Lower extremity arterial ultrasound standardized examination system and method based on ai vertical large model

Abstract

The application discloses a lower limb artery ultrasound standardized examination system and method based on an AI vertical large model, which comprises the following steps: S1, standard and error images are collected, labeled and processed, and an image dataset is constructed; S2, image preprocessing is performed, and a multi-modal training sample set is constructed; S3, an AI vertical large model is constructed, and the model is trained by fusing image and track features; S4, the model is deployed to a front-end examination device, real-time scoring and augmented reality guidance are performed, and a structured report is output; S5, the report and images are uploaded to a cloud quality control platform, consistency analysis and quality evaluation are performed, and expert review is triggered; and S6, based on quality control feedback, AI model parameters are updated locally, uploaded to an aggregation node for unified integration, and distributed collaborative optimization is completed. The application aims to realize intelligent guidance of a lower limb artery ultrasound examination process, standardized image collection and automatic generation of structured diagnosis, and improve examination quality and efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence assisted medical imaging and ultrasonic diagnosis automation, and particularly relates to a lower limb artery ultrasonic standardized examination system and method based on an AI vertical large model. BACKGROUND

[0002] With the rising incidence of peripheral arterial disease worldwide, lower limb artery ultrasonography, 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 process of patients with chronic diseases such as diabetes, atherosclerosis and vascular stenosis, timely and accurate acquisition of lower limb artery structure and blood flow parameters is of key significance for formulating subsequent treatment plans.

[0003] However, in current clinical practice, lower limb artery ultrasonography has several technical and application challenges. First, the quality of ultrasonic images is highly dependent on the experience level of the operator, and there are significant differences in probe angle, pressure and scanning path during image acquisition by different operators, which easily causes unstable image quality and makes it difficult to meet the standardization requirements. Second, the image acquisition process lacks real-time quality feedback and operation guidance mechanism, and beginners or non-professionals have difficulty in accurately identifying whether the acquisition meets the standard when completing the examination, resulting in repeated scanning or missing information. Third, ultrasonic image data is generally stored in an unstructured manner, and traditional examination reports rely on manual input, which has the problems of strong subjectivity, low efficiency and large heterogeneity of results, which is not conducive to remote quality control and subsequent AI modeling utilization.

[0004] In recent years, artificial intelligence and deep learning have made significant progress in medical image analysis. Existing research has attempted to use convolutional neural networks for automatic segmentation, lesion identification and blood flow evaluation of ultrasonic images, and some algorithms have achieved good performance in laboratory environment. However, most researches still remain in the offline processing stage, lack of deep integration with ultrasonic examination operation process, and fail to solve the key links in the complete closed loop of "image acquisition - scoring - guidance - diagnosis - feedback". For example, existing systems only focus on image post-processing quality scoring, cannot provide timely intervention suggestions for image generation process, and cannot achieve report structured output and diagnosis closed loop optimization.

[0005] In terms of data processing, the traditional training mode relies on centralized modeling of all samples on a central server, which faces privacy protection and data island problems in the environment of ultrasound equipment deployed in multiple terminals. Patient image data is highly sensitive, and direct uploading to the cloud for model training may violate data compliance requirements and make it 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 practical integration path, model synchronization efficiency, and collaborative update strategy in ultrasound diagnosis systems still lack systematic design and engineering verification.

[0006] Therefore, how to provide an AI vertical large model-based lower extremity artery ultrasound standardized examination system and method is a problem that those skilled in the art urgently need to solve. SUMMARY

[0007] One object of the present application is to provide an AI vertical large model-based lower extremity artery ultrasound standardized examination system and method. The present application integrates multiple modal data processing, deep image understanding, reinforcement learning guidance, structured diagnosis generation, quality control analysis, and federated learning optimization, and is committed to solving key problems such as reliance on manual experience for image acquisition, unstable quality, lack of real-time guidance for operation, unstructured report results, low quality control efficiency, and inability to close loop for model optimization in the existing lower extremity artery ultrasound examination process.

[0008] The AI vertical large model-based lower extremity artery ultrasound standardized examination method according to an embodiment of the present application comprises the following steps:

[0009] S1, collecting lower extremity artery ultrasound image data containing standard images and error images, labeling the lower extremity artery ultrasound image data, and constructing an image data set;

[0010] S2, preprocessing the image data set, extracting image features of blood vessel boundaries and plaque structures using a U-Net network, modeling probe operation sequences based on a Manus framework, and constructing a multi-modal training sample set;

[0011] S3, constructing an AI vertical large model comprising a perception layer, a decision layer, and an output layer, wherein the perception layer extracts image features based on a ResNet-50 network, the decision layer models probe trajectories using a long short-term memory network and combines a proximal policy optimization algorithm for action scoring and error correction feedback, and the output layer generates structured diagnosis results and is trained using the multi-modal training sample set;

[0012] S4, integrating the trained AI vertical large model into a front-end examination device, real-time collecting examination images and operation data, inputting the AI vertical large model for quality scoring and guidance feedback, and generating a structured report;

[0013] S5. Upload the structured report and inspection images to the cloud quality control platform for consistency analysis and quality assessment, output quality control feedback results, and trigger remote expert review when the cloud quality control platform identifies difficult or abnormal situations, completes the review and returns the review report;

[0014] S6. Based on the quality control feedback results, local parameters are updated for large-scale AI vertical models. Distributed training is completed using a federated learning mechanism, and the updated results are uploaded to the aggregation node for unified integration and optimization, thus building a collaborative update process with privacy protection capabilities.

[0015] Optionally, the lower extremity arterial ultrasound image data specifically includes standard images and erroneous images, and image data with multiple angles, different pressures and cross-sectional positions and blood flow parameters.

[0016] Optionally, the annotation processing of lower extremity arterial ultrasound image data refers to adding corresponding probe angle, pressure value, cross-sectional position and hemodynamic parameters to each image.

[0017] Optionally, S2 specifically includes:

[0018] S21. Normalize the constructed image dataset, map the pixel values ​​of each image to the [0,1] interval, and perform uniform size resampling on image data with inconsistent sizes to form standard format image data that can be input into the network.

