A Method and System for Low-Dose CT Image Quality Optimization Based on Generative Adversarial Networks

Abstract

This application discloses a method and system for optimizing the quality of low-dose CT images based on generative adversarial networks (GANs), belonging to the field of image optimization technology. The method includes: optimizing a preset loss function to obtain an adapted weighted loss function; constructing an image quality optimization channel for low-dose CT images using a GAN; acquiring low-dose CT images of the lungs of a target user and evaluating and determining the image optimization difficulty index; performing quality optimization on the low-dose lung CT images to generate several initial optimized CT images and evaluating and determining the initial generation bias characteristics; and aggregating the several initial optimized CT images to output the CT image quality optimization result. This solves the technical problem of insufficient optimization accuracy in lesion areas due to the degradation of low-dose CT image quality in existing methods.

Description

Technical Field

[0001] This application relates to the field of image optimization technology, specifically to a method and system for optimizing the quality of low-dose CT images based on generative adversarial networks. Background Technology

[0002] With the continuous development of medical imaging technology, CT scans play an important role in the early screening, diagnosis, and efficacy evaluation of lung diseases because they can provide high-resolution tomographic images.

[0003] However, traditional standard-dose CT scans expose patients to a certain dose of ionizing radiation, and repeated or prolonged exposure may increase potential health risks such as cancer. Low-dose CT technology reduces radiation dose by lowering tube current, tube voltage, or shortening scan time, but this introduces a large amount of noise and artifacts into the images, reducing image contrast and spatial resolution. This makes it difficult to clearly display the fine structures of the lungs, thus affecting the accuracy of clinical diagnosis and the formulation of treatment plans. Summary of the Invention

[0004] This application provides a method and system for optimizing the quality of low-dose CT images based on generative adversarial networks, which solves the technical problem of insufficient optimization accuracy of lesion areas due to the decline in the quality of existing low-dose CT images.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows:

[0006] In a first aspect, this application provides a method for optimizing the quality of low-dose CT images based on generative adversarial networks, the method comprising:

[0007] The lung region weight map of chronic obstructive pulmonary disease is obtained by analysis, and the preset loss function is optimized to obtain an appropriate weighted loss function;

[0008] Based on the aforementioned adaptive weighted loss function, an image quality optimization channel for low-dose CT images is constructed by combining a generative adversarial network.

[0009] Acquire low-dose CT images of the lungs of the target user, and determine the image optimization difficulty index based on the actual radiation dose and the characteristics of the user's condition;

[0010] Based on the image optimization difficulty index, the image quality optimization channel is activated to perform quality optimization on the low-dose CT image of the lungs, generating several initial optimized CT images, and evaluating and determining the initial generation deviation characteristics.

[0011] Based on the lung region weight map and the initial generation deviation features, an adaptation aggregation strategy is generated to aggregate the several initial optimized CT images and output the CT image quality optimization results.

[0012] Secondly, this application provides a low-dose CT image quality optimization system based on generative adversarial networks, including:

[0013] The function optimization module is used to analyze and obtain the lung region weight map of chronic obstructive pulmonary disease, and optimize the preset loss function to obtain an adapted weighted loss function.

[0014] The model building module is used to construct an image quality optimization channel for low-dose CT images based on the adaptive weighted loss function and in combination with a generative adversarial network.

[0015] The image acquisition module is used to acquire low-dose CT images of the lungs of the target user and to assess and determine the image optimization difficulty index based on the actual radiation dose to the target user and the characteristics of the user's condition.

[0016] The image optimization module is used to activate the image quality optimization channel based on the image optimization difficulty index, optimize the quality of the low-dose CT image of the lungs to generate several initial optimized CT images, and evaluate and determine the initial generation deviation characteristics.

[0017] The result output module is used to generate an adaptation aggregation strategy based on the lung region weight map and the initial generation deviation features, perform image aggregation on the several initial optimized CT images, and output the CT image quality optimization results.

[0018] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0019] This application provides a method and system for optimizing low-dose CT image quality based on generative adversarial networks (GANs). First, by analyzing and obtaining a lung region weight map for chronic obstructive pulmonary disease (COPD), a preset loss function is optimized to obtain an adaptive weighted loss function. This allows the loss function to differentiate the weights based on the importance of different lung regions, thus focusing on areas highly correlated with the disease during optimization. Second, based on this adaptive weighted loss function, an image quality optimization channel is constructed using a GAN to adapt to different data distributions and optimization needs. Subsequently, the generation bias characteristics of the initial images are evaluated, and an adaptive aggregation strategy is generated using the lung region weight map to aggregate the initial optimized images, ultimately outputting high-quality CT image optimization results. From loss function optimization and network model construction to image acquisition, difficulty assessment, multi-branch optimization, and result aggregation, a complete and targeted image quality optimization process is formed.

[0020] Through the above technical solution, this application can effectively improve the display clarity of fine lung structures in low-dose CT images, reduce noise and artifacts, improve the optimization accuracy of lesion areas, provide imaging evidence for clinical diagnosis, and at the same time minimize the radiation dose received by patients, ensuring medical safety while guaranteeing diagnostic accuracy. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the low-dose CT image quality optimization method based on generative adversarial networks provided in the embodiments of this application.

[0023] Figure 2 This is a schematic diagram of the structure of a low-dose CT image quality optimization system based on generative adversarial networks provided in an embodiment of this application.

[0024] The components represented by each number in the attached diagram are explained below:

[0025] Function optimization module 11, model building module 12, image acquisition module 13, image optimization module 14, and result output module 15. Detailed Implementation

[0026] This application provides a method and system for optimizing the quality of low-dose CT images based on generative adversarial networks, which addresses the technical problem of insufficient optimization accuracy in lesion areas due to the decline in the quality of existing low-dose CT images.

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0029] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0030] Example 1, as Figure 1 As shown, embodiments of this application provide a method for optimizing the quality of low-dose CT images based on generative adversarial networks, including:

[0031] S10: Analyze and obtain the lung region weight map of chronic obstructive pulmonary disease, and optimize the preset loss function to obtain an appropriate weighted loss function;

[0032] In this embodiment, firstly, high-dose CT image data of the lungs of patients with chronic obstructive pulmonary disease (COPD) and corresponding clinical diagnostic information are collected to construct a lung image database containing different lesion degrees and types. Then, lung regions are segmented from the high-dose CT images in this database.

[0033] Furthermore, for each predefined lung region type, its importance weight in COPD diagnosis and disease assessment is determined by combining clinical diagnostic information and imaging feature analysis. The region weights are then assigned to the corresponding segmented regions in the form of pixel values, thereby generating a lung region weight map. The value of each pixel in this weight map represents the degree of importance that location should be assigned in image quality optimization.

[0034] Furthermore, the preset loss function can be a commonly used image generation loss function, such as the mean squared error loss function, the negative function of the structural similarity index loss function, or perceptual loss. Based on the acquired lung region weight map, the preset loss function is weighted and optimized to achieve adaptive adjustment of the loss function, resulting in an adaptive weighted loss function.

