Ultrasonic report quality inspection method suitable for schools and hospitals, electronic equipment and program product

Abstract

The invention provides an ultrasonic report quality inspection method suitable for schools and hospitals, electronic equipment and a program product, and relates to the technical field of computers. According to the method, unstructured ultrasonic reports are converted into standardized structured report slices, automatic classification of the reports is achieved based on conclusion content, then a targeted predefined category report description template is called for matching verification, and finally a quality inspection result is automatically generated. According to the mode, a standardized and targeted automatic quality inspection process is realized, a unified quality inspection judgment standard is determined, the problems of subjective deviation and standard non-uniformity of traditional manual auditing and omissions and misjudgments caused by fatigue are effectively avoided, and the accuracy and result consistency of ultrasonic report quality inspection are greatly improved; meanwhile, the method supports batch processing of the reports to be subjected to quality inspection, can efficiently complete quality inspection work of massive ultrasonic reports, remarkably improves the quality inspection efficiency, reduces the labor intensity of manual review, and meets the auditing requirements of massive ultrasonic reports in a medical examination scene.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a method, electronic device, and program product for quality inspection of ultrasound reports applicable to school hospitals. Background Technology

[0002] Within the current school healthcare system, the school hospital faces immense pressure in handling centralized physical examinations. Due to the highly concentrated examination schedule, large daily patient volume, and tight report deadlines, the traditional review of ultrasound reports relies entirely on manual review by physicians. Reviewers must read each report individually, using their professional knowledge to determine whether the descriptive content is logically consistent with the diagnostic conclusions and whether the wording is standardized.

[0003] During peak health checkup periods, the sheer volume of reports generated in a short time makes this highly manpower-dependent approach a bottleneck in the workflow, leading to delays in report issuance. Furthermore, the repetitive and demanding nature of the review process easily causes visual and mental fatigue for doctors, making it difficult to maintain consistent judgment standards and significantly increasing the risk of oversights and misjudgments.

[0004] Therefore, the existing technology system is unable to achieve efficient, standardized, and traceable automated quality inspection of large-capacity ultrasound reports while ensuring quality, and it is difficult to meet the dual requirements of efficiency and consistency in report review during peak periods of physical examinations at the university hospital. Summary of the Invention

[0005] The purpose of this application is to provide a method, electronic device, and program product for ultrasound report quality inspection applicable to schools and hospitals, so as to improve the existing technology that cannot meet the dual requirements of modern medical institutions for report quality and operational efficiency.

[0006] In a first aspect, embodiments of this application provide a method for quality inspection of ultrasound reports applicable to school hospitals, the method comprising: Obtain multiple ultrasound reports awaiting quality inspection; Extract the descriptive and conclusion texts of each examination item from each ultrasound report to be inspected, and generate a structured report slice corresponding to each examination item; The multiple ultrasound reports to be inspected are classified according to the conclusion text in the structured report slices to obtain multiple categories of ultrasound reports to be inspected. For each type of ultrasound report to be inspected, the corresponding structured report slice is matched and verified with the predefined category report description template to obtain the corresponding verification result; Based on the verification results, the quality inspection result for each ultrasound report to be inspected is determined.

[0007] In the above implementation process, unstructured ultrasound reports are transformed into standardized structured report slices, and reports are automatically classified based on their conclusions. Then, targeted predefined category report description templates are used for matching and verification, ultimately generating automated quality inspection results. This approach achieves a standardized and targeted automated quality inspection process. It not only establishes unified quality inspection judgment standards, effectively avoiding the subjective biases, inconsistent standards, and oversights and misjudgments caused by fatigue inherent in traditional manual review, but also significantly improves the accuracy and consistency of ultrasound report quality inspection results. Furthermore, this method supports batch processing of reports awaiting quality inspection, enabling efficient completion of quality inspection work on massive numbers of ultrasound reports, significantly improving inspection efficiency, reducing the workload of manual review, and adapting to the large-scale ultrasound report review needs in medical examination scenarios.

[0008] Optionally, the multiple types of ultrasound reports to be inspected include abnormal ultrasound reports and normal ultrasound reports. For each type of ultrasound report to be inspected, the corresponding structured report slice is matched and verified against a predefined category report description template to obtain the corresponding verification result, including: For the normal ultrasound report, the corresponding structured report slice is matched and verified with the predefined normal report description template to obtain the corresponding first verification result; For the abnormal ultrasound report, the corresponding structured report slice is matched and verified with the predefined abnormal report description template to obtain the corresponding second verification result.

[0009] In the above implementation process, the ultrasound reports to be inspected are accurately divided into two categories: normal and abnormal. Predefined, exclusive description templates are matched to each category for targeted verification, achieving differentiated and refined processing of ultrasound report quality inspection and effectively avoiding the coarseness of a uniform verification model. Simultaneously, standardized verification templates are set for normal and abnormal reports, allowing the structured report slices of the inspection items to accurately match the matching templates. This establishes unified and standardized quality inspection judgment standards for both categories of reports, significantly improving the accuracy and objectivity of the inspection results and effectively reducing the subjective bias and oversights caused by inconsistent standards in traditional manual review.

[0010] Optionally, for the normal ultrasound report, matching and verifying the corresponding structured report slice with a predefined normal report description template to obtain the corresponding first verification result includes: For the normal ultrasound report, obtain the normal report description template corresponding to each examination item; Insert the normal report description template into the preset main prompt word frame to generate the corresponding target prompt word; Based on preset consistency verification rules and preset judgment conditions, the structured report slices of the target prompt words and corresponding check items are verified to obtain the corresponding first verification result; The preset consistency verification rules include at least one of the following: word consistency, punctuation consistency, sequence consistency, and option validity; the preset judgment conditions include: whether the description text in the structured report slice is consistent with the normal report description template, whether the conclusion text uses the standardized options provided in the normal report description template, and whether the use of each text attribute is standardized.

[0011] In the above implementation process, a standardized and refined hierarchical verification system was constructed for normal ultrasound reports. First, a dedicated normal report description template was matched according to the examination items. Then, a unified target prompt word was generated by combining the main prompt word framework. This made the verification process of normal reports a fixed standard and completely avoided the problems of inconsistent verification standards and arbitrary processes. At the same time, relying on multi-dimensional consistency verification rules including word, punctuation, order, and option legality, as well as clear judgment conditions for description text matching degree, standardized use of conclusion options, and text attribute standardization, the system achieved accurate verification of all elements of the examination items. It can meticulously identify various non-standard details in the descriptions and conclusions of normal reports, so that the verification judgment has a clear basis and ensures the consistency and objectivity of the quality inspection results of all normal ultrasound reports.

[0012] Optionally, the step of matching and verifying the corresponding structured report slice with a predefined abnormal report description template for the abnormal ultrasound report to obtain a corresponding second verification result includes: Identify the lesion type corresponding to each abnormal ultrasound report; The abnormal ultrasound reports are classified according to the lesion type to obtain multiple report subsets corresponding to different lesion types. Retrieve the lesion report description template corresponding to each lesion type from the template library; According to the preset review rules, the structured report slices of each report in the report subset are verified using the lesion report description template to obtain the corresponding second verification result; The preset review rules include: whether the description structure is consistent, whether the logical relationship is consistent, and whether the numerical range is consistent.

[0013] In the above implementation process, a refined verification strategy guided by lesion type was adopted for abnormal ultrasound reports. First, the lesion type was accurately identified and classified into exclusive report subsets. Then, corresponding lesion report description templates were matched for each subset to carry out targeted verification. This ensured that the verification standards for abnormal reports were highly compatible with the professional description specifications of different lesions, solving the problem of the lack of specificity and poor adaptability of the unified verification of abnormal reports. At the same time, the verification was carried out based on preset review rules that focused on the description structure, logical relationship and numerical range. This can accurately identify key issues such as the mismatch between the description and the conclusion structure in the report, the contradiction in the diagnostic logic, and the lesion numerical description exceeding the standard range, which greatly improves the accuracy of abnormal report verification.