[0019] S22. Input standard format image data into the U-Net network. The encoder extracts low-level and high-level semantic features, while the decoder performs upsampling and feature fusion, outputting a segmentation mask image including blood vessel boundaries, plaque contours, and luminal structures. The cross-entropy loss function is then used to supervise the construction of the image feature set, denoted as S22. ,in Indicates the first Feature vectors extracted from image data, , where n is the total number of image data in the image dataset;

[0020] S23. Synchronously record the operator's probe operation information during the image acquisition process, including the probe position coordinate sequence. , direction angle sequence Contact pressure sequence with probe ,in, Indicates a time step. Let t be the position of the probe in the horizontal direction at the t-th time step. Let be the position of the probe along the vertical axis at time step t. Let be the depth of the probe in the vertical direction at the t-th time step. is an angle in the horizontal rotation direction at the tth time step, is an angle in the vertical tilt direction at the tth time step, is a contact pressure value of the probe acting on the skin surface at the tth time step;

[0021] S24, inputting the probe position coordinate sequence, the direction angle sequence, and the probe contact pressure sequence into a time sequence modeling module constructed by the Manus framework, modeling the dynamic behavior of the probe operation through an LSTM network, extracting time correlation features, and outputting a probe trajectory feature vector ;

[0022] S25, fusing the image feature set , the probe trajectory feature vector , and the labeled label information to construct a multi-modal training sample set, and each sample is represented as a triple.

[0023] Optionally, the S3 specifically comprises:

[0024] S31, constructing an AI vertical large model structure, dividing it into a perception layer, a decision layer, and an output layer, respectively used for image feature extraction, probe trajectory modeling, and structured diagnosis generation;

[0025] S32, inputting the image feature set to the ResNet-50 network in the perception layer to extract deep semantic features and generate a deep semantic feature vector;

[0026] S33, inputting the probe trajectory feature vector to the long short-term memory network LSTM to extract time-dependent features and generate a time-dependent feature vector;

[0027] S34, concatenating the deep semantic feature vector and the time-dependent feature vector to form a fusion vector as an input representation for strategy learning and diagnosis generation;

[0028] S35, defining a task decomposition type instant reward function ;

[0029] S36, constructing a structure perception advantage function based on the task decomposition reward ;

[0030] S37, constructing a state perception shear range adjustment mechanism to jointly adjust the shear coefficient according to the score fluctuation and the reward change ;

[0031] S38, constructing a proximal policy optimization objective function based on the structure perception advantage function and the shear coefficient ​ ;

[0032] S39, the fusion vector is input into the policy network, and a policy distribution is output , and an action is sampled from the policy distribution , for probe direction, angle and pressure guidance generation;

[0033] S310, the fusion vector is synchronously input into a multi-layer perceptron in the output layer, and a structured diagnosis result is output , the result including information of vessel stenosis degree, plaque state and blood flow velocity;

[0034] S311, a joint training loss function is constructed ;

[0035] S312, a multi-modal training sample set is used to jointly optimize the parameter set through back propagation and gradient descent until the loss function converges, completing the training process of the AI vertical large model.

[0036] Optionally, the S4 specifically comprises:

[0037] S41, the trained AI vertical large model is loaded and integrated into the front-end examination device, and real-time reception of ultrasound images and operation inputs is performed;

[0038] S42, during the examination process, patient lower extremity artery ultrasound images and probe operation data are collected and synchronously input into the AI vertical large model for analysis and processing;

[0039] S43, the AI vertical large model automatically scores the quality of the currently collected images, and determines whether the current probe operation meets the standard image acquisition requirements according to the score result;

[0040] S44, if the image quality score result is lower than a set threshold, an image quality insufficient prompt is triggered, the prompt information is automatically fed back to the operator interface, a guidance prompt is generated in the examination interface through an augmented reality device, and visual information of probe adjustment direction, angle or pressure is displayed to the operator;

[0041] S45, after the image quality meets the requirements, a structured report containing standard diagnosis elements is automatically generated based on the output of the AI vertical large model.

[0042] Optionally, the S5 specifically comprises:

[0043] S51, the structured report generated by the AI vertical large model and the corresponding examination images are uploaded to the cloud quality control platform through the front-end examination device, completing the remote transmission of the examination data;

[0044] S52, receiving and parsing the structured report and checking the image in the cloud quality control platform, analyzing the content consistency between the two, including the matching of the text conclusion and the image diagnosis basis;

[0045] S53, based on the image standard, parameter integrity and report element accuracy set by the platform, performing quality evaluation operation of the image and text data, forming quality evaluation result;

[0046] S54, the cloud quality control platform combines the consistency analysis result and the quality evaluation result, outputs the standardized quality control feedback result, and identifies whether there are suspicious or substandard items;

[0047] S55, when the quality control feedback identifies difficult cases or there are image-text inconsistency, diagnosis deviation and information missing abnormal situations, the cloud quality control platform automatically triggers the remote expert review process;

[0048] S56, the qualified remote expert completes the review work of the uploaded report and image within the preset time, and outputs the final review report, and the review conclusion is returned to the original inspection terminal through the platform.

[0049] Optionally, the S6 specifically comprises:

[0050] S61, based on the quality control feedback result output by the cloud quality control platform, identifying the AI vertical large model in the front-end inspection device that needs to be updated, and starting the AI vertical large model local update process;

[0051] S62, without uploading the original image data and report information, using the label data marked in the quality control feedback result to perform incremental training on the local AI vertical large model, and generating the AI vertical large model parameters optimized locally;

[0052] S63, after completing the local training, uploading the optimized AI vertical large model parameters to the federal aggregation node in an encrypted form, without involving any transmission of original image, trajectory or structured text information in the process;

[0053] S64, in the federal aggregation node, receiving the local AI vertical large model parameters from multiple front-end devices, and performing unified integration and weight optimization according to the terminal participation and sample feature distribution, to generate new global AI vertical large model parameters;

[0054] S65, synchronously distributing the updated global AI vertical large model parameters to the participating terminals, and replacing the original parameter configuration, to realize the unified version iteration of the AI vertical large model in each front-end inspection device;

[0055] S66, repeat the local training, parameter uploading, aggregated updating and AI vertical large model synchronization process within the set federal training period to form a distributed model collaborative updating process with privacy protection capability and continuous self-optimization.

[0056] The lower limb artery ultrasound standardized examination system based on an AI vertical large model according to an embodiment of the application comprises the following modules:

[0057] An image and operation data acquisition module is configured to acquire image and operation data in the lower limb artery ultrasound examination process.

[0058] An AI vertical large model module is configured to perform image feature extraction, trajectory modeling, quality scoring, guidance feedback and structured diagnosis generation.

[0059] An augmented reality guidance module is configured 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 a threshold.

[0060] A structured report generation module is configured to automatically generate a structured examination report containing key diagnostic elements according to the output result of the AI vertical large model.

[0061] A cloud-side quality control analysis module is configured to receive the structured report and examination image, perform consistency analysis and quality evaluation, and output quality control feedback results.

[0062] A remote expert review module is configured to trigger an expert remote review process and output a review report when difficult or abnormal cases are identified in the quality control feedback.

[0063] A federal learning updating module is configured 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 federal aggregation mechanism.

[0064] The lower limb artery ultrasound standardized examination system based on an AI vertical large model according to an embodiment of the application has the following beneficial effects:

[0065] The lower limb artery ultrasound standardized examination system and method based on an AI vertical large model according to the application realize automatic, standardized and intelligent closed-loop examination of the whole process, and have achieved significant beneficial effects compared with the prior art.