[0035] The analysis includes obtaining a weighted map of lung regions in patients with chronic obstructive pulmonary disease (COPD), including:

[0036] Images of several lung regions were collected, constrained by the characteristics of chronic obstructive pulmonary disease.

[0037] The images of several lung regions are segmented according to a preset lung region type to obtain several sets of lung region blocks. The preset lung region type includes emphysema region, air retention region, normal lung parenchyma region, small airway perivascular region, and perivascular region.

[0038] The correlation between the aforementioned sets of lung regions and chronic obstructive pulmonary disease was evaluated, and several mean correlation values ​​were calculated.

[0039] The ratio of the mean correlation degree corresponding to any region type to the mean correlation degree of the plurality of regions is used as the region weight, and a lung region weight map is constructed based on the mean correlation degree of the plurality of regions.

[0040] In this embodiment, firstly, using typical characteristics of chronic obstructive pulmonary disease (COPD), such as airflow limitation, airway inflammation, and alveolar destruction, as core constraints, a large number of lung region images containing different pathological manifestations are systematically collected from a constructed lung imaging database. The images cover COPD patient cases at different disease stages, with different smoking histories, ages, and genders to ensure the diversity and representativeness of the data.

[0041] Secondly, based on predefined lung region types—namely, emphysematous regions, air trapping regions, normal lung parenchyma regions, peri-small airway regions, and perivascular regions—image segmentation is performed on several acquired lung region images. Deep learning-based segmentation models, such as U-Net, can be used, combined with medical image annotations for model training and optimization, to achieve the division of each lung region, thereby obtaining several sets of lung region blocks corresponding to different predefined region types.

[0042] Furthermore, for each lung region cluster, the correlation between the region cluster and chronic obstructive pulmonary disease is assessed by analyzing its imaging characteristics, such as density, texture, shape, and volume ratio; COPD clinical diagnostic indicators, such as FEV1 / FVC, mMRC score, and CAT score; and the correlation with pathophysiological changes. The mean correlation score is then calculated.

[0043] Finally, the average correlation score for any preset lung region type is divided by the sum of the average correlation scores for all preset lung region types; the resulting ratio is the region weight for that region type. The region weight values ​​are then assigned according to the spatial location of the corresponding segmented region in the original image, thus constructing a complete lung region weight map. The value of each pixel reflects the importance level of the lung region at that location in COPD-related image quality optimization.

[0044] Furthermore, the preset loss function is optimized to obtain an adaptive weighted loss function, including:

[0045] Configure a preset loss function, wherein the preset loss function consists of adversarial loss, pixel-level reconstruction loss and perceptual loss, which are used to constrain the overall image distribution, pixel accuracy and deep feature fidelity, respectively;

[0046] Based on the lung region weight map, the pixel-level reconstruction loss and perception loss in the preset loss function are spatially weighted and compensated to generate compensated pixel-level reconstruction loss and compensated perception loss.

[0047] By combining the adversarial loss, the compensated pixel-level reconstruction loss, and the compensated perception loss, an adaptive weighted loss function is constructed.

[0048] In this embodiment, firstly, a preset loss function is configured, which consists of three parts: adversarial loss, pixel-level reconstruction loss, and perceptual loss.

[0049] Among them, the adversarial loss is used to constrain the consistency between the generated image and the real high-dose CT image in the overall data distribution, and to encourage the generator to generate images that are closer to the real distribution; the pixel-level reconstruction loss, such as the mean square error loss, is used to measure the difference between the generated image and the real image at the pixel value level, and to ensure the basic pixel accuracy of the image; the perceptual loss extracts the deep features of the image through a pre-trained deep convolutional neural network and calculates the distance between the generated image and the real image in the deep feature space, so as to ensure the fidelity of the generated image in deep features such as structure and texture.

[0050] Secondly, spatial weighted compensation is applied to the pixel-level reconstruction loss and perceptual loss in the preset loss function based on the lung region weight map. Specifically, the weight value in the lung region weight map is multiplied by the pixel error at the corresponding position in the pixel-level reconstruction loss to obtain the compensated pixel-level reconstruction loss. This ensures that the pixel error in areas with high importance in the weight map, such as emphysema areas and areas around small airways, accounts for a larger proportion in the loss calculation. For the perceptual loss, the deep feature map is also weighted using the lung region weight map to enhance the contribution of feature differences in important areas to the perceptual loss, thus generating a compensated perceptual loss.

[0051] For example, suppose the weight value in the lung region weight map is w(x,y), where (x,y) represents the image pixel coordinates. For emphysematous regions and regions around small airways, w(x,y) takes a value of 1.2-1.5; for air retention regions and regions around blood vessels, w(x,y) takes a value of 0.8-1.0; and for normal lung parenchyma regions, w(x,y) takes a value of 0.5-0.7.

[0052] When calculating the compensation pixel-level reconstruction loss, the original pixel-level reconstruction loss is multiplied by the weight map, that is, the compensation pixel-level reconstruction loss L=Σ(w(x,y)×(G(x,y)-T(x,y))²), where G(x,y) is the pixel value of the generated image and T(x,y) is the pixel value of the real high-dose CT image.

[0053] Finally, the adversarial loss, the compensation for pixel-level reconstruction loss, and the compensation for perceptual loss are linearly fused according to a preset proportional coefficient to construct an adaptive weighted loss function. This adaptive weighted loss function can dynamically adjust the weight of each loss according to the difference in importance of different regions of the lung, thereby giving higher optimization priority to key diagnostic regions and improving the optimization accuracy of lesion areas during the image quality optimization process.

[0054] S20: Based on the adaptive weighted loss function, a generative adversarial network is combined to construct an image quality optimization channel for low-dose CT images;

[0055] In this embodiment, the generative adversarial network consists of a generator and a discriminator. The image quality optimization channel built based on the adaptive weighted loss function in this embodiment is characterized by targeted design and training optimization of the network structure of the generator and discriminator.

[0056] Low-dose CT images from a lung imaging database serve as input to the generator, while corresponding high-dose CT images act as ground truth labels. The generator produces optimized images based on the input low-dose CT images, and a discriminator distinguishes between the generated images and the ground truth high-dose CT images. An adaptive weighted loss function is used to calculate the loss between the generated and ground truth images, as well as the discriminator's discrimination loss. Backpropagation is then used to continuously update the network parameters of the generator and discriminator, thereby constructing an image quality optimization channel.

[0057] Specifically, step S20 in the method includes:

[0058] Using chronic obstructive pulmonary disease as a constraint, a set of low-dose CT images was collected, and high-quality standard-dose CT images corresponding to different sets of low-dose CT images were obtained as optimized CT images to obtain a set of optimized CT images.

[0059] The sample low-dose CT image set and the sample optimized CT image set are used as the sample training set, and K-fold cross-partitioning is performed to obtain K training data, where K is an integer greater than or equal to 5;

[0060] Using the sample low-dose CT images as input data and the sample optimized CT images as label data, a generative adversarial network is trained using the K training data based on the adaptive weighted loss function until both the generator and discriminator converge, generating K image quality optimization branches, which are then combined to obtain the image quality optimization channel.