[0014] Optionally, the step of classifying the multiple ultrasound reports to be inspected based on the conclusion text in the structured report slices to obtain multiple categories of ultrasound reports to be inspected, including: The conclusion text in the structured report slice corresponding to each examination item is traversed, and anomaly identification is performed according to the preset lesion identification rules to obtain the identification results. The identification results include lesion presence and lesion absence. The lesion identification rules include whether the conclusion in the conclusion text is an empty field, whether it contains predefined negative words, and / or whether there is no specific lesion description. If the identification result indicates the presence of lesions, the ultrasound report to be inspected will be classified as an abnormal ultrasound report. If the identification result is no lesion, the ultrasound report to be inspected will be classified as a normal ultrasound report.

[0015] In the above implementation process, based on the examination items as the basic dimension, anomaly identification is carried out by traversing the conclusion text of the structured report slices. This allows the classification judgment of ultrasound reports to be reduced to the level of specific examination items, realizing the refinement and accuracy of the classification work. This lays a solid foundation for accurate classification from the source for subsequent differential verification. At the same time, the preset lesion identification rules are formulated around objective dimensions such as empty fields in the conclusion text, predefined negative words, and specific lesion descriptions. This provides the report classification with clear and unified quantitative judgment criteria, ensuring the objectivity and consistency of the classification results.

[0016] Optionally, before obtaining multiple ultrasound reports to be inspected, the process further includes: Obtain multiple initial ultrasound reports awaiting quality inspection; The intent of each initial ultrasound report to be inspected is identified using a large language model, resulting in a set of ultrasound reports to be inspected with the intent type of refusal and a set of ultrasound reports to be inspected with the intent type of normal inspection. The set of ultrasound reports to be inspected includes multiple ultrasound reports to be inspected.

[0017] In the above implementation process, before formally carrying out the quality inspection of ultrasound reports, the initial ultrasound reports to be inspected are identified by a large language model, and the rejection and normal inspection reports are accurately classified. This realizes the pre-screening and precise allocation of resources for the quality inspection work, effectively separating the rejection reports that do not need to be further structured for quality inspection, avoiding the waste of quality inspection resources and ineffective workload caused by them entering the subsequent process. This allows the core quality inspection work such as subsequent slicing, classification, and verification to focus on the normal inspection reports that really need to be reviewed, which greatly improves the operational efficiency of the overall quality inspection process.

[0018] Optionally, the quality inspection results include those that pass inspection and those pending manual review. After determining the quality inspection result for each ultrasound report pending quality inspection based on the verification results, the process further includes: Ultrasound reports marked as awaiting manual review are output to the physician's terminal.

[0019] In the above process, after completing the automated quality inspection of the ultrasound report and determining the inspection results as either passed or pending manual review, the report pending manual review is accurately output to the physician's terminal. This achieves efficient connection and precise linkage between automated quality inspection and manual review, allowing physicians to focus on reports requiring manual review directly without having to sift through a massive number of reports to identify problematic ones. This significantly reduces the physician's ineffective workload and improves the relevance and efficiency of manual review.

[0020] Optionally, after outputting the ultrasound report marked as requiring manual review to the physician's terminal, the process further includes: Collect feedback results from manual review of ultrasound reports awaiting manual verification; Based on the feedback from the manual review, the predefined category report description template was adjusted.

[0021] In the above implementation process, after the report to be manually reviewed is pushed to the physician's terminal and the feedback results of the manual review are collected, the predefined category report description templates are dynamically adjusted according to the feedback to build a closed-loop ultrasound report quality inspection system. This allows the template optimization to have real and professional clinical review basis, making the report description templates of each category more in line with the actual writing standards of ultrasound reports and the professional diagnosis and treatment requirements of medical clinical practice, greatly improving the adaptability and accuracy of the templates, and thus effectively improving the accuracy of subsequent automated verification.

[0022] Secondly, embodiments of this application provide an ultrasound report quality inspection device, the device comprising: The report acquisition module is used to acquire multiple ultrasound reports to be inspected. The slice generation module is used to extract the descriptive text and conclusion text of each examination item in each ultrasound report to be inspected, and generate a structured report slice corresponding to each examination item. The classification module is used to classify the multiple ultrasound reports to be inspected based on the conclusion text in the structured report slices, and obtain multiple categories of ultrasound reports to be inspected. The verification module is used to match and verify the corresponding structured report slices with the predefined category report description templates for each type of ultrasound report to be inspected, and obtain the corresponding verification results. The quality inspection module is used to determine the quality inspection result of each ultrasound report to be inspected based on the verification result.

[0023] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps of the method provided in the first aspect above are performed.

[0024] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method provided in the first aspect above.

[0025] Fifthly, embodiments of this application provide a computer program product, including computer program instructions, which, when read and executed by a processor, perform the steps of the method provided in the first aspect above.

[0026] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart of an ultrasound report quality inspection method provided in this application embodiment; Figure 2 A detailed flowchart illustrating the implementation of an ultrasound report quality inspection method provided in this application embodiment; Figure 3 A structural block diagram of an ultrasonic report quality inspection device provided in this application embodiment; Figure 4This is a schematic diagram of the structure of an electronic device for performing an ultrasound report quality inspection method, provided as an embodiment of this application. Detailed Implementation

[0029] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0030] It should be noted that the terms "system" and "network" in the embodiments of this invention can be used interchangeably. "Multiple" refers to two or more; therefore, in the embodiments of this invention, "multiple" can also be understood as "at least two". "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0031] It should also be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0032] This application provides a method for quality inspection of ultrasound reports applicable to school hospitals. This method transforms unstructured ultrasound reports into standardized structured report slices, automatically classifies reports based on their conclusions, and then uses targeted predefined category report description templates for matching and verification, ultimately automatically generating quality inspection results. This approach achieves a standardized and targeted automated quality inspection process, establishing unified quality inspection judgment standards and effectively avoiding the subjective biases, inconsistent standards, and oversights caused by fatigue inherent in traditional manual review. This significantly improves the accuracy and consistency of ultrasound report quality inspection results. Furthermore, this method supports batch processing of reports awaiting quality inspection, enabling efficient completion of quality inspection work on massive numbers of ultrasound reports, significantly improving inspection efficiency, reducing the workload of manual review, and adapting to the large-volume ultrasound report review needs in medical examination scenarios.

[0033] Please refer to Figure 1 , Figure 1 A flowchart of an ultrasound report quality inspection method applicable to school hospitals, provided for embodiments of this application, is included in the following steps: Step S110: Obtain multiple ultrasound reports to be inspected.

[0034] The quality inspection system can retrieve ultrasound reports awaiting quality inspection in batches from hospital information systems, health checkup center platforms, or image archiving systems. These reports are typically in free text format, containing basic information about the examinee, examination findings (description), and diagnostic opinions (conclusion). For example, an abdominal ultrasound report might include the descriptive text "The liver is of normal size and shape, with a smooth capsule, homogeneous parenchymal echoes, and clear intrahepatic blood vessels," and the conclusion text "No obvious abnormalities were found in the liver."

[0035] Step S120: Extract the descriptive text and conclusion text of each examination item in each ultrasound report to be inspected, and generate a structured report slice corresponding to each examination item.

[0036] The quality control system can employ natural language processing (NLP) technology, such as large language models pre-trained in the medical field, to identify various examination items in the report (e.g., organs or parts such as the liver, gallbladder, pancreas, spleen, kidneys, bladder / ureter, prostate, uterus, breast, thyroid, neck, and heart). For each examination item, the system locates and extracts its descriptive and concluding texts from the original report and organizes them into structured data units, such as standardized JSON format data, called structured report slices. For example, for the "gallbladder" examination item, the following slice might be generated: JSON { "organ": "gallbladder", Description: "Normal size and shape, smooth walls, good sound transmission within the cavity," Conclusion: No abnormalities were found in the gallbladder. } For examinations not mentioned, null slices are generated. This process transforms unstructured clinical text into standardized data that can be processed by machines, laying the foundation for subsequent automated analysis.

[0037] Step S130: Classify multiple ultrasound reports to be inspected according to the conclusion text in the structured report slice to obtain multiple types of ultrasound reports to be inspected.

[0038] After obtaining the structured report slides, the report can be intelligently categorized based on the conclusions of the structured slides. The categorization logic can be based on predefined medical knowledge rules. For example, if the conclusions of all examination items are empty values ​​or contain negative expressions such as "no obvious abnormalities" or "normal," the report is categorized as "Category A - Overall Normal Report." If the conclusion of any examination item contains specific abnormality detection terms (such as "cyst," "nodule," "thickening," etc.), the report is categorized as "Category B - Abnormality Detection Report." If the conclusion section explicitly states "patient refused examination" or "examination not completed," it is categorized as "Category C - Refusal / Incomplete Report." For example, a report with only the thyroid examination conclusion of "hypoechoic nodule in the right lobe of the thyroid" would be categorized as Category B.

