A carotid ultrasound image generation method based on a generative adversarial network

Abstract

The application discloses a carotid ultrasound image generation method based on a generative adversarial network, comprising: acquiring data and constructing prior information; constructing an image generation model based on the generative adversarial network, wherein the image generation model comprises a conditional generator, a discriminator group, an anatomical consistency constraint subnetwork and a quality evaluation and screening module; inputting the data and the prior information into the image generation model, and jointly training the conditional generator, the discriminator group and the anatomical consistency constraint subnetwork; inputting target structure prior and target imaging style prior into the trained image generation model to generate an initial carotid ultrasound image; and performing quality detection and screening on the initial carotid ultrasound image through the quality evaluation and screening module. The application solves the pain points of carotid ultrasound data scarcity and poor cross-domain adaptation, and has the advantages of anatomical consistency, clinical usability and closed-loop optimization.

Description

Technical Field

[0001] This invention relates to the field of ultrasound image generation, and more specifically, to a method for generating carotid ultrasound images based on generative adversarial networks. Background Technology

[0002] In the screening and follow-up of cardiovascular and cerebrovascular diseases, carotid ultrasound is widely used for assessing indicators such as intima-media thickness (IMT) and plaque morphology due to its advantages of being non-invasive, real-time, and low-cost. However, there have long been prominent issues at the data level for algorithm development and clinical intelligent applications: compliance collection and privacy regulations restrict cross-center and cross-device data flow; differences in probes from different manufacturers, imaging presets, and operating techniques lead to significant inter-domain differences in image grayscale distribution and speckle texture, making it difficult to form a large-scale unified training dataset; and the low proportion of clinical positive samples and class imbalance further exacerbate the lack of data effectiveness. Traditional clinical collection and prospective specialized collection are inefficient, costly, and rely heavily on senior physicians for annotation, which is highly subjective. To address the issues of data scarcity and domain shift, traditional image enhancement can only perform minor perturbations in the near-neighbor space of the original image, failing to cover new anatomical variations and lesion phenotypes, and easily disrupting ultrasound speckle texture statistics; physics-based simulation methods require complex parameter settings, limiting generation efficiency, realism, and diversity. In recent years, although Generative Adversarial Networks (GANs) have been used for medical image synthesis, when directly applied to the field of carotid ultrasound, they are prone to pattern collapse due to data scarcity. Furthermore, the generation of images lacks anatomical consistency and clinical usability due to the focus on pixel-level statistical optimization alone. At the same time, the lack of targeted quality assessment and closed-loop optimization mechanisms makes it difficult to meet the requirements of downstream algorithm training for data scale, fidelity, and clinical adaptability, which seriously restricts the development of intelligent applications of carotid ultrasound. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, such as the scarcity of carotid ultrasound image data and significant differences across device domains, lack of anatomical consistency and clinical measurement usability of generated images, susceptibility to pattern collapse, incomplete quality assessment system, and lack of data closed-loop iterative optimization mechanism, this invention provides a carotid ultrasound image generation method based on generative adversarial networks.

[0004] The technical solution of this invention is as follows:

[0005] A method for generating carotid artery ultrasound images based on generative adversarial networks includes the following steps:

[0006] S1. Acquire data and construct prior information. The data includes a small number of real carotid ultrasound images. The prior information includes structural prior and imaging style prior. The structural prior is a binary or multi-class mask of the vascular cavity and the intima-media boundary. The imaging style prior is a style code that characterizes the device, probe, or gain domain information.

[0007] S2. Construct an image generation model based on generative adversarial networks. The image generation model includes a conditional generator, a discriminator group, an anatomical consistency constraint subnetwork, and a quality assessment and screening module.

[0008] S3. Input the data and the prior information into the image generation model, and jointly train the conditional generator, the discriminator group and the anatomical consistency constraint sub-network. The joint training process introduces a multi-dimensional loss function to optimize the model parameters.

