Lower limb artery ultrasonic standardization examination system and method based on AI vertical large model

Abstract

The invention discloses a lower limb artery ultrasonic standardization examination system and method based on an AI vertical large model, and the method comprises the following steps: S1, collecting standard and wrong images, carrying out the labeling processing, and constructing an image data set; s2, carrying out image preprocessing, and constructing a multi-modal training sample set; s3, constructing an AI vertical large model, and fusing the image and trajectory features for training; s4, deploying the model to a front-end inspection device, performing real-time scoring and augmented reality guidance, and outputting a structured report; s5, uploading the report and the image to a cloud quality control platform, executing consistency analysis and quality evaluation, and triggering expert re-checking; and S6, based on quality control feedback, locally updating AI model parameters, uploading the AI model parameters to aggregation nodes for unified integration, and completing distributed collaborative optimization. The invention aims to realize intelligent guidance, image standardized acquisition and structured diagnosis automatic generation of the lower limb artery ultrasonic examination process, and improve the examination quality and efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence-assisted medical imaging and ultrasonic diagnosis automation, and particularly relates to a lower extremity artery ultrasound standardized examination system and method based on an AI vertical large model. Background Art

[0002] With the increasing incidence of peripheral artery disease globally, lower extremity artery ultrasound, as a non-invasive, safe, and real-time imaging examination method, plays an increasingly important role in clinical diagnosis, disease monitoring, and treatment evaluation. Especially in the management of patients with chronic diseases such as diabetes, atherosclerosis, and vascular stenosis, timely acquisition of accurate lower extremity artery structure and blood flow parameters is crucial for formulating subsequent treatment plans.

[0003] However, in current clinical practice, there are multiple technical and application challenges in lower extremity artery ultrasound examinations. Firstly, the quality of ultrasound images highly depends on the operator's experience level. There are significant differences in probe angles, pressures, and scanning paths among different operators during image acquisition, which easily leads to unstable image quality and is difficult to meet the standardized requirements. Secondly, there is a lack of real-time quality feedback and operation guidance mechanisms during the image acquisition process. Beginners or non-professionals are difficult to accurately identify whether the acquisition is up to standard when completing the examination, resulting in repeated scans or missing information. Thirdly, ultrasound image data is generally stored in an unstructured manner, and traditional examination reports rely on manual entry, with problems such as strong subjectivity, low efficiency, and large result heterogeneity, which is not conducive to remote quality control and subsequent AI modeling utilization.

[0004] In recent years, significant progress has been made in the field of medical image analysis with artificial intelligence and deep learning. Existing research has attempted to use convolutional neural networks for automatic segmentation, lesion recognition, and blood flow assessment of ultrasound images, and some algorithms have achieved good performance in laboratory environments. However, most research still remains in the offline processing stage, lacking deep integration with the ultrasound examination operation process and failing to solve the key links in the complete closed loop of "image acquisition - scoring - guidance - diagnosis - feedback". For example, existing systems only focus on the quality scoring of post-image processing, unable to provide timely intervention suggestions for the image generation process, nor can they achieve structured report output and diagnostic closed-loop optimization.

[0005] In terms of data processing, the traditional training mode relies on centralized modeling of all samples on the central server, which faces the problems of privacy protection and data silos in the environment of ultrasound devices deployed on multiple terminals. Patient image data is highly sensitive. Directly uploading it to the cloud for model training may violate data compliance requirements and it is difficult to form a sustainable iterative optimization mechanism. Although federated learning has gradually gained attention in recent years as a new path for privacy-preserving machine learning, its actual integration path, model synchronization efficiency, and collaborative update strategy in ultrasound diagnostic systems still lack systematic design and engineering verification.

[0006] Therefore, how to provide a lower extremity artery ultrasound standardized examination system and method based on an AI vertical large model is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose a lower extremity artery ultrasound standardized examination system and method based on an AI vertical large model. The present invention integrates artificial intelligence technologies such as multi-modal data processing, deep image understanding, reinforcement learning guidance, structured diagnosis generation, quality control analysis, and federated learning optimization, aiming to solve key problems in the existing lower extremity artery ultrasound examination process, such as image acquisition relying on manual experience, unstable quality, lack of real-time guidance in operation, unstructured report results, low quality control efficiency, and inability to close the loop in model optimization.

[0008] According to the lower extremity artery ultrasound standardized examination method based on the AI vertical large model of the embodiment of the present invention, the following steps are included: S1. Collect lower extremity artery ultrasound image data including standard images and error images, perform annotation processing on the lower extremity artery ultrasound image data, and construct an image data set; S2. Preprocess the image data set, use the U-Net network to extract image features of blood vessel boundaries and plaque structures, model the probe operation sequence based on the Manus framework, and construct a multi-modal training sample set; S3. Construct an AI vertical large model including a perception layer, a decision layer, and an output layer. The perception layer extracts image features based on the ResNet-50 network, the decision layer uses a long short-term memory network to model the probe trajectory and combines a proximal policy optimization algorithm for action scoring and error correction feedback, the output layer generates a structured diagnosis result, and uses the multi-modal training sample set for training; S4. Integrate the trained AI vertical large model into the front-end examination device, collect inspection images and operation data in real time, input them into the AI vertical large model for quality scoring and guidance feedback, and generate a structured report; S5. Upload the structured report and inspection images to the cloud quality control platform for consistency analysis and quality assessment, output the quality control feedback results. When the cloud quality control platform identifies difficult or abnormal situations, trigger remote expert review, complete the review, and return the review report. S6. Based on the quality control feedback results, perform local parameter updates on the AI vertical large model, complete distributed training using the federated learning mechanism, and upload the updated results to the aggregation node for unified integration and optimization to build a collaborative update process with privacy protection capabilities.

[0009] Optionally, the lower limb artery ultrasound image data specifically includes image data with multiple angles, different pressures, section positions, and blood flow parameters, including standard images and error images.

[0010] Optionally, the annotation process for the lower limb artery ultrasound image data refers to adding corresponding annotation label information of the probe angle, pressure value, section position, and hemodynamic parameters to each image.

[0011] Optionally, the specific steps of S2 include: S21. Perform normalization processing on the constructed image dataset, uniformly map the pixel values of each image to the interval [0, 1], and perform unified size resampling on the image data with inconsistent sizes to form standard format image data that can be input into the network. S22. Input the standard format image data into the U-Net network. Use the encoder part to extract low-level and high-level image semantic features, and the decoder part to perform upsampling and feature fusion to output a segmentation mask image including blood vessel boundaries, plaque contours, and lumen structures. Use the label mask obtained by manual annotation in the image dataset for cross-entropy loss function supervision to construct an image feature set denoted as , where represents the feature vector extracted from the -th image data, , and n is the total number of image data in the image dataset. S23. Synchronously record the probe operation information of the operator during the image acquisition process, including the probe position coordinate sequence , the direction angle sequence , and the probe contact pressure sequence , where represents the time step, is the position of the probe in the horizontal axis direction at the t-th time step, is the position of the probe in the vertical axis direction at the t-th time step, is the depth of the probe in the vertical direction at the t-th time step, is the angle in the horizontal rotation direction at the t-th time step, is the angle in the vertical tilt direction at the t-th time step, is the contact pressure value of the probe acting on the skin surface at the t-th time step; S24. Input the probe position coordinate sequence, direction angle sequence, and probe contact pressure sequence into the time series modeling module constructed by the Manus framework. Model the dynamic behavior of the probe operation through the LSTM network, extract time correlation features, and output the probe trajectory feature vector ; S25. Combine the image feature set , the probe trajectory feature vector and the annotation label information to construct a multi-modal training sample set, and each sample is represented as a triple.