[0066] Secondly, the AI vertical large model constructed by the application not only has a deep understanding ability of images and operation trajectories, but also models and optimizes feedback of the probe operation path through a reinforcement learning mechanism, so that the operator can obtain real-time error correction and guidance prompts in the image acquisition process, thereby improving the image compliance rate and reducing the number of repeated scans. On this basis, the system automatically generates a structured ultrasound diagnosis report, which not only reduces the subjectivity and workload of manual recording, but also improves the standardization and standardization of report content, and provides a good foundation for subsequent remote diagnosis, review and medical research.

[0067] Further, the application can effectively find problems such as inconsistency between images and report content, missing key information or suspicious diagnostic conclusions, and output standardized quality control feedback results in a timely manner, and support automatic triggering of remote expert review process in the presence of abnormal or difficult cases, realize remote medical cooperation and multi-level quality control closed loop, and provide a reliable quality assurance mechanism for areas with uneven medical resources.

[0068] In terms of model optimization, the application introduces a federated learning mechanism, performs local parameter fine-tuning on the terminal model based on the quality control feedback results, uploads the encrypted updated parameters to the aggregation node after training, performs unified integration and synchronous distribution, thereby realizing the collaborative update and continuous optimization of the global model while protecting the data privacy without uploading the original image or diagnostic data. This mechanism breaks the bottleneck of the traditional "centralized" training mode, enabling the AI model to have self-adaptive evolution ability in multiple terminals and multiple scenarios, and adapting to the application needs of different populations, devices and environments.

[0069] In summary, the application forms a complete intelligent ultrasound examination solution in terms of image acquisition standardization, operation intelligent feedback, report structured generation, remote quality control review and privacy protection model optimization, which not only effectively improves the examination efficiency and diagnostic accuracy, but also provides an efficient, safe and sustainable optimization technical path for primary medical care, remote ultrasound and intelligent auxiliary diagnosis, and has good clinical application prospect and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0070] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application are used to explain the application, and do not constitute a limitation on the application. In the drawings:

[0071] Figure 1 The flowchart of the lower extremity arterial ultrasound standardized examination method based on the AI vertical large model proposed by the application;

[0072] Figure 2A structure diagram of a lower limb artery ultrasound standardized examination system based on an AI vertical large model according to the present application. DETAILED DESCRIPTION

[0073] The present application will now be further described in greater detail in connection with the accompanying drawings. These drawings are simplified schematic diagrams which only show the basic structure of the present application in a schematic manner, and thus only show the components relevant to the present application.

[0074] REFERENCE Figure 1 The lower limb artery ultrasound standardized examination method based on the AI vertical large model comprises the following steps:

[0075] S1, collect lower limb artery ultrasound image data containing standard images and error images, perform labeling processing on the lower limb artery ultrasound image data, and construct an image data set;

[0076] S2, pre-process the image data set, extract image features of blood vessel boundaries and plaque structures using a U-Net network, model probe operation sequences based on a Manus framework, and construct a multi-modal training sample set;

[0077] S3, construct an AI vertical large model comprising a perception layer, a decision layer, and an output layer, wherein the perception layer extracts image features based on a ResNet-50 network, the decision layer models probe trajectories using a long short-term memory network and combines a proximal policy optimization algorithm to perform action scoring and error correction feedback, and the output layer generates structured diagnostic results and is trained using the multi-modal training sample set;

[0078] S4, integrate the trained AI vertical large model into a front-end examination device, real-time collect examination images and operation data, input the AI vertical large model to perform quality scoring and guidance feedback, and generate a structured report;

[0079] S5, upload the structured report and the examination images to a cloud quality control platform, perform consistency analysis and quality evaluation, output quality control feedback results, and trigger remote expert review when the cloud quality control platform identifies difficult or abnormal cases, complete the review and return a review report;

[0080] S6, based on the quality control feedback results, update the local parameters of the AI vertical large model, complete distributed training using a federated learning mechanism, upload the updated results to an aggregation node, perform unified integration and optimization, and construct a collaborative update process with privacy protection capability.

[0081] The application realizes a complete closed-loop process of ultrasound image acquisition, quality scoring, guidance feedback, structured report generation, cloud quality control and model self-optimization by constructing a lower limb artery ultrasound standardized examination method based on an AI vertical large model, and has significant beneficial effects. The system fuses image features and probe trajectories, uses deep learning and reinforcement learning algorithms to realize real-time scoring of image quality and intelligent guidance of operator operation behavior, and effectively improves the image standardized acquisition rate. Through structured diagnostic result output, report writing is simplified, and accuracy and consistency are improved; the cloud quality control platform supports text and image consistency analysis and remote expert review, and enhances the diagnostic reliability. In terms of model updating, the federal learning mechanism is used to realize local fine-tuning and parameter aggregation of each terminal model, which improves the model generalization ability while ensuring data privacy, forming a collaborative optimization system with privacy protection capability. The method significantly improves the standardization and intelligent diagnosis efficiency of ultrasound examination, and is suitable for multi-terminal and multi-scene deployment.

[0082] In the embodiment, the lower limb artery ultrasound image data specifically includes image data with multiple angles, different pressures, and cross-sectional positions and blood flow parameters.

[0083] In the embodiment, the labeling processing of the lower limb artery ultrasound image data refers to adding corresponding probe angle, pressure value, cross-sectional position and blood flow hemodynamic parameter labeling tag information for each image.

[0084] In the embodiment, S2 specifically includes:

[0085] S21, normalizing the constructed image data set, uniformly mapping each image pixel value to the [0, 1] interval, and performing uniform size resampling on image data of inconsistent sizes to form standard format image data that can be input into the network;

[0086] S22, inputting the standard format image data into the U-Net network, using the encoder part to extract low-level and high-level image semantic features, and using the decoder part to perform upsampling and feature fusion, outputting segmentation mask images 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 refers to gradually extracting edge texture, local structure and high-level semantic information of the image through the convolution operation and the maximum pooling operation of the multiple layers stacked in the encoder, and in the decoder, the feature map resolution is restored by using the upsampling method, and the feature map of the corresponding encoder layer is spliced with the decoder feature map through the jump connection method, then the joint features are extracted through convolution fusion, and the label mask obtained by artificial labeling in the image data set is used to supervise the cross-entropy loss function, and an image feature set is constructed, denoted as wherein represents the feature vectors extracted from the amplitude image data, n is the total number of image data in the image data set;

[0087] S23, synchronizing the probe operation information of the operator in the image acquisition process, including probe position coordinate sequence , direction angle sequence , and probe contact pressure sequence , wherein, represents the time step, is the position of the probe in the horizontal axis direction at the tth time step, is the position of the probe in the vertical axis direction at the tth time step, is the depth of the probe in the vertical direction at the tth time step, is the angle in the horizontal rotation direction at the tth time step, is the angle in the vertical tilt direction at the tth time step, is the contact pressure value of the probe acting on the skin surface at the tth time step;

