Thyroid ultrasound full-process intelligent quality control method
By using a collaborative quality control model and deep analysis and visual language model at the edge and cloud, the problem of incomplete quality control coverage for thyroid ultrasound was solved, achieving real-time quality control and reporting consistency throughout the entire process, thus improving the quality and interpretability of thyroid ultrasound examinations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-03-31
AI Technical Summary
In existing technologies, the quality control of thyroid ultrasound is incomplete, it is difficult to balance real-time performance with complex analysis, there is insufficient verification of report consistency, and there is a lack of closed-loop iterative optimization.
Deploy quality control models and decision logic modules at the edge, and combine them with cloud-based deep analysis and visual language models to achieve full-process quality control before, during, and after inspection. Through the collaborative work of quality control models, deep analysis models, and visual language models, real-time quality control and report verification are performed, and the model is updated through the training set to achieve closed-loop optimization.
It achieves full-process quality control coverage, improves the real-time performance, accuracy, and interpretability of thyroid ultrasound examinations, enhances examination quality and report consistency, and has sustainable iteration capabilities.
Smart Images
Figure CN121767747A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical ultrasound information technology and artificial intelligence technology, and in particular to an intelligent quality control method for the entire process of thyroid ultrasound. Background Technology
[0002] Medical ultrasound examinations are characterized by their real-time nature, reliance on operator experience, and significant impact on results from equipment parameters and scanning procedures. Thyroid ultrasound examinations are performed frequently and are highly repetitive in clinical practice; the main sources of quality variation are:
[0003] 1) Incomplete pre-examination information or inconsistencies between identity and medical orders led to procedural errors;
[0004] 2) Incomplete standard sections during inspection, non-standard parameter settings, and artifacts and noise effects lead to insufficient image usability;
[0005] 3) Missing fields, inconsistent terminology, and inconsistencies between image, measurement, and text descriptions in post-inspection reports lead to a decline in report quality;
[0006] 4) The follow-up management lacks a closed-loop mechanism, making it difficult to continuously improve and trace quality control.
[0007] Existing technologies already include ultrasonic quality control or quality screening schemes targeting "single links," for example:
[0008] A scheme that uses convolutional neural networks to classify image quality into high / medium / low and select qualified images for subsequent processing. This type of scheme focuses on image quality level classification and screening, mainly serving the quality control of subsequent model input, but usually does not cover the consistency quality control of pre- and post-inspection reports, and also lacks a cross-stage closed-loop optimization mechanism.
[0009] In scenarios where the probe is fixed or multiple areas are acquired, an image classification model is used to classify images into qualified / unqualified categories and adjust the probe position accordingly. This type of solution focuses on ensuring the availability of the acquisition process, but it does not form an end-to-end quality control system from pre-inspection to post-inspection report generation and review. Summary of the Invention
[0010] In view of this, the present invention provides an intelligent quality control method for the entire process of thyroid ultrasound, in order to solve the problems of incomplete quality control coverage, difficulty in balancing real-time performance and complex analysis, insufficient report consistency verification, and lack of closed-loop iterative optimization in the existing technology.
[0011] The present invention provides an intelligent quality control method for the entire process of thyroid ultrasound, comprising:
[0012] A quality control model and a decision logic module are deployed at the edge. The quality control model performs inspection quality control tasks on real-time frame images generated during the thyroid ultrasound examination. During the execution of the inspection quality control tasks, the decision logic module determines whether it is necessary to call the deep analysis model deployed in the cloud according to a first preset rule. If it is not necessary, the quality control model continues to execute the inspection quality control tasks. If it is necessary to call the deep analysis model, a request is sent to the cloud and the data required for analysis by the deep analysis model is uploaded. The deep analysis model performs deep analysis and returns the analysis results to the edge.
[0013] After a thyroid ultrasound examination, the quality control model performs a report quality control task on the examination report. After the report quality control task is performed, the decision logic module determines whether it is necessary to call the visual language model deployed in the cloud to analyze the examination report according to the second preset rule. If so, a request is sent to the cloud and the data required for analysis by the visual language model is uploaded. The visual language model analyzes and modifies the examination report and returns the modified report to the edge.
[0014] It also includes constructing different training sets for the quality control model, deep analysis model, and visual language model based on the data generated from the inspection quality control task and the report quality control task, and training and updating the model based on the constructed training sets.