[0009] S4. Input the target structure prior and the target imaging style prior into the trained image generation model to generate an initial carotid ultrasound image;

[0010] S5. The initial carotid ultrasound image is subjected to quality detection and screening through the quality assessment and screening module, and a high-fidelity carotid ultrasound image that meets the preset standard is output. The high-fidelity carotid ultrasound image is accompanied by quality grading results and uncertainty estimation.

[0011] S6. Establish a data closed-loop iteration mechanism to return the selected high-fidelity carotid ultrasound images and manually reviewed edge samples to the training dataset for iterative optimization of the model.

[0012] Furthermore, in one embodiment, the conditional generator adopts a hybrid architecture combining multi-scale U-Net, residual blocks, and self-attention, with encoding-decoding as the backbone. The encoding end uses residual convolutional groups with spectral normalization, and introduces channel attention and non-local self-attention operators at each scale. The decoding end adopts a skip connection structure and injects conditional information at the skip connection points.

[0013] Furthermore, in one embodiment, the conditional injection method of the conditional generator includes: using multi-path embedding for the structural prior, one path concatenating the structural prior with a Gaussian blurred version of the structural prior as the input first layer, and another path adjusting the feature distribution of each layer through conditional normalization; introducing a style mapping network into the imaging style prior, encoding the device domain, probe frequency band, or dynamic range as latent variables, and injecting them into the mid-to-high-level features through style convolution.

[0014] Furthermore, in one embodiment, the discriminator group includes three scale PatchGAN discriminators, covering patch-level, vascular cross-section-level, and full-frame-level regions respectively. Each discriminator receives the structural prior and the imaging style prior as conditions, and uses projection discrimination or conditional stitching to achieve the discrimination function.

[0015] Furthermore, in one embodiment, the anatomical consistency constraint subnetwork includes a structural consistency branch and a measurement consistency branch. The structural consistency branch performs semantic segmentation on the generated image and outputs a predicted structural prior, calculating the Dice loss or boundary distance loss with the input structural prior. The measurement consistency branch automatically measures the intima-media thickness based on a differentiable measuring head, constraining the intima-media thickness of the generated image to be within a clinically critical value range.

[0016] Furthermore, in one embodiment, the multidimensional loss function includes adversarial loss, L1 loss, perceptual loss, feature matching loss, anatomical consistency loss, inner-middle membrane thickness measurement consistency loss, gradient constraint loss, and gray-level histogram distribution alignment loss. The total loss is obtained by weighting and summing the various losses using weight coefficients.

[0017] Furthermore, in one embodiment, the joint training adopts a small-sample training strategy that combines transfer learning and self-supervised pre-training. Self-supervised pre-training is first performed on large-scale general ultrasound data or B-mode ultrasound frames of adjacent anatomical sites, and then fine-tuning is performed on target carotid artery ultrasound data.

[0018] Furthermore, in one embodiment, the evaluation indicators of the quality assessment and screening module are divided into three categories, including general image quality indicators, ultrasound specific indicators, and anatomical-measurement consistency indicators. The general image quality indicators are FID or LPIPS. The ultrasound specific indicators include histogram divergence, speckle contrast factor, and texture co-occurrence matrix statistics. The anatomical-measurement consistency indicators include structural segmentation accuracy, inner-middle membrane thickness measurement error, and cavity surface continuity.

[0019] Furthermore, in one embodiment, step S5 includes a human-machine collaborative review process. The quality assessment and screening module presets a quality threshold. Images higher than the quality threshold are automatically passed, while edge samples lower than the quality threshold enter the manual rapid review interface. The review results are fed back to the data closed-loop iteration mechanism.

[0020] Furthermore, in one embodiment, the method also includes a step of verification in conjunction with downstream tasks, incorporating the high-fidelity carotid ultrasound image into a real training set at a preset ratio, training downstream models for segmentation, detection, or measurement, and adjusting the training parameters or sampling strategy of the image generation model based on the performance evaluation results of the downstream models.