[0039] Step S140: For each type of ultrasound report to be inspected, match and verify the corresponding structured report slice with the predefined category report description template to obtain the corresponding verification result; The quality inspection system can maintain a template knowledge base, predefining description templates and rules for each type of report, such as: For Category A reports, call the normal report template, verify whether the description uses standard and normal terminology (such as normal size and shape, smooth capsule, etc.), whether the conclusion strictly matches the preset options (such as no obvious abnormalities), and check the standardization of numerical units and punctuation.

[0040] For Category B reports, the system first identifies the specific abnormality type (e.g., "nodule" or "calcification"), and then calls the corresponding abnormality report template. This template includes essential elements of the abnormality description (e.g., location, size, echogenicity), logical association rules (e.g., if "hypoechoic" appears in the description, the conclusion should include a corresponding assessment), and numerical range validation (e.g., nodule size must be in "mm"). For example, the template for "thyroid nodule" will validate whether it fully includes key descriptive elements such as "location, size, boundary, shape, and echogenicity."

[0041] For Category C reports, check whether they contain a compliant statement of refusal to inspect record.

[0042] During the verification process, the system compares the structured slice of each inspection item with the corresponding template item by item, and can output a verification report with detailed results including complete match, partial match (such as non-standard terminology), and non-match (such as missing key elements).

[0043] Step S150: Based on the verification results, determine the quality inspection result for each ultrasound report to be inspected.

[0044] After obtaining the verification results of each inspection item in each ultrasound report to be inspected, the quality inspection system can generate a final quality inspection conclusion by combining the verification results of all inspection items. For example, if all inspection items are completely matched, the quality inspection is marked as passed; if there are partially matched items, it is marked as recommended for review and the specific non-standard content is indicated; if there are non-matched items (such as missing descriptions of key abnormalities), the quality inspection is marked as failed and modification guidelines are generated.

[0045] All subsequent results, along with the original report, structured slides, and verification details, can be output to the quality control platform for final confirmation by the reviewing physician or returned for modification.

[0046] In the above implementation process, unstructured ultrasound reports are transformed into standardized structured report slices, and reports are automatically classified based on their conclusions. Then, targeted predefined category report description templates are used for matching and verification, ultimately generating automated quality inspection results. This approach achieves a standardized and targeted automated quality inspection process. It not only establishes unified quality inspection judgment standards, effectively avoiding the subjective biases, inconsistent standards, and oversights and misjudgments caused by fatigue inherent in traditional manual review, but also significantly improves the accuracy and consistency of ultrasound report quality inspection results. Furthermore, this method supports batch processing of reports awaiting quality inspection, enabling efficient completion of quality inspection work on massive numbers of ultrasound reports, significantly improving inspection efficiency, reducing the workload of manual review, and adapting to the large-scale ultrasound report review needs in medical examination scenarios.

[0047] Based on the above embodiments, the multiple ultrasound reports to be inspected can be ultrasound reports that need further quality inspection after screening. For example, multiple initial ultrasound reports to be inspected can be obtained first, and then the intent of each initial ultrasound report to be inspected can be identified using a large language model to obtain a set of ultrasound reports to be rejected with the intent type of rejection and an opportunity for ultrasound reports to be inspected with the intent type of normal inspection. The set of ultrasound reports to be inspected includes multiple ultrasound reports to be inspected.

[0048] Intent recognition refers to the process of using large language models to analyze the text content of ultrasound reports and determine whether the report and internal examination items belong to "rejection" or "normal examination".

[0049] Refusal ultrasound reports are divided into full refusal reports and partial refusal reports. A full refusal report means that all examination items in the report include a description of refusal, while a partial refusal report means that the report contains both refusal items and normal examination items.

[0050] The set of ultrasound reports awaiting quality inspection is a collection of reports that, after intent identification and screening, consist of some rejected reports and normal inspection reports, and requires further structured quality inspection.

[0051] The quality inspection system can first export multiple initial ultrasound reports to be inspected in batches from hospital information systems (such as physical examination CRM systems). These reports cover basic information such as examinee ID, examination date, and examining physician, as well as complete ultrasound examination descriptions and conclusion texts. For example, 300 ultrasound reports generated by a school hospital in a single day can be exported as initial objects to be inspected.

[0052] The next step is to identify and screen the intent. A large language model is used to identify the intent of each initial ultrasound report to be inspected. The intent description results in the report text are analyzed item by item to distinguish the report type.

[0053] To address the textual characteristics of ultrasound reports, a large language model prompt word was designed specifically for intent recognition. This prompt word explicitly instructs the model to identify the examination intent from the report and defines two key outputs: Intent to refuse examination: This refers to reports that explicitly state the failure to complete the assessment of the scheduled examination item was due to patient reasons (such as refusal to undergo examination or inability to cooperate) or equipment / technical limitations. For example, if the intent descriptions for all examination items in the report contain refusal information, it is considered a complete refusal report and is included in the refusal ultrasound report collection.

[0054] Normal inspection intent: This means the report has completed the investigation and description of the predetermined inspection items, regardless of whether the conclusion is normal or not. For example, if some inspection items in the report contain rejected items and some items are normal inspection items, it is judged as a partially rejected inspection report; if all inspection items in the report do not contain rejected items, it is judged as a normal inspection report.

[0055] The prompt will provide clear judgment criteria and output format requirements, such as: "Please determine whether this ultrasound report is a refusal report. If the report explicitly mentions refusal, failure to perform examination, patient non-cooperation, or similar statements, resulting in one or more major organs not being described, it is judged as a refusal report. Otherwise, it is judged as a normal examination. Please only output 'refusal report' or 'normal examination'."

[0056] The quality inspection system combines each initial report text with the aforementioned prompts and calls the large language model API for batch asynchronous processing. The model reads the full report and, based on its understanding of the semantics of the medical text, determines its core intent.

[0057] The quality control system analyzes the text results returned by the large language model. If the result is a rejection, the report is categorized into the rejected ultrasound report set. For example, a report might conclude, "Due to the patient's severe cough, the liver, gallbladder, and pancreas areas were not clearly visualized, and the examination could not be completed." The large language model will recognize this as an "intent to reject the examination," and the report will be filtered into the rejected report set. If the result is a normal examination, the report is categorized into the ultrasound report set awaiting quality control. For example, even if a report describes "multiple liver cysts," because it completes the examination and description of the liver, the model will still determine it as an "intent to perform a normal examination," and it will proceed to the subsequent quality control process.

[0058] After the system completes the intent identification of all initial reports, it generates two distinct report sets: Rejected ultrasound reports collection: Reports in this collection will directly enter a special processing flow, usually only for record archiving or simple format compliance checks, without in-depth quality checks on organ-level description-conclusion consistency.

[0059] Ultrasound reports awaiting quality inspection: This collection contains reports that require and are suitable for further in-depth quality inspection. They will be sent to subsequent quality inspection modules.

[0060] In the above implementation process, before formally carrying out the quality inspection of ultrasound reports, the initial ultrasound reports to be inspected are identified by a large language model, and the rejection and normal inspection reports are accurately classified. This realizes the pre-screening and precise allocation of resources for the quality inspection work, effectively separating the rejection reports that do not need to be further structured for quality inspection, avoiding the waste of quality inspection resources and ineffective workload caused by them entering the subsequent process. This allows the core quality inspection work such as subsequent slicing, classification, and verification to focus on the normal inspection reports that really need to be reviewed, which greatly improves the operational efficiency of the overall quality inspection process.

[0061] Based on the above embodiments, in classifying multiple ultrasound reports to be inspected, the reports can be divided into two categories: abnormal and non-abnormal. Specifically, the conclusion text in the structured report slice corresponding to each examination item can be traversed, and abnormality identification can be performed according to preset lesion identification rules to obtain identification results. The identification results include lesion presence and lesion absence. The lesion identification rules include whether the interface in the conclusion text is an empty field, whether it contains predefined negative words, and / or whether there is no specific lesion description, etc. If the identification result is lesion presence, the ultrasound report to be inspected is classified as an abnormal ultrasound report; if the identification result is lesion absence, the ultrasound report to be inspected is classified as a normal ultrasound report.