[0012] Optionally, the specific steps of S3 are as follows: S31. Construct the AI vertical large model structure, which is divided into a perception layer, a decision layer, and an output layer, for image feature extraction, probe trajectory modeling, and structured diagnosis generation respectively; S32. Input the image feature set into the ResNet-50 network in the perception layer to extract deep semantic features and generate deep semantic feature vectors;

[0013] S33. Input the probe trajectory feature vector into the long short-term memory network LSTM to extract time-dependent features and generate time-dependent feature vectors;

[0014] S34. Concatenate the deep semantic feature vector and the time-dependent feature vector to form a fusion vector, which is used as the input representation for policy learning and diagnosis generation; S35. Define a task decomposition-based immediate reward function ; S36. Construct a structure-aware advantage function based on the task decomposition reward ; S37. Construct a state-aware clipping range adjustment mechanism to jointly adjust the clipping coefficient according to the score fluctuation and reward change ; S38. Based on the structure-aware advantage function and the clipping coefficient , construct the proximal policy optimization objective function ; S39. Input the fusion vector into the policy network, output the policy distribution , and sample an action from it for probe direction, angle, and pressure guidance generation; S310. Input the fusion vector Synchronously input to the multi-layer perceptron in the output layer to output a structured diagnostic result The result includes information on the degree of vascular stenosis, plaque status, and blood flow velocity; S311. Construct a combined training loss function ; S312. Use a multi-modal training sample set , and jointly optimize the parameter set through backpropagation and gradient descent until the loss function converges to complete the training process of the AI vertical large model.

[0015] Optionally, the S4 specifically includes: S41. Load and integrate the trained AI vertical large model into the front-end inspection device to receive ultrasonic images and operation inputs in real time; S42. During the inspection process, collect the ultrasonic images of the patient's lower limb arteries and the probe operation data, and synchronously input them into the AI vertical large model for analysis and processing; S43. The AI vertical large model automatically scores the quality of the currently collected images, and judges whether the current probe operation meets the requirements for obtaining standard images according to the scoring results; S44. If the image quality scoring result is lower than the set threshold, trigger a prompt for insufficient image quality. The prompt information is automatically fed back to the operator interface, and visual information on the probe adjustment direction, angle, or pressure is generated in the inspection interface through the augmented reality device and displayed to the operator; S45. After the image quality meets the requirements, automatically generate a structured report containing standard diagnostic elements based on the output of the AI vertical large model.

[0016] Optionally, the S5 specifically includes: S51. Upload the structured report generated by the AI vertical large model and the corresponding inspection images to the cloud quality control platform through the front-end inspection device to complete the remote transmission of inspection data; S52. Receive and parse the structured report and inspection images in the cloud quality control platform, and analyze the content consistency between the two, including the matching of text conclusions and image diagnostic bases; S53. Based on the dimensions of image standards, parameter integrity, and report element accuracy set by the platform, perform quality assessment operations on the graphic and text data to form a quality evaluation result; S54. The cloud quality control platform combines the consistency analysis result and the quality assessment result to output a standardized quality control feedback result and mark whether there are suspicious or non-compliant items; S55. When difficult cases or abnormal situations such as inconsistent graphics and texts, diagnostic deviations, and information omissions are identified in the quality control feedback, the cloud quality control platform automatically triggers the remote expert review process. S56. Qualified remote experts complete the review of the uploaded reports and images within the preset time, and output the final review report. The review conclusion is returned to the original examination terminal through the platform.

[0017] Optionally, the S6 specifically includes: S61. Based on the quality control feedback results output by the cloud quality control platform, identify the AI vertical large model in the front-end inspection device that needs to be updated, and start the local update process of the AI vertical large model. S62. Without the need to upload the original image data and report information, use the labeled tag data in the quality control feedback results to perform incremental training on the local AI vertical large model, and generate the locally optimized AI vertical large model parameters. S63. After completing the local training, upload the optimized AI vertical large model parameters to the federated aggregation node in an encrypted form, and no transmission of any original images, trajectories, or structured text information is involved in the process. S64. In the federated aggregation node, receive the local AI vertical large model parameters from multiple front-end devices, and perform unified integration and weight optimization according to the participation of each terminal and the sample feature distribution, and generate new global AI vertical large model parameters. S65. Synchronously distribute the updated global AI vertical large model parameters to the participating terminals, and replace the original parameter configuration to achieve the unified version iteration of the AI vertical large model in each front-end inspection device. S66. Repeat the local training, parameter upload, aggregation update, and AI vertical large model synchronization processes within the set federated training cycle to form a distributed model collaborative update process with privacy protection capabilities and continuous self-optimization.

[0018] A lower limb artery ultrasound standardized inspection system based on an AI vertical large model according to an embodiment of the present invention includes the following modules: An image and operation data acquisition module, which is used to acquire images and operation data during the lower limb artery ultrasound inspection process. An AI vertical large model module, which is used to perform image feature extraction, trajectory modeling, quality scoring, guidance feedback, and structured diagnosis generation. An augmented reality guidance module, which is used to generate visual prompt information through an augmented reality device to guide the operator to adjust the probe direction and angle when the image quality score is lower than the threshold. A structured report generation module, which is used to automatically generate a structured inspection report containing key diagnostic elements according to the output results of the AI vertical large model. A cloud quality control analysis module, which is used to receive structured reports and examination images, perform consistency analysis and quality assessment, and output quality control feedback results; A remote expert review module, which is used to trigger an expert remote review process after identifying difficult or abnormal situations in the quality control feedback and output a review report; A federated learning update module, which is used to perform local parameter fine-tuning on the AI vertical large model in each terminal based on the quality control feedback results and complete distributed model optimization through a federated aggregation mechanism.

[0019] The beneficial effects of the present invention are as follows: By constructing a lower extremity artery ultrasound standardized examination system and method based on an AI vertical large model, the present invention realizes the automation, standardization and intelligent closed-loop of the entire examination process, and achieves remarkable beneficial effects compared with the prior art. First of all, the present invention breaks the high dependence on manual experience in traditional ultrasound examinations. Through multi-modal data fusion, behavior modeling and model scoring mechanisms, it realizes the automatic evaluation of image quality and operation guidance feedback, significantly reduces the influence of operator level differences on image acquisition results, improves the ability of beginners or grass-roots operators to obtain high-quality images, and effectively improves the standardization and controllability of the examination process.

