / 9 count), and 60 seconds (1 / 3 count), with reconstruction parameters identical to those for standard acquisition time images (SCPET). After retrospective reconstruction, over 200,000 pairs of low-count and standard-count PET images were obtained.

[0066] Each standard count PET sample image was paired with simulated low-count PET sample images taken at 15 seconds (1 / 12 count), 20 seconds (1 / 9 count), and 30 seconds (1 / 6 count) to form a dataset. After anonymizing the original data, the PET image size was adjusted from 288×288 to 256×256, and then the .dcm format PET cross-sectional images were converted to .npy format.

[0067] The network training adopted a supervised training method, arranging the required data sequentially according to the cross-sectional order, accumulating data case by case, and ensuring that the LCPET data corresponded to the SCPET data in the same order and used the same naming format. The dataset was divided into training set, validation set, and test set in an 8:1:1 ratio.

[0068] It should be noted that the quality of PET images in clinical practice mainly depends on the coincidence photon count, which is positively correlated with the radioactive tracer dose and acquisition time. Low-count PET (LCPET) in clinical practice corresponds to low tracer dose and short scan time. Therefore, this invention can be used for quality enhancement and evaluation of low-count PET images and low-quality PET images in all situations, including low-dose PET images, fast-scan PET images, and low-quality PET images caused by equipment limitations.

[0069] In one embodiment, the network architecture of MI-CycleWGAN is as follows: Figure 2 As shown. The MI-CycleWGAN provided in this embodiment is based on the CycleGAN architecture, and the generator, discriminator and loss function in CycleGAN have been improved. It can complete the many-to-one generation task, that is, it can generate an SCPET image based on multiple LCPET images with different counting ratios.

[0070] Specifically, Figure 2 In this context, A and B are defined as the LCPET image domain and SCPET image domain, respectively; generator G... A Represents a mapping from A to B; generator G B This represents the mapping from B to A. Two discriminators D. A and D B Used to determine whether the input image comes from an image generated by the generator or a real image; specifically, the discriminator D A Used for generator G B Multiple reconstructed LCPET images and multiple original LCPET images are distinguished; discriminator D B Used for generator G A The reconstructed SCPET image was compared with the original SCPET image.

[0071] In this embodiment, since the generator adopts a multi-input mode, there is no longer an ontology mapping loss (L) in the MI-CycleWGAN network. identity Furthermore, this embodiment introduces a perceptual loss based on the CycleGAN network to enable the generator to retain more image detail information during the process of recovering low-count PET images to standard-count PET images. This invention incorporates a perceptual loss (L... peptTherefore, the loss function of the MI-CycleWGAN network includes adversarial loss (L...). adv ), periodic consistency loss (L) cycle ) and perceived loss (L pept The formula for calculating the total loss function is as follows:

[0072] Therefore, the total network loss is defined as in formula (1):

[0073] L=λ cycle L cycle +L adv +λ pept L pept (1)

[0074] by Figure 2 Taking the upper part of the text as an example, let X be set. A X represents multiple original LCPET images with different counting ratios. B This represents the corresponding original SCPET image. The calculation process for each loss is as follows:

[0075] (1) Perceived loss serves as a guiding signal for the generator. X A After generator G A Get G A (X A ), i.e., the first generated image; calculate G A (X A ) and X B Perceptual loss L between pept1 Similarly, X B After generator G B Get G B (X B ), that is, the second generated image; calculate G B (X B ) and X A Perceptual loss L between pept2 ;

[0076] In this embodiment, the VGG-19 network is used as the feature extractor. The VGG-19 contains 16 convolutional layers and 3 fully connected layers. Feature maps are extracted from the output of the 16th convolutional layer. The difference between the generated image and the target image in the feature space is calculated. The perceptual loss serves as a guiding signal for the generator, prompting the generator to output an image that is closer to the target image. The calculation of the perceptual loss is shown in formula (2):

[0077]

[0078] Where N is the batch size, φ is the pre-trained VGG-19 network, and G(x) i ) and y iThese are the generated image and the real image, respectively.