[0015] Furthermore, the intelligent quality control method for the entire thyroid ultrasound process includes examination quality control tasks, including pre-examination quality control tasks and in-examination quality control tasks;
[0016] The pre-examination quality control tasks include: obtaining the examination task sheet from the quality control model, completing consistency verification based on patient identification information and examination arrival information, retrieving medical history and generating examination preparation prompts and recording the prompt completion status;
[0017] The quality control task during the examination includes: using a quality control model to detect quality defects in real-time frames acquired during thyroid ultrasound examinations, and outputting image quality scores, defect types, defect labels, defect type confidence levels, and lesion candidate regions. The defect types include inappropriate imaging parameters, artifacts, noise, and blur. The imaging parameters include at least one of grayscale statistics, contrast index, sharpness index, noise index, and artifact index.
[0018] The quality control task during the examination also includes identifying necessary views from the image sequence acquired by the quality control model from the thyroid ultrasound examination, determining whether the necessary views are complete, and prompting for supplementary acquisition of missing items. The necessary views include transverse and longitudinal views of the left and right lobes of the thyroid gland, isthmus view, nodule measurement view, and color blood flow view.
[0019] The in-inspection quality control task also includes generating an in-inspection quality control log from the quality control model. The in-inspection quality control log includes: timestamp, quality score, defect type, prompt, and operator response status.
[0020] Furthermore, the first preset rule includes triggering the invocation of the deep analysis model when any of the following conditions are met:
[0021] The following conditions are met: the quality score of the real-time frame output is lower than the set score; the confidence of the maximum defect type of the real-time frame output is lower than the set probability value; the defect type meets the predefined criteria for complex cases; a missing necessary view is detected; the imaging parameters deviate from the preset parameter range; and artifacts are present.
[0022] The artifact complexity refers to motion artifacts, sound occlusion, reverberation, side lobes, or noise enhancement that appear in B-mode grayscale frames or color blood flow frames.
[0023] Furthermore, the data required for the deep analysis model to perform the analysis is determined based on the triggering reason:
[0024] When the triggering reason is that the quality score is lower than the set score, the confidence level of the maximum defect type is lower than the set probability value, the defect type meets the predefined criteria for complex cases, a required view is detected as missing, or the imaging parameters deviate from the preset parameter range, the uploaded data includes keyframe indexes and structured features. The keyframes are frames selected from the image sequence acquired by ultrasound examination within a preset time window before and after the triggering time for depth analysis. The structured features include the quality score, defect type and confidence level, imaging parameter vector and its deviation, and candidate region map information.
[0025] When the triggering reason is the presence of complex artifacts, the uploaded data includes the anonymized original image or short video clip.
[0026] Furthermore, the report quality control task includes: checking the completeness of fields in the inspection report, the units and dimensions of data fields, whether the data is within the allowable range, the standardization of terminology, and the logical consistency between fields; it also includes checking the consistency between fields in the inspection report, structured fields in the inspection quality control log, and corresponding image evidence; and locating the detected non-conformities to the corresponding fields in the report and outputting revision prompts.
[0027] Furthermore, the second preset rule includes triggering the invocation of the visual language model when any of the following occurs:
[0028] The inspection report was found to be substandard during the quality control of the report; the inspection report involved predefined multiple lesions or difficult-to-diagnose indicators.
[0029] Furthermore, the data required for the visual language model to perform analysis is determined based on the triggering reason:
[0030] When the triggering reason is that an inspection report is found to be non-compliant during the execution of a report quality control task, the uploaded data includes: fields in the inspection report, structured fields in the inspection quality control log, non-compliance location information, and image evidence corresponding to the non-compliance.
[0031] When the triggering reason is that the examination report involves predefined multiple lesions or difficult indications, the uploaded data includes: fields in the examination report, structured fields in the quality control log during the examination, a set of measurement results organized by lesion number, a set of views confirmed during the examination, and image evidence corresponding to the lesion number.
[0032] Furthermore, the report processing tasks performed by the visual language model include:
[0033] The system generates modification suggestions for non-compliance in the report quality control task, generates explanatory suggestions for predefined multiple lesions or difficult indicators involved in the inspection report, and generates a modified report; then it sends the generated suggestions, the modified report, and the revised evidence to the edge terminal.
[0034] Furthermore, the training set for the quality control model consists of samples composed of images acquired from thyroid ultrasound examinations and corresponding defect labels for those images;
[0035] The training set for the depth analysis model consists of images captured by thyroid ultrasound examination and corresponding depth analysis labels for those images. The depth analysis labels include at least a defect type label and / or a defect location label, wherein the defect location label represents the defect region in the form of a bounding box, a segmentation mask, or key points.