[0062] The core of the classification process is to traverse and analyze the conclusion text in each examination slice. The system automatically analyzes each conclusion text according to preset lesion identification rules built upon a medical knowledge base. This rule base mainly includes the following logical judgment rules: 1) Empty Field Judgment: If the conclusion text is null or an empty string, it means that the report did not provide any conclusive opinion on that examination item. In routine physical examinations, the absence of mention is usually assumed to be "no abnormalities found," so this rule initially classifies such cases as no lesion clues.

[0063] 2) Negative word matching: If the conclusion text contains words from a predefined list of negative medical terms, such as "no obvious abnormalities," "normal," "no abnormal echoes detected," or "no space-occupying lesions," then the examination item is directly determined to have no lesions. The system maintains a continuously updated negative word dictionary to ensure the coverage and accuracy of the identification.

[0064] 3) Specific Lesion Description Detection: This is the core rule for identifying lesions. The system uses natural language processing technology to detect whether the conclusion text contains descriptive words or phrases indicating specific abnormalities or diseases, such as "cyst," "nodule," "calcification," "thickening," "rich blood flow signal," and "fatty liver." Once such keywords or expressions highly related to their meaning are detected, the system immediately determines that the examination item has a lesion.

[0065] Suppose an ultrasound report's liver section conclusion is "fatty liver (moderate)". During system iteration, rules 1 and 2 are not satisfied. However, when executing rule 3, the keyword "fatty liver" is detected, and the liver examination item is marked as having lesions. Meanwhile, the report's gallbladder section conclusion is "no abnormalities found in the gallbladder", which is then marked as having no lesions according to rule 2.

[0066] After completing a comprehensive analysis of all examination items in a single report, the system summarizes the identification results. The classification decision follows a simple yet strict principle: if any examination item is identified as having a lesion, the entire report is classified as an abnormal ultrasound report; conversely, only when all examination items are identified as having no lesions is the report classified as a normal ultrasound report.

[0067] As in the previous example, because the "liver" item was identified as having a lesion, even though the "gallbladder" item was normal, the entire report would still be classified into the abnormal ultrasound report set. Conversely, if the conclusion for all organs (such as liver, gallbladder, pancreas, spleen, kidney, etc.) in a report is "no abnormalities found" or a similar statement, then it would be classified into the normal ultrasound report set.

[0068] After classification, the two types of report sets will be routed to different downstream quality inspection pipelines: normal ultrasound reports will enter the standardized rapid verification channel, which mainly compares them with the standard normal template for consistency; while abnormal ultrasound reports will enter the logical consistency in-depth review channel, which requires calling more complex abnormal knowledge templates to strictly verify the medical logic between their descriptions and conclusions.

[0069] In the above implementation process, based on the examination items as the basic dimension, anomaly identification is carried out by traversing the conclusion text of the structured report slices. This allows the classification judgment of ultrasound reports to be reduced to the level of specific examination items, realizing the refinement and accuracy of the classification work. This lays a solid foundation for accurate classification from the source for subsequent differential verification. At the same time, the preset lesion identification rules are formulated around objective dimensions such as empty fields in the conclusion text, predefined negative words, and specific lesion descriptions. This provides the report classification with clear and unified quantitative judgment criteria, ensuring the objectivity and consistency of the classification results.

[0070] Based on the above embodiments, during verification, for normal ultrasound reports, the corresponding structured report slices are matched and verified with a predefined normal report description template to obtain the corresponding first verification result; for abnormal ultrasound reports, the corresponding structured report slices are matched and verified with a predefined abnormal report description template to obtain the corresponding second verification result.

[0071] The system sequentially retrieves reports and their structured report slices from the normal ultrasound report collection. For these reports, the system calls upon the normal report description template library for matching and verification. The normal report description template refers to the pre-defined standard description and conclusion paradigms for the lesion-free state of various organs in the template library, covering unified specifications for words, punctuation, and content order. Each template corresponds to an examination item (e.g., "Normal Liver Template," "Normal Gallbladder Template"), and specifies the following elements: 1. Standard Description Structure and Keywords: This section specifies the standard phrases, order, and terminology to be used when describing normal morphology. For example, a "normal liver template" might specify that the description must include, in order, four core phrases: normal size and shape, smooth capsule, homogeneous parenchymal echogenicity, and clear intrahepatic vascular course.

[0072] 2. Standard conclusion options: Provides a list of allowed conclusion phrases indicating normality, such as no obvious abnormalities or normal.

[0073] 3. Format and unit standards: including the correct use of punctuation marks and units of measurement (such as mm and cm).

[0074] The verification process is performed on a per-item basis. The system automatically compares a structured report slice (containing description and conclusion fields) of a specific inspection item in the report with the corresponding normal template in the template library.

[0075] When performing comparison and verification, the following items can be compared and verified: Exact match of description text: This checks whether the description text in the slice exactly matches the keywords and phrases in the order specified in the template. Any additional modifiers, synonym substitutions, or disordered order may result in a mismatch.

[0076] Validity of the conclusion text: Check whether the conclusion text in the slice strictly belongs to one of the standard conclusion options provided by the template.

[0077] Standardization check: Verify that the number format, units, and punctuation conform to the template specifications.

[0078] After traversing all the check item slices of a report, the system summarizes the verification results of each check item and generates the first verification result of the report. This result is usually output in a structured form (such as JSON), clearly listing the verification status (pass / fail) of each check item, as well as the specific reason for failure (such as missing descriptive elements or use of non-standard terminology).

[0079] For reports in the abnormal ultrasound report set, the system performs a more complex logical consistency check. First, based on the specific abnormality type identified in the report (e.g., "liver cyst," "thyroid nodule"), the system dynamically retrieves the corresponding abnormality report description template from the abnormality report description template library. An abnormality report description template refers to a standard description template categorized by lesion type in the template library, containing matching rules between descriptive elements such as lesion location, size, and morphology and the corresponding diagnostic conclusion. One template corresponds to one abnormality type (e.g., "liver cyst template," "thyroid nodule (TI-RADS 3) template"). Each template contains the following core rules: 1. Description Element Rules: These rules specify the key elements that must be included when describing the abnormality. For example, a liver cyst template might require the description to explicitly mention "location," "size," "shape," "boundary," "internal echo," and "posterior acoustic enhancement."

[0080] 2. Logical Relationship Rules: Define the necessary medical logical relationship between descriptive elements and conclusions. For example, if the description mentions "multiple microcalcifications seen within the nodule," the conclusion should suggest a risk of malignancy or recommend further examination (such as a biopsy).

[0081] 3. Numerical and Unit Compliance Rules: Specify the unit (mm or cm) for key measurements (such as cyst size) and a reasonable range of values.

[0082] 4. Structured correspondence rules: Ensure that the structure of the description section clearly corresponds to the conclusion section, and avoid contradictory situations where multiple lesions are described but only one is mentioned in the conclusion.

[0083] The verification process is also performed item by item. The system intelligently compares the sliced ​​content of the abnormal check item with the corresponding abnormal template. This process not only performs text matching, but also focuses on semantic and logical review, mainly including: Element completeness review: Check whether the description text contains all the key elements required by the template.

[0084] Logical consistency review: Using a rule engine, analyze whether there are logical contradictions or omissions in the template definition between the description content and the conclusion content.

[0085] Numerical compliance review: Extract the measurement values ​​in the description and check whether their units and value ranges comply with medical common sense and template regulations.

[0086] After completing a thorough verification of all abnormal items in the report, the system generates a second verification result for the report. This result is more detailed than the first verification result, listing not only the problematic items but also specifying whether the issue is due to missing key descriptive elements, logical contradictions between description and conclusion, or incorrect numerical units, providing clear guidance for reviewing physicians to verify and revise the report.

[0087] In the above implementation process, the ultrasound reports to be inspected are accurately divided into two categories: normal and abnormal. Predefined, exclusive description templates are matched to each category for targeted verification, achieving differentiated and refined processing of ultrasound report quality inspection and effectively avoiding the coarseness of a uniform verification model. Simultaneously, standardized verification templates are set for normal and abnormal reports, allowing the structured report slices of the inspection items to accurately match the matching templates. This establishes unified and standardized quality inspection judgment standards for both categories of reports, significantly improving the accuracy and objectivity of the inspection results and effectively reducing the subjective bias and oversights caused by inconsistent standards in traditional manual review.