[0021] According to the above-described scheme, the beneficial effects of this invention are as follows: Through the core design of dual-prior construction, multi-module collaborative generation, and closed-loop optimization, it specifically addresses the core pain points of existing technologies in carotid ultrasound image generation, such as data scarcity, poor cross-device domain adaptation, lack of anatomical consistency and clinical usability of generated images, imperfect quality control system, and lack of iterative optimization mechanism: First, by leveraging the strong conditional driving force of structural priors and imaging style priors, combined with a joint training strategy, only a small number of real carotid ultrasound images are needed to stably generate large-scale images, significantly reducing the dependence on massive amounts of real labeled data. At the same time, through the domain information encoding of style priors, it effectively adapts to the image style characteristics of different devices and centers, alleviating the algorithm generalization problem caused by inter-domain differences; Second, by introducing an anatomical consistency constraint sub-network in the image generation model, it ensures that the vascular anatomy of the generated images conforms to clinical reality, and the measurement accuracy of key clinical indicators such as intima-media thickness is consistent with the actual situation. The solution achieves several key improvements: First, it closely resembles real-world images, addressing the shortcomings of traditional generative adversarial networks (GANs) that generate images that are "visually realistic but clinically useless." Second, it relies on multi-dimensional detection through quality assessment and screening modules to accurately screen high-fidelity, structurally abnormal, or texture-distorted effective samples, reducing noise interference in downstream algorithm training and ensuring data utilization efficiency. Third, through a data closed-loop iteration mechanism, the selected high-quality samples and manually reviewed edge samples are fed back into the training process, enabling continuous optimization of model performance and improving the stability and reliability of data generation. Fourth, the overall solution significantly improves the generation efficiency and supply scale of carotid ultrasound images, effectively reducing the comprehensive cost of data collection and annotation. Furthermore, the generated images come with quality grading and traceability information, meeting compliant sharing requirements while providing high-quality, large-scale adaptive data for downstream intelligent algorithms such as segmentation and detection, thus providing strong support for cross-center collaboration and the implementation of intelligent ultrasound applications in primary healthcare institutions. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a method for generating carotid ultrasound images based on generative adversarial networks in this embodiment.

[0024] Figure 2 This is a flowchart of the data processing of the image generation model in this embodiment;

[0025] Figure 3 This is a sample illustration of a carotid artery ultrasound image in this embodiment. Figure 1 ;

[0026] Figure 4 This is a sample illustration of a carotid artery ultrasound image in this embodiment. Figure 2 . Detailed Implementation

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

[0028] In the description of the embodiments, it should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0029] In the description of the embodiments, it should be noted that when a component / part is referred to as being "set on" another component / part, it can be directly set on the other component / part or there may be an intervening component / part. When a component / part is referred to as being "connected / joined" to another component / part, it can be directly connected / joined to the other component / part or there may be an intervening component / part. The term "connected / joined" as used herein can include mechanical physical connections / joinings. The term "comprising / including" as used herein refers to the presence of a feature, step, or component / part, but does not exclude the presence or addition of one or more other features, steps, or components / parts. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. Furthermore, in the description of this application, the terms "first," "second," etc., are used for descriptive purposes and to distinguish similar objects only; there is no order between them, nor should they be construed as indicating or implying relative importance. Additionally, in the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0031] I. Data Acquisition and Prior Construction

[0032] (a) Selection of real data

[0033] We selected 120 clinical carotid ultrasound grayscale images as the training baseline, with a uniform image resolution of 512×512 pixels, covering the long and short axis sections of the common carotid artery, carotid bifurcation, internal carotid artery, and external carotid artery. The data included healthy individuals, individuals with intimal-media thickening, and individuals with plaque formation (including hypoechoic, isoechoic, hyperechoic, and mixed-echoic plaques), with a positive sample rate of 30%, covering common clinical pathological scenarios. All data underwent de-identification processing, and patient privacy information was removed, complying with medical ethics and data compliance requirements.