[0088] S24, inputting the probe position coordinate sequence, direction angle sequence and probe contact pressure sequence into the timing modeling module constructed by the Manus framework, modeling the dynamic behavior of the probe operation through the LSTM network, extracting time correlation features, and outputting the probe trajectory feature vector , the modeling of the dynamic behavior of the probe operation through the LSTM network and the extraction of the time correlation features refer to inputting the multi-dimensional time sequence data of the position coordinates, direction angles and contact pressures of the probe in the examination process as input sequences into the long short-term memory network, modeling the state changes between different time steps of the operation actions by using the gating mechanism, capturing the potential time-dependent relationships and behavior patterns of the operator in the process of moving, adjusting and pressing the probe, and finally extracting the trajectory feature vector reflecting the stability, continuity and standardization of the probe operation;

[0089] S25, fusing the image feature set , the probe trajectory feature vector , and the labeled label information to construct a multi-modal training sample set, and each sample is represented as a triple.

[0090] The application introduces an image feature extraction and probe operation modeling dual-channel fusion strategy in the multi-modal training sample construction process, significantly improves the global understanding and behavior discrimination ability of the AI model for the ultrasound examination scene, and has good practical value and technical advantages. Through image normalization and size resampling, the consistency preprocessing of images of different sources is realized, and the format standardization and training stability of the input network are ensured;U-Net network is used to extract multi-level semantic features including blood vessel boundary, plaque structure and the like, and the segmentation and identification ability of the model to the lesion morphology is enhanced. At the same time, the Manus framework and LSTM network are combined to model the dynamic operation information of the probe in the image acquisition process, accurately describe the behavior trajectory and time sequence characteristics of the operator, and effectively supplement the lack of spatial behavior understanding of traditional image processing methods. Finally, through the fusion of image features, operation trajectory and label information, high-quality multi-modal training samples are constructed, providing sufficient and accurate joint perception basis for the training of subsequent AI vertical large models, improving the robustness and operation feedback ability of the model.

[0091] In the embodiment, the S3 specifically includes:

[0092] S31, constructing an AI vertical large model structure, which is divided into a perception layer, a decision layer and an output layer, respectively used for image feature extraction, probe trajectory modeling and structured diagnosis generation;

[0093] S32, the image feature set is input into the ResNet-50 network in the perception layer, deep semantic features are extracted, and a deep semantic feature vector is generated;

[0094] S33, the probe trajectory feature vector is input into the long short-term memory network LSTM, time-dependent features are extracted, and a time-dependent feature vector is generated;

[0095] S34, the deep semantic feature vector and the time-dependent feature vector are spliced to form a fusion vector as the input representation of strategy learning and diagnosis generation;

[0096] S35, defining a task decomposition type instant reward function :

[0097] ;

[0098] Wherein, represents the image quality score, represents the consistency score of the probe path and the standard trajectory, represents the score gain after the operator responds, 、 and are weight coefficients;

[0099] The practical significance of the task-decomposition-based immediate reward function formula lies in breaking down the quality assessment objective in the ultrasound image acquisition process into multiple dimensions with practical operational significance, thereby achieving refined guidance and evaluation of operator behavior. This reward function divides the immediate feedback signal into three parts: image quality score, probe path consistency score with standard trajectory score, and operator response score to system guidance score, comprehensively covering key factors such as image clarity, operational standardization, and effective guidance execution. By organically integrating these sub-objectives, the system can perceive the impact of each operator's action on the image results in real time and dynamically adjust the feedback and scoring mechanism accordingly. Compared to the traditional single reward model, this task decomposition strategy is closer to actual clinical procedures, effectively improving the model's ability to identify non-standard operations and its sensitivity to image acquisition quality, thus providing a solid foundation for building highly reliable guidance strategies and accurately outputting structured diagnostic results. This design ensures that the system has a clear optimization direction during the training phase and realistic and feasible operational intervention capabilities during the application phase.

[0100] S36. Construct a structure-aware advantage function based on task decomposition rewards. The advantage estimation expression is as follows:

[0101] ;

[0102] in, This represents the estimated value of the structure-aware advantage function at time step t. Current state The state value function output, For the next state The state-value function is modeled by a neural network, with a fusion vector as input and a predicted reward for the corresponding state as output. and Representing states respectively and The corresponding image's structural sharpness scoring function value, As a discount factor, As the dominant smoothing factor, This represents the estimated value of the structure-aware advantage function at the (t+1)th time step;

[0103] The practical significance of the structure-aware dominance function lies in constructing a more accurate measure of action dominance that better reflects the actual needs of ultrasound examinations by combining information on changes in image quality with state value assessment. In traditional reinforcement learning, dominance functions typically rely solely on rewards and state values ​​for calculation, neglecting the significant impact of changes in image structure on model judgment. However, in ultrasound image acquisition scenarios, factors such as image structural clarity and edge stability directly affect diagnostic quality. This invention introduces changes in image structural clarity as part of the dominance function, effectively reflecting the degree of improvement in image structure before and after a particular operator's action, thereby enhancing the model's ability to identify "beneficial actions." This dominance function not only considers the immediate reward and value difference brought by the current action but also incorporates the trend of changes at the image structural level, making the model more inclined to retain actions that consistently improve image quality during policy learning. This design, which integrates image perception and reinforcement learning values, significantly enhances the behavioral discriminative power and training stability of the policy network, serving as a key foundation for achieving high-quality image-guided generation and action policy learning.

[0104] S37. Construct a state-aware shearing range adjustment mechanism to jointly adjust the shearing coefficient based on score fluctuations and reward changes. :

[0105] ;

[0106] in, This is the initial shear range. For rating fluctuation factor, To reward the change factor, The standard deviation of the image rating. For the front Average step reward The immediate reward value from the previous time step;

[0107] The practical significance of the state-aware clipping range adjustment mechanism is to introduce dynamic constraints of image quality fluctuation and operation stability into the policy optimization process in reinforcement learning, making the policy update more stable and adaptive. During the ultrasound image acquisition process, the image quality score will fluctuate with the change of the operator's gesture, and the influence of the operation behavior on the model reward will also change with the scene and the operator's level. The mechanism perceives the standard deviation of the image score and the reward change trend, dynamically adjusts the amplitude of policy update, avoids excessive or unreliable parameter update when the score fluctuates greatly or the reward changes dramatically, and thus reduces the problems of policy degradation and unstable training. Compared with the traditional method of fixed clipping range, the clipping coefficient of the invention has state adaptive ability, which can flexibly contract or expand the policy update boundary according to the current environmental changes, ensuring the learning efficiency of the model in the high confidence state, and enhancing the safety and robustness of convergence in the low quality state. This mechanism plays a key adjustment role in the model training process, effectively improving the adaptability of the policy network to the actual scene of ultrasound image acquisition.