[0036] The training set for the visual language model consists of samples composed of image evidence associated with fields in the inspection report, structured fields corresponding to the inspection report, and supervision labels corresponding to the structured fields. The supervision labels include at least field integrity labels, terminology standardization labels, unit and dimension consistency labels, consistency labels between fields in the inspection report, structured fields in the inspection quality control log, and image evidence, or logical conflict field location labels.
[0037] Furthermore, the quality control model includes a lightweight convolutional neural network or a lightweight visual Transformer; the deep analysis model includes a deep neural network for detection, segmentation, or quality regression; and the visual language model includes a visual encoder coupled with a language model.
[0038] The beneficial effects of this invention are:
[0039] This invention presents an intelligent quality control method for the entire process of thyroid ultrasound, which integrates quality control throughout the pre-examination, during-examination, post-examination, and follow-up stages. Through cloud-edge collaboration and a visual language model, it achieves real-time prompts, in-depth verification, and closed-loop optimization. This method balances the real-time performance, accuracy, interpretability, and sustainable iteration capabilities of the quality control process, effectively improving the quality of thyroid ultrasound examinations. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the decision-making process for edge computing and cloud computing.
[0041] Figure 2 This is a schematic diagram of the cloud-based deep analysis process.
[0042] Figure 3 This is the overall quality control process diagram. Detailed Implementation
[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0044] The intelligent quality control method for the entire thyroid ultrasound process in this embodiment includes:
[0045] A quality control model and decision logic module are deployed at the edge. The edge can be an ultrasound device, an ultrasound workstation, or a local server. The quality control model performs inspection quality control tasks on real-time frame images generated during the thyroid ultrasound examination. During the execution of the inspection quality control tasks, the decision logic module determines whether it is necessary to call the deep analysis model deployed in the cloud according to a first preset rule. If not, the quality control model continues to execute the inspection quality control tasks; if it is necessary to call the deep analysis model, a request is sent to the cloud and the data required for analysis by the deep analysis model is uploaded. The deep analysis model performs deep analysis and returns the analysis results to the edge.
[0046] After a thyroid ultrasound examination, the quality control model performs a report quality control task on the examination report. After the report quality control task is performed, the decision logic module determines whether it is necessary to call the visual language model deployed in the cloud to analyze the examination report according to the second preset rule. If so, a request is sent to the cloud and the data required for analysis by the visual language model is uploaded. The visual language model analyzes and modifies the examination report and returns the modified report to the edge.
[0047] It also includes constructing different training sets for the quality control model, deep analysis model, and visual language model based on the data generated from the inspection quality control task and the report quality control task, and training and updating the model based on the constructed training sets.
[0048] The quality control tasks described in this embodiment include pre-inspection quality control tasks and in-inspection quality control tasks.
[0049] The pre-examination quality control tasks include: obtaining the examination task sheet from HIS (Hospital Information Management System) or RIS (Radiology Information System) by the quality control model, completing consistency verification based on patient identification information and examination arrival information, retrieving medical history and generating examination preparation prompts (such as body position, neck exposure, removal of metal jewelry, etc.) and recording the completion status of the prompts;
[0050] The in-process quality control task includes: using a quality control model to detect quality defects in real-time frames acquired during thyroid ultrasound examinations, and outputting image quality scores, defect types, defect labels, defect type confidence levels, and lesion candidate regions. The defect types include inappropriate imaging parameters, artifacts, noise, and blur. The imaging parameters include at least one of grayscale statistics, contrast indices, sharpness indices, noise indices, and artifact indices. Grayscale statistics include the mean grayscale value, grayscale variance, grayscale histogram distribution characteristics, and grayscale dynamic range proportion within the image or lesion candidate region. Contrast indices include the contrast between the lesion candidate region and the background region. Sharpness indices include sharpness measures based on gradient energy, Laplacian variance, or edge intensity. Noise indices include the grayscale variance and signal-to-noise ratio of the background region. Artifact indices include measures of motion artifacts, acoustic occlusion, reverberation, and the intensity or coverage ratio of side lobes.