[0034] (ii) A priori construction of structure

[0035] The defined structural prior is a binary or multi-class mask of the vascular lumen and the intima-media boundary. Its core function is to provide clear anatomical structural constraints and solve the problem of disordered image structure in existing technologies. The specific implementation is as follows:

[0036] Precision standard sample preparation: 20 images were selected as precision standard samples. Two physicians with more than 10 years of experience in carotid ultrasound diagnosis used the ITK-SNAP medical image annotation tool to manually delineate the outline of the vascular cavity, the intima-media boundary, and the plaque area (if present) to form multiple masks - the vascular cavity was labeled as label 1, the intima-media layer as label 2, the plaque as label 3, and the calcification point as label 4, which served as the "gold standard" for structural priors.

[0037] Weakly labeled sample preparation: The remaining 100 images adopt the "semi-automatic segmentation + manual review" mode. First, the pre-trained U-Net segmentation model automatically outputs the preliminary segmentation results. Then, the physician reviews only 10% of the key frames and the segmented blurred areas. After correcting the boundary deviation, a binary or multi-class mask is formed to reduce the labeling cost.

[0038] Mask optimization: Morphological processing was performed on all masks, and tiny noise points were removed by erosion and dilation operations using 3×3 convolution kernels. Gaussian smoothing was then used to optimize boundary continuity, ensuring that the anatomical structure is consistent with clinical reality.

[0039] (III) Imaging Style Prior Construction

[0040] The defined imaging style prior is a style code that characterizes device, probe, or gain domain information, used to address cross-device domain differences. The specific implementation is as follows:

[0041] Style parameter acquisition: Record the imaging parameters of each real image, including the device model (GE Logiq E9, Philips EPIQ7), probe frequency band (7–12MHz linear array probe), gain (40–60dB), dynamic range (50–70dB), and time gain compensation (TGC) curve parameters.

[0042] Style feature quantification: Extract quantitative features such as image grayscale histogram statistics (mean, variance, skewness), speckle contrast factor, and texture co-occurrence matrix parameters (energy, entropy, correlation) to capture the essence of imaging style of different devices.

[0043] Style encoding generation: A fusion of "parameter encoding + feature encoding" is adopted. Discrete parameters are converted into vectors through one-hot encoding. Continuous parameters are normalized and then concatenated with quantized feature vectors to form a 128-dimensional initial style vector. Then, a nonlinear transformation is performed through the style mapping network M to output a 64-dimensional style code z_s, thereby realizing the condensation and standardization of style information.

[0044] II. Data Preprocessing and Augmentation

[0045] To improve model robustness and data utilization, the following steps were taken to specifically address data noise and distribution issues:

[0046] Basic preprocessing: The gray values ​​of all ultrasound images are normalized to the [0,1] range, and extreme gray noise is removed by 1–99 quantile intensity clipping; the structural prior mask is pixel-level aligned with the corresponding ultrasound image, and the offset mask is translated and rotated for fine adjustment to ensure accurate structural constraints.

[0047] Multi-dimensional augmentation:

[0048] Geometric augmentation: Performs slight rotation (±10°), scaling (0.9–1.1 times), horizontal flipping, and elastic deformation (α=5, σ=0.5) to simulate changes in clinical position and probe angle;

[0049] Physics-inspired augmentation: Simulating ultrasound imaging characteristics, including TGC simulation, Gaussian noise and multiplicative noise injection, and point diffusion kernel changes, closely matching the real imaging physics process;

[0050] Adversarial augmentation: A DiffAugment strategy is employed to perform mild affine transformations, color perturbations, and CutMix operations on the input features of the generator and discriminator during training, reducing the risk of overfitting with small samples. Weakly labeled samples are used only for unpaired reconstruction and consistency loss calculation, avoiding invalid labels from interfering with training.

[0051] III. Image Generation Model Construction

[0052] (a) Construction of conditional generators

[0053] The conditional generator employs a hybrid architecture combining multi-scale U-Net, residual blocks, and self-attention. Its core solution addresses the issues of imprecise texture modeling and insufficient integration of conditional information. The specific implementation is as follows:

[0054] Overall architecture: The generator uses encoding-decoding as its main structure, with channel numbers of 64, 128, 256, and 512 respectively. Figure 1 As shown.