[0088] Based on the above embodiments, in the matching and verification method for normal ultrasound reports, a normal report description template corresponding to each examination item can be obtained for each normal ultrasound report. This normal report description template is then inserted into a preset main prompt word framework to generate corresponding target prompt words. Finally, according to preset consistency verification rules and preset judgment conditions, the structured report slices of the target prompt words and corresponding examination items are verified to obtain the corresponding first verification result. The preset consistency verification rules include at least one of the following: word consistency, punctuation consistency, sequence consistency, and option legality. The preset judgment conditions include: whether the description text in the structured report slice is consistent with the normal report description template, whether the conclusion text uses the standardized options provided in the normal report description template, and whether the use of each text attribute is standardized.

[0089] The system can first retrieve and obtain the corresponding normal report description template for each examination item (such as liver and gallbladder) from the template library based on the examination items included in the current normal ultrasound report to be verified. This template is a structured data object that clearly defines the standard description text for the examination item, the list of allowed standard conclusion options, and other format specifications.

[0090] Next, the system dynamically inserts these specific template contents into a preset main prompt word framework, thereby generating target prompt words for each inspection item. The main prompt word framework is a general instruction template whose function is to clearly convey the specific review tasks and review standards to the large language model. For example, a main prompt word framework might be: "Please rigorously review the following ultrasound report fragment. The standard template requires that the description section must be entirely: '[Insert description template]'. The conclusion section must strictly use one of the following options: [Insert conclusion option list]. Please make judgments according to the following rules: ...".

[0091] Subsequently, the system calls the Large Language Model API, taking the generated target prompt words and the actual report slice text as input. The Large Language Model's output is a structured analysis of whether the slice content conforms to the template. The system's preset consistency verification rules are the core dimensions guiding the model's analysis, mainly including: Word consistency: The actual description text must be completely consistent with the template in terms of keywords and medical terminology, and synonym substitution or addition is not allowed.

[0092] Punctuation consistency: The use of punctuation marks must be strictly consistent with the template. For example, if the template uses "、" as a separator for parallel phrases, then "," or ";" cannot be used in the report.

[0093] Order consistency: The order in which the key points appear in the description must be consistent with the template.

[0094] Option validity: The conclusion text must be selected verbatim from the standardized option list provided by the template.

[0095] The system automatically adjudicates the analysis results output by the large language model based on preset judgment conditions, generating the first verification result for that check item. The judgment conditions are a set of logical judgments, for example: The conditions for passing the test may include: the description text is completely identical to the template in terms of words, punctuation, and order; the conclusion text completely matches one of the standard options; no text or numbers outside the template appear; and all special rule requirements in the template are met.

[0096] The conditions for failing the test correspond to any violation of the above rules, and the specific type of violation will be recorded, such as the description lacking "clear course of intrahepatic blood vessels", the conclusion using the non-standard expression "no abnormalities found", the description containing the phrase "no abnormalities found below the ribs" or using English commas.

[0097] By traversing all the examination items in the report and summarizing the verification results of each examination item, the system finally generates a complete first verification result report for the normal ultrasound report.

[0098] In the above implementation process, a standardized and refined hierarchical verification system was constructed for normal ultrasound reports. First, a dedicated normal report description template was matched according to the examination items. Then, a unified target prompt word was generated by combining the main prompt word framework. This made the verification process of normal reports a fixed standard and completely avoided the problems of inconsistent verification standards and arbitrary processes. At the same time, relying on multi-dimensional consistency verification rules including word, punctuation, order, and option legality, as well as clear judgment conditions for description text matching degree, standardized use of conclusion options, and text attribute standardization, the system achieved accurate verification of all elements of the examination items. It can meticulously identify various non-standard details in the descriptions and conclusions of normal reports, so that the verification judgment has a clear basis and ensures the consistency and objectivity of the quality inspection results of all normal ultrasound reports.

[0099] Based on the above embodiments, in the method of matching and verifying abnormal ultrasound reports, the lesion type corresponding to each abnormal ultrasound report can also be identified. Then, the abnormal ultrasound reports are classified according to the lesion type to obtain multiple report subsets corresponding to multiple lesion types. Then, the lesion report description template corresponding to each lesion type is called from the template library. According to the preset review rules, the structured report slice of each report in the report subset is verified using the lesion report description template to obtain the corresponding second verification result.

[0100] The system receives a collection of categorized abnormal ultrasound reports. Each report has been preprocessed into a structured report slice containing one or more abnormal examination items (such as liver, thyroid), with each slice explicitly containing descriptive and conclusion text for that site.

[0101] First, the system analyzes each abnormal report individually. For each abnormal examination item's structured slice, key lesion descriptive features are extracted using natural language processing technology to determine the specific lesion type. Lesion types are abnormal categories defined according to clinical medical standards, such as: cystic lesions: such as liver cysts, kidney cysts, and breast cysts; solid nodules: such as thyroid nodules (TI-RDRIS category 3) and breast nodules (BI-RADS category 4A); diffuse changes: such as fatty liver (moderate) and sonographic changes in chronic liver disease; other specific abnormalities: such as gallbladder polyps, kidney stones, and prostate calcification.

[0102] The identification process combines keyword matching, medical entity recognition models, and predefined lesion classification rules. For example, for a slice with the conclusion text "hypoechoic nodule in the right lobe of the thyroid gland, with clear borders, regular shape, and TI-RADS 3", the system will extract the features "hypoechoic", "nodule", and "TI-RADS 3" and classify it as the lesion type "thyroid nodule (TI-RADS 3)".

[0103] The system divides the entire set of abnormal reports into multiple report subsets based on the identified lesion types. Each subset contains all reports involving the same specific lesion type. A single report may belong to multiple subsets simultaneously due to containing multiple different abnormalities. For example, report A, which contains both liver cysts and kidney stones, will be assigned to the liver cyst report subset and the kidney stone report subset, respectively. This division ensures that subsequent specialized quality checks can be performed based on the unique characteristics of each lesion.

[0104] The system maintains a lesion report description template library, which stores the corresponding lesion report description template for each predefined lesion type. This template is a structured knowledge rule package, typically containing: Required descriptive elements list: Specifies the key information that must be included when describing the lesion. For example, a liver cyst template may require the description to clearly mention the location, number, size, shape, borders, internal echoes, and posterior acoustic enhancement.

[0105] Rules for logical relationships between elements: Define the medical logical constraints between elements. For example, if the number is described as multiple, the size should describe the largest one or the range; if the internal echo is cloudy, the conclusion should consider the possibility of a complex cyst.

[0106] Conclusion guidelines: These guidelines specify the core diagnostic opinion, classification (e.g., TI-RADS / BI-RADS classification), and recommendations that the conclusion should include. For example, the "TI-RADS 4 for thyroid nodules" template requires the conclusion to clearly state the classification and provide corresponding follow-up or biopsy recommendations based on the classification.

[0107] Numerical and unit compliance rules: Limit the units, formats, and reasonable ranges of key measurements (such as nodule size).

[0108] Based on the lesion type corresponding to each report subset, the appropriate lesion report description template is accurately retrieved from the template library. Each report subset is iterated through, and for each report within it, a structured slice of the examination items related to that lesion type is extracted. Then, according to preset review rules, automated verification is performed using the retrieved lesion template. The preset review rules are a set of configurable verification logic, mainly including: Element completeness review: Check whether the description text covers all the required elements required by the template.

[0109] Text consistency review: Verify whether the descriptions of each element in the description conform to the terminology standards specified in the template.

[0110] Logical consistency review: Use a rule engine to determine whether the medical logical relationships between descriptive elements and between description and conclusion conform to the template definition.

[0111] Numerical and format compliance review: Verify whether the units of measurement values ​​are correct and within a reasonable range.

[0112] The verification process is driven by a rule engine, comparing the report slice content with each template rule. For example, for a report categorized under the liver cyst subset, its liver slice description is "an anechoic area of ​​approximately 20x18mm in size, with clear borders and posterior acoustic enhancement," and the conclusion is "liver cyst." After the system calls the "liver cyst template," the following review will be performed: Completeness of elements: The examination found that the description included "location" (right lobe of liver), "size" (20x18mm), "border" (clear), "internal echo" (anechoic), and "posterior echo" (enhanced), and the elements were complete, so it passed.