[0020] Secondly, the AI vertical large model constructed by the present invention not only has the ability to deeply understand images and operation trajectories, but also models and optimizes the feedback of the probe operation path through a reinforcement learning mechanism, enabling the operator to obtain real-time error correction and guidance prompts during the image acquisition process, reducing the number of repeated scans while improving the image compliance rate. On this basis, the system automatically generates a structured ultrasound diagnosis report, which not only reduces the subjectivity and workload of manual records, but also improves the standardization and normalization of the report content, providing a good foundation for subsequent remote diagnosis, review and medical research.

[0021] Furthermore, by deploying a cloud quality control platform, the present invention performs consistency analysis and graphic and text quality assessment on the uploaded structured reports and original image data, can effectively discover problems such as inconsistent images and report content, missing key information or suspicious diagnosis conclusions, timely output standardized quality control feedback results, and support the automatic triggering of a remote expert review process in case of abnormalities or difficult situations, realizing remote medical collaboration and multi-level quality control closed-loop, and providing a reliable quality guarantee mechanism for areas with unbalanced medical resources.

[0022] In terms of model optimization, the present invention introduces a federated learning mechanism, locally fine-tunes the parameters of the terminal model based on the quality control feedback results, and uploads the encrypted updated parameters to the aggregation node after training is completed for unified integration and synchronous distribution, so as to achieve the collaborative update and continuous optimization of the global model while protecting data privacy without uploading the original images or diagnostic data. This mechanism breaks through the bottleneck of the traditional "centralized" training mode, enabling the AI model to have the ability of adaptive evolution in multiple terminals and scenarios and adapt to the application requirements of different populations, devices and environments.

[0023] In summary, the present invention has formed a complete intelligent ultrasound examination solution in aspects such as image acquisition standardization, intelligent operation feedback, structured report generation, remote quality control review, and privacy protection model optimization. It not only effectively improves the examination efficiency and diagnostic accuracy, but also provides an efficient, safe and sustainable optimization technical path for primary healthcare, remote ultrasound, and intelligent assisted diagnosis, and has good clinical application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is a flowchart of the lower limb artery ultrasound standardization examination method based on the AI vertical large model proposed by the present invention; Figure 2 is a schematic structural diagram of the lower limb artery ultrasound standardization examination system based on the AI vertical large model proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0026] Refer to Figure 1 , the lower limb artery ultrasound standardization examination method based on the AI vertical large model includes the following steps: S1. Collect lower limb artery ultrasound image data including standard images and error images, perform annotation processing on the lower limb artery ultrasound image data, and construct an image data set; S2. Preprocess the image data set, use the U-Net network to extract the image features of the vascular boundary and plaque structure, model the probe operation sequence based on the Manus framework, and construct a multimodal training sample set; S3. Construct an AI vertical large model including a perception layer, a decision-making layer, and an output layer. The perception layer extracts image features based on the ResNet-50 network. The decision-making layer uses a long short-term memory network to model the probe trajectory and combines the proximal policy optimization algorithm for action scoring and error correction feedback. The output layer generates structured diagnostic results and is trained using a multi-modal training sample set. S4. Integrate the trained AI vertical large model into the front-end inspection device, and in real-time collect inspection images and operation data and input them into the AI vertical large model for quality scoring and guiding feedback to generate a structured report. S5. Upload the structured report and inspection images to the cloud quality control platform for consistency analysis and quality assessment, and output the quality control feedback result. When the cloud quality control platform identifies difficult or abnormal situations, it triggers remote expert review, completes the review, and returns the review report. S6. Based on the quality control feedback result, perform local parameter updates on the AI vertical large model, complete distributed training using the federated learning mechanism, and upload the updated results to the aggregation node for unified integration and optimization to build a collaborative update process with privacy protection capabilities.

[0027] Through the construction of a standardized lower extremity artery ultrasound inspection method based on an AI vertical large model, the present invention realizes a complete closed-loop process of ultrasound image acquisition, quality scoring, guiding feedback, structured report generation, cloud quality control, and model self-optimization, and has significant beneficial effects. The system integrates image features and probe trajectories, and uses deep learning and reinforcement learning algorithms to achieve real-time scoring of image quality and intelligent guidance of the operator's operation behavior, effectively improving the standardized acquisition rate of images. Through the output of structured diagnostic results, it simplifies report writing and improves accuracy and consistency; the cloud quality control platform supports graphic consistency analysis and remote expert review, enhancing the credibility of diagnosis. In terms of model update, the federated learning mechanism is used to achieve local fine-tuning and parameter aggregation of each terminal model, improving the generalization ability of the model while protecting data privacy, and forming a collaborative optimization system with privacy protection capabilities. This method significantly improves the standardization degree and intelligent diagnosis efficiency of ultrasound inspection, and is applicable to multi-terminal and multi-scenario deployment.

[0028] In this embodiment, the lower extremity artery ultrasound image data specifically includes image data with multi-angles, different pressures, section positions, and blood flow parameters, including standard images and error images.

[0029] In this embodiment, the annotation processing of the lower extremity artery ultrasound image data refers to adding annotation label information of the corresponding probe angle, pressure value, section position, and hemodynamic parameters to each image.

[0030] In this embodiment, the specific steps of S2 include: S21. Normalize the constructed image dataset, uniformly map the pixel values of each image to the interval [0, 1], and perform resampling with a unified size on the image data with inconsistent sizes to form image data in a standard format that can be input into the network; S22. Input the standard format image data into the U-Net network. Use the encoder part to extract low-level and high-level image semantic features, and the decoder part to perform upsampling and feature fusion to output a segmentation mask image including blood vessel boundaries, plaque contours, and lumen structures. The use of the encoder part to extract low-level and high-level image semantic features and the decoder part to perform upsampling and feature fusion means that through the convolutional operations and max-pooling operations stacked in multiple layers in the encoder, gradually extract the edge texture, local structure, and high-level semantic information of the image, and in the decoder, sequentially use the upsampling method to restore the resolution of the feature map. At the same time, splice the feature map of the corresponding encoder layer with the decoder feature map through skip connection, then extract joint features through convolutional fusion, and use the label mask obtained by manual annotation in the image dataset for cross-entropy loss function supervision to construct an image feature set denoted as , where represents the feature vector extracted from the th image data, and n is the total number of image data in the image dataset; S23. Synchronously record the probe operation information of the operator during the image acquisition process, including the probe position coordinate sequence , the direction angle sequence , and the probe contact pressure sequence , where represents the time step, is the position of the probe in the horizontal axis direction at the t-th time step, is the position of the probe in the vertical axis direction at the t-th time step, is the depth of the probe in the vertical direction at the t-th time step, is the angle in the horizontal rotation direction at the t-th time step, is the angle in the vertical tilt direction at the t-th time step, is the contact pressure value of the probe acting on the skin surface at the t-th time step; S24. Input the probe position coordinate sequence, direction angle sequence, and probe contact pressure sequence into the time series modeling module constructed by the Manus framework, model the dynamic behavior of the probe operation through the LSTM network, extract time correlation features, and output the probe trajectory feature vector The operation dynamic behavior of the probe is modeled by an LSTM network, and the extraction of time-correlation features means that the multi-dimensional time-series data of the position coordinates, orientation angles, and contact pressures of the probe during the inspection process are used as input sequences and fed into the long short-term memory network. The gating mechanism is used to model the state changes of the operation actions between different time steps, capture the potential time dependencies and behavior patterns of the operator during the process of moving, adjusting, and pressing the probe, and finally extract the trajectory feature vectors reflecting the stability, continuity, and standardization of the probe operation; S25. Combine the image feature set , the probe trajectory feature vector and the labeled tag information to construct a multi-modal training sample set, and each sample is represented as a triple.