[0051] In this embodiment, the ultrasound image frame detected The image quality score is obtained by calculating and fusing five types of imaging quality parameters:
[0052] 1) Calculate image frames Candidate regions for internal lesions (denoted as) Gray-scale statistics (dynamic range / exposure) within:
[0053]
[0054] in Grayscale quantiles are used to reflect the effective dynamic range (overexposure / underexposure will cause...). (Become smaller).
[0055] 2) Calculate the contrast ratio (RMS contrast ratio)
[0056]
[0057] in For the mean and standard deviation, Prevent division by zero.
[0058] 3) Calculate the sharpness index (existing classic algorithm: Laplace variance VoL)
[0059]
[0060] in This is the Laplacian operator. The clearer the image and the sharper the edges, the better. Generally, the larger the size, the more significantly the blur will decrease. .
[0061] 4) Calculate noise levels:
[0062] Gaussian smoothing is obtained first residual Take again:
[0063]
[0064] MAD is the median absolute deviation, used for robust estimation of noise intensity.
[0065] 5) Calculate the artifact quantification index:
[0066] In this example, the strong acoustic shadow / occlusion coverage ratio is an artifact quantification metric. It calculates the proportion of low-echo strips / occlusion areas within the ROI, for example:
[0067] in Use the grayscale quantile of the ROI or a fixed threshold; the larger the proportion, the more severe the artifacts.
[0068] 6) Normalize and fuse the results into an image quality score (0–100):
[0069] Normalizing the above five types of imaging quality parameters yields Then, the image quality score Q is obtained by weighting the results.
[0070]
[0071] in And the sum is 1; the greater the noise and artifacts, the worse the quality, therefore use , Enter the weighted items.
[0072] The quality control task during the examination also includes identifying necessary views from the image sequence acquired by the quality control model from the thyroid ultrasound examination, determining whether the necessary views are complete, and prompting for supplementary acquisition of missing items. The necessary views include transverse and longitudinal views of the left and right lobes of the thyroid gland, isthmus view, nodule measurement view, and color blood flow view.
[0073] The in-inspection quality control task also includes generating an in-inspection quality control log from the quality control model. The in-inspection quality control log includes: timestamp, quality score, defect type, prompt, and operator response status.
[0074] In this example, the first preset rule includes triggering the invocation of the deep analysis model when any of the following conditions are met:
[0075] The following conditions are met: the quality score of the real-time frame output is lower than the set score; the confidence level of the maximum defect type of the real-time frame output is lower than the set probability value; the defect type meets the predefined criteria for complex cases; a required view is missing; the imaging parameters deviate from the preset parameter range; and artifacts are present.
[0076] The artifact complexity refers to motion artifacts, sound occlusion, reverberation, side lobes, or noise enhancement that appear in B-mode grayscale frames or color blood flow frames.
[0077] In this embodiment, the data required for the deep analysis model to perform analysis is determined based on the triggering reason:
[0078] When the triggering reason is that the quality score is lower than the set score, the confidence level of the maximum defect type is lower than the set probability value, the defect type meets the predefined criteria for complex cases, a required view is detected as missing, or the imaging parameters are detected as deviating from the preset parameter range, the uploaded data includes keyframe indexes and structured features. The keyframes are frames selected from the image sequence acquired by ultrasound examination within a preset time window before and after the triggering time for depth analysis. The structured features include the quality score, defect type and confidence level, imaging parameter vector and its deviation, and candidate region map information.
[0079] When the triggering reason is the presence of complex artifacts, the uploaded data includes the anonymized original image or short video clip.
[0080] In this embodiment, the report quality control task includes: checking the completeness of fields in the inspection report, the units and dimensions of data fields, whether the data is within the allowable range, the standardization of terminology, and the logical consistency between fields; it also includes checking the consistency between fields in the inspection report, structured fields in the inspection quality control log, and corresponding image evidence; and locating the detected non-conformities to the corresponding fields in the report and outputting revision prompts.
[0081] In this embodiment, the second preset rule includes triggering the invocation of the visual language model when any of the following occurs:
[0082] The inspection report was found to be substandard during the quality control of the report; the inspection report involved predefined multiple lesions or difficult-to-diagnose indicators.
[0083] In this embodiment, the data required for the visual language model to perform analysis is determined based on the triggering reason:
[0084] When the triggering reason is that an inspection report is found to be non-compliant during the execution of a report quality control task, the uploaded data includes: fields in the inspection report, structured fields in the inspection quality control log, non-compliance location information, and image evidence corresponding to the non-compliance.