[0113] Textual consistency: The term "anechoic area" is consistent with the typical description of a cyst, passing.

[0114] Logical consistency: The descriptive features (anechoic, clear borders, posterior enhancement) are highly consistent with the conclusion "liver cyst", thus passing the test.

[0115] Numerical compliance: The unit "20x18mm" is correct, pass.

[0116] If a report lacks a description of "posterior echo" or concludes with "liver mass" without specifying "cyst," the corresponding item will fail the verification.

[0117] Finally, the verification results for all lesion types (all abnormal examination items) in each report are summarized to generate a second verification result for that report. This result details the verification status (pass / fail) of each abnormal examination item and pinpoints the specific issues that failed, such as missing "boundary" descriptions, conclusions not including TI-RADS grading, or misuse of "cm" as the unit for lesion size.

[0118] In the above implementation process, a refined verification strategy guided by lesion type was adopted for abnormal ultrasound reports. First, the lesion type was accurately identified and classified into exclusive report subsets. Then, corresponding lesion report description templates were matched for each subset to carry out targeted verification. This ensured that the verification standards for abnormal reports were highly compatible with the professional description specifications of different lesions, solving the problem of the lack of specificity and poor adaptability of the unified verification of abnormal reports. At the same time, the verification was carried out based on preset review rules that focused on the description structure, logical relationship and numerical range. This can accurately identify key issues such as the mismatch between the description and the conclusion structure in the report, the contradiction in the diagnostic logic, and the lesion numerical description exceeding the standard range, which greatly improves the accuracy of abnormal report verification.

[0119] Based on the above embodiments, the quality inspection result of each ultrasound report to be inspected is determined according to the first and second verification results. The quality inspection result is the final status label generated by the system for each report, which typically includes quality inspection passed, quality inspection failed, and pending manual review.

[0120] For normal ultrasound reports, the quality inspection result can be determined based on the first verification result. If all the verification results of the inspection items in the report are passed, the quality inspection result is determined to be passed. If a small number of the inspection items in the report are not passed, the quality inspection result is determined to be not passed. If the number of inspection items in the report that are not passed reaches a set threshold, the quality inspection result is determined to be pending manual review.

[0121] For abnormal ultrasound reports, the quality inspection result can be determined based on the second verification result. If all verification results for all items in the report are passed, the quality inspection result is considered passed. If a small number of verification results for a few items are failed, the quality inspection result is considered failed. If a set threshold number of verification results for a number of items are failed, the quality inspection result is considered pending manual review. Alternatively, if at least one verification result for a verification item is failed, the quality inspection result can be either pending manual review, in which case the quality inspection result includes two possibilities: passed or pending manual review.

[0122] After obtaining the quality inspection results for each ultrasound report to be inspected, the ultrasound reports marked as awaiting manual review can be output to the doctor's terminal.

[0123] The quality control system can automatically filter out all reports labeled as requiring manual review, forming a set of reports awaiting review. These reports are those identified by the automated process as having potential problems or uncertainties that require physician intervention for judgment. Examples include reports with minor inconsistencies between description and conclusion that do not meet the criteria for direct rejection, reports using non-standard but potentially reasonable terminology, or complex cases where the system cannot make a clear pass / fail decision based on its rules.

[0124] In some implementations, the system can assemble a corresponding review task data package for each ultrasound report awaiting manual review. This data package does not simply contain the original report text, but rather a structured view integrating multi-dimensional information, designed to provide physicians with sufficient decision-making context. It typically includes: Full text of the original report: available for physicians to view in its entirety.

[0125] Structured report slices: Present the description and conclusions of specific examination items (such as liver, thyroid) where the problem occurred in a clear table or highlighted format.

[0126] Detailed quality inspection problem summary: clearly points out the specific problem identified by the system, the rules violated, and the reasons for uncertainty.

[0127] Confidence score for automated quality inspection: The system can provide a numerical value or level to indicate the degree of uncertainty in its judgment, assisting physicians in deciding the priority of review.

[0128] Subsequently, the system pushes the review tasks to the doctor's terminal through a preset output interface. The doctor's terminal typically refers to the professional workstation software used by physicians in their daily work, the doctor's workstation module of a hospital information system (HIS), or a dedicated quality inspection and review web platform. The core form of the push is to write the aforementioned review task data package into the doctor's terminal's backend database in a structured record format (such as JSON data), and generate a corresponding to-do list on the front-end interface.

[0129] In terms of front-end presentation, the system typically creates a functional module called "Report Quality Inspection Review" or similar within the doctor's terminal. This module's interface is often designed as a clear task list or multi-dimensional table view. Each row represents a task to be reviewed, and the columns include: examinee ID, examination date, examination site, problem type summary, quality inspection confidence level, and receipt time. The doctor can click on any task to bring up or jump to a details page to view the complete review task data package.

[0130] To ensure timely certification, the system can be integrated with a message notification mechanism. When a new report awaiting review is generated, the system will send a notification to the physician or physician team responsible for the review through integration with the hospital's internal communication system.

[0131] The results of the doctor's review (pass or return) on the terminal will be fed back to the quality inspection system in real time, forming a closed loop of business process, and may be used for subsequent template optimization and model training.

[0132] In some implementations, feedback from manual review of ultrasound reports awaiting manual verification can be collected, and then the predefined category report description templates can be adjusted based on this feedback. The overall implementation process can be as follows: Figure 2 As shown.

[0133] The system retrieves all manual review feedback within a set period (e.g., daily or weekly) in batches, forming a feedback dataset. Physicians' manual review feedback can include statements such as "verification correct," "verification error," and "human error." For example, when a physician processes an ultrasound report marked "pending review," a clear decision must be made. The interface provides standardized label options, such as "verification correct" (accepting the original report, the system determines it's a false alarm), "verification error" (confirming a medical logical contradiction or diagnostic error in the report's description and conclusion), and "human error" (confirming a standard error in terminology, format, units, etc., but without logical contradictions). After the physician selects the appropriate label and submits, the system stores the manual review result along with the corresponding report's unique identifier (ID), the problem item's location information (e.g., specific organ and field), and a timestamp in the database, forming a traceable feedback record.

[0134] The system is configured with a scheduled task, such as automatically triggering at the beginning of each month, to aggregate all manually reviewed results generated within the past month, forming a set of manually reviewed results. The system then iterates through this set, and based on the tags attached to each result, categorizes the corresponding original reports awaiting review into three preset sets: The "Check Correction" report collection contains all reports that doctors have marked as "check correct." These cases represent "false positives" in the system, meaning misjudgments caused by overly strict rules or incomplete rule coverage.

[0135] Check Error Report Collection: This collection contains all reports that doctors have identified as "check errors." These cases represent real-world clinical logical errors and serve as high-quality learning samples.

[0136] Human-caused writing errors collection: This section contains all reports identified by doctors as "human-caused writing errors." These cases reveal common but non-standard writing habits in clinical practice.

[0137] The system performs the following data cleaning process for each set: 1) Text extraction: For each report in the set, locate the specific inspection item (organ) that initially caused the quality inspection problem (leading it to enter the "pending review" state), extract its descriptive text and conclusion text, and merge them into a descriptive conclusion text unit.

[0138] 2) Organ-level text normalization: All extracted text units are grouped by organ and standardized to eliminate surface differences and preserve core semantic patterns. a. Number Normalization: Replace all numbers in the text that represent specific measurements, dimensions, or quantities with a uniform variable symbol (such as X). For example, "cyst approximately 5mm in size" and "cyst approximately 8mm in size" are both normalized to "cyst approximately Xmm in size". This allows the template to learn to describe structure rather than specific numerical values.

[0139] b. Synonym normalization: Based on a medical thesaurus, different but semantically identical terms are mapped to standard expressions. For example, "calcification spot" and "calcification foci" are uniformly mapped to the standard expression "calcification spot / calcification foci"; "anechoic area" and "cystic dark area" are mapped to "anechoic area / cystic dark area".

[0140] c. Location Normalization: Standardize the vocabulary representing location to a standard format. For example, unify "left leaf," "right side," and "left / right" into "left / right," so that the template does not depend on specific locations for pattern learning.