[0031] In the process of constructing the multi-modal training samples of the present invention, a dual-channel fusion strategy of image feature extraction and probe operation modeling is introduced, which significantly improves the global understanding and behavior discrimination ability of the AI model for the ultrasound inspection scenario, and has good practical value and technical advantages. Through image normalization and size resampling, the consistent preprocessing of images from different sources is realized, ensuring the format standardization and training stability of the input network; the U-Net network is used to extract multi-level semantic features including blood vessel boundaries, plaque structures, etc., enhancing the model's segmentation and recognition ability of lesion morphology. At the same time, combined with the Manus framework and the LSTM network, the dynamic operation information of the probe during the image acquisition process is modeled in time series, accurately depicting the operator's behavior trajectory and time series features, effectively supplementing the lack of traditional image processing methods in understanding spatial behavior. Finally, through the fusion of image features, operation trajectories, and label information, high-quality multi-modal training samples are constructed, providing a sufficient and accurate joint perception basis for the training of subsequent AI vertical large models, and improving the robustness and operation feedback ability of the model.

[0032] In this embodiment, the specific steps of S3 are as follows: S31. Construct the structure of the AI vertical large model, which is divided into a perception layer, a decision layer, and an output layer, for image feature extraction, probe trajectory modeling, and structured diagnosis generation respectively; S32. Input the image feature set into the ResNet-50 network in the perception layer to extract deep semantic features and generate deep semantic feature vectors;

[0033] S33. Input the probe trajectory feature vector into the long short-term memory network LSTM to extract time-dependent features and generate time-dependent feature vectors;

[0034] S34. Concatenate the deep semantic feature vector and the time-dependent feature vector to form a fusion vector, which serves as the input representation for policy learning and diagnostic generation; S35. Define a task-decomposed immediate reward function : ; where represents the image quality score, represents the consistency score between the probe path and the standard trajectory, represents the score gain after the operator's response, , and are weight coefficients; The practical significance of the task-decomposed immediate reward function formula is to split the quality evaluation objectives in the ultrasonic image acquisition process into multiple dimensions with practical operational significance, so as to achieve refined guidance and evaluation of the operator's behavior. This reward function divides the immediate feedback signal into three parts: the image quality score, the consistency score between the probe path and the standard trajectory, and the score of the operator's response effect to the system guidance, comprehensively covering key factors such as whether the acquired image is clear, whether the operation is standardized, and whether the guidance is effectively executed. By organically integrating these sub-goals, the system can real-time perceive the impact of each operator's operation behavior on the image result, and accordingly dynamically adjust the feedback and scoring mechanism. Compared with the traditional single reward mode, this task decomposition strategy is closer to the actual clinical process, can effectively improve the model's recognition ability for non-standard operations and the sensitivity to the image acquisition quality, thus providing a solid foundation for constructing a highly reliable guidance strategy and accurately outputting structured diagnostic results. This design ensures that the system has a clear optimization direction in the training stage and has a practical and feasible operation intervention ability in the application stage.

[0035] S36. Based on the task-decomposed reward, construct a structure-aware advantage function, and the advantage estimation expression is as follows: ; where represents the estimated value of the structure-aware advantage function at the t-th time step, is the output of the state value function for the current state , is the output of the state value function for the next state , and the state value function is modeled by a neural network, with the input being the fusion vector and the output being the estimated return value for the corresponding state, and respectively represent the structure clarity score function values of the images corresponding to the states and , is the discount factor, is the advantage smoothing factor, represents the estimated value of the structure-aware advantage function at the (t + 1)-th time step; The practical significance of the structure-aware advantage function is to construct a measure of action advantage that is closer to the actual task requirements of ultrasound examination by combining image quality change information and state value evaluation. In traditional reinforcement learning, the advantage function is usually calculated only based on rewards and state values, ignoring the important influence of the structural changes of the image itself on the model's judgment. In the scenario of ultrasound image acquisition, factors such as the clarity of the image structure and the stability of the edges directly affect the diagnostic quality. The present invention introduces the change amount of image structure clarity as part of the advantage function, which can effectively reflect the improvement degree of the image structure before and after an operator's certain action, thereby enhancing the model's recognition ability of "beneficial operations". This advantage function not only considers the immediate reward and value difference brought by the current action, but also integrates the change trend at the image structure level, making the model more inclined to retain those operation behaviors that can stably improve the image quality during policy learning. This design that integrates image perception and reinforcement learning value significantly enhances the behavior discrimination ability and training stability of the policy network, and is the key basis for realizing high-quality image-guided generation and operation policy learning.

[0036] S37. Construct a state-aware shear range adjustment mechanism to jointly adjust the shear coefficient according to the score fluctuation and reward change : ; wherein, is the initial shear range, is the score fluctuation factor, is the reward change factor, is the standard deviation of the image score, is the previous step reward mean, is the immediate reward value at the previous time step; The practical significance of the state-aware clipping range adjustment mechanism lies in introducing dynamic constraints of image quality fluctuations and operation stability into the policy optimization process in reinforcement learning, making the policy update more stable and adaptable. During the ultrasonic image acquisition process, the image quality score fluctuates with the changes in the operator's gestures, and at the same time, the impact of the operation behavior on the model reward also varies with the scenario and the operator's level. This mechanism dynamically adjusts the amplitude of the policy update by sensing the standard deviation of the image score and the trend of reward changes, avoiding the model from making excessive or unreliable parameter updates when the score fluctuates greatly or the reward changes violently, thereby reducing the problems of policy degradation and training instability. Compared with the traditional method of a fixed clipping range, the clipping coefficient of the present invention has state adaptability, and can flexibly shrink or relax the policy update boundary according to the current environmental changes, which not only ensures the learning efficiency of the model in a high-confidence state, but also enhances the safety and robustness of convergence in a low-quality state. This mechanism plays a key regulatory role in the model training process, effectively improving the adaptability of the policy network to the actual scenario of ultrasonic image acquisition.