[0055] Encoding end design:

[0056] Basic structure: Each layer consists of residual convolutional groups with spectral normalization, including two 3×3 convolutional layers, a BatchNorm normalization layer, and a LeakyReLU activation function. Spectral normalization stabilizes the adversarial training process by constraining the spectral norm of the convolutional kernel weights.

[0057] Attention operators: Channel attention and nonlocal self-attention operators are embedded after each scale residual convolution group. Channel attention highlights key structural features and suppresses noise by calculating channel weights. Nonlocal self-attention captures long-distance texture dependencies and ensures global consistency of blood vessel morphology.

[0058] Decoding end design:

[0059] Skip-connection structure: By splicing the corresponding scale feature maps of the encoder with skip connections, high-resolution local details are transmitted and upsampling information is avoided;

[0060] Conditional injection: Conditional information is injected into the skip-connect feature splicing stage. The structural prior modulates the feature distribution through the SPADE conditional normalization layer to ensure that the generated image structure is aligned with the mask. The style encoding z_s incorporates mid-to-high-level features through the StyleConv style convolutional layer to control the image grayscale distribution and speckle characteristics.

[0061] Detail Enhancement: The final layer of the decoding end is connected in parallel detail enhancement subnet, which consists of two layers of 3×3 dilated convolution (dilation rates 2 and 4) and a Laplacian residual reconstruction layer. It learns the high-frequency correction amount ΔX and finally outputs X̂=X_base+ΔX, accurately restoring the fine echoes in the neighborhood of the inner elastic membrane and the details of the fiber cap edge, avoiding over-smoothing.

[0062] (II) Construction of the discriminator group

[0063] To address the requirements for judging the authenticity and consistency of generated images, a three-scale PatchGAN discriminator group (D1, D2, D3) is constructed:

[0064] Scale division of labor: The receptive fields are 70×70, 140×140 and global size, respectively, corresponding to plaque level, blood vessel section level and whole frame level discrimination - D1 focuses on the realism of local texture, D2 focuses on the rationality of blood vessel section structure, and D3 judges the overall anatomical consistency and style unity.

[0065] Conditional discrimination: Each discriminator adopts a projection discrimination method, incorporating structural prior S and style coding C to improve the sensitivity of "structure-texture consistency" discrimination and avoid generating images that are "visually similar but structurally abnormal";

[0066] Stable training: Introducing R1 gradient penalty and spectral normalization, and using feature matching loss optimization, the generator learns the multi-layer feature statistical distribution of the discriminator, effectively reducing the probability of mode collapse.

[0067] (III) Construction of Auxiliary Modules

[0068] Anatomical consistency constraint subnetwork A: It includes a structural consistency branch A_seg and a measurement consistency branch A_imt, which participate in joint optimization in the form of frozen parameters. A_seg adopts a lightweight U-Net architecture to predict maskŜ for semantic segmentation output of generated images, and calculates Dice loss and Hausdorff distance loss with input prior S to ensure structural consistency. A_imt automatically measures IMT based on a differentiable measuring head, constraining the measured value to fall within the clinical key value range (0.5–1.2 mm), and adding boundary smoothness, thickness continuity and abnormal mutation penalty terms to ensure the clinical measurement usability.

[0069] The quality assessment and screening module Q adopts a "multi-dimensional index fusion + discriminator reuse calibration" approach. The assessment indicators cover general indicators such as FID and LPIPS, ultrasound-specific indicators such as histogram divergence and speckle contrast factor, and anatomical-measurement consistency indicators such as structural segmentation accuracy and IMT measurement error. The discriminator intermediate layer features are reused as a scorer. After temperature calibration (τ=0.1), the score is weighted and fused with the Monte Carlo Dropout uncertainty estimate of A_imt to generate a quality score of 1.0, which is used to screen high-fidelity samples.