[0108] S38, based on structure-aware advantage function and clipping coefficient , construct the proximal policy optimization objective function :

[0109] ;

[0110] wherein, is the policy ratio, represents the statistical expectation calculation of multiple samples, is the minimum value operation, is the clipping operation on the policy ratio ;

[0111] The practical significance of the policy optimization objective function is to effectively control the policy update range in the reinforcement learning process, thereby ensuring the stability of model training and the reliability of policy output. In the ultrasound image acquisition scenario, the operator's behavior has certain uncertainty. If the policy network updates too fast or fluctuates too much during training, it is easy to cause the model to collapse or fail. Therefore, the objective function sets a clipping range to limit the change range between the new and old policies, so that each update is within a controllable range, avoiding the model from deviating from the optimal direction due to abnormal local rewards in the early training stage. At the same time, the 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 invention further introduces a state-aware clipping mechanism, which dynamically adjusts the clipping range according to image quality fluctuations and reward changes, achieving adaptive regulation of policy change strength. This mechanism can accelerate the learning speed when the image quality is good or the operation is stable, and reduce the update strength when the image fluctuates sharply or the behavior is uncertain, thereby improving the robustness and generalization ability of the policy network, which is the core component of ensuring the intelligent evolution of the guide policy.

[0112] S39, the fusion vector Input policy network, output policy distribution , and sample actions from it , used for probe direction, angle and pressure guidance generation;

[0113] S310, the fusion vector is input into the multi-layer perceptron in the output layer synchronously, and a structured diagnosis result is output , the result contains information of blood vessel stenosis degree, plaque state and blood flow velocity;

[0114] S311, construct a joint training loss function :

[0115] ;

[0116] wherein, is a structured label, CrossEntropy is a classification loss, is a loss balancing coefficient;

[0117] The actual significance of the joint training loss function is to realize the synergistic optimization of the policy guidance capability and the structured diagnosis accuracy, to build a unified training target, and to improve the overall performance of the model. During the lower extremity artery ultrasound examination process, the system needs to learn a reasonable probe guiding strategy through reinforcement learning, and also needs to accurately output the structured diagnosis results such as blood vessel stenosis and plaque state through supervised learning. The loss function weights and fuses the policy optimization target and the structured classification target, so that the model can continuously improve the quality score of the action strategy and the operation feedback capability during the training process, and also ensure that the output diagnosis results have medical accuracy and integrity. Through this multi-task joint optimization design, the model can learn “how to operate” and “how to judge” at the same time, thereby having stronger intelligent guidance and diagnosis decision-making capability. At the same time, the adjustable weight coefficient is introduced, which can flexibly adjust the proportion of policy learning and diagnosis classification in the total loss according to the task demand, and realize the dynamic regulation and control of the training emphasis at different stages. This mechanism effectively solves the problem of model bias caused by single optimization target, and is the key foundation to ensure the unity of the intelligent guidance and diagnosis accuracy of the model.

[0118] S312, using a multi-modal training sample set , optimizing the parameter set through the combination of back propagation and gradient descent until the loss function converges, completing the training process of the AI vertical large model.

[0119] The present application realizes the deep fusion of lower extremity artery ultrasound image understanding, operator behavior modeling and structured diagnosis generation by constructing an AI vertical large model with perception layer, decision layer and output layer, and has significant beneficial effects. The model fuses image features and probe trajectory information, uses ResNet-50 and LSTM network to extract spatial and temporal features respectively, and through the task decomposition reward mechanism and the structure perception advantage function, the model's scoring ability for operation quality and sensitivity to image structure are enhanced. The state perception shear coefficient and the structure definition factor are introduced to realize the adaptive adjustment of the strategy network to the image fluctuation and the operation behavior during the training process, and to improve the training stability and feedback accuracy. On the basis of reinforcement learning strategy optimization, combined with structured label supervised learning to build a joint loss function, the guiding strategy and diagnosis accuracy are considered, and the integrity and consistency of the output results are ensured. The model has intelligent scoring, self-correction and structured output capability, which provides a core intelligent engine for realizing the whole process automation and standardization of ultrasound examination.

[0120] In the embodiment, the S4 specifically comprises:

[0121] S41, loading and integrating the trained AI vertical large model to the front-end examination device, and real-time receiving ultrasound images and operation inputs;

[0122] S42, in the inspection process, the patient's lower extremity artery ultrasound image and probe operation data are collected and input synchronously to an AI vertical large model for analysis and processing;

[0123] S43, the AI vertical large model automatically scores the quality of the currently collected image, and determines whether the current probe operation meets the standard image acquisition requirements according to the score result; the automatic scoring of the AI vertical large model refers to inputting the real-time collected ultrasound image and the corresponding probe operation trajectory into the trained AI vertical large model, extracting the key features of the image clarity, edge integrity and plaque visibility by the perception layer, and comprehensively analyzing the stability, angle change and pressure control behavior indicators of the operation trajectory by the decision layer, and outputting a quantitative score value for measuring whether the current image meets the standardized collection requirements;

[0124] S44, if the image quality score result is lower than the set threshold, an image quality insufficient prompt is triggered, the prompt information is automatically fed back to the operator interface, and the guidance prompt is 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;

[0125] 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.

[0126] The application realizes real-time intelligent feedback and automatic generation of structured diagnosis in the lower extremity artery ultrasound image collection process by integrating the trained AI vertical large model into the front-end inspection device, which has significant beneficial effects. The system can receive image and probe operation data in real time during the inspection process, and automatically score the image quality by the AI model to accurately determine 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 of the probe direction, angle or pressure, significantly improving the standardization of image collection and the accuracy of operator operation. After the image quality meets the requirements, the system can automatically generate a structured report containing key diagnostic elements, reducing manual intervention and improving work efficiency and result consistency. The process realizes an intelligent closed loop of image collection, quality evaluation, operation correction and result output, greatly reduces the dependence on manual experience, and is especially suitable for primary medical care and remote assistance scenes, improving the standardization level and clinical application efficiency of ultrasound examination.