[0037] S38. Based on the structure-aware advantage function and the clipping coefficient , construct the proximal policy optimization objective function : ; where is the policy ratio, represents the calculation of the statistical expectation for multiple samples, is the minimum operation, is the clipping operation on the policy ratio ; The practical significance of the policy optimization objective function lies in effectively controlling the amplitude of policy updates during the reinforcement learning process, thereby ensuring the stability of model training and the reliability of policy output. In the scenario of ultrasonic image acquisition, the operator's behavior has a certain degree of uncertainty. If the policy network updates too quickly or fluctuates too much during training, it is easy to cause the model to experience policy collapse or failure. Therefore, by setting a clipping range, this objective function restricts the change amplitude between the old and new policies, making each update within a controllable range and preventing the model from deviating from the optimal direction due to abnormal local rewards in the early training stage. At the same time, this function combines the advantage function calculated in the previous stage to evaluate the effect of the current action, balancing exploration and stability while ensuring learning efficiency. The present invention further introduces a state-aware clipping mechanism to dynamically adjust the clipping range according to image quality fluctuations and reward changes, achieving adaptive regulation of the intensity of policy changes. This mechanism enables the model to accelerate the learning speed when the image quality is good or the operation is stable, and reduce the update intensity when the image fluctuates violently or the behavior is uncertain, thereby enhancing the robustness and generalization ability of the policy network, which is the core component for ensuring the intelligent evolution of the guiding policy.

[0038] S39. Input the fusion vector into the policy network to output a policy distribution and sample an action from it for the generation of probe direction, angle, and pressure guidance; S310. Synchronously input the fusion vector into the multi-layer perceptron in the output layer to output a structured diagnostic result that includes information on the degree of vascular stenosis, plaque status, and blood flow velocity; S311. Construct a joint training loss function : ; where is the structured label, CrossEntropy is the classification loss, is the loss balance coefficient; The practical significance of the joint training loss function lies in achieving the collaborative optimization of the policy guidance ability and the structured diagnosis accuracy, constructing a unified training objective, and improving the overall performance of the model. During the lower extremity arterial ultrasound examination, the system needs to learn a reasonable probe guidance strategy through reinforcement learning and accurately output structured diagnosis results such as vascular stenosis and plaque status through supervised learning. This loss function weights and fuses the policy optimization objective and the structured classification objective, enabling the model to continuously improve the quality score and operation feedback ability of the action strategy during the training process and ensuring that the output diagnosis results have medical accuracy and integrity. Through this multi-task joint optimization design, the model can learn both "how to operate" and "how to judge", thus possessing stronger intelligent guidance and diagnostic decision-making abilities. At the same time, an adjustable weight coefficient is introduced, which can flexibly adjust the proportion of policy learning and diagnostic classification in the total loss according to task requirements, realizing the dynamic regulation of the training focus in different stages. This mechanism effectively solves the problem that a single optimization objective is prone to causing model bias and is the key foundation for ensuring the unity of the intelligent guidance and diagnostic accuracy of the model of the present invention.

[0039] S312. Use a multi-modal training sample set , and jointly optimize the parameter set through backpropagation and gradient descent until the loss function converges to complete the training process of the AI vertical large model.

[0040] The present invention realizes the deep integration of lower extremity arterial ultrasound image understanding, operator behavior modeling, and structured diagnosis generation by constructing an AI vertical large model with a perception layer, a decision-making layer, and an output layer, and has significant beneficial effects. This model fuses image features and probe trajectory information, uses ResNet-50 and LSTM networks to extract spatial and temporal features respectively, and enhances the model's scoring ability for operation quality and sensitivity to image structures through a task decomposition reward mechanism and a structure-aware advantage function. By introducing a state-aware shear coefficient and a structure clarity factor, the policy network is adaptively adjusted to image fluctuations and operation behaviors during the training process, improving training stability and feedback accuracy. On the basis of optimizing the reinforcement learning strategy, a joint loss function is constructed by combining structured label supervised learning, taking into account both the guidance strategy and diagnostic accuracy, and ensuring the integrity and consistency of the output results. This model has the abilities of intelligent scoring, self-correction, and structured output, providing a core intelligent engine for realizing the full automation and standardization of the entire ultrasound examination process.

[0041] In this embodiment, the specific steps of S4 include: S41. Load and integrate the trained AI vertical large model into the front-end examination device to receive ultrasound images and operation inputs in real time; S42. During the inspection process, collect the lower limb artery ultrasound images of the patient and the probe operation data, and synchronously input them into the AI vertical large model for analysis and processing; S43. The AI vertical large model automatically scores the quality of the currently collected images, and judges whether the current probe operation meets the standard image acquisition requirements according to the scoring results. The AI vertical large model automatically scoring the quality of the currently collected images means that the real-time collected ultrasound images and the corresponding probe operation trajectories are used as inputs and sent into the trained AI vertical large model. The perception layer extracts the key features of the image clarity, edge integrity and plaque visibility, and combines with the decision layer to comprehensively analyze the behavioral indicators of the operation trajectory stability, angle change and pressure control, and outputs a quantitative score value to measure whether the current image meets the standardized acquisition requirements; S44. If the image quality scoring result is lower than the set threshold, trigger a prompt for insufficient image quality. The prompt information is automatically fed back to the operator interface, and visual guidance prompts are generated in the inspection interface through the augmented reality device to display the visual information of the probe adjustment direction, angle or pressure to the operator; S45. After the image quality meets the requirements, based on the output of the AI vertical large model, automatically generate a structured report containing standard diagnostic elements.

[0042] In the present invention, by integrating the trained AI vertical large model into the front-end inspection device, real-time intelligent feedback and automatic generation of structured diagnosis during the lower limb artery ultrasound image acquisition process are realized, which has significant beneficial effects. The system can receive the image and probe operation data in real time during the inspection process, and the AI model automatically scores the image quality to accurately judge whether the current operation of the operator meets the standard. When the image quality score is lower than the set threshold, the system can immediately trigger a prompt and generate visual guidance information through the augmented reality device to clearly show the operator the adjustment suggestions for the probe direction, angle or pressure, significantly improving the standardization degree of image acquisition and the accuracy of the operator's operation. After the image quality meets the standard, the system can automatically generate a structured report containing key diagnostic elements, reducing manual intervention and improving work efficiency and result consistency. This process realizes an intelligent closed loop of image acquisition, quality assessment, operation correction and result output, greatly reducing the dependence on manual experience, especially suitable for primary medical and remote assistance scenarios, and improving the standardization level and clinical application efficiency of ultrasound inspection.