[0079] It should be noted that the perceptual loss can also be achieved through other pre-trained models, such as using ResNet, Inception, and DenseNet as feature extractors to extract feature maps of generated and real images respectively, and then calculate the perceptual loss according to the above formula (2).

[0080] (2)X B After generator G B Get G B (X B ), calculate G B (X B ) and X A The counter-loss L adv1 Similarly, X A After generator G A Get G A (X A ), calculate G A (X A ) and X B The counter-loss L adv2 ;

[0081] In this embodiment, Wasserstein distance is used instead of JS divergence in the original GAN ​​network to alleviate the instability problem caused by JS divergence when comparing data distribution, thereby improving the convergence and generation quality of the network. The adversarial loss function adopts the Wasserstein distance formula with gradient penalty term, as shown in formula (3):

[0082]

[0083] Where x represents the distribution of the real data P r The sampled sample, z, represents the sampled sample from the prior noise distribution. The first two terms are Wasserstein's distance estimates, and the last term is the gradient penalty term for network regularization. λ is the weight hyperparameter of the gradient penalty term.

[0084] (3) The goal of the periodic consistency loss is to ensure that LDPET and SDPET can be converted to each other, guaranteeing that the input image, after being processed by the two generators, is as close as possible to the original image, thus ensuring the network maintains consistency and stability when performing image conversion tasks. The periodic consistency loss consists of two parts: forward periodic consistency loss L... cycle1 and backward periodic consistency loss L cycle2 X A After generator G A Then through generator G B Get G B (GA (X A )), calculate G B (G A (X A )) and X A loss L cycle1 ;X B After generator G B Then through generator G A Get G A (G B (X B )), calculate G A (G B (X B )) and X B loss L cycle2 The calculation formula is:

[0085]

[0086] In one embodiment, the two generators G of MI-CycleWGAN A and G B The network architecture is the same, such as Figure 3 As shown. To capture the changes in PET images with different counting ratios and thus obtain more intermediate state information for image generation, the generator structure is designed with multi-channel input. The basic architecture of the generator adopts U-Net, which generally includes a symmetrical encoder and decoder structure, and uses skip connections to pass features between corresponding layers to enhance detail preservation.

[0087] In this embodiment, the encoder's encoding process includes: inputting a three-channel LCPET image, with each channel having an input image size of 256×256×1; extracting preliminary features through a convolutional layer (which is also the first downsampling operation, such as...). Figure 3 (As shown by the downward green arrow in the image), this convolutional layer has a kernel size of 4, a stride of 2, and padding of 1; subsequently, the input image undergoes four linearly connected downsampling operations (such as...). Figure 3 (As shown by the downward red and purple arrows in the middle), and perform batch normalization, extract features and gradually reduce the size of the feature map. The number of channels of the feature map gradually increases during downsampling, specifically 64, 128, 256, 512 and 512. The output feature map is retained for use in skip connections in subsequent decoding paths. Figure 3 The number next to the feature map indicates the size of the feature map. Combined with the number of channels, we can see that during the downsampling process, the size (width × height × number of channels) of the feature map is 128×128×64, 64×64×128, 32×32×256, 16×16×512, and 8×8×512, respectively.

[0088] The decoding process of the decoder includes: the feature map output from the last layer of the encoder undergoes 6 layer-by-layer upsampling operations (e.g. Figure 3 (As shown in the diagram, five upward arrows and one adjacent left-pointing yellow arrow) the spatial dimensions of the feature map are gradually restored through transposed convolution. The output of each upsampling operation is normalized and ReLU activated, and then concatenated with the output of the corresponding downsampling operation in the encoder through skip connections, thereby preserving more local features.

[0089] Furthermore, after the upsampling phase, attention mechanism and AdaIN (adaptive instance normalization) are used to compress the channels to complete noise reduction.