[0127] In the embodiment, the S5 specifically includes:

[0128] S51, the structured report generated by the AI vertical large model and the corresponding inspection image are uploaded to the cloud quality control platform through the front-end inspection device to complete the remote transmission of the inspection data;

[0129] S52, receiving and analyzing the structured report and the inspection image in the cloud quality control platform, and analyzing the content consistency between the two, including the matching of the text conclusion and the image diagnostic basis, wherein the analysis of the content consistency between the two refers to that after the cloud quality control platform receives the structured report and the corresponding inspection image, first, the image content is segmented and feature extracted, the diagnostic basis of the blood vessel boundary, plaque position and blood flow direction is recognized, and the key diagnostic items of the stenosis degree, plaque state and blood flow parameter recorded in the structured report are analyzed, through the image-text matching algorithm, each item of conclusion in the report and whether there is corresponding anatomical structure or lesion feature in the image are compared and analyzed, whether the image and the text correspond, the content is complete or not is judged, and a consistency score or a marked item with deviation is output, which is used for quality control feedback and expert review;

[0130] S53, performing quality evaluation operation of image and text data based on the image standard, parameter integrity and report element accuracy dimensions set by the platform, and forming quality evaluation result, wherein the performing quality evaluation operation of image and text data refers to that the cloud quality control platform performs quality check on the uploaded inspection image and structured report according to the pre-set evaluation index system, wherein the image quality evaluation includes image definition, anatomical structure integrity, noise interference degree and scanning coverage range; the parameter integrity evaluation includes whether all the required diagnostic items such as stenosis ratio, blood flow velocity and plaque type are included; the report element accuracy evaluation compares whether the image content and the report conclusion are consistent, whether there is expression error or missing item. The platform scores the above dimensions by rule engine or intelligent scoring model, forms a standardized image and text quality evaluation result, and serves as the basis for quality control feedback and model updating;

[0131] S54, outputting the standardized quality control feedback result by the cloud quality control platform combined with the consistency analysis result and the quality evaluation result, and identifying whether there is suspicious or substandard item;

[0132] S55, when the difficult cases or abnormal conditions of image-text inconsistency, diagnostic deviation and information missing are identified in the quality control feedback, the cloud quality control platform automatically triggers the remote expert review process;

[0133] S56, the qualified remote experts complete the review work of the uploaded report and image within the pre-set time, and output the final review report, and the review conclusion is returned to the original inspection terminal through the platform.

[0134] The application realizes remote quality control and expert review process of lower limb arterial ultrasound examination results by introducing a cloud quality control platform, significantly improves the automation, standardization and professional level of diagnosis quality management, and has outstanding beneficial effects. The structured report and corresponding examination image can be automatically uploaded to the quality control platform by the front-end examination device, avoiding manual transmission omission or format disorder. The platform can effectively identify the matching degree between the report conclusion and the image content, the element integrity and the image compliance through consistency analysis and quality evaluation of the image and text data, so as to 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 a limited time, outputs the final review opinion, and returns it to the examination terminal in time, forming an effective closed loop. This mechanism not only improves the quality control efficiency and accuracy, but also enhances the automatic identification and correction ability of low-quality or non-standard examination in the medical process, which is suitable for multi-terminal deployment and remote medical cooperation scenarios, and helps to build a high-quality and traceable ultrasound examination service system.

[0135] In the embodiment, S6 specifically includes:

[0136] S61, based on the quality control feedback result output by the cloud quality control platform, identifying the AI vertical large model in the front-end examination device that needs to be updated, and starting the AI vertical large model local update process;

[0137] S62, without uploading the original image data and report information, using the labeled label data in the quality control feedback result to perform incremental training on the local AI vertical large model, to generate AI vertical large model parameters optimized locally, wherein the incremental training of the local AI vertical large model using the labeled label data in the feedback result refers to extracting the label information of the images and report items marked as "quality abnormality" or "diagnostic deviation" in the feedback result returned from the cloud quality control platform based on the existing parameters of the AI vertical large model in the local device, including the structural regions that need to be re-identified in the images or the diagnostic conclusions that need to be corrected in the report, using these labels as supervision signals and matching with local historical image and operation trajectory features to construct a small batch training sample set, inputting into the AI vertical large model for short-period fine-tuning training, and updating part of the model parameters through limited steps of back propagation, so as to improve the adaptability of the AI vertical large model to image features and operation behaviors in the local scene, and realize rapid local optimization of the performance of the AI vertical large model without re-training in its entirety;

[0138] S63, after completing the local training, uploading the optimized AI vertical large model parameters to the federal aggregation node in an encrypted form, without involving any transmission of original images, trajectories or structured text information in the process;

[0139] S64, in the federal aggregation node, receiving local AI vertical large model parameters from multiple front-end devices, and performing 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;

[0140] S65, synchronously distributing the updated global AI vertical large model parameters to the participating terminals, and replacing the original parameter configuration, to realize unified version iteration of the AI vertical large model in each front-end inspection device;

[0141] S66, repeating the local training, parameter uploading, aggregation updating and AI vertical large model synchronization process within the set federal training period, to form a distributed model collaborative updating process with privacy protection capability and continuous self-optimization.

[0142] The application realizes privacy protection, self-optimization and efficient synchronization of the lower extremity artery ultrasound intelligent diagnosis model in a multi-terminal environment by constructing an AI vertical large model collaborative updating process based on a federal learning mechanism, which has significant beneficial effects. The system identifies the model instances that need to be updated according to the feedback results of the cloud quality control platform, and performs incremental training locally without uploading the original image or report data, effectively avoiding patient privacy leakage and data security risks. After local optimization, only the model parameters are uploaded to the federal aggregation node in an encrypted form, realizing lightweight and secure data exchange. The aggregation node can integrate parameters according to the participation weight of each terminal and the sample feature distribution to generate a unified global model, ensuring the fairness and generalization ability of model updating. The updated global model will be automatically synchronized to all terminals, realizing consistent improvement of AI capability between systems. By setting the training period, the system can periodically perform local updating and global aggregation to continuously improve model performance and adapt to different scenarios and operator characteristics, thus building a distributed intelligent optimization system with privacy protection, automatic learning and multi-terminal collaboration.

[0143] Reference Figure 2 , the lower extremity artery ultrasound standardized examination system based on AI vertical large model includes the following modules:

[0144] An image and operation data acquisition module is used to acquire image and operation data during lower extremity artery ultrasound examination;

[0145] An AI vertical large model module is used to perform image feature extraction, trajectory modeling, quality scoring, guided feedback and structured diagnosis generation;

[0146] An augmented reality guidance module is used to generate visual prompt information through an augmented reality device when the image quality score is below a threshold, guiding the operator to adjust the probe direction and angle;

[0147] a structured report generation module configured to automatically generate a structured examination report containing key diagnostic elements based on the output of the AI vertical large model;

[0148] a cloud-side quality control analysis module configured to receive the structured report and the examination image, perform consistency analysis and quality evaluation, and output quality control feedback results;

[0149] a remote expert review module configured to trigger a remote expert review process when difficult or abnormal cases are identified in the quality control feedback, and output a review report;

[0150] a federated learning update module configured 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.