[0043] In this embodiment, the specific steps of S5 include: S51. Upload the structured report generated by the AI vertical large model and the corresponding inspection images to the cloud quality control platform through the front-end inspection device to complete the remote transmission of the inspection data; S52. Receive and parse the structured report and examination images in the cloud quality control platform, and analyze the content consistency between the two, including the matching of text conclusions and image diagnosis bases. The analysis of the content consistency between the two means that after the cloud quality control platform receives the structured report and the corresponding examination images, it first segments and extracts features from the image content, identifies the diagnosis bases of blood vessel boundaries, plaque positions, and blood flow directions, and at the same time parses the key diagnosis items of stenosis degree, plaque status, and blood flow parameters recorded in the structured report. Through the text-image matching algorithm, it compares and analyzes each conclusion in the report with whether there are corresponding anatomical structures or lesion features in the image, judges whether the text and image correspond and whether the content is complete, and outputs a consistency score or marks the items with deviations for quality control feedback and expert review; S53. Based on the dimensions of image standards, parameter integrity, and report element accuracy set by the platform, perform quality assessment operations on the text-image data to form a quality evaluation result. The execution of the quality assessment operations on the text-image data means that the cloud quality control platform conducts quality inspections on the uploaded examination images and structured reports respectively according to the preset evaluation index system. Among them, the image quality assessment includes image clarity, anatomical structure integrity, noise interference degree, and scanning coverage; the parameter integrity assessment includes whether all required diagnosis items such as stenosis ratio, blood flow velocity, and plaque type are included; the report element accuracy assessment compares whether the image content and the report conclusion match, whether there are expression errors or missing items. The platform comprehensively scores the above dimensions through a rule engine or an intelligent scoring model to form a standardized text-image quality evaluation result, which is used as the basic basis for quality control feedback and model update; S54. The cloud quality control platform combines the consistency analysis result and the quality assessment result to output a standardized quality control feedback result and identify whether there are suspicious or non-compliant items; S55. When difficult cases or abnormal situations such as text-image inconsistency, diagnostic deviation, and information loss are identified in the quality control feedback, the cloud quality control platform automatically triggers the remote expert review process; S56. Qualified remote experts complete the review work on the uploaded report and images within the preset time and output a final review report. The review conclusion is returned to the original examination terminal through the platform.

[0044] By introducing a cloud quality control platform, the present invention realizes the remote quality control and expert review process of lower extremity artery ultrasound examination results, significantly improves the automation, standardization and professional level of diagnostic quality management, and has prominent beneficial effects. The structured report and the corresponding examination images can be automatically uploaded to the quality control platform by the front-end examination device, avoiding omission or format confusion caused by manual transmission. By performing consistency analysis and quality assessment on the graphic and text data, the platform can effectively identify the matching degree, element integrity and image compliance between the report conclusion and the image content, and thus output standardized quality control feedback. If suspicious results or abnormal information are found, the system can automatically trigger the remote expert review process without manual intervention, ensuring the timeliness and authority of the quality control process. The expert completes the review within the specified time, outputs the final review opinion, and promptly transmits it back to the examination terminal to form an effective closed loop. This mechanism not only improves the quality control efficiency and accuracy, but also enhances the ability to automatically identify and correct low-quality or non-standard examinations in the medical process, is applicable to multi-terminal deployment and remote medical collaboration scenarios, and helps to build a high-quality and traceable ultrasound examination service system.

[0045] In this embodiment, S6 specifically includes: S61. Based on the quality control feedback results output by the cloud quality control platform, identify the AI vertical large model in the front-end examination device that needs to be updated, and start the local update process of the AI vertical large model; S62. Without the need to upload the original image data and report information, use the label data marked in the quality control feedback results to perform incremental training on the local AI vertical large model, and generate locally optimized AI vertical large model parameters. The use of the label data marked in the feedback results to perform incremental training on the local AI vertical large model means that on the basis of calling the existing parameters of the AI vertical large model in the local device, extract the label information of the images and report items marked as "quality anomaly" or "diagnostic deviation" in the feedback results returned from the cloud quality control platform, including the structural areas in the images that need to be re-identified or the diagnostic conclusions that need to be corrected in the reports. Use these labels as supervision signals to match with the local historical image and operation trajectory features, construct a mini-batch training sample set, input it into the AI vertical large model for short-cycle fine-tuning training, and update some model parameters through a limited number of steps of backpropagation, so as to improve the adaptability of the AI vertical large model to the image features and operation behaviors in the local scenario, and realize the rapid local optimization of the performance of the AI vertical large model without re-training completely; S63. After completing the local training, upload the optimized AI vertical large model parameters to the federated aggregation node in an encrypted form, and no transmission of any original image, trajectory or structured text information is involved in the process; S64. In the federated aggregation node, receive the local AI vertical large model parameters from multiple front-end devices, and perform unified integration and weight optimization according to the participation of each terminal and the sample feature distribution to generate new global AI vertical large model parameters; S65. Synchronously distribute the updated global AI vertical large model parameters to the participating terminals and replace the original parameter configuration to achieve unified version iteration of the AI vertical large model in each front-end inspection device; S66. Repeatedly execute the processes of local training, parameter upload, aggregation update, and AI vertical large model synchronization within the set federated training cycle to form a distributed model collaborative update process with privacy protection ability and continuous self-optimization.

[0046] Through the construction of an AI vertical large model collaborative update process based on the federated learning mechanism, the present invention realizes privacy protection, autonomous optimization, and efficient synchronization of the lower limb artery ultrasound intelligent diagnosis model in a multi-terminal environment, with significant beneficial effects. The system identifies the model instances to be updated based on the feedback results of the cloud quality control platform and performs incremental training locally without uploading the original image or report data, effectively avoiding the risks of patient privacy leakage and data security. After local optimization, only the model parameters are uploaded to the federated aggregation node in encrypted form to achieve lightweight and secure data exchange. The aggregation node can perform parameter integration according to the participation weights of each terminal and the sample feature distribution to generate a unified global model, ensuring the fairness and generalization ability of model updates. The updated global model will be automatically synchronized to all terminals to achieve a consistent improvement in the AI capabilities between systems. By setting the training cycle, the system can periodically perform local updates and global aggregations to continuously improve the model performance, adapt to different scenarios and operator characteristics, and construct a distributed intelligent optimization system with privacy protection, automatic learning, and multi-terminal collaboration.

[0047] Reference Figure 2 , the lower limb artery ultrasound standardized inspection system based on the AI vertical large model includes the following modules: An image and operation data acquisition module for acquiring images and operation data during the lower limb artery ultrasound inspection process; An AI vertical large model module for performing image feature extraction, trajectory modeling, quality scoring, guidance feedback, and structured diagnosis generation; An augmented reality guidance module for generating visual prompt information through an augmented reality device to guide the operator to adjust the probe direction and angle when the image quality score is lower than the threshold; A structured report generation module for automatically generating a structured inspection report containing key diagnostic elements according to the output results of the AI vertical large model; Cloud quality control analysis module, which is used to receive structured reports and inspection images, perform consistency analysis and quality assessment, and output quality control feedback results; Remote expert review module, which is used to trigger the expert remote review process after identifying difficult or abnormal situations in the quality control feedback and output a review report; Federated learning update module, which is used to perform local parameter fine-tuning on the AI vertical large model in each terminal based on the quality control feedback results and complete distributed model optimization through the federated aggregation mechanism.

[0048] Example 1:

[0049] To verify the feasibility of the present invention in implementation, the present invention was applied to the ultrasound department of a large Class III Grade A hospital in a certain city, and 10 real lower extremity artery ultrasound examinations were carried out for outpatient patients. The tests covered two types of operators, primary physicians and advanced physicians, aiming to evaluate the comprehensive capabilities of the AI vertical large model in real-time scoring, guiding feedback, and structured report generation. The test scenarios included daily outpatient screening and preoperative assessment. The patients' ages ranged from 45 to 72 years old, including those with atherosclerotic symptoms and those in high-risk groups for preventive examinations.