[0090] Specifically, AdaIN is used instead of the ontology mapping loss (L identity The ontology mapping loss aims to limit the generator's autonomous generation. However, due to variations in the network structure, calculating the ontology mapping loss is not feasible. Therefore, AdaIN is introduced into the network to adjust the mean and standard deviation of the feature maps, making the feature distribution of low-count PET images closer to that of standard-count images. This not only reduces noise interference in low-count PET images but also adaptively adjusts image contrast and tissue details, thereby improving the quality and biological consistency of the generated images. Overall, by combining the channel attention mechanism and AdaIN, this model effectively enhances attention to key regions, improves the accuracy of image conversion, reduces artifacts in low-count PET images, and improves the quality and biological credibility of the generated images.

[0091] As one possible implementation, this embodiment introduces a Squeeze-and-Excitation (SE) attention mechanism into the tail structure of the generator to enhance the network's ability to model key features. The SE module performs global average pooling on the spatial features of each channel through a "squeeze" operation to extract channel-level global information, and then generates importance weights for each channel through an "excitation" operation. This mechanism adaptively enhances feature channels highly correlated with anatomical structures and lesion regions while suppressing irrelevant or redundant information, thereby improving the network's ability to identify key regions in low-count PET images. The introduction of the SE module not only significantly enhances the model's selective attention capability during feature extraction but also effectively improves the accuracy of detail restoration and structure preservation during image generation. Combined with the U-Net architecture and subsequent residual modules, the SE mechanism further strengthens the expressive power of the feature transfer path, providing the generator with a more robust and biologically consistent image construction capability. Experimental results show that after introducing the SE attention mechanism, the generated images exhibit significant improvements in structural similarity, detail preservation, and artifact suppression.

[0092] Furthermore, multiple residual blocks are introduced into the generator structure to enhance feature extraction capabilities. Each residual block consists of two convolutional layers and a ReLU activation function layer. Specifically, the last two downsampling layers of the encoder and the last two downsampling layers of the decoder are both set as residual blocks. The residual modules are integrated into the generator network, combined with a U-Net-based architecture, to preserve more image details and improve the network's feature representation capabilities.

[0093] Specifically, residual blocks effectively extract high-level features, enhancing the model's learning ability. With the support of skip connections, the decoder can effectively recover local features from low-count images and combine them with global features to denoise low-count PET images, obtaining the final denoised PET image. Simultaneously, AdaIN, combined with residual connections, preserves key anatomical information and enhances the stability and generalization ability of the generative adversarial network.

[0094] The generator provided in this invention introduces multi-channel input and, through the synergistic effect of the U-Net architecture, AdaIN normalization and attention mechanism, and residual module, effectively improves the model's ability to model noise distribution and image structure, significantly enhancing feature representation and image reconstruction performance. Experiments show that this generator structure improves the detail quality and subjective perception effect of the generated image while maintaining image structure consistency.

[0095] In one embodiment, the discriminator in the SCPET image domain takes either a generated SCPET image or a real SCPET image as input and determines whether it is genuine. Similarly, the discriminator in the LCPET image domain takes either a generated LCPET image or a real LCPET image as input and determines whether it is genuine.

[0096] In this embodiment, the discriminator includes three parallel sub-discriminators, each employing a patchGAN structure and spectral normalization. Correspondingly, for the same generated image, it is downsampled to three different resolutions, and the generated images at the three different resolutions are input into the three sub-discriminators respectively. The patchGAN structure is as follows: Figure 4 As shown, patchGAN is used to progressively reduce the spatial dimensionality of the image and extract high-dimensional features through multiple convolutional layers. The discriminator is designed as a multi-scale discriminator consisting of three sub-discriminators at different scales. Each sub-discriminator employs spectral normalization to stabilize the training process and improve its discriminative ability at different scales. By progressively downsampling the input image using average pooling, the sub-discriminators can extract multi-scale features, enabling the discriminator to comprehensively evaluate the generated image at different scales and capture more details.