[0151] Embodiment 1:

[0152] To verify the feasibility of the present application in implementation, the present application is applied to the ultrasonography department of a large third-grade hospital in a certain city, and 10 real lower extremity arterial ultrasound examination tests are carried out for outpatients, covering two types of operators, primary physicians and advanced physicians, aiming to evaluate the comprehensive ability of the AI vertical large model in real-time scoring, guidance feedback and structured report generation. The test scenarios include daily outpatient screening and preoperative assessment, and the patients are aged between 45 and 72 years old, including both those with symptoms of atherosclerosis and those with high-risk groups for preventive examination.

[0153] The system is deployed in a conventional portable ultrasound device, and the AI model automatically receives image frames and probe behavior trajectories while the operator is performing ultrasound examination, and scores the image quality in real time, and when the image score does not reach the set threshold (80 points), the operator is suggested to adjust the probe angle, position or pressure through visual prompt. In most operations, guidance is only needed once, which can significantly improve image clarity.

[0154] For example, in the examination numbered CASE_002, the original image score was only 55 points, and the system prompted "left deflection 3° and reduce pressure", and after the operator corrected the operation, the score jumped to 81 points, and the clarity of the blood vessel boundary and plaque outline in the image was significantly improved. Another example, CASE_003, operated by a primary physician, scored 66 points in the first image, and after optimization prompted by the system, it reached 92 points, saving about 31 seconds in image acquisition time. Except for one case with a slight drop (CASE_004 score from 80 to 78), the quality scores of the remaining examinations were positively improved.

[0155] All checks in the image quality, the system automatically generates a structured ultrasound report, covering the degree of vascular stenosis, plaque morphology and blood flow velocity three elements, and automatically upload to the hospital quality control platform to complete the text consistency analysis. In this batch of tests, the consistency score of the structured report is generally higher than 90 points, the highest is 95 points, the average consistency score is 92.6, which shows that the diagnostic report and the image content are highly matched.

[0156] In terms of image acquisition efficiency, the system statistics show that 10 cases save an average of about 61 seconds, and the efficiency is improved obviously. Especially for junior doctors under the condition of low proficiency, they can also quickly achieve standard images under the guidance of the system, significantly reducing the learning threshold and the rate of misoperation.

[0157] Table 1 Comparison of image quality and efficiency before and after AI guidance

[0158]

[0159] According to the data in Table 1, it can be clearly seen that the system of the present application has significant effects in improving image quality, enhancing report consistency and optimizing image acquisition efficiency. The data table records the core evaluation indicators of 10 real lower extremity artery ultrasound examination cases, covering image score, structured report quality and time efficiency, reflecting the application value of AI guidance mechanism in actual operation.

[0160] From the comparison results of image quality scores, the average image score before AI system guidance is 69.3, and after guidance, it is improved to an average of 88.1, with a significant overall improvement, which shows that the system effectively improves the operator's behavior deviation in angle control, pressure adjustment and position selection through real-time scoring and augmented reality guidance mechanism. Among them, the score of case number CASE_005 is improved from 59 to 91, with an improvement of 32 points, which shows that even in the case of low initial operation quality, the system also has strong guidance correction ability.

[0161] In terms of structured report consistency score, the scores of 10 cases are all above 85, with an average of 92.6, the highest is 95, and the lowest is also above 88. This shows that the structured report generated by the AI system has good integrity and accuracy in covering diagnostic elements, and is highly consistent with the image content collected, greatly reducing the subjective errors and omissions that may be brought by manual recording.

[0162] In terms of image acquisition efficiency, the AI-guided image acquisition time is generally shortened, with a time saving of between 29 seconds and 113 seconds. In particular, CASE_004 and CASE_008, the time saving is 113 seconds and 101 seconds respectively, indicating that during the image acquisition process, the system guidance significantly reduces the time waste caused by repeated image acquisition, substandard images or operation deviation. Overall, the average image acquisition time of the 10 cases is saved by about 65 seconds, which has practical significance for improving the examination turnover efficiency of the department and reducing the workload of physicians.

[0163] In addition, the operator composition includes "primary physician A", "primary physician B", "advanced physician C" and "advanced physician D". From the overall distribution, the image quality score of primary physicians is improved more obviously after AI guidance, further verifying that the system has a strong operation assistance value for beginners, which is helpful for standardized training and rapid capacity building.

[0164] In summary, the table data fully verifies that in the process of lower extremity arterial ultrasound examination, the present application significantly improves the image quality and operation standardization, and at the same time improves the examination efficiency and the standardization level of the report, which has good clinical practicability and popularization prospect.