[0050] The system was deployed in a conventional portable ultrasound device. While the operator was performing an ultrasound examination, the AI model automatically received image frames and probe behavior trajectories, scored the image quality in real-time, and when the image score did not reach the set threshold (80 points), it suggested that the operator adjust the probe angle, position, or pressure through a visual prompt. In most operations, only 1 guidance was required to significantly improve the image clarity.

[0051] For example, in the examination numbered CASE_002, the original image score was only 55 points. The system prompted "deflect 3° to the left and reduce the pressure". After the operation was corrected, the score jumped to 81 points, and the clarity of the blood vessel boundary and plaque contour in the image was significantly improved. In another case, CASE_003, operated by a primary physician, the score in the first image acquisition was 66 points, and it reached 92 points after being optimized by the system prompt. The image acquisition time was saved by about 31 seconds. Except for one case with a slight decline (the score of CASE_004 dropped from 80 to 78), the quality scores of the other examinations all showed a positive increase.

[0052] After the image quality of all examinations reached the standard, the system automatically generated a structured ultrasound report covering three major elements: the degree of blood vessel stenosis, plaque morphology, and blood flow velocity, and automatically uploaded it to the hospital quality control platform to complete the graphic-text consistency analysis. In this batch of tests, the consistency scores of the structured reports were generally higher than 90 points, the highest reached 95 points, and the average consistency score was 92.6, indicating a high degree of matching between the diagnostic report and the image content.

[0053] In terms of image acquisition efficiency, system statistics show that for 10 cases, the average time saved is about 61 seconds, and the efficiency improvement is obvious. Especially for junior physicians with low proficiency, they can also quickly obtain standard images under the guidance of the system, significantly reducing the learning threshold and the rate of incorrect operations.

[0054] Table 1 Comparison of Image Quality and Efficiency Before and After AI Guidance

[0055] Based on the data in Table 1, it can be clearly seen the remarkable effects of the system of the present invention in aspects such as improving image quality, enhancing report consistency, and optimizing image acquisition efficiency. This data table records the core evaluation indicators of 10 real cases of lower limb artery ultrasound examinations, covering image scores, structured report quality, and time efficiency, reflecting the application value of the AI guidance mechanism in actual operations.

[0056] From the comparison results of the image quality scores, the average image score before the guidance of the AI system was 69.3 points, and it increased to an average of 88.1 points after the guidance. The overall increase was obvious, indicating that the system effectively improved the operator's behavioral deviations in angle control, pressure adjustment, and position selection through real-time scoring and augmented reality guidance mechanisms. Among them, the score of case number CASE_005 increased from 59 points to 91 points, with an increase of up to 32 points, indicating that even in the case of low initial operation quality, the system has a powerful guidance and correction ability.

[0057] In terms of the consistency score of structured reports, the scores of all 10 examinations were above 85 points, with an average of 92.6 points, a maximum of 95 points, and a minimum of no less than 88 points. This shows that the structured reports generated by the AI system have good integrity and accuracy in covering diagnostic elements, and are highly consistent with the content of the collected images, significantly reducing the subjective errors and omissions that may be brought by manual records.

[0058] In terms of image acquisition efficiency, the image acquisition time after AI guidance was generally shortened, and the time saved ranged from 29 seconds to 113 seconds. Especially for CASE_004 and CASE_008, the time saved was 113 seconds and 101 seconds respectively, indicating that during the image acquisition process, the system guidance significantly reduced the time waste caused by repeated image acquisitions, unqualified images, or operation deviations. Overall, the average image acquisition time saved for 10 cases was about 65 seconds, which has practical significance for improving the inspection turnover efficiency of the department and reducing the workload of physicians.

[0059] In addition, the operator composition includes "Junior Physician A", "Junior Physician B", "Advanced Physician C", and "Advanced Physician D". From the overall distribution, the improvement in the image quality score of junior physicians after AI guidance is more obvious, further verifying that this system has extremely strong operation assistance value for beginners, which is helpful for standardized training and rapid ability improvement.

[0060] In summary, the tabular data fully verifies that during the lower limb artery ultrasound examination, the present invention significantly improves the image quality and operation standardization through real-time quality scoring, intelligent guidance prompts, and structured report generation. At the same time, it improves the examination efficiency and the standardization level of reports, and has good clinical practicability and popularization prospects.

[0061] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, making equivalent replacements or changes, should be covered within the protection scope of the present invention.

Claims

1. A standardized inspection method for lower extremity artery ultrasound based on an AI vertical large model, characterized in that, It includes the following steps: S1. Collect lower extremity artery ultrasound image data including standard images and error images, perform annotation processing on the lower extremity artery ultrasound image data, and construct an image dataset; S2. Preprocess the image dataset, use the U-Net network to extract image features of blood vessel boundaries and plaque structures, model the probe operation sequence based on the Manus framework, and construct a multimodal training sample set; S3. Construct an AI vertical large model including a perception layer, a decision-making layer, and an output layer. The perception layer extracts image features based on the ResNet-50 network, the decision-making layer uses a long short-term memory network to model the probe trajectory and combines the proximal policy optimization algorithm for action scoring and error correction feedback, the output layer generates a structured diagnostic result, and uses the multimodal training sample set for training; S4. Integrate the trained AI vertical large model into the front-end inspection device, collect inspection images and operation data in real time, input them into the AI vertical large model for quality scoring and guidance feedback, and generate a structured report; S5. Upload the structured report and inspection images to the cloud quality control platform, perform consistency analysis and quality assessment, output the quality control feedback result, and trigger remote expert review when the cloud quality control platform identifies difficult or abnormal situations, complete the review and return the review report; S6. Based on the quality control feedback result, perform local parameter update on the AI vertical large model, complete distributed training using the federated learning mechanism, and upload the update result to the aggregation node for unified integration and optimization, and construct a collaborative update process with privacy protection capabilities.

2. The lower limb artery ultrasound standardized examination method based on the AI vertical large model according to claim 1, characterized in that, The lower extremity artery ultrasound image data specifically includes image data with multi-angles, different pressures, section positions, and blood flow parameters, including standard images and error images.

3. The lower limb artery ultrasound standardized examination method based on the AI vertical large model according to claim 1, wherein, The annotation processing of the lower extremity artery ultrasound image data refers to adding annotation label information of the corresponding probe angle, pressure value, section position, and hemodynamic parameters to each image.