[0061] In this embodiment, firstly, based on the clinical diagnostic needs and imaging characteristics of chronic obstructive pulmonary disease (COPD), a systematic set of low-dose CT images is collected, including COPD patients of different genders, age groups, smoking histories, disease severity, and comorbidities. For each low-dose CT image, a corresponding high-quality standard-dose CT image is obtained as the optimized CT image, constructing a one-to-one optimized CT image set. The low-dose CT image set and the optimized CT image set together constitute the training set for model training.

[0062] Secondly, to improve the stability and generalization ability of the model training and avoid overfitting, the constructed sample training set is subjected to K-fold cross-partitioning, where K is set to an integer greater than or equal to 5, such as K=5 or K=10. Specifically, the sample training set is randomly and uniformly divided into K mutually exclusive subsets, and the low-dose CT images in each subset maintain a one-to-one correspondence with their corresponding high-quality standard-dose CT images.

[0063] Next, for each of the K training datasets, a generative adversarial network is trained independently to generate K independent image quality optimization branches. In each independent training session, the low-dose CT images from that dataset are used as input data for the generator, while the corresponding optimized CT images are used as label data and the real sample inputs for the discriminator, i.e., high-quality standard-dose CT images.

[0064] During training, the generator aims to learn the mapping relationship from low-dose CT images to high-quality CT images, generating optimized images that are as close as possible to the labeled data. The discriminator is responsible for distinguishing the optimized images generated by the generator from real high-quality standard-dose CT images. The entire training process uses the adaptive weighted loss function obtained in step S10 as the main optimization objective function. The generator and discriminator are trained alternately, and the network parameters are continuously adjusted using the backpropagation algorithm. Training iterations continue until the image quality generated by the generator no longer significantly improves, and the discriminator can hardly distinguish between generated and real images, that is, both the generator and discriminator have reached convergence. At this point, the generative adversarial network corresponding to the training data is trained and forms an image quality optimization branch.

[0065] Finally, the K image quality optimization branches that have been trained and converged independently, obtained through K-fold cross-training, are combined to form an image quality optimization channel with a multi-branch structure.

[0066] Furthermore, each branch has the ability to independently optimize the quality of low-dose CT images. Subsequently, based on the specific features of the input image or optimization requirements, appropriate branches can be selected for optimization, or the output results of multiple branches can be aggregated, thereby further improving the overall optimization performance and robustness.

[0067] S30: Acquire low-dose CT images of the lungs of the target user, and determine the image optimization difficulty index based on the actual radiation dose and the characteristics of the user's condition;

[0068] In this embodiment, low-dose CT images of the target user's lungs are acquired, ensuring that the images contain complete lung anatomy and that scanning parameters, such as tube current, tube voltage, and pitch, are recorded completely for subsequent analysis of the actual radiation dose. The actual radiation dose can be obtained through the CT equipment's built-in dose reporting system, such as the volumetric CT dose index and dose-length product, and corrected in conjunction with the patient's body shape parameters to obtain an effective radiation dose value that better reflects the individual. The assessment of the user's condition characteristics needs to integrate the target user's clinical information and the types of lesions identified in previous imaging examinations.

[0069] Based on the actual radiation dose and the user's medical condition characteristics, an image optimization difficulty index assessment model is constructed. This model divides the actual radiation dose into several levels, with different dose levels corresponding to a basic difficulty coefficient. Generally, the lower the radiation dose, the higher the image noise and the more obvious the artifacts, and the greater the basic difficulty coefficient.

[0070] Meanwhile, based on the characteristics of the user's condition, weighting factors are set for different lesion types, lesion severity and complications. The basic difficulty coefficient is weighted and summed with the weighting factors of each condition to obtain a preliminary difficulty index. Then, through preset normalization processing, it is mapped to the integer range of 0-1 to finally determine the image optimization difficulty index of the target user.

[0071] The image optimization difficulty index is determined based on the target user's actual radiation dose and the user's medical condition characteristics, including:

[0072] Acquire basic vital signs and lung pathological characteristics of the target user, as well as the actual radiation dose during the CT scan of the target user;

[0073] The basic vital signs data and lung disease pathological features are input into a pre-constructed difficulty assessment model to evaluate and determine the difficulty coefficient of the first image optimization.

[0074] The ratio of the recommended standard radiation dose to the actual radiation dose is used as the second image optimization difficulty coefficient;

[0075] The first image optimization difficulty coefficient and the second image optimization difficulty coefficient are processed without dimensions and weighted to output the image optimization difficulty index.

[0076] In this embodiment, firstly, basic vital signs data of the target user are acquired, including age, gender, height, weight, and body mass index (BMI). These basic vital signs data affect the noise level and artifact distribution of CT images; for example, images of obese individuals typically have higher noise levels. Simultaneously, the user's lung pathological characteristics are collected, such as the severity of emphysema, airway wall thickening, and the extent of air trapping. Furthermore, the actual radiation dose parameters during the CT scan are acquired, such as the volumetric CT dose index and dose-length product, and scanning conditions such as the scanning equipment model, tube voltage, tube current, pitch, and slice thickness are recorded.

[0077] Secondly, the collected basic vital signs data and lung disease pathological features are input into a pre-constructed difficulty assessment model. This model is based on a neural network and trained using historical case data. The model assigns different weights to various features. For example, users with severe emphysema or diffuse small airway lesions have more complex textures and blurred boundaries in their images, increasing the optimization difficulty and corresponding to higher weights. Advanced age and excessively high or low BMI are also assigned corresponding influence weights due to differences in basic image quality. Through model calculation, the first image optimization difficulty coefficient, reflecting the impact of user pathological features and individual differences on image optimization, is output.

[0078] Next, the difficulty coefficient for the second image optimization is determined. The recommended standard radiation dose refers to the standard dose value recommended by the industry or determined based on clinical studies, while ensuring the image quality required for COPD diagnosis. The ratio of the recommended standard radiation dose to the actual radiation dose of the target user is used as the difficulty coefficient for the second image optimization. If the actual radiation dose is lower than the recommended standard radiation dose, the ratio is greater than 1, indicating that the image quality may be compromised at the current dose, requiring stronger optimization capabilities, and the difficulty coefficient increases accordingly. If the actual dose is close to or higher than the recommended standard dose, the ratio is close to or less than 1, and the difficulty coefficient decreases accordingly.

[0079] Finally, the first and second image optimization difficulty coefficients are processed in a dimensionless manner, for example, by mapping them to the range of 0-1 using the Min-Max normalization method to eliminate the difference in dimensions.