[0141] 3) Deduplication and Target Set Generation: Deduplication is performed on the normalized text within each report set to remove identical records, ultimately yielding the target conclusion text set. For example, after deduplication, the "Checked without error" set may yield dozens of different but standardized text patterns that are accepted by doctors and describe "liver cysts".

[0142] Subsequently, the system uses the cleaned report set to optimize the category report description templates in the module library, including the aforementioned abnormal report description template (lesion report description template) and normal report description template.

[0143] Learning from the set of error-free reports: This set reveals legitimate, clinically accepted variations of expression not covered by the current template. The system analyzes these variations and adds them as supplementary options or synonym rules to the template for the corresponding organ. For example, if the set contains multiple instances of the physician-approved phrase "clear intrahepatic vascular texture" (the original template requires "clear course"), the system adds "clear texture" as a legitimate synonym to the description rules for the "normal liver template," thereby reducing future false positives.

[0144] Learning from a collection of manually written error reports: This collection reveals frequently occurring error patterns. Based on these patterns, the system can strengthen validation rules or add typical error warnings to templates. For example, if the collection repeatedly shows "Unit misuse: writing 'cm' instead of 'mm'", the system can add stricter unit validation and error messages to the numerical rules of the corresponding template.

[0145] Learning from the error report set: This set provides real-world examples of logical inconsistencies. The system can analyze the erroneous correspondence between descriptions and conclusions in these cases, and strengthen or add logical consistency rules to the template to prevent similar errors from passing. For example, if multiple cases are found where descriptions of "microcalcifications" are found but conclusions of "benign" are incorrectly judged by doctors, the system can strengthen the rule in the "Thyroid Nodule Template": "If the description includes microcalcifications, the TI-RADS classification in the conclusion should not be lower than 4A."

[0146] In the above implementation process, after completing the automated quality inspection of ultrasound reports and determining the results of those that passed the inspection and those awaiting manual review, the reports awaiting manual review are accurately output to the physician's terminal. This achieves efficient connection and precise linkage between automated quality inspection and manual review, allowing physicians to focus directly on reports requiring manual review without having to sift through a massive number of reports to identify problematic ones. This significantly reduces physicians' ineffective workload and improves the relevance and efficiency of manual review. After pushing the reports awaiting manual review to the physician's terminal and collecting feedback from the manual review, the predefined category report description templates are dynamically adjusted based on the feedback. This constructs a closed-loop ultrasound report quality inspection system, ensuring that template optimization has a real and professional clinical review basis. This makes the report description templates of each category more in line with the actual writing standards of ultrasound reports and the professional diagnostic and treatment requirements of medical practice, greatly improving the adaptability and accuracy of the templates, and thus effectively improving the accuracy of subsequent automated verification.

[0147] In some implementations, an adaptive large language model feedback and parameter optimization module can be integrated into the output of the quality inspection system to construct an intelligent closed loop from "human review" to "model evolution." The implementation of this module begins with the automatic capture and structuring of feedback data: when a report is marked as "awaiting human review" by the system and undergoes final review by a physician, the system not only records the human judgment result (e.g., "checked without error," "checked without error"), but also automatically associates and archives the complete context of the report, including the original report text, the intermediate judgments previously made by the large language model (e.g., intent recognition results, conclusion classification), the physician's corrected standard text, and the attribution error type (e.g., classification error, semantic misunderstanding). These paired "model initial output - human standard answer" data are stored in real-time in a dedicated incremental feedback database.

[0148] Based on this database, the system automatically triggers an incremental fine-tuning process for the large language model on a weekly basis (or when a certain data volume threshold is reached). This process first extracts all recent misclassified samples from the database to construct an incremental training dataset focused on quality control tasks (such as intent recognition and lesion classification). Subsequently, using transfer learning techniques, lightweight supervised fine-tuning is performed on the pre-trained large model using this dataset, updating only some model layers or adapter parameters. This low-cost, high-efficiency approach specializes the model to the current reporting characteristics and review standards of medical institutions, thereby specifically reducing the misclassification rate for ambiguous or rare expressions.

[0149] To further improve system reliability and human-machine collaboration efficiency, the fine-tuned model is required to add a confidence score (range 0-1) to its output. This score is calculated based on the model's internal uncertainty regarding its own judgment. The system sets a threshold (e.g., 0.7). When the model's confidence score for a report falls below this threshold, regardless of its classification result, the report will be automatically prioritized and routed to the "awaiting human review" queue. This mechanism achieves proactive quality control based on uncertainty, preventing downstream errors that may be caused by low-confidence judgments and intelligently guiding human effort towards cases that most require review.

[0150] Finally, the system's embedded error pattern analysis unit periodically mines the feedback database to identify high-frequency or typical model misclassification patterns (e.g., frequently misclassifying descriptions like "unclear display due to gas interference" as "rejection"). These analytical conclusions not only guide the data sampling strategy for the next round of model fine-tuning but also update the system's knowledge template library. For example, if the analysis finds that the model repeatedly misclassifies a certain atypical description of "fatty liver," the system can automatically add a new recognition rule or example for that type of description in the anomaly report description template library, thereby enabling the rule engine and the AI ​​model to co-evolve in iteration.

[0151] Through the aforementioned closed-loop mechanism, the quality inspection system transforms from a static automated executor into an intelligent agent capable of continuously learning from practice and dynamically optimizing its core AI components and knowledge base, ultimately achieving a spiral increase in quality inspection accuracy and system adaptability.

[0152] Please refer to the above method embodiments. Figure 3 , Figure 3 This is a structural block diagram of an ultrasound report quality inspection device 200 provided in an embodiment of this application. The device 200 may be a module, program segment, or code on an electronic device. It should be understood that the device 200 corresponds to the above method embodiment and is capable of performing the various steps involved in the method embodiment. The specific functions of the device 200 can be found in the description above. To avoid repetition, detailed descriptions are appropriately omitted here.

[0153] Optionally, the device 200 includes: The report acquisition module 210 is used to acquire multiple ultrasound reports to be inspected. The slice generation module 220 is used to extract the descriptive text and conclusion text of each examination item in each ultrasound report to be inspected, and generate a structured report slice corresponding to each examination item. The classification module 230 is used to classify the multiple ultrasound reports to be inspected according to the conclusion text in the structured report slices, and obtain multiple types of ultrasound reports to be inspected. The verification module 240 is used to match and verify the corresponding structured report slice with the predefined category report description template for each type of ultrasound report to be inspected, and obtain the corresponding verification result. The quality inspection module 250 is used to determine the quality inspection result of each ultrasound report to be inspected based on the verification result.

[0154] Optionally, the multiple types of ultrasound reports to be inspected include abnormal ultrasound reports and normal ultrasound reports. The verification module 240 is used to match and verify the corresponding structured report slice with a predefined normal report description template for the normal ultrasound report to obtain a corresponding first verification result; and to match and verify the corresponding structured report slice with a predefined abnormal report description template for the abnormal ultrasound report to obtain a corresponding second verification result.

[0155] Optionally, the verification module 240 is used to obtain a normal report description template corresponding to each examination item for the normal ultrasound report; insert the normal report description template into a preset main prompt word frame to generate a corresponding target prompt word; and verify the target prompt word and the structured report slice of the corresponding examination item according to preset consistency verification rules and preset judgment conditions to obtain the corresponding first verification result. The preset consistency verification rules include at least one of the following: word consistency, punctuation consistency, sequence consistency, and option validity; the preset judgment conditions include: whether the description text in the structured report slice is consistent with the normal report description template, whether the conclusion text uses the standardized options provided in the normal report description template, and whether the use of each text attribute is standardized.

[0156] Optionally, the verification module 240 is used to identify the lesion type corresponding to each abnormal ultrasound report; classify the abnormal ultrasound reports according to the lesion type to obtain multiple report subsets corresponding to lesion types; call the lesion report description template corresponding to each lesion type from the template library; and verify the structured report slice of each report in the report subset according to the preset review rules using the lesion report description template to obtain the corresponding second verification result. The preset review rules include: whether the description structure is consistent, whether the logical relationship is consistent, and whether the numerical range is consistent.