4. The lower limb artery ultrasound standardized examination method based on the AI vertical large model according to claim 3, characterized in that, The specific content of S2 includes: S21. Perform normalization processing on the constructed image dataset, uniformly map the pixel values of each image to the [0,1] interval, and perform unified size resampling on the image data with inconsistent sizes to form standard format image data that can be input into the network; S22. Input the standard format image data into the U-Net network, extract the low-level and high-level image semantic features using the encoder part, perform upsampling and feature fusion in the decoder part, output the segmentation mask image including the blood vessel boundary, plaque contour, and lumen structure, and use the label mask obtained from the manual annotation in the image dataset for cross-entropy loss function supervision to construct the image feature set denoted as , where represents the feature vector extracted from the th image data, , and n is the total number of image data in the image dataset; S23. Synchronously record the probe operation information of the operator during image acquisition, including the probe position coordinate sequence , the direction angle sequence , and the contact pressure sequence with the probe , where represents the time step, is the position of the probe in the horizontal axis direction at the t-th time step, is the position of the probe in the vertical axis direction at the t-th time step, is the depth of the probe in the vertical direction at the t-th time step, is the angle in the horizontal rotation direction at the t-th time step, is the angle in the vertical tilt direction at the t-th time step, is the contact pressure value exerted by the probe on the skin surface at the t-th time step; S24. Input the probe position coordinate sequence, direction angle sequence, and probe contact pressure sequence into the time series modeling module constructed by the Manus framework. Model the dynamic behavior of the probe operation through the LSTM network, extract the time correlation features, and output the probe trajectory feature vector ; S25. Combine the image feature set , the probe trajectory feature vector and the annotation label information to construct a multi-modal training sample set, and each sample is represented as a triple.

5. The lower limb artery ultrasound standardized examination method based on the AI vertical large model according to claim 4, wherein, The specific content of S3 includes: S31. Construct an AI vertical large model, divide it into a perception layer, a decision-making layer, and an output layer, which are respectively used for image feature extraction, probe trajectory modeling, and structured diagnosis generation; S32. Input the image feature set into the ResNet-50 network in the perception layer to extract deep semantic features and generate deep semantic feature vectors; S33. Input the probe trajectory feature vector into the long short-term memory network (LSTM) to extract time-dependent features and generate a time-dependent feature vector; S34. Concatenate the deep semantic feature vector and the time-dependent feature vector to form a fusion vector, which is used as the input representation for policy learning and diagnosis generation; S35. Define the task-decomposed immediate reward function ; S36. Construct a structure-aware advantage function based on task decomposition rewards ; S37. Build a state-aware shear range adjustment mechanism to jointly adjust the shear coefficient according to the score fluctuation and reward change ; S38. Proximal Policy Optimization objective function is constructed based on structure-aware advantage function and shear coefficient , ; S39. Input the fusion vector into the policy network to output the policy distribution , and sample an action from it for the generation of probe direction, angle, and pressure guidance; S310. Input the fusion vector synchronously into the multi-layer perceptron in the output layer to output a structured diagnosis result , where the result includes information on the degree of vascular stenosis, plaque status, and blood flow velocity; S311. Construct a joint training loss function ; S312. Use a multi-modal training sample set , and jointly optimize the parameter set through backpropagation and gradient descent until the loss function converges to complete the training process of the AI vertical large model.

6. The standardized lower limb artery ultrasound examination method based on the AI vertical large model according to claim 5, wherein, The specific content of S4 includes: S41. Load and integrate the trained AI vertical large model into the front-end inspection device, and receive ultrasound images and operation inputs in real time; S42. During the inspection process, collect lower extremity artery ultrasound images and probe operation data of the patient, and synchronously input them into the AI vertical large model for analysis and processing; S43. The AI vertical large model automatically scores the quality of the currently collected image, and judges whether the current probe operation meets the requirements for obtaining standard images according to the scoring result; S44. If the image quality score result is lower than the set threshold, an insufficient image quality prompt is triggered, and the prompt information is automatically fed back to the operator interface. A guide prompt is generated in the inspection interface through the augmented reality device, and visual information of the probe adjustment direction, angle or pressure is displayed to the operator; S45. After the image quality meets the requirements, a structured report containing standard diagnostic elements is automatically generated based on the output of the AI ​​vertical large model.

7. The lower limb artery ultrasound standardized examination method based on the AI vertical large model according to claim 6, characterized in that, The S5 specifically includes: S51. Upload the structured report and the corresponding inspection image generated by the AI ​​vertical large model to the cloud quality control platform through the front-end inspection device to complete the remote transmission of the inspection data; S52. Receive and parse the structured report and inspection images in the cloud quality control platform, and analyze the consistency of the contents between the two, including the matching of the text conclusion and the image diagnosis basis; S53. Based on the image standards, parameter integrity and report element accuracy dimensions set by the platform, perform quality assessment operations on the graphic data to form a quality assessment result; S54. The cloud-based quality control platform combines the consistency analysis results with the quality assessment results to output standardized quality control feedback results and identify whether there are any suspicious or substandard items; S55. When difficult cases or abnormal situations such as discrepancies between images and texts, diagnostic deviations, and missing information are identified in the quality control feedback, the cloud-based quality control platform automatically triggers the remote expert review process; S56. Qualified remote experts complete the review of the uploaded reports and images within the preset time and output the final review report. The review conclusion is returned to the original inspection terminal through the platform.

8. The standardized lower limb artery ultrasound examination method based on the AI vertical large model according to claim 7, characterized in that, The S6 specifically includes: S61. Based on the quality control feedback results output by the cloud quality control platform, identify the AI ​​vertical large model in the front-end inspection device that needs to be updated, and start the AI ​​vertical large model local update process; S62. Without uploading the original image data and report information, incremental training is performed on the local AI vertical large model using the label data marked in the quality control feedback results to generate locally optimized AI vertical large model parameters; S63. After completing local training, the optimized AI vertical large model parameters are uploaded to the federation aggregation node in encrypted form. The process does not involve the transmission of any original images, trajectories or structured text information. S64. In the federated aggregation node, local AI vertical big model parameters are received from multiple front-end devices, and unified integration and weight optimization are performed according to the participation of each terminal and the distribution of sample characteristics to generate new global AI vertical big model parameters; S65. Synchronously distribute the updated global AI vertical large model parameters to the participating terminals and replace the original parameter configuration to achieve unified version iteration of the AI ​​vertical large model in each front-end inspection device; S66. Repeat the local training, parameter uploading, aggregation update and AI vertical large model synchronization process within the set federated training cycle to form a distributed model collaborative update process with privacy protection capabilities and continuous self-optimization.

9. The lower limb artery ultrasound standardized examination system based on the AI vertical large model is applied to the lower limb artery ultrasound standardized examination method based on the AI vertical large model described in any one of claims 1 to 8, and is characterized in that, Includes the following modules: Image and operation data acquisition module, used to acquire images and operation data during lower limb artery ultrasound examination; AI Vertical Large Model Module, used to perform image feature extraction, trajectory modeling, quality scoring, guiding feedback, and structured diagnosis generation; Augmented Reality Guiding Module, used to generate visual prompt information through an augmented reality device to guide the operator to adjust the probe direction and angle when the image quality score is lower than the threshold; Structured Report Generation Module, used to automatically generate a structured examination report containing key diagnostic elements based on the output results of the AI Vertical Large Model; Cloud Quality Control Analysis Module, used to receive the structured report and examination images, perform consistency analysis and quality assessment, and output the quality control feedback results; Remote Expert Review Module, used to trigger the expert remote review process after identifying difficult or abnormal situations in the quality control feedback and output the review report; Federated Learning Update Module, used to perform local parameter fine-tuning on the AI Vertical Large Model in each terminal based on the quality control feedback results and complete distributed model optimization through the federated aggregation mechanism.

Images