[0080] Then, based on clinical experience and model training results, reasonable weights are assigned to the two dimensionless coefficients, such as a weight of 0.6 for the first coefficient and 0.4 for the second coefficient, and a weighted summation is performed. The result is then rounded to produce an image optimization difficulty index in the integer range of 0-1, where 0 represents extremely low optimization difficulty and 1 represents extremely high optimization difficulty, thus quantifying the optimization challenge of low-dose CT images for the target user.

[0081] S40: Activate the image quality optimization channel based on the image optimization difficulty index, perform quality optimization on the low-dose CT image of the lungs to generate several initial optimized CT images, and evaluate and determine the initial generation deviation characteristics;

[0082] In this embodiment, the image quality optimization channel is activated based on the image optimization difficulty index. Specifically, once the image optimization difficulty index is determined, a suitable number or a specific combination of branches is selected from the K image quality optimization branches included in the image quality optimization channel and activated according to the numerical range of the index.

[0083] Secondly, each activated image quality optimization branch receives low-dose lung CT images of the target user as input in parallel and processes them independently to generate several initial optimized CT images corresponding to the number of branches. Each initial optimized CT image reflects the corresponding branch's ability to suppress noise and artifacts in low-dose CT images and its ability to recover lesion features, but there may be some generation deviations due to differences in the focus or convergence state of each branch during training.

[0084] Next, the initial generation deviation features are evaluated and determined. The degree and specific manifestation of the deviation between the initial optimized CT image and the expected features of the ideal optimized result or the real high-dose CT image are analyzed, thereby forming a comprehensive description of the initial generation deviation features.

[0085] The process involves optimizing the quality of the low-dose CT images of the lungs to generate several initial optimized CT images, including:

[0086] Multiply the ratio of the image optimization difficulty index to the historical maximum image optimization difficulty index recorded within the historical time range by K and round down to obtain the number of adaptive optimization branches selected, P, where P is greater than or equal to 2 and less than or equal to K.

[0087] P optimization branches are randomly selected from the K image quality optimization branches in the image quality optimization channel to optimize the quality of the low-dose CT images of the lungs, and several initial optimized CT images are output.

[0088] In this embodiment, firstly, to determine the number P of suitable optimization branches, the historical maximum image optimization difficulty index is introduced as a reference. This historical maximum image optimization difficulty index is the maximum value obtained after statistical analysis of the image optimization difficulty indices of all target users in the past during system operation, representing the highest optimization difficulty level processed.

[0089] Furthermore, the image optimization difficulty index of the current target user is divided by the historical maximum image optimization difficulty index to obtain a relative difficulty ratio value, which reflects the position of the current task difficulty in the historical difficulty spectrum.

[0090] Subsequently, this ratio is multiplied by the total number of branches K in the image quality optimization channel, and the product is rounded to obtain the number of branches selected for the adaptation optimization, P. The value of P is constrained to be greater than or equal to 2 and less than or equal to K. Selecting too few branches may lead to excessive reliance on a single model, lacking diversity and robustness, while selecting all K branches may waste computational resources and reduce optimization efficiency.

[0091] By using a dynamic calculation method, when the current image optimization difficulty index is close to the historical maximum, P will approach K, thereby mobilizing more branches to participate in optimization to cope with high-difficulty tasks; when the current difficulty is low, P will decrease accordingly, improving processing efficiency while ensuring optimization effect.

[0092] For example, if the historical maximum image optimization difficulty index is normalized to 1.0, the current image optimization difficulty index is 0.6, and K=10, then P=0.6×10=6, that is, 6 optimization branches are selected.

[0093] Next, after determining the value of P, P optimization branches are randomly selected from the K image quality optimization branches of the image quality optimization channel. Random selection avoids systematic biases that might be introduced by fixed selection of certain branches, ensuring that each branch has an equal opportunity to participate in optimization tasks of varying difficulty, thereby improving the overall generalization ability of the multi-branch set. The selected P optimization branches will work simultaneously in parallel, each independently receiving low-dose CT images of the target user's lungs as input data.

[0094] Specifically, the generative adversarial network model within each branch has already learned and solidified the mapping rules from low-dose CT to high-quality CT during the training phase. At this point, based on the features of the input image, it uses its internal network parameters to perform forward propagation calculations, suppressing noise, eliminating artifacts, and enhancing and restoring key features such as lung texture and lesion details. After their respective processing flows, each optimization branch outputs an independent initial optimized CT image. Therefore, P branches will output P initial optimized CT images, which together constitute the basic dataset for subsequent bias assessment and result aggregation.

[0095] Further, the initial generation bias characteristics are evaluated and determined, including:

[0096] Several initial optimized CT images are segmented according to the preset lung region type, and several sets of initial optimized CT image blocks are output.

[0097] Consistency evaluation is performed on the several initial optimized CT image block sets respectively, and several image consistency coefficients are output as initial generation deviation features.

[0098] In this embodiment, the initial optimized CT image is first segmented according to a preset lung region type. The preset lung region type is determined based on lung anatomy and common sites of COPD lesions, including emphysema regions, air trapping regions, normal lung parenchyma regions, small airway perivascular regions, and perivascular regions. A deep learning-based multi-organ segmentation model, such as 3DU-Net, is used to automatically segment each initial optimized CT image. This segmentation model is pre-trained on a CT image dataset containing lung annotations and can identify and delineate the boundaries of each preset region, segmenting the initial optimized CT image into several independent initial optimized CT image blocks. All image blocks together constitute the initial optimized CT image block set corresponding to the initial optimized CT image.

[0099] For example, after segmentation, an initial optimized CT image can be divided into a set of 15 image blocks, including 12 lung segment image blocks, 1 tracheal image block, 1 main bronchus image block, and 1 hilar region image block.

[0100] Secondly, a consistency evaluation was performed on several initial optimized CT image block sets. The consistency evaluation used image blocks at corresponding locations in all initial optimized CT image block sets as evaluation units. For image blocks within each evaluation group, quantitative analysis was performed from three dimensions: grayscale distribution consistency, texture feature consistency, and structural morphology consistency.

[0101] Furthermore, this process was performed on all evaluation groups for all preset lung region types, resulting in several region consistency coefficients. Finally, the arithmetic mean of all region consistency coefficients was calculated to obtain the image consistency coefficient corresponding to the initial optimized CT image patch set. The image consistency coefficients of all initial optimized CT image patch sets collectively constitute the initial generation bias feature. The closer the value is to 1, the higher the consistency of each initial optimized CT image in the corresponding lung region, and the smaller the generation bias; conversely, it indicates a more significant generation bias.

[0102] Furthermore, a consistency evaluation is performed on each of the several initial optimized CT image block sets, and several image consistency coefficients are output, including:

[0103] A first initial optimized CT image block set is randomly selected from the plurality of initial optimized CT image block sets, and tissue registration and alignment are performed on the first initial optimized CT image block set to obtain a first aligned CT image block set;

[0104] Calculate the local standard deviation, range, and signal-to-noise ratio of the first aligned CT image block set to generate a first grayscale statistical evaluation map;

[0105] The gradient consistency and texture similarity of the first aligned CT image block set are extracted to form a first structural feature evaluation map;

[0106] The first grayscale statistical evaluation map and the first structural feature evaluation map are pixel-level weighted summation to generate a first comprehensive inconsistency heatmap.