[0097] like Figure 4As shown, the input is a single-channel image of 256×256×1. First, the input image is processed through a 4×4 convolutional layer (stride=2, padding=1), expanding the number of channels to 64, and using LeakyReLU(0.2) as the activation function. Then, it goes through a second 4×4 convolutional layer (stride=2, padding=1), increasing the number of channels to 128, and undergoes batch normalization, before applying LeakyReLU(0.2). A third 4×4 convolutional layer (stride=2, padding=1) continues to extract deeper features, reducing the feature map size to 32×32 and increasing the number of channels to 256. The fourth convolutional layer has a stride of 1 (stride=1, padding=1), making the feature map size 31×31×512, further enhancing feature extraction capabilities, and still uses BatchNormalization and LeakyReLU(0.2) for processing. Finally, after a 4×4 convolutional layer (stride=1, padding=1), a single-channel feature map (size 30×30×1) is output, which is used to predict the realism of local regions (patches). The entire network uses multi-scale feature extraction, enabling the discriminator to effectively distinguish the realism of generated images and enhancing training stability.

[0098] To verify the effectiveness of the present invention, the present invention also provides the following experimental data.

[0099] Experimental environment: The present invention was implemented on a computer using an NVIDIA GeForce RTX4090 GPU, based on the Python 3.12, PyTorch 2.3.0, and cuda12.1 platform. LCPET images with different counting ratios and their corresponding SCPET images from the training set were input into the MI-CycleWGAN model for supervised training.

[0100] First, the network structure was optimized through preliminary experiments. The model was validated using small-sample preliminary experiments, and a series of ablation experiments were conducted to evaluate the impact of introducing different downsampling times and different numbers of residual blocks on network performance. Experimental results showed that the generator architecture using 5 downsampling times performed optimally on image quantization metrics. Subsequently, a clinically oriented comprehensive evaluation method was used to further optimize the model parameters.

[0101] During the training phase, the Adam optimizer and the CosineAnnealingWarmRestarts learning rate scheduler were used. In the experiments, parameters T0 = 10 and Tmult = 2 were set; the learning rate underwent frequent decay in the initial period, and then the period length was gradually increased with each restart. This allowed the model to converge quickly in the early stages of training, while also maintaining the learning rate within a suitable range in the later stages through gradually increasing period lengths. This promoted better exploration of the loss space by the model in different training phases, achieving better generalization performance. The MI-CycleWGAN denoising model was trained using mini-batches consisting of 32 image patches. The gradient penalty term's weight hyperparameter λ was set to 10, and the perceptual loss weight λ... pept Set to 1.5, weight λ of the periodic consistency loss function cycle The hyperparameters were set to 4, and all were determined based on experimental results with a cycle of 60. The learning rate was chosen as 1e-4, and the two exponential decay rates for momentum estimation were β1 = 0.5 and β2 = 0.999, respectively.

[0102] For the trained MI-CycleWGAN model, its generator network G is used to convert LCPET images into SCPET images. A The LCPET images that require quality enhancement or are low-quality PET images caused by equipment or other reasons are input into the generator network G. A In this process, the corresponding SCPET image, i.e., a high-quality PET image, can be output.

[0103] This invention designs a comprehensive, clinically-oriented evaluation scheme. The purpose of LCPET image quality enhancement is to obtain images that meet clinical diagnostic needs while minimizing radiation dose and shortening scan time. Therefore, this invention integrates clinical practice into the model evaluation process, focusing on the most critical aspects of clinical diagnosis: lesion detectability and quantitative indicators. Unlike previous studies that often used only some common image quantification metrics, this invention provides a comprehensive qualitative and quantitative evaluation of the model-generated images from both lesion contrast and image quality perspectives, such as... Figure 5 As shown, it mainly includes the following steps:

[0104] First, a clinical evaluation was conducted. The image quality was visually assessed by multiple experienced diagnostic physicians using a five-point Likert scale, categorized into five groups: (1) unexplainable, (2) poor, (3) adequate, (4) good, and (5) excellent. The detectability of the lesion was then evaluated using a binary assessment: acceptable or unacceptable.

[0105] Next, quantitative index analysis was performed. This was conducted separately for the overall image and key anatomical regions to assess the differences and similarities between generated and real SCPET images in terms of tracer uptake. First, the overall image was quantitatively evaluated using the Normalized Root Mean Square Error (NRMSE), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR), as shown in the formulas below. Then, a joint histogram analysis of the generated and real SCPET images was performed to describe their voxel-level correlation in tracer uptake, and this was compared and evaluated using linear regression parameters. (See [link to relevant documentation]). Figure 7 The combined histogram shown is the fitted curve for the above data.