[0165] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change within the technical range disclosed by the present application according to the technical scheme and the inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. An AI-based vertical large model lower extremity arterial ultrasound standardized examination system, characterized in that, Comprise the following modules: An image and operation data acquisition module for acquiring image and operation data during lower extremity arterial ultrasound examination; 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 when the image quality score is below a threshold, guiding the operator to adjust the probe direction and angle; A structured report generation module for automatically generating a structured examination report containing key diagnostic elements based on the output of the AI vertical large model; A cloud quality control analysis module for receiving structured reports and examination images, performing consistency analysis and quality evaluation, and outputting quality control feedback results; A remote expert review module for triggering expert remote review procedures when difficult or abnormal cases are identified in the quality control feedback, and outputting review reports; A federated learning update module for locally fine-tuning the AI vertical large model in each terminal based on the quality control feedback results, and completing distributed model optimization through federated aggregation mechanism; The lower extremity arterial ultrasound standardized examination system running process specifically comprises the following steps: S1, acquire lower extremity arterial ultrasound image data containing standard images and error images, label the lower extremity arterial ultrasound image data, and construct an image dataset; S2, preprocess the image dataset, extract image features of blood vessel boundaries and plaque structures using U-Net network, model probe operation sequences based on 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, wherein the perception layer extracts image features based on ResNet-50 network, the decision layer models probe trajectory using long short-term memory network and combines near-end policy optimization algorithm for action scoring and error correction feedback, and the output layer generates structured diagnosis results and is trained using the multi-modal training sample set; S4, integrate the trained AI vertical large model into the front-end examination device, real-time acquire examination images and operation data, input the AI vertical large model for quality scoring and guidance feedback, and generate structured reports; S5, upload the structured reports and examination images to the cloud quality control platform, perform consistency analysis and quality evaluation, output quality control feedback results, and trigger remote expert review when the cloud quality control platform identifies difficult or abnormal cases, complete review and return review reports; S6, based on the quality control feedback results, update the local parameters of the AI vertical large model, complete distributed training using federated learning mechanism, upload the update results to the aggregation node, perform unified integration and optimization, and construct a collaborative update process with privacy protection capability; S2 specifically comprises: S21, normalize the constructed image dataset, uniformly map the pixel values of each image to the [0, 1] interval, and perform uniform size resampling on image data of 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 by using the encoder part, perform upsampling and feature fusion by using the decoder part, output the segmentation mask image including the blood vessel boundary, plaque contour and lumen structure, and use the label mask obtained by manual labeling in the image data set to supervise the cross-entropy loss function, and build an image feature set denoted as , wherein represents the feature vector extracted from the i-th image data in the image data set, n is the total number of image data in the image data set.​ S23, synchronously record the probe operation information of the operator in the image acquisition process, including a sequence of probe position coordinates , a sequence of direction angles , and a sequence of contact pressure values with the probe , wherein, denotes a time step, is the position of the probe in the horizontal axis direction at the tth time step, is the position of the probe in the vertical axis direction at the tth time step, is the depth of the probe in the vertical direction at the tth time step, is the angle in the horizontal rotation direction at the tth time step, is the angle in the vertical tilt direction at the tth time step, is the contact pressure value of the probe acting on the skin surface at the tth time step; S24, input the probe position coordinate sequence, the direction angle sequence and the probe contact pressure sequence into the time sequence 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, the image feature set , probe track feature vector and label information Fusion is carried out to construct a multi-modal training sample set, and each sample is represented as a triple The S3 specifically comprises: S31, construct an AI vertical large model, divide it into a perception layer, a decision layer and an output layer, and respectively use them for image feature extraction, probe trajectory modeling and structured diagnosis generation; S32, the image feature set ResNet-50 network in the perception layer, extract deep semantic features, and generate a deep semantic feature vector; S33, the probe track feature vector input to a long short-term memory network LSTM, extract time-dependent features, and generate a time-dependent feature vector; S34, splice the deep semantic feature vector and the time-dependent feature vector to form a fusion vector as an input representation for strategy learning and diagnosis generation; S35、define task decomposition type instant reward function ; ； wherein, represents an image quality score, represents a consistency score of the probe path with a standard trajectory, represents a score gain after the operator response, , and are weight coefficients; S36, constructing a structure perception advantage function based on task decomposition reward ; ； wherein, represents the structure-aware advantage function estimate at the t-th time step, is the state value function output for the current state is the state value function output for the next state is the state value function output for the next state is the state value function output for the next state and represent the structure sharpness score function value for the state and represent the structure sharpness score function value for the state is the discount factor, is the advantage smoothing factor, represents the structure-aware advantage function estimate at the t+1-th time step; S37, a state-aware shearing range adjustment mechanism is constructed to adjust the shearing coefficient according to the combination of score fluctuation and reward change ; ； wherein, is an initial shear range, is a score fluctuation factor, is a reward change factor, is an image score standard deviation, is a previous step reward mean, is an immediate reward value of the previous time step; S38, based on structure perception advantage function and shear coefficient , construct proximal policy optimization objective function ; ； wherein, is a policy ratio, denotes a statistical expectation computation over a plurality of samples, is a min operation, is a policy ratio a clipping operation; S39, fuse the vectors inputting a policy network, outputting a policy distribution and sampling an action from it for probe direction, angle, and pressure guidance generation; S310、fuse the vectors synchronize input to a multi-layer perceptron in an output layer, output structured diagnosis results , results include information on the degree of vascular stenosis, plaque state and blood flow velocity; S311、constructing a joint training loss function ; ； wherein, is a structured label, CrossEntropy is a classification loss, is a loss balancing coefficient; S312. Using a multimodal training sample set The parameter set is optimized through backpropagation and gradient descent until the loss function is obtained. Convergence completes the training process of a large-scale AI vertical model; The S4 specifically comprises: S41, load and integrate the trained AI vertical large model into the front-end examination device, and real-time receive ultrasound images and operation inputs; S42, during the examination process, collect patient lower extremity artery ultrasound images and probe operation data, and synchronously input them to 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 score result; S44, if the image quality score result is lower than the set threshold, an insufficient image quality prompt is triggered, the prompt information is automatically fed back to the operator interface, a guidance prompt is generated in the examination interface through an augmented reality device, and visual information of probe adjustment direction, angle or pressure is displayed to the operator; S45, after the image quality meets the requirements, a structured report containing standard diagnosis elements is automatically generated based on the output of the AI vertical large model.

2. The AI-based vertical large model-based lower extremity arterial ultrasound standardized examination system according to claim 1, characterized in that, The lower extremity artery ultrasound image data specifically includes image data with multiple angles, different pressures, and cross-sectional positions and blood flow parameters, including standard images and error images.

3. The AI-based vertical large model-based lower extremity arterial ultrasound standardized examination system according to claim 1, characterized in that, The labeling processing of the lower extremity artery ultrasound image data refers to adding corresponding probe angle, pressure value, cross-sectional position and blood flow hemodynamic parameter label information to each image.

4. The AI-based vertical large model-based lower extremity arterial ultrasound standardized examination system according to claim 3, characterized in that, The S5 specifically comprises: S51, upload the structured report generated by the AI vertical large model and the corresponding examination image to the cloud quality control platform through the front-end examination device, complete the remote transmission of the examination data; S52, receive and analyze the structured report and the examination image in the cloud quality control platform, and analyze the content consistency between the two, including the matching of text conclusions and image diagnosis basis; S53, based on the image standards, parameter integrity and report element accuracy dimensions set by the platform, perform quality evaluation operation on the image and text data to form quality evaluation results; S54, the cloud quality control platform outputs standardized quality control feedback results in combination with the consistency analysis results and the quality evaluation results, and identifies whether there are suspicious or substandard items; S55, when the quality control feedback identifies difficult cases or there are inconsistencies between images and texts, diagnosis deviations and information missing, the cloud quality control platform automatically triggers a remote expert review process; S56, the qualified remote expert completes the review of the uploaded report and image within the preset time, and outputs the final review report, and the review conclusion is returned to the original examination terminal through the platform.

5. The AI-based vertical large model-based lower extremity arterial ultrasound standardized examination system according to claim 4, characterized in that, The S6 specifically comprises: 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 AI vertical large model local update process; S62, under the premise of not uploading the original image data and report information, the local AI vertical large model is incrementally trained using the labeled label data in the quality control feedback result, to generate AI vertical large model parameters optimized locally; S63, after completing the local training, the optimized AI vertical large model parameters are uploaded to the federal aggregation node in an encrypted form, without involving any transmission of original image, trajectory or structured text information; S64, in the federal aggregation node, the local AI vertical large model parameters from multiple front-end devices are received, and unified integration and weight optimization are performed according to the participation of each terminal and the sample feature distribution, to generate new global AI vertical large model parameters; S65, the updated global AI vertical large model parameters are synchronously distributed to the participating terminals, and replace the original parameter configuration, to realize the unified version iteration of the AI vertical large model in each front-end inspection device; S66, the local training, parameter uploading, aggregation updating and AI vertical large model synchronization process are repeatedly performed within the set federal training period, to form a distributed model collaborative updating process with privacy protection capability and continuous self-optimization.

Images