[0107] The first image consistency coefficient is obtained by calculating the global average value of the first integrated inconsistency heatmap and added to the plurality of image consistency coefficients.

[0108] In this embodiment, firstly, one of several initial optimized CT image block sets is randomly selected as the first initial optimized CT image block set. Since different initial optimized CT image block sets originate from different optimization branches, the spatial positions of the image blocks within them may differ. Therefore, tissue registration and alignment processing is performed on the first initial optimized CT image block set.

[0109] Specifically, the grayscale-based multimodal registration algorithm takes one reference image block, such as a normal lung parenchyma region image block, as the benchmark, and performs spatial transformations on other image blocks in the set, including translation, rotation, scaling, and shearing. By maximizing mutual information or minimizing mean square error and other similarity metrics, all image blocks are aligned in terms of anatomical structure, thereby obtaining the first aligned CT image block set.

[0110] Secondly, the local standard deviation, range, and signal-to-noise ratio (SNR) of the first aligned CT image block set are calculated to generate the first grayscale statistical evaluation map. For each voxel in the first aligned CT image block set, the standard deviation of the grayscale value is calculated within its local neighborhood of a preset size, such as a 3×3×3 voxel window, to reflect the degree of dispersion of grayscale distribution in that region. A larger standard deviation indicates more severe local grayscale fluctuations, which may indicate the presence of noise or artifacts. Simultaneously, the range within the neighborhood is calculated, i.e., the maximum value minus the minimum value, to further quantify the dynamic range of grayscale changes. The SNR is calculated as the ratio of the signal mean to the noise standard deviation, where the signal mean is the average grayscale of the region of interest, such as the average grayscale of normal lung tissue, and the noise standard deviation is the standard deviation of background noise in the air region or far from the lesion region. The three indicators are normalized to the range of 0-1, such as by taking the reciprocal of the local standard deviation, range, and SNR. A pixel-level weighted sum is then performed according to preset weights to generate the first grayscale statistical evaluation map. High-value regions in this map represent areas with strong inconsistencies in grayscale statistical characteristics.

[0111] Next, the gradient consistency and texture similarity of the first aligned CT image patch set are extracted to form the first structural feature evaluation map. Gradient consistency is achieved by calculating the coefficient of variation of isotropic gradient magnitudes. First, the Sobel operator or Prewitt operator is used to calculate the gradient of each voxel in the x, y, and z directions, synthesizing a gradient magnitude image. Then, the coefficient of variation of gradient magnitudes is calculated in the local neighborhood. The larger the coefficient of variation, the worse the consistency of gradient direction and intensity. Texture similarity is achieved by extracting multiple texture features, such as energy, entropy, contrast, and correlation, using the gray-level co-occurrence matrix to calculate the degree of difference in texture features at corresponding positions between different image patches, for example, by measuring Euclidean distance or cosine distance. The gradient consistency index and texture similarity difference index are also normalized to the range of 0-1 and assigned appropriate weights, such as gradient consistency 0.5 and texture similarity difference 0.5, and then pixel-level weighted summation is performed to generate the first structural feature evaluation map. In this map, high-value areas represent areas with strong inconsistencies in anatomical structure and texture details.

[0112] Then, the first grayscale statistical evaluation map and the first structural feature evaluation map are summed at the pixel level to generate the first comprehensive inconsistency heatmap. Based on the different levels of emphasis on grayscale and structural information in clinical diagnosis, a weight of 0.4 is assigned to the grayscale statistical evaluation map, and a weight of 0.6 is assigned to the structural feature evaluation map. The corresponding pixel values ​​of the two maps are multiplied by their respective weights and then summed to obtain the first comprehensive inconsistency heatmap. This heatmap uses color depth or numerical value to reflect the degree of comprehensive inconsistency in each region of the first aligned CT image block set. Regions with higher numerical values ​​in the heatmap exhibit greater deviations from the ideal state in terms of grayscale statistics and structural features.

[0113] Finally, the global average of the first comprehensive inconsistency heatmap is calculated to obtain the first image consistency coefficient, which is then added to several image consistency coefficients. The global average is calculated by taking the arithmetic mean of all voxel values ​​in the heatmap. The lower the average, the lower the overall inconsistency of the first initial optimized CT image block set, i.e., the higher the consistency. This first image consistency coefficient serves as the quantitative evaluation result of the randomly selected initial optimized CT image block set.

[0114] For example, if the global average value of the first integrated inconsistency heatmap is 0.25, then the first image consistency coefficient is 1-0.25=0.75, indicating that the overall consistency of the initial optimized CT image block set is high.

[0115] Furthermore, by performing a similar operation on all initial optimized CT image block sets, i.e. selecting and calculating their image consistency coefficients respectively, several image consistency coefficients are finally obtained, reflecting the generation deviation between the optimization results of each branch.

[0116] S50: Generate an adaptation aggregation strategy based on the lung region weight map and the initial generation deviation features, perform image aggregation on the several initial optimized CT images, and output the CT image quality optimization results.

[0117] In this embodiment of the application, firstly, a lung region weight map is generated. Based on the differences in importance of different lung regions, the weight map assigns corresponding weight values ​​to each preset lung region.

[0118] Secondly, the image consistency coefficients in the initially generated bias features are fused with the lung region weight map to construct a region bias weighting matrix. Based on the region bias weighting matrix and the consistency coefficients of each region in the initially generated bias features, an adaptation aggregation strategy is generated. The adaptation aggregation strategy consists of two parts: region priority ranking and aggregation weight allocation.

[0119] Then, image aggregation is performed on several initial optimized CT images according to the adaptation aggregation strategy. Based on region priority, starting with the region with the highest priority, pixel-level weighted fusion is performed on the corresponding image blocks in each initial optimized CT image according to the aggregation weight of that region. The final output is a CT image quality optimization result that incorporates the advantages of multiple branches.

[0120] The process of generating an adaptation aggregation strategy based on the lung region weight map and initial generation bias features includes:

[0121] A first lung region type is randomly selected from the preset lung region types. Based on the lung region weight map and several image consistency coefficients, the first lung region weight and the first region image consistency coefficient corresponding to the first lung region type are obtained.

[0122] Based on a pre-built regional feature-aggregation scheme library, a first aggregation scheme is determined by matching the weight of the first lung region with the consistency coefficient of the first region image. Then, multiple aggregation schemes corresponding to multiple lung region types are analyzed in sequence to generate an adaptive aggregation strategy. The aggregation scheme includes at least fast averaging, weighted averaging, optimal selection, median filtering, consistency weighting, feature-level fusion, conservative smoothing, physical constraint arbitration, and expert intervention mode.

[0123] In this embodiment of the application, firstly, one is randomly selected from the preset lung region types and defined as the first lung region type, such as the emphysema region.