[0085] When the triggering reason is that the examination report involves predefined multiple lesions or difficult indications, the uploaded data includes: fields in the examination report, structured fields in the quality control log during the examination, a set of measurement results organized by lesion number, a set of views confirmed during the examination, and image evidence corresponding to the lesion number.
[0086] In this embodiment, the report processing task performed by the visual language model includes:
[0087] The system generates modification suggestions for non-compliance in the report quality control task, generates explanatory suggestions for predefined multiple lesions or difficult indicators involved in the inspection report, and generates a modified report; then it sends the generated suggestions, the modified report, and the revised evidence to the edge terminal.
[0088] In this embodiment, the training set for the quality control model consists of samples composed of images acquired from thyroid ultrasound examinations and corresponding defect labels for those images.
[0089] The training set for the depth analysis model consists of images captured by thyroid ultrasound examination and corresponding depth analysis labels for those images. The depth analysis labels include at least a defect type label and / or a defect location label, wherein the defect location label represents the defect region in the form of a bounding box, a segmentation mask, or key points.
[0090] The training set for the visual language model consists of samples composed of image evidence associated with fields in the inspection report, structured fields corresponding to the inspection report, and supervision labels corresponding to the structured fields. The supervision labels include at least field integrity labels, terminology standardization labels, unit and dimension consistency labels, consistency labels between fields in the inspection report, structured fields in the inspection quality control log, and image evidence, or logical conflict field location labels.
[0091] Furthermore, in this embodiment, the intelligent quality control method for the entire thyroid ultrasound process also includes, after the report quality control task is completed, the quality control model generates a follow-up task for the doctor to the patient based on the structured fields in the final examination report, and records the follow-up feedback.
[0092] In this embodiment, the quality control model includes a lightweight convolutional neural network or a lightweight visual Transformer; the deep analysis model includes a deep neural network for detection, segmentation, or quality regression; and the visual language model includes a visual encoder coupled with a language model.
[0093] In this embodiment, the lightweight convolutional neural network or lightweight visual Transformer can be MobileNet, EfficientNet, or Swin-Transformer; however, the quality control model is not limited to the aforementioned models.
[0094] In this embodiment, the deep neural network may be U-Net, DeepLab, Mask R-CNN, or YOLO series; however, the deep analysis model is not limited to the aforementioned models.
[0095] In this embodiment, the visual language model may be BLIP-2, LLaVA, or Qwen-VL; however, the visual language model is not limited to the aforementioned models.
[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A fully intelligent quality control method for thyroid ultrasound, characterized in that, include: A quality control model and a decision logic module are deployed at the edge. The quality control model performs inspection quality control tasks on real-time frame images generated during the thyroid ultrasound examination. During the execution of the inspection quality control tasks, the decision logic module determines whether it is necessary to call the deep analysis model deployed in the cloud according to a first preset rule. If it is not necessary, the quality control model continues to execute the inspection quality control tasks. If it is necessary to call the deep analysis model, a request is sent to the cloud and the data required for analysis by the deep analysis model is uploaded. The deep analysis model performs deep analysis and returns the analysis results to the edge. After a thyroid ultrasound examination, the quality control model performs a report quality control task on the examination report. After the report quality control task is performed, the decision logic module determines whether it is necessary to call the visual language model deployed in the cloud to analyze the examination report according to the second preset rule. If so, a request is sent to the cloud and the data required for analysis by the visual language model is uploaded. The visual language model analyzes and modifies the examination report and returns the modified report to the edge. It also includes constructing different training sets for the quality control model, deep analysis model, and visual language model based on the data generated from the inspection quality control task and the report quality control task, and training and updating the model based on the constructed training sets.
2. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 1, characterized in that: The aforementioned quality control tasks include pre-inspection quality control tasks and in-inspection quality control tasks; The pre-examination quality control tasks include: obtaining the examination task sheet from the quality control model, completing consistency verification based on patient identification information and examination arrival information, retrieving medical history and generating examination preparation prompts and recording the prompt completion status; The quality control task during the examination includes: using a quality control model to detect quality defects in real-time frames acquired during thyroid ultrasound examinations, and outputting image quality scores, defect types, defect labels, defect type confidence levels, and lesion candidate regions. The defect types include inappropriate imaging parameters, artifacts, noise, and blur. The imaging parameters include at least one of grayscale statistics, contrast index, sharpness index, noise index, and artifact index. The quality control task during the examination also includes identifying necessary views from the image sequence acquired by the quality control model from the thyroid ultrasound examination, determining whether the necessary views are complete, and prompting for supplementary acquisition of missing items. The necessary views include transverse and longitudinal views of the left and right lobes of the thyroid gland, isthmus view, nodule measurement view, and color blood flow view. The in-inspection quality control task also includes generating an in-inspection quality control log from the quality control model. The in-inspection quality control log includes: timestamp, quality score, defect type, prompt, and operator response status.
3. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 2, characterized in that: The first preset rule includes triggering the invocation of the deep analysis model when any of the following conditions are met: The following conditions are met: the quality score of the real-time frame output is lower than the set score; the confidence of the maximum defect type of the real-time frame output is lower than the set probability value; the defect type meets the predefined criteria for complex cases; a missing necessary view is detected; the imaging parameters deviate from the preset parameter range; and artifacts are present. The artifact complexity refers to motion artifacts, sound occlusion, reverberation, side lobes, or noise enhancement that appear in B-mode grayscale frames or color blood flow frames.
4. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 3, characterized in that: The data required for analysis by the deep analysis model is determined based on the triggering reason: When the triggering reason is that the quality score is lower than the set score, the confidence level of the maximum defect type is lower than the set probability value, the defect type meets the predefined criteria for complex cases, a required view is detected as missing, or the imaging parameters deviate from the preset parameter range, the uploaded data includes keyframe indexes and structured features. The keyframes are frames selected from the image sequence acquired by ultrasound examination within a preset time window before and after the triggering time for depth analysis. The structured features include the quality score, defect type and confidence level, imaging parameter vector and its deviation, and candidate region map information. When the triggering reason is the presence of complex artifacts, the uploaded data includes the anonymized original image or short video clip.
5. The intelligent quality control method for the entire process of thyroid ultrasound according to any one of claims 1-4, characterized in that: The report quality control tasks include: checking the completeness of fields in the inspection report, the units and dimensions of data fields, whether the data is within the allowable range, the standardization of terminology, and the logical consistency between fields; it also includes checking the consistency between fields in the inspection report, structured fields in the inspection quality control log, and corresponding image evidence; and locating the detected non-conformities to the corresponding fields in the report and outputting revision prompts.
6. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 5, characterized in that: The second preset rule includes triggering the invocation of the visual language model when any of the following occurs: The inspection report was found to be substandard during the quality control of the report; the inspection report involved predefined multiple lesions or difficult-to-diagnose indicators.
7. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 6, characterized in that: The data required for analysis by the visual language model is determined based on the triggering reason: When the triggering reason is that an inspection report is found to be non-compliant during the execution of a report quality control task, the uploaded data includes: fields in the inspection report, structured fields in the inspection quality control log, non-compliance location information, and image evidence corresponding to the non-compliance. When the triggering reason is that the examination report involves predefined multiple lesions or difficult indications, the uploaded data includes: fields in the examination report, structured fields in the quality control log during the examination, a set of measurement results organized by lesion number, a set of views confirmed during the examination, and image evidence corresponding to the lesion number.
8. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 7, characterized in that: The report processing tasks performed by the visual language model include: The system generates modification suggestions for non-compliance of the report quality control task test, generates explanatory suggestions for predefined multiple lesions or difficult indicators involved in the inspection report, and generates the modified report. The generated recommendations, revised reports, and evidence cited in the revisions are then sent to the edge.
9. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 8, characterized in that: The training set for the quality control model consists of samples composed of images acquired from thyroid ultrasound examinations and corresponding defect labels for those images. The training set for the depth analysis model consists of images captured by thyroid ultrasound examination and corresponding depth analysis labels for the images. The depth analysis labels include at least: defect type labels and / or defect location labels, wherein the defect location labels represent the defect region in the form of bounding boxes, segmentation masks, or key points. The training set for the visual language model consists of samples composed of image evidence associated with fields in the inspection report, structured fields corresponding to the inspection report, and supervision labels corresponding to the structured fields. The supervision labels include at least field integrity labels, terminology standardization labels, unit and dimension consistency labels, consistency labels between fields in the inspection report, structured fields in the inspection quality control log, and image evidence, or logical conflict field location labels.
10. The intelligent quality control method for the entire process of thyroid ultrasound according to claim 1, characterized in that: The quality control model includes a lightweight convolutional neural network or a lightweight visual Transformer; the deep analysis model includes a deep neural network for detection, segmentation, or quality regression; and the visual language model includes a visual encoder coupled with a language model.