[0070] IV. Model Training Process

[0071] (a) Training parameter settings

[0072] Optimizer and Loss Function: The AdamW optimizer is used, with a generator learning rate of 1e-4 and a discriminator learning rate of 4e-4. Momentum parameters β1=0.5 and β2=0.999. Hinge loss is used for adversarial loss, and the total loss is a weighted sum of multi-dimensional losses, as shown in the formula:

[0073] The loss weights are λ_rec=10, λ_perc=1, λ_fm=10, λ_anat=5, λ_imt=2, λ_grad=1, and λ_hist=1, respectively.

[0074] Training strategy: Batch size is 16, and progressive training is adopted. First, train at 256×256 resolution for 100,000 steps to converge texture statistical features; then increase to 512×512 resolution for 150,000 steps. In this stage, the weights of λ_grad and λ_imt are adjusted to 2 and 3, respectively, to enhance the edge sharpness and IMT measurement consistency. ADA adversarial augmentation and generator EMA strategy (EMA coefficient 0.999) are enabled to improve model stability and generalization ability.

[0075] (II) Implementation of the training process

[0076] Parameter initialization: Initialize the parameters of the generator, discriminator group and the anatomy consistency constraint subnetwork. The parameters of A use pre-trained weights and freeze the first 50% of the layers to avoid parameter drift.

[0077] Phased training:

[0078] Phase 1 (10,000–50,000 steps): Fix the discriminator and train only the generator to learn the conditional information injection method so that the generated image structure contour basically meets the constraints.

[0079] Phase 2 (50,000–250,000 steps): Alternately train the generator and discriminator (2:1 training ratio), gradually unfreeze 50%–80% of A's parameters, and strengthen the consistency constraints between anatomy and measurement.

[0080] Phase 3 (250,000–250,000 steps): Fine-tune all parameters, focusing on optimizing the scoring accuracy of the Q module to make the quality score highly correlated with human evaluation. During training, observe changes in generated images, such as… Figure 3 and Figure 4 As shown, the training effectiveness is verified, and the model reaches a stable state after 250,000 steps.

[0081] V. Image Generation and Quality Control

[0082] (a) Large-scale data generation

[0083] Prior library expansion: The structural prior library is expanded by combining morphological transformations and pathological morphologies to cover multiple perspectives and multiple pathological morphologies. Each structural prior sample is associated with 3–5 style vector clusters {z_s} to ensure generation diversity.

[0084] Conditional sampling and generation: Sampling is performed from the (S, z_s) space according to the target distribution strategy, including lesion type equalization sampling, IMT range coverage sampling and device domain proportion adaptation sampling; the sampling results are input into the trained generator to generate initial carotid ultrasound images in batches, and latent variable jitter (±0.05) and truncation technique (0.7) are enabled to balance diversity and stability.

[0085] (II) Quality Control and Storage

[0086] Automatic quality control: The quality assessment module Q is run to screen high-fidelity images with a quality score ≥ 0.7; the structure mask and IMT measurement values ​​are automatically derived through A_seg and A_imt to form "image + derived label" pairs, which can be optionally superimposed with a Doppler synthesis head to expand application scenarios.

[0087] Human-machine collaborative review: Samples with a quality score ≥ 0.7 are automatically entered into the database; marginal samples with a score < 0.7 are checked by physicians for IMT measurement values ​​and obvious artifacts. Samples that pass the review are entered into the database, and unqualified samples are marked and returned to the training dataset, dynamically updating the threshold and weight of the Q module.

[0088] Data Management: Generated samples are stored in a dedicated database, along with metadata such as source S hash value, z_s, loss value, and quality score. A distribution monitoring dashboard is established to track sample distribution and adjust sampling strategies in a timely manner to ensure comprehensive data coverage.