[0124] Secondly, the weight value corresponding to the first lung region type is extracted from the lung region weight map and recorded as the first lung region weight. This weight value reflects the relative importance of this region in clinical diagnosis. For example, if the emphysema region is assigned a weight of 0.8 in the weight map, it indicates that it should receive higher attention and priority processing in the image quality optimization results.

[0125] Meanwhile, from the several regional consistency coefficients of the initially generated deviation features, the regional consistency coefficient corresponding to the first lung region type is extracted and denoted as the first region image consistency coefficient. This coefficient quantifies the consistency level of the output results of each optimization branch in this region. For example, if the first region image consistency coefficient is 0.75, it means that the optimization results of each branch in this region have a moderate degree of consistency.

[0126] Secondly, based on a pre-built regional feature-aggregation scheme library, a first aggregation scheme is determined by matching the weight of the first lung region with the consistency coefficient of the first region image. This regional feature-aggregation scheme library was built during the system training and debugging phase using experimental data and clinical feedback. It stores the mapping relationship between different lung region weights, different combinations of region image consistency coefficients, and the corresponding optimal aggregation scheme.

[0127] For example, for regions with high weight and high consistency, such as the first lung region with a weight > 0.8 and an image consistency coefficient > 0.85, a "weighted average" or "consistency weighting" scheme is recommended to utilize the consensus information of each branch's results, smooth out minor differences, and improve the stability of the results. For regions with high weight but low consistency, such as the first lung region with a weight > 0.8 and an image consistency coefficient < 0.6, an "optimal selection" combined with an "expert intervention mode" or "physical constraint arbitration" is recommended, that is, selecting the branch results that best match the preset anatomical prior or physical model in that region. The results are used as the main reference, and preset expert rules can be introduced for correction to avoid the fusion of erroneous information. For regions with medium weight and medium consistency, such as the first lung region with a weight between 0.5 and 0.8 and an image consistency coefficient between 0.6 and 0.85, the "feature-level fusion" or "conservative smoothing" scheme is recommended. While preserving the detailed features of each branch, obvious conflict areas are moderately smoothed. For low-weight regions, regardless of the consistency level, computationally efficient aggregation schemes such as "fast averaging" or "median filtering" can be used to balance the overall optimization efficiency.

[0128] Furthermore, by querying this mapping relationship, the most suitable first aggregation scheme can be matched for the first lung region type. Subsequently, following the same method, other regions in the preset lung region types, such as air retention regions, normal lung parenchyma regions, small airway perivascular regions, and perivascular regions, are analyzed in turn. That is, each region is selected as the first lung region type, its corresponding region weight and region consistency coefficient are extracted, and their respective aggregation schemes are matched.

[0129] The aggregation schemes corresponding to all lung region types are organized and sorted, for example, by arranging the regions and their corresponding aggregation schemes in descending order of lung region weight, thus generating a complete adaptive aggregation strategy. This strategy clarifies the specific processing schemes and priority order to be used for different lung regions during the image aggregation stage.

[0130] In summary, compared to existing technologies, this application constructs a multi-branch optimization network, utilizes sub-networks with different optimization objectives to generate diverse initial optimized CT images, and performs region segmentation and consistency evaluation on these images to form quantified initial generation bias features. Simultaneously, by combining a lung region weight map, a pre-constructed region feature-aggregation scheme library is used to dynamically match the optimal aggregation scheme for lung regions with different importance and consistency levels, achieving region priority ranking and aggregation weight allocation. Finally, an adaptive aggregation strategy is used to perform pixel-level weighted fusion of the multi-branch initial optimization results, solving the problem of local image quality degradation caused by a single optimization objective.

[0131] In summary, the embodiments of this application have at least the following technical effects:

[0132] This application provides a low-dose CT image quality optimization method based on generative adversarial networks (GANs). First, by analyzing and obtaining a lung region weight map for chronic obstructive pulmonary disease (COPD), a preset loss function is optimized to obtain an adaptive weighted loss function. This allows the loss function to differentiate the weights based on the importance of different lung regions, thus focusing on areas highly correlated with the disease during optimization. Second, based on this adaptive weighted loss function, an image quality optimization channel is constructed using a GAN to adapt to different data distributions and optimization needs. Subsequently, the generation bias characteristics of the initial images are evaluated, and an adaptive aggregation strategy is generated using the lung region weight map to aggregate the initial optimized images, ultimately outputting high-quality CT image optimization results. From loss function optimization and network model construction to image acquisition, difficulty assessment, multi-branch optimization, and result aggregation, a complete and targeted image quality optimization process is formed.

[0133] Through the above technical solution, this application can effectively improve the display clarity of fine lung structures in low-dose CT images, reduce noise and artifacts, improve the optimization accuracy of lesion areas, provide imaging evidence for clinical diagnosis, and at the same time minimize the radiation dose received by patients, ensuring medical safety while guaranteeing diagnostic accuracy.

[0134] Example 2, as Figure 2 As shown, based on the same inventive concept as the low-dose CT image quality optimization method based on generative adversarial networks provided in Embodiment 1, this application also provides a low-dose CT image quality optimization system based on generative adversarial networks, including:

[0135] The function optimization module 11 is used to analyze and obtain the lung region weight map of chronic obstructive pulmonary disease, and optimize the preset loss function to obtain the appropriate weighted loss function.

[0136] Model building module 12 is used to construct an image quality optimization channel for low-dose CT images based on the adaptive weighted loss function and in combination with a generative adversarial network.

[0137] Image acquisition module 13 is used to acquire low-dose CT images of the lungs of the target user and to assess and determine the image optimization difficulty index based on the actual radiation dose of the target user and the characteristics of the user's condition.

[0138] The image optimization module 14 is used to activate the image quality optimization channel based on the image optimization difficulty index, optimize the quality of the low-dose CT image of the lungs to generate several initial optimized CT images, and evaluate and determine the initial generation deviation characteristics.

[0139] The result output module 15 is used to generate an adaptation aggregation strategy based on the lung region weight map and the initial generation deviation features, perform image aggregation on the several initial optimized CT images, and output the CT image quality optimization results.

[0140] Furthermore, in one embodiment of the application, the analysis to obtain a weighted map of lung regions in chronic obstructive pulmonary disease includes:

[0141] Images of several lung regions were collected, constrained by the characteristics of chronic obstructive pulmonary disease.

[0142] The images of several lung regions are segmented according to a preset lung region type to obtain several sets of lung region blocks. The preset lung region type includes emphysema region, air retention region, normal lung parenchyma region, small airway perivascular region, and perivascular region.

[0143] The correlation between the aforementioned sets of lung regions and chronic obstructive pulmonary disease was evaluated, and several mean correlation values ​​were calculated.

[0144] The ratio of the mean correlation degree corresponding to any region type to the mean correlation degree of the plurality of regions is used as the region weight, and a lung region weight map is constructed based on the mean correlation degree of the plurality of regions.