[0106]

[0107] Where, μ x and σ x To generate the mean and variance of image x, μ y and σ y Let σ be the mean and variance of the true image y. xy Let x be the covariance of images x and y. C1 and C2 represent stability constants introduced to prevent the denominator from being zero; their values ​​are related to the range of pixel values ​​and are generally small. MAX is the maximum possible pixel value in the image. MSE is the mean square error. PSNR is used to measure the similarity between images; a higher value indicates less distortion, and the unit is usually dB. ij To generate the grayscale value of image x at pixel i,j; y ij Let y be the grayscale value of the i-th and j-th pixels of the real image; N and M are the dimensions (rows and columns) of the image.

[0108] Next, key anatomical regions were evaluated, including lesion areas and clinically relevant normal tissue areas. Normal tissue areas included the brain, left and right lungs, liver (right lobe, 3cm diameter ROI, avoiding major blood vessels), thoracic aorta (1cm diameter ROI within the thoracic aorta, avoiding the vessel wall, representing the mediastinal blood pool), and abdominal aorta. ROIs were drawn within these key anatomical regions. Considering the clinical diagnostic requirements for the quantitative accuracy of standardized uptake values ​​(SUV values), the mean SUVmean and maximum SUVmax value for each ROI were calculated. Using SCPET images as a reference standard, Bland-Altman plots were used to evaluate the consistency of SUV values. Experimental results are as follows: Figure 6 , Figure 7 , Figure 8 As shown in the figure and Table 1. In Table 1, the prefix S represents SCPET, the prefix G represents the generated PET, and the prefix L represents LCPET.

[0109] Table 1

[0110]

[0111]

[0112] It should be noted that the current protocol selects some typical normal tissue areas that are of great clinical interest: the brain and left and right lungs, liver, mediastinal blood pool, and abdominal aorta. It is also possible to add representative anatomical areas, such as bone marrow, basal ganglia, and spleen, based on the actual situation, and to increase the evaluation of their quantitative indicators.

[0113] In summary, this invention systematically optimizes the generator structure. Through the synergistic integration of multi-channel input, residual modules, AdaIN normalization, and the SE attention mechanism, it significantly enhances the model's feature modeling and image generation capabilities. Multi-channel input expands the information dimension, improving the model's perception of complex tissue structures; the residual module enhances the expression and transmission of deep features; AdaIN guides the feature distribution to align with standard count images, improving the consistency of image style and details; and the SE attention mechanism further strengthens the response of key regions and suppresses redundant interference. Supported by the U-Net backbone architecture, these improvements complement each other, jointly constructing a more robust and biologically consistent low-count PET image generation model. Experimental results show that the combined use of the above modules outperforms the original CycleGAN structure in terms of image quality, structure restoration, and artifact suppression, verifying the effectiveness and necessity of comprehensive structural optimization.

[0114] Based on the same inventive concept, such as Figure 9 As shown, this embodiment of the invention also provides a low-count PET image quality enhancement system based on a fused multi-input recurrent consistent generative adversarial network, including: a data acquisition and preprocessing module, a network model construction module, a training optimization module, and an image quality enhancement module.

[0115] Specifically, the data acquisition and preprocessing module is used to acquire a PET image set, which includes several training sample pairs. Each training sample pair includes SCPET image samples and corresponding LCPET image samples; SCPET is standard count PET, and LCPET is low count PET. The network model construction module is used to construct a fused multi-input cyclic consistent generative adversarial network, including: expanding the inputs of the generator and discriminator in the LCPET image domain to multiple input channels, expanding the output of the generator in the SCPET image domain to multiple output channels, where one input channel or output channel corresponds to one PET image, and the count ratios of the multiple PET images corresponding to multiple input channels or output channels are different; and designing a total loss function, which includes perceptual loss, [missing information - likely a typo, should be removed]. The network employs loss tolerance and periodic consistency loss; the perceptual loss refers to the difference in feature space between the first generated image and the SCPET image sample, and the difference in feature space between the second generated image and the LCPET image sample. The first generated image is the image reconstructed by the generator in the LCPET image domain from the LCPET image sample; the second generated image is the image reconstructed by the generator in the SCPET image domain from the SCPET image sample. The training optimization module is used to train the fused multi-input cyclic consistent generative adversarial network using the image set based on the total loss function, to obtain a trained image quality enhancement model. The image quality enhancement module is used to input the low-count PET image to be enhanced into the trained generator in the LCPET image domain to obtain the enhanced PET image.