[0089] VI. Downstream Task Linkage Verification and Parameter Adjustment

[0090] The core solution to the problem of insufficient adaptability between generated data and downstream tasks is to adjust the training parameters or sampling strategy of the generated model based on the performance of the downstream model. The specific implementation is as follows:

[0091] Linked verification design:

[0092] Downstream tasks include carotid ultrasound image segmentation, IMT measurement regression, and plaque detection and classification. The evaluation metrics are Dice coefficient, MAE, and sensitivity / specificity.

[0093] Control group: Downstream models were trained using only 60% of available real data;

[0094] Experimental group: The generated images were mixed in at a ratio of "synthetic:real = 2:1" to train the downstream model;

[0095] Performance comparison: A / B testing was conducted on a 40% independent real validation set to quantify the gain effect of the generated data.

[0096] Targeted adjustment strategy:

[0097] Insufficient improvement in segmentation accuracy: Increase λ_anat to 8, increase the sampling ratio of the structure prior mask, and strengthen the structural consistency constraint;

[0098] Excessive IMT measurement error: Increase λ_imt to 5, increase the sampling frequency of IMT anomaly range (>1.0mm), and optimize the measurement head constraint effect;

[0099] Insufficient sensitivity in plaque detection: Increase the proportion of difficult positive samples such as hypoechoic plaques from 15% to 30%, and adjust the positive sample quality threshold of the Q module to 0.75;

[0100] Poor cross-device generalization: Increase the training weights of style samples from different devices, optimize the expressive power of style encoding z_s, and improve cross-domain adaptation.

[0101] VII. Model Deployment and Inference

[0102] (a) Lightweight Model

[0103] To adapt to clinical deployment scenarios and reduce hardware dependence, the following lightweight optimizations were implemented:

[0104] Structural pruning: A channel pruning strategy is adopted to remove 20% of low-contribution channels in the generator and anatomy consistency constraint subnetwork, thereby reducing the number of parameters;

[0105] Knowledge distillation: Transferring the feature representation capabilities of complex models to lightweight models while ensuring that the generation quality is not significantly reduced;

[0106] Module simplification: The quality assessment module is simplified into a lightweight CNN scorer for a single forward inference, reducing inference time. The lightweight model can be deployed on ordinary GPUs (NVIDIA Tesla T4) or high-end CPUs (Intel Core i9), and batch inference speeds of up to 120 frames per second for 512×512 resolution images can meet the needs of large-scale data supply.

[0107] (II) Controllable Interface Design

[0108] After deployment, a visual and controllable generation interface is provided, supporting users to precisely configure parameters:

[0109] Target structure parameters: Upload a custom mask or select the target structure from a priori library (e.g., "carotid bifurcation + medium plaque").

[0110] Style adaptation parameters: Select the target device domain ID or directly enter the style vector z_s;

[0111] Clinical parameters: Set IMT range (e.g., 0.6–0.8 mm), plaque echogenicity level;

[0112] Output configuration parameters: Set resolution (256×256, 512×512, 768×768), number of frames generated (1–1000 frames), and random seed (for sample reproduction). Generated images have built-in "synthesis / enhancement" metadata tags and watermarks. A quality threshold (≥0.7) and an upper limit of uncertainty (≤0.3) can be set. Samples that do not meet the standards are not output to the training library, ensuring data quality.

[0113] The present invention has been described above with reference to the accompanying drawings. Obviously, the implementation of the present invention is not limited to the above-described manner. Any improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A method for generating carotid artery ultrasound images based on generative adversarial networks, characterized in that, Includes the following steps: S1. Acquire data and construct prior information. The data includes a small number of real carotid ultrasound images. The prior information includes structural prior and imaging style prior. The structural prior is a binary or multi-class mask of the vascular cavity and the intima-media boundary. The imaging style prior is a style code that characterizes the device, probe, or gain domain information. S2. Construct an image generation model based on generative adversarial networks. The image generation model includes a conditional generator, a discriminator group, an anatomical consistency constraint subnetwork, and a quality assessment and screening module. S3. Input the data and the prior information into the image generation model, and jointly train the conditional generator, the discriminator group and the anatomical consistency constraint sub-network. The joint training process introduces a multi-dimensional loss function to optimize the model parameters. S4. Input the target structure prior and the target imaging style prior into the trained image generation model to generate an initial carotid ultrasound image; S5. The initial carotid ultrasound image is subjected to quality detection and screening through the quality assessment and screening module, and a high-fidelity carotid ultrasound image that meets the preset standard is output. The high-fidelity carotid ultrasound image is accompanied by quality grading results and uncertainty estimation. S6. Establish a data closed-loop iteration mechanism to return the selected high-fidelity carotid ultrasound images and manually reviewed edge samples to the training dataset for iterative optimization of the model.

2. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 1, characterized in that, The conditional generator adopts a hybrid architecture combining multi-scale U-Net, residual blocks, and self-attention, with an encoder-decoder backbone. The encoder uses residual convolutional groups with spectral normalization and introduces channel attention and non-local self-attention operators at each scale. The decoder adopts a skip connection structure and injects conditional information at the skip connection points.

3. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 2, characterized in that, The conditional injection method of the conditional generator includes: multi-path embedding for the structural prior, one path concatenating the structural prior with the Gaussian blurred version of the structural prior as the input first layer, and another path adjusting the feature distribution of each layer through conditional normalization; introducing a style mapping network into the imaging style prior, encoding the device domain, probe frequency band or dynamic range as latent variables, and injecting them into the mid-to-high-level features through style convolution.

4. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 1, characterized in that, The discriminator group includes three scales of PatchGAN discriminators, covering patch-level, vascular cross-section-level, and full-frame-level regions respectively. Each discriminator receives the structural prior and the imaging style prior as conditions, and uses projection discrimination or conditional stitching to achieve the discrimination function.

5. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 1, characterized in that, The anatomical consistency constraint subnetwork includes a structural consistency branch and a measurement consistency branch. The structural consistency branch performs semantic segmentation on the generated image and outputs a predicted structural prior, calculating the Dice loss or boundary distance loss with the input structural prior. The measurement consistency branch automatically measures the intima-media thickness based on a differentiable measuring head, constraining the intima-media thickness of the generated image to be within the clinically critical value range.

6. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 1, characterized in that, The multidimensional loss function includes adversarial loss, L1 loss, perceptual loss, feature matching loss, anatomical consistency loss, inner-middle membrane thickness measurement consistency loss, gradient constraint loss, and gray-level histogram distribution alignment loss. The total loss is obtained by weighting and summing the various losses through weight coefficients.

7. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 1, characterized in that, The joint training adopts a small-sample training strategy that combines transfer learning and self-supervised pre-training. Self-supervised pre-training is first performed on large-scale general ultrasound data or B-mode ultrasound frames of adjacent anatomical sites, and then fine-tuning is performed on target carotid artery ultrasound data.

8. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 1, characterized in that, The evaluation indicators of the quality assessment and screening module are divided into three categories, including general image quality indicators, ultrasound specific indicators, and anatomical-measurement consistency indicators. The general image quality indicators are FID or LPIPS. The ultrasound specific indicators include histogram divergence, speckle contrast factor, and texture co-occurrence matrix statistics. The anatomical-measurement consistency indicators include structural segmentation accuracy, inner-middle membrane thickness measurement error, and cavity continuity.

9. The method for generating carotid ultrasound images based on generative adversarial networks according to claim 1, characterized in that, Step S5 includes a human-machine collaborative review process. The quality assessment and screening module presets a quality threshold. Images that are higher than the quality threshold are automatically passed, while edge samples that are lower than the quality threshold enter the manual quick review interface. The review results are fed back to the data closed-loop iteration mechanism.

10. The method for generating carotid artery ultrasound images based on generative adversarial networks according to claim 1, characterized in that, It also includes a step of linkage verification with downstream tasks, in which the high-fidelity carotid ultrasound images are integrated into a real training set at a preset ratio to train downstream models for segmentation, detection or measurement, and the training parameters or sampling strategies of the image generation model are adjusted according to the performance evaluation results of the downstream models.

Images