[0145] Furthermore, in one embodiment, optimizing the preset loss function to obtain an adaptive weighted loss function includes:

[0146] Configure a preset loss function, wherein the preset loss function consists of adversarial loss, pixel-level reconstruction loss and perceptual loss, which are used to constrain the overall image distribution, pixel accuracy and deep feature fidelity, respectively;

[0147] Based on the lung region weight map, the pixel-level reconstruction loss and perception loss in the preset loss function are spatially weighted and compensated to generate compensated pixel-level reconstruction loss and compensated perception loss.

[0148] By combining the adversarial loss, the compensated pixel-level reconstruction loss, and the compensated perception loss, an adaptive weighted loss function is constructed.

[0149] In one embodiment, the model building module 12 is specifically used for:

[0150] Using chronic obstructive pulmonary disease as a constraint, a set of low-dose CT images was collected, and high-quality standard-dose CT images corresponding to different sets of low-dose CT images were obtained as optimized CT images to obtain a set of optimized CT images.

[0151] The sample low-dose CT image set and the sample optimized CT image set are used as the sample training set, and K-fold cross-partitioning is performed to obtain K training data, where K is an integer greater than or equal to 5;

[0152] Using the sample low-dose CT images as input data and the sample optimized CT images as label data, a generative adversarial network is trained using the K training data based on the adaptive weighted loss function until both the generator and discriminator converge, generating K image quality optimization branches, which are then combined to obtain the image quality optimization channel.

[0153] Furthermore, in one embodiment of the application, the image optimization difficulty index is determined based on the actual radiation dose to the target user and the user's medical condition characteristics, including:

[0154] Acquire basic vital signs and lung pathological characteristics of the target user, as well as the actual radiation dose during the CT scan of the target user;

[0155] The basic vital signs data and lung disease pathological features are input into a pre-constructed difficulty assessment model to evaluate and determine the difficulty coefficient of the first image optimization.

[0156] The ratio of the recommended standard radiation dose to the actual radiation dose is used as the second image optimization difficulty coefficient;

[0157] The first image optimization difficulty coefficient and the second image optimization difficulty coefficient are processed without dimensions and weighted to output the image optimization difficulty index.

[0158] Furthermore, the low-dose CT images of the lungs are quality optimized to generate several initial optimized CT images, including:

[0159] Multiply the ratio of the image optimization difficulty index to the historical maximum image optimization difficulty index recorded within the historical time range by K and round down to obtain the number of adaptive optimization branches selected, P, where P is greater than or equal to 2 and less than or equal to K.

[0160] P optimization branches are randomly selected from the K image quality optimization branches in the image quality optimization channel to optimize the quality of the low-dose CT images of the lungs, and several initial optimized CT images are output.

[0161] Further, the initial generation bias characteristics are evaluated and determined, including:

[0162] Several initial optimized CT images are segmented according to the preset lung region type, and several sets of initial optimized CT image blocks are output.

[0163] Consistency evaluation is performed on the several initial optimized CT image block sets respectively, and several image consistency coefficients are output as initial generation deviation features.

[0164] Furthermore, in one embodiment of the application, a consistency evaluation is performed on the plurality of initial optimized CT image block sets respectively, and a plurality of image consistency coefficients are output, including:

[0165] A first initial optimized CT image block set is randomly selected from the plurality of initial optimized CT image block sets, and tissue registration and alignment are performed on the first initial optimized CT image block set to obtain a first aligned CT image block set;

[0166] Calculate the local standard deviation, range, and signal-to-noise ratio of the first aligned CT image block set to generate a first grayscale statistical evaluation map;

[0167] The gradient consistency and texture similarity of the first aligned CT image block set are extracted to form a first structural feature evaluation map;

[0168] The first grayscale statistical evaluation map and the first structural feature evaluation map are pixel-level weighted summation to generate a first comprehensive inconsistency heatmap.

[0169] The first image consistency coefficient is obtained by calculating the global average value of the first integrated inconsistency heatmap and added to the plurality of image consistency coefficients.

[0170] Furthermore, in one embodiment, an adaptation aggregation strategy is generated based on the lung region weight map and the initial generation bias features, including:

[0171] A first lung region type is randomly selected from the preset lung region types. Based on the lung region weight map and several image consistency coefficients, the first lung region weight and the first region image consistency coefficient corresponding to the first lung region type are obtained.

[0172] Based on a pre-built regional feature-aggregation scheme library, a first aggregation scheme is determined by matching the weight of the first lung region with the consistency coefficient of the first region image. Then, multiple aggregation schemes corresponding to multiple lung region types are analyzed in sequence to generate an adaptive aggregation strategy. The aggregation scheme includes at least fast averaging, weighted averaging, optimal selection, median filtering, consistency weighting, feature-level fusion, conservative smoothing, physical constraint arbitration, and expert intervention mode.

[0173] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0174] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0175] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for optimizing the quality of low-dose CT images based on generative adversarial networks, characterized in that the method... include: The lung region weight map of chronic obstructive pulmonary disease is obtained by analysis, and the preset loss function is optimized to obtain an appropriate weighted loss function; Based on the aforementioned adaptive weighted loss function, an image quality optimization channel for low-dose CT images is constructed using a generative adversarial network, including: Using chronic obstructive pulmonary disease as a constraint, a set of low-dose CT images was collected, and high-quality standard-dose CT images corresponding to different sets of low-dose CT images were obtained as optimized CT images to obtain a set of optimized CT images. The sample low-dose CT image set and the sample optimized CT image set are used as the sample training set, and K-fold cross-partitioning is performed to obtain K training data, where K is an integer greater than or equal to 5; Using the sample low-dose CT images as input data and the sample optimized CT images as label data, based on the adaptive weighted loss function, the generator adversarial network is trained using the K training data until both the generator and the discriminator converge, generating K image quality optimization branches, which are then combined to obtain the image quality optimization channel. Acquire low-dose CT images of the lungs of the target user, and determine the image optimization difficulty index based on the actual radiation dose and the characteristics of the user's condition; Based on the image optimization difficulty index, the image quality optimization channel is activated to perform quality optimization on the low-dose CT image of the lungs, generating several initial optimized CT images, and evaluating and determining the initial generation deviation characteristics. Based on the lung region weight map and the initial generation deviation features, an adaptation aggregation strategy is generated to aggregate the several initial optimized CT images and output the CT image quality optimization results. The evaluation determines the initial generation bias characteristics, including: Several initial optimized CT images are segmented according to the preset lung region type, and several sets of initial optimized CT image blocks are output. Consistency evaluation is performed on the several initial optimized CT image block sets respectively, and several image consistency coefficients are output as initial generation deviation features; The consistency evaluation is performed on each of the several initial optimized CT image block sets, and several image consistency coefficients are output, including: A first initial optimized CT image block set is randomly selected from the plurality of initial optimized CT image block sets, and tissue registration and alignment are performed on the first initial optimized CT image block set to obtain a first aligned CT image block set; Calculate the local standard deviation, range, and signal-to-noise ratio of the first aligned CT image block set to generate a first grayscale statistical evaluation map; The gradient consistency and texture similarity of the first aligned CT image block set are extracted to form a first structural feature evaluation map; The first grayscale statistical evaluation map and the first structural feature evaluation map are pixel-level weighted summation to generate a first comprehensive inconsistency heatmap. The first image consistency coefficient is obtained by calculating the global average value of the first integrated inconsistency heatmap and added to the plurality of image consistency coefficients.