[0116] It should be noted that the low-count PET image quality enhancement system based on a fused multi-input cyclic consistent generative adversarial network provided in this embodiment of the invention is for implementing the above method. Its specific functions can be referred to in the above method embodiments, and will not be repeated here.

[0117] Figure 10 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 10As shown, the electronic device may include: a processor 1001, a communication interface 1002, a memory 1003, and a communication bus 1004. The processor 1001, communication interface 1002, and memory 1003 communicate with each other via the communication bus 1004. The processor 1001 can call logical instructions in the memory 1003 to execute a low-count PET image quality enhancement method based on a fused multi-input cyclic consistent generative adversarial network. This method includes: acquiring a PET image set, which includes several training sample pairs, each training sample pair including SCPET image samples and three corresponding LCPET image samples with different count ratios; SCPET is standard-count PET, and LCPET is low-count PET; constructing a fused multi-input cyclic consistent generative adversarial network, including: expanding the inputs of the generator and discriminator in the LCPET image domain to three input channels, and expanding the output of the generator in the SCPET image domain to three output channels. One input channel or output channel corresponds to one PET image, and the three input channels or output channels correspond to three PET images. The counting ratios between them are different; a total loss function is designed, which includes perceptual loss, adversarial loss, and periodic consistency loss; the perceptual loss includes the difference in feature space between the first generated image and the SCPET image sample and the difference in feature space between the second generated image and the LCPET image sample; the first generated image is the image reconstructed by the generator of the LCPET image domain from the LCPET image sample; the second generated image is the image reconstructed by the generator of the SCPET image domain from the SCPET image sample; based on the total loss function, the fused multi-input cyclic consistent generative adversarial network is trained using the image set to obtain a trained image quality enhancement model; the low-count PET image to be enhanced is input into the trained generator of the LCPET image domain to obtain the enhanced PET image.

[0118] Furthermore, when the logical instructions in the aforementioned memory 1003 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0119] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer is able to execute the low-count PET image quality enhancement method based on a fused multi-input cyclic consistent generative adversarial network provided in the above-described method embodiments.

[0120] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the low-count PET image quality enhancement method based on a fused multi-input cyclic consistent generative adversarial network provided in the above-described method embodiments.

[0121] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network, characterized in that, include: Obtain a PET image set, which includes several training sample pairs. Each training sample pair includes SCPET image samples and three corresponding LCPET image samples with different counting ratios. SCPET is standard count PET, and LCPET is low count PET. A multi-input cyclic consistent generative adversarial network (GAN) is constructed, comprising: expanding the inputs of the generator and discriminator in the LCPET image domain to three input channels, and expanding the output of the generator in the SCPET image domain to three output channels. Each input channel or output channel corresponds to one PET image, and the counting ratios of the three PET images corresponding to the three input channels or output channels are different. The generator adopts a U-Net structure, and residual blocks are introduced in the encoder and decoder, and an attention mechanism and adaptive instance normalization are used at the end of the decoder. The discriminator includes three parallel sub-discriminators, each of which adopts a patchGAN structure and uses spectral normalization. Correspondingly, for the same generated image, it is downsampled to three different resolutions, and the generated images at the three different resolutions are respectively input into the three sub-discriminators. Design a total loss function, which includes perceptual loss, adversarial loss, and periodic consistency loss; the perceptual loss includes the difference between the first generated image and the SCPET image sample in the feature space and the difference between the second generated image and the LCPET image sample in the feature space; the first generated image is the image reconstructed from the LCPET image sample by the generator in the LCPET image domain; the second generated image is the image reconstructed from the SCPET image sample by the generator in the SCPET image domain. Based on the total loss function, the image set is used to train the fused multi-input cyclic consistent generative adversarial network to obtain a trained image quality enhancement model; The low-count PET image to be enhanced is input into the generator of the trained LCPET image domain to obtain the enhanced PET image.