[0157] Optionally, the classification module 230 is used to traverse the conclusion text in the structured report slice corresponding to each examination item, perform anomaly identification according to preset lesion identification rules, and obtain identification results. The identification results include lesion presence and lesion absence. The lesion identification rules include whether the conclusion in the conclusion text is an empty field, whether it contains predefined negative words, and / or whether there is no specific lesion description. If the identification result is lesion presence, the ultrasound report to be inspected is classified as an abnormal ultrasound report. If the identification result is lesion absence, the ultrasound report to be inspected is classified as a normal ultrasound report.

[0158] Optionally, the device 200 further includes: An intent recognition module is used to acquire multiple initial ultrasound reports to be inspected; and to perform intent recognition on each initial ultrasound report to be inspected using a large language model to obtain a set of ultrasound reports to be inspected with the intent type of refusal and a set of ultrasound reports to be inspected with the intent type of normal examination. The set of ultrasound reports to be inspected includes multiple ultrasound reports to be inspected.

[0159] Optionally, the quality inspection results include "inspection passed" and "awaiting manual review," and the device 200 further includes: The quality inspection output module is used to output ultrasound reports marked as awaiting manual review to the physician's terminal.

[0160] Optionally, the quality inspection module 250 is also used to collect feedback results from manual review of ultrasound reports to be manually reviewed; and to adjust the predefined category report description template based on the feedback results from manual review.

[0161] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0162] Please refer to Figure 4 , Figure 4This application provides a schematic diagram of an electronic device for performing an ultrasound report quality inspection method. The electronic device may include: at least one processor 310, such as a CPU; at least one communication interface 320; at least one memory 330; and at least one communication bus 340. The communication bus 340 is used to establish communication between these components. In this embodiment, the communication interface 320 is used for signaling or data communication with other node devices. The memory 330 may be a high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 330 may also be at least one storage device located remotely from the aforementioned processor. The memory 330 stores computer-readable instructions, which, when executed by the processor 310, cause the electronic device to perform the aforementioned method process.

[0163] Understandable. Figure 4 The structure shown is for illustrative purposes only; the electronic device may also include components that are more advanced than those shown. Figure 4 The more or fewer components shown, or having the same Figure 4 The different configurations shown. Figure 4 The components shown can be implemented using hardware, software, or a combination thereof.

[0164] This application provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it performs the method process executed by the electronic device in the above method embodiments.

[0165] This embodiment discloses a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can perform the methods provided in the above-described method embodiments, such as including: Obtain multiple ultrasound reports awaiting quality inspection; Extract the descriptive and conclusion texts of each examination item from each ultrasound report to be inspected, and generate a structured report slice corresponding to each examination item; The multiple ultrasound reports to be inspected are classified according to the conclusion text in the structured report slices to obtain multiple categories of ultrasound reports to be inspected. For each type of ultrasound report to be inspected, the corresponding structured report slice is matched and verified with the predefined category report description template to obtain the corresponding verification result; Based on the verification results, the quality inspection result for each ultrasound report to be inspected is determined.

[0166] In summary, this application provides a method, electronic device, and program product for ultrasound report quality inspection applicable to schools and hospitals. This method converts unstructured ultrasound reports into standardized structured report slices, automatically classifies reports based on conclusion content, and then calls targeted predefined category report description templates for matching and verification, ultimately automatically generating quality inspection results. This approach achieves a standardized and targeted automated quality inspection process, establishing unified quality inspection judgment standards and effectively avoiding the subjective bias, inconsistent standards, and oversights and misjudgments caused by fatigue inherent in traditional manual review. This significantly improves the accuracy and consistency of ultrasound report quality inspection results. Furthermore, this method supports batch processing of reports awaiting quality inspection, enabling efficient completion of quality inspection work on massive numbers of ultrasound reports, significantly improving inspection efficiency, reducing the workload of manual review, and adapting to the large-scale ultrasound report review needs in medical examination scenarios.

[0167] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0168] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0169] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0170] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0171] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for quality inspection of ultrasound reports applicable to school hospitals, characterized in that, The method includes: Obtain multiple ultrasound reports awaiting quality inspection; Extract the descriptive and conclusion texts of each examination item from each ultrasound report to be inspected, and generate a structured report slice corresponding to each examination item; The multiple ultrasound reports to be inspected are classified according to the conclusion text in the structured report slices to obtain multiple categories of ultrasound reports to be inspected. For each type of ultrasound report to be inspected, the corresponding structured report slice is matched and verified with the predefined category report description template to obtain the corresponding verification result; Based on the verification results, the quality inspection result for each ultrasound report to be inspected is determined.

2. The method according to claim 1, characterized in that, The various types of ultrasound reports to be inspected include abnormal ultrasound reports and normal ultrasound reports. For each type of ultrasound report to be inspected, the corresponding structured report slice is matched and verified with a predefined category report description template to obtain the corresponding verification result, including: For the normal ultrasound report, the corresponding structured report slice is matched and verified with the predefined normal report description template to obtain the corresponding first verification result; For the abnormal ultrasound report, the corresponding structured report slice is matched and verified with the predefined abnormal report description template to obtain the corresponding second verification result.

3. The method according to claim 2, characterized in that, For the normal ultrasound report, the corresponding structured report slice is matched and verified with a predefined normal report description template to obtain the corresponding first verification result, including: For the normal ultrasound report, obtain the normal report description template corresponding to each examination item; Insert the normal report description template into the preset main prompt word frame to generate the corresponding target prompt word; Based on preset consistency verification rules and preset judgment conditions, the structured report slices of the target prompt words and corresponding check items are verified to obtain the corresponding first verification result; The preset consistency verification rules include at least one of the following: word consistency, punctuation consistency, sequence consistency, and option validity; the preset judgment conditions include: whether the description text in the structured report slice is consistent with the normal report description template, whether the conclusion text uses the standardized options provided in the normal report description template, and whether the use of each text attribute is standardized.

4. The method according to claim 2, characterized in that, For the abnormal ultrasound report, the corresponding structured report slice is matched and verified with a predefined abnormal report description template to obtain the corresponding second verification result, including: Identify the lesion type corresponding to each abnormal ultrasound report; The abnormal ultrasound reports are classified according to the lesion type to obtain multiple report subsets corresponding to different lesion types. Retrieve the lesion report description template corresponding to each lesion type from the template library; According to the preset review rules, the structured report slices of each report in the report subset are verified using the lesion report description template to obtain the corresponding second verification result; The preset review rules include: whether the description structure is consistent, whether the logical relationship is consistent, and whether the numerical range is consistent.

5. The method according to claim 1, characterized in that, The process involves classifying the multiple ultrasound reports to be inspected based on the conclusion text in the structured report slices, resulting in multiple categories of ultrasound reports to be inspected, including: The conclusion text in the structured report slice corresponding to each examination item is traversed, and anomaly identification is performed according to the preset lesion identification rules to obtain the identification results. The identification results include lesion presence and lesion absence. The lesion identification rules include whether the conclusion in the conclusion text is an empty field, whether it contains predefined negative words, and / or whether there is no specific lesion description. If the identification result indicates the presence of lesions, the ultrasound report to be inspected will be classified as an abnormal ultrasound report. If the identification result is no lesion, the ultrasound report to be inspected will be classified as a normal ultrasound report.

6. The method according to claim 1, characterized in that, Before obtaining multiple ultrasound reports to be inspected, the process also includes: Obtain multiple initial ultrasound reports awaiting quality inspection; The intent of each initial ultrasound report to be inspected is identified using a large language model, resulting in a set of ultrasound reports to be inspected with the intent type of refusal and a set of ultrasound reports to be inspected with the intent type of normal inspection. The set of ultrasound reports to be inspected includes multiple ultrasound reports to be inspected.

7. The method according to claim 1, characterized in that, The quality inspection results include those that have passed inspection and those pending manual review. After determining the quality inspection result for each ultrasound report pending review based on the verification results, the process further includes: Ultrasound reports marked as awaiting manual review are output to the physician's terminal.

8. The method according to claim 7, characterized in that, After outputting the ultrasound report marked as requiring manual review to the physician's terminal, the process also includes: Collect feedback results from manual review of ultrasound reports awaiting manual verification; Based on the feedback from the manual review, the predefined category report description template was adjusted.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer-readable instructions that, when executed by the processor, perform the method as described in any one of claims 1-8.

10. A computer program product, characterized in that, It includes computer program instructions, which, when read and executed by a processor, perform the method as described in any one of claims 1-8.

Images