2. The method for optimizing low-dose CT image quality based on generative adversarial networks according to claim 1, characterized in that, Analysis to obtain a weighted map of lung regions in chronic obstructive pulmonary disease, including: Images of several lung regions were collected, constrained by the characteristics of chronic obstructive pulmonary disease. The images of several lung regions are segmented according to a preset lung region type to obtain several sets of lung region blocks. The preset lung region type includes emphysema region, air retention region, normal lung parenchyma region, small airway perivascular region, and perivascular region. The correlation between the aforementioned sets of lung regions and chronic obstructive pulmonary disease was evaluated, and several mean correlation values ​​were calculated. The ratio of the mean correlation degree corresponding to any region type to the mean correlation degree of the plurality of regions is used as the region weight, and a lung region weight map is constructed based on the mean correlation degree of the plurality of regions.

3. The method for optimizing low-dose CT image quality based on generative adversarial networks according to claim 1, characterized in that, Optimize the preset loss function to obtain an adaptive weighted loss function, including: Configure a preset loss function, wherein the preset loss function consists of adversarial loss, pixel-level reconstruction loss and perceptual loss, which are used to constrain the overall image distribution, pixel accuracy and deep feature fidelity, respectively; Based on the lung region weight map, the pixel-level reconstruction loss and perception loss in the preset loss function are spatially weighted and compensated to generate compensated pixel-level reconstruction loss and compensated perception loss. By combining the adversarial loss, the compensated pixel-level reconstruction loss, and the compensated perception loss, an adaptive weighted loss function is constructed.

4. The method for optimizing low-dose CT image quality based on generative adversarial networks according to claim 1, characterized in that, The image optimization difficulty index is determined based on the target user's actual radiation dose and the user's medical condition characteristics, including: Acquire basic vital signs and lung pathological characteristics of the target user, as well as the actual radiation dose during the CT scan of the target user; The basic vital signs data and lung disease pathological features are input into a pre-constructed difficulty assessment model to evaluate and determine the difficulty coefficient of the first image optimization. The ratio of the recommended standard radiation dose to the actual radiation dose is used as the second image optimization difficulty coefficient; The first image optimization difficulty coefficient and the second image optimization difficulty coefficient are processed without dimensions and weighted to output the image optimization difficulty index.

5. The method for optimizing low-dose CT image quality based on generative adversarial networks according to claim 1, characterized in that, Based on the image optimization difficulty index, the image quality optimization channel is activated to perform quality optimization on the low-dose lung CT image, generating several initial optimized CT images, including: Multiply the ratio of the image optimization difficulty index to the historical maximum image optimization difficulty index recorded within the historical time range by K and round down to obtain the number of adaptive optimization branches selected, P, where P is greater than or equal to 2 and less than or equal to K. P optimization branches are randomly selected from the K image quality optimization branches in the image quality optimization channel to optimize the quality of the low-dose CT images of the lungs, and several initial optimized CT images are output.

6. The method for optimizing low-dose CT image quality based on generative adversarial networks according to claim 1, characterized in that, An adaptation aggregation strategy is generated based on the lung region weight map and the initial generation bias features, including: A first lung region type is randomly selected from the preset lung region types. Based on the lung region weight map and several image consistency coefficients, the first lung region weight and the first region image consistency coefficient corresponding to the first lung region type are obtained. Based on a pre-built regional feature-aggregation scheme library, a first aggregation scheme is determined by matching the weight of the first lung region with the consistency coefficient of the first region image. Then, multiple aggregation schemes corresponding to multiple lung region types are analyzed in sequence to generate an adaptive aggregation strategy. The aggregation scheme includes at least fast averaging, weighted averaging, optimal selection, median filtering, consistency weighting, feature-level fusion, conservative smoothing, physical constraint arbitration, and expert intervention mode.

7. A low-dose CT image quality optimization system based on generative adversarial networks, characterized in that, The method for performing the low-dose CT image quality optimization based on generative adversarial networks as described in any one of claims 1-6 includes: The function optimization module is used to analyze and obtain the lung region weight map of chronic obstructive pulmonary disease, and optimize the preset loss function to obtain an adapted weighted loss function. The model building module, used to construct an image quality optimization channel for low-dose CT images based on the adaptive weighted loss function and combined with a generative adversarial network, includes: Using chronic obstructive pulmonary disease as a constraint, a set of low-dose CT images was collected, and high-quality standard-dose CT images corresponding to different sets of low-dose CT images were obtained as optimized CT images to obtain a set of optimized CT images. The sample low-dose CT image set and the sample optimized CT image set are used as the sample training set, and K-fold cross-partitioning is performed to obtain K training data, where K is an integer greater than or equal to 5; Using the sample low-dose CT images as input data and the sample optimized CT images as label data, based on the adaptive weighted loss function, the generator adversarial network is trained using the K training data until both the generator and the discriminator converge, generating K image quality optimization branches, which are then combined to obtain the image quality optimization channel. The image acquisition module is used to acquire low-dose CT images of the lungs of the target user and to assess and determine the image optimization difficulty index based on the actual radiation dose to the target user and the characteristics of the user's condition. The image optimization module is used to activate the image quality optimization channel based on the image optimization difficulty index, optimize the quality of the low-dose CT image of the lungs to generate several initial optimized CT images, and evaluate and determine the initial generation deviation characteristics. The result output module is used to generate an adaptation aggregation strategy based on the lung region weight map and the initial generation deviation features, perform image aggregation on the several initial optimized CT images, and output the CT image quality optimization results. The evaluation determines the initial generation bias characteristics, including: Several initial optimized CT images are segmented according to the preset lung region type, and several sets of initial optimized CT image blocks are output. Consistency evaluation is performed on the several initial optimized CT image block sets respectively, and several image consistency coefficients are output as initial generation deviation features; The consistency evaluation is performed on each of the several initial optimized CT image block sets, and several image consistency coefficients are output, including: A first initial optimized CT image block set is randomly selected from the plurality of initial optimized CT image block sets, and tissue registration and alignment are performed on the first initial optimized CT image block set to obtain a first aligned CT image block set; Calculate the local standard deviation, range, and signal-to-noise ratio of the first aligned CT image block set to generate a first grayscale statistical evaluation map; The gradient consistency and texture similarity of the first aligned CT image block set are extracted to form a first structural feature evaluation map; The first grayscale statistical evaluation map and the first structural feature evaluation map are pixel-level weighted summation to generate a first comprehensive inconsistency heatmap. The first image consistency coefficient is obtained by calculating the global average value of the first integrated inconsistency heatmap and added to the plurality of image consistency coefficients.

Images