2. The method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network according to claim 1, characterized in that, The total loss function L is: in, This is due to periodic consistency loss; To combat the loss, the Wasserstein distance formula with a gradient penalty term is used; In order to perceive loss, G represents the difference between the first generated image and the SCPET image sample in the feature space. A (X A ) and X B Perceptual loss between them The difference between the second generated image and the LCPET image sample in the feature space is G. B (X B ) and X A Perceptual loss between them ; and The corresponding weights are N, where N is the batch size. For a given feature extractor, and These are the generated image and the real image, respectively. A and B are defined as the LCPET image domain and the SCPET image domain, respectively. X A X represents multiple original LCPET images with different counting ratios. B This represents the corresponding original SCPET image, generator G. A Represents a mapping from A to B, with generator G. B This represents the mapping from B to A.

3. The method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network according to claim 1, characterized in that, The training process also includes: clinical evaluation of the enhanced PET images, quantitative index analysis, and key anatomical region analysis to optimize model parameters; wherein, the key anatomical regions include lesion areas and normal tissue areas.

4. The method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network according to claim 3, characterized in that, The clinical evaluation specifically includes: visual assessment of image quality using a five-point Likert scale; and binary evaluation of lesion detectability.

5. The method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network according to claim 3, characterized in that, The quantitative index analysis specifically includes: quantitatively evaluating the overall image based on normalized root mean square error, structural similarity index, and peak signal-to-noise ratio; analyzing the joint histogram of the generated image and the real SCPET image; and comparing and evaluating the results by combining linear regression parameters.

6. The method for enhancing the quality of low-count PET images based on a fused multi-input cyclic consistent generative adversarial network according to claim 3, characterized in that, The analysis of key anatomical regions specifically includes: drawing ROIs in key anatomical regions; calculating the average and maximum standard uptake values ​​for each ROI based on the quantitative accuracy requirements of standard uptake values ​​in clinical diagnosis; and using SCPET images as a reference standard and Bland-Altman plots to evaluate the consistency of standard uptake values.

7. A low-count PET image quality enhancement system based on a fused multi-input cyclic consistent generative adversarial network, characterized in that, include: The data acquisition and preprocessing module is used to acquire a PET image set, which includes several training sample pairs. Each training sample pair includes SCPET image samples and three corresponding LCPET image samples with different count ratios. SCPET is standard count PET, and LCPET is low count PET. The network model construction module is used to construct a fused multi-input cyclic consistency generative adversarial network, including: expanding the inputs of the generator and discriminator in the LCPET image domain to three input channels, and expanding the output of the generator in the SCPET image domain to three output channels. Each input channel or output channel corresponds to one PET image, and the counting ratios of the three PET images corresponding to the three input channels or output channels are different. The module also designs a total loss function, which includes perceptual loss, adversarial loss, and cyclic consistency loss. The perceptual loss refers to the difference between the first generated image and the SCPET image sample in the feature space, and the difference between the second generated image and the LCPET image sample in the feature space. The first generated image is an image reconstructed from an LCPET image sample by a generator in the LCPET image domain; the second generated image is an image reconstructed from an SCPET image sample by a generator in the SCPET image domain; the generator adopts a U-Net structure and introduces residual blocks in the encoder and decoder, and uses an attention mechanism and adaptive instance normalization at the end of the decoder; the discriminator includes three parallel sub-discriminators, each of which adopts a patchGAN structure and uses spectral normalization; correspondingly, for the same generated image, it is downsampled to three different resolutions, and the generated images of the three different resolutions are respectively input into the three sub-discriminators; The training optimization module is used to train the fused multi-input cyclic consistent generative adversarial network based on the total loss function and the image set, so as to obtain a trained image quality enhancement model. The image quality enhancement module is used to input the low-count PET image to be enhanced into the generator of the trained LCPET image domain to obtain the enhanced PET image.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

Images