Multidimensional quality control methods, devices, equipment, and readable media for medical imaging reports

Abstract

This application relates to a multi-dimensional quality control method, apparatus, device, and readable medium for medical imaging reports. The method includes: acquiring the medical imaging report to be quality controlled; inputting the medical imaging report into a large-scale medical quality control model to parse the report into multiple logical domains; the large-scale model is trained using training samples containing specific medical errors and fine-tuned with all parameters; performing multi-dimensional semantic logic verification on the medical imaging report using the large-scale model based on internalized medical logic rules to obtain verification results; identifying logical conflict points in the medical imaging report based on the verification results; and outputting a structured quality control report based on the logical conflict points. This application solves the technical problem of high false negative rates in existing image report quality control by applying a large-scale medical quality control model trained with training samples containing specific medical errors and fine-tuned with all parameters.

Description

Technical Field

[0001] This application relates to the field of medical artificial intelligence technology, and in particular to a multi-dimensional quality control method, device, equipment and readable medium for medical image reports. Background Technology

[0002] With the rapid development of medical imaging technology and the acceleration of medical digitalization, the diagnostic and treatment tasks undertaken by medical imaging departments such as radiology, CT, and MRI are becoming increasingly heavy, and the number of imaging diagnostic reports generated daily is exploding. The accuracy of medical imaging reports is a key basis for clinical decision-making. Any minor logical omission or descriptive conflict (such as errors in recording left and right anatomical locations, or contradictions between diagnostic conclusions and descriptions of objective signs) may lead to missed or misdiagnosed diagnoses, thereby negatively impacting the patient's treatment plan.

[0003] Currently, quality control of medical imaging reports mainly relies on manual sampling and rule-based or regular expression-based automated quality control systems. In rule-based automated quality control, the system typically uses a predefined keyword library or complex matching logic to identify potential problems in the text. However, these technologies suffer from the following limitations: their ability to understand the deep semantics of medical language and their logical reasoning capabilities are weak. Due to the high complexity and implicit contextual logic of medical language, such systems cannot understand the deep semantics in imaging reports, struggle to correlate disparate information points, and fail to identify deep logical errors such as contradictions between organ resection history and imaging descriptions, inconsistencies in anatomical localization, and logical conflicts between findings and diagnostic conclusions. This results in a high rate of missed detections in imaging report quality control, failing to meet the accuracy requirements of clinical practice for imaging report quality control.

[0004] There is currently no effective solution to the problem of high false negative rates in the quality control of existing image reports. Summary of the Invention

[0005] This application provides a multidimensional quality control method, device, equipment, and readable medium for medical image reports to solve the technical problem of high missed detection rate in existing image report quality control.

[0006] According to one aspect of the embodiments of this application, this application provides a multi-dimensional quality control method for medical image reports, including: acquiring the medical image report to be quality controlled; inputting the medical image report into a large-scale medical quality control model to parse the medical image report into multiple logical domains using the large-scale medical quality control model, wherein the large-scale medical quality control model is a large model trained with training samples containing specific medical errors and fine-tuned with all parameters; using the large-scale medical quality control model to perform reasoning based on internalized medical logic rules to perform multi-dimensional semantic logic verification on the medical image report and obtain verification results; identifying logical conflict points in the medical image report based on the verification results; and outputting a structured quality control report based on the logical conflict points.

[0007] Optionally, before training the medical quality control model using training samples containing specific medical errors, the method further includes generating training samples containing specific medical errors through reverse data construction. Specifically, this includes: collecting correct reports and performing anonymization and structuring processing on the correct reports to obtain a correct report corpus. The structuring processing includes dividing the correct reports into multiple logical domains, including at least one of a metadata domain, a clinical information domain, an imaging findings domain, and a diagnostic conclusion domain. Using a reverse injection method, corresponding logical error features are automatically injected into the correct report corpus based on preset medical knowledge of various error types to generate text samples to be quality controlled containing specific medical errors. The text samples to be quality controlled after error injection are used as input instructions, and the corresponding error type, original error fragment, and correction suggestions are encapsulated as structured data as output targets to construct instruction data pairs. All instruction data pairs are used as training samples containing specific medical errors.

[0008] Optionally, the injected logical error features include at least one of the following: identifying descriptions of past surgical history or organ resection in the correct report and inserting organ morphology descriptions that contradict the resection status into the imaging findings domain to construct an organ resection inconsistency error; identifying anatomical location descriptions of the same lesion in different paragraphs of the correct report and causing a conflict between the location of the imaging findings domain and the diagnostic conclusion domain by altering locative words or anatomical terms to construct an anatomical location error; identifying the examination equipment type corresponding to each examination item in the correct report and inserting imaging terminology descriptions that do not match the imaging principle of the current examination equipment type into the imaging findings domain to construct an equipment type error.

[0009] Optionally, training the medical quality control large model through full parameter fine-tuning includes: selecting a lightweight pre-trained large model with parameter magnitudes within a preset range as the base model, and unfreezing all weight parameters of the base model; setting loss weight coefficients for directional words, negation words, and anatomical terms to construct a logically weighted loss function; training the base model with the goal of optimizing the logically weighted loss function, so that all weight parameters of the base model participate in gradient updates, so that the base model learns long-distance, cross-domain medical logical dependencies, until the model converges to obtain the medical quality control large model.

[0010] Optionally, using a large-scale medical quality control model to perform multi-dimensional semantic and logical verification of medical imaging reports includes using the large-scale medical quality control model to perform metadata consistency verification of medical imaging reports. Specifically, this includes: using the large-scale medical quality control model to parse gender metadata, age metadata, and examination equipment metadata in the medical imaging report, and extracting the physiological characteristics and technical terms involved in the medical imaging report; calculating the matching degree between gender metadata, age metadata, and examination equipment metadata and physiological characteristics and various technical terms respectively; based on the matching degree, identifying errors in the description of specific organs due to gender mismatch and errors in the use of imaging terms due to examination equipment mismatch.

[0011] Optionally, using a large-scale medical quality control model to perform multi-dimensional semantic logic verification on medical imaging reports includes using the large-scale medical quality control model to perform anatomical-fact consistency verification on medical imaging reports. Specifically, this includes: parsing the anatomical location terms, anatomical terms, and measurement values ​​of lesions in medical imaging reports, and identifying the corresponding anatomical structural units; verifying whether the spatial description logic of anatomical location terms and anatomical terms in the context is contradictory, and comparing the measurement values ​​with the preset physiological constant range of anatomical structural units, so as to lock in anatomical location contradiction errors or physiological value exceeding the limit errors.

[0012] Optionally, using a large-scale medical quality control model to perform multi-dimensional semantic logic verification on medical imaging reports includes using the large-scale medical quality control model to perform bidirectional implication verification of the medical imaging reports based on what is seen and what is concluded. Specifically, this includes: describing positive signs in the field of what is seen in the scanned images and searching whether the diagnostic conclusion field contains the corresponding diagnostic opinion to perform positive implication verification. If the diagnostic conclusion field does not contain the corresponding diagnostic opinion, it is identified as an error of missing diagnostic conclusion. For each diagnostic opinion in the diagnostic conclusion field, backtracking to the field of what is seen in the images to find whether there is any imaging evidence supporting the diagnostic opinion is performed to perform reverse implication verification. If there is no imaging evidence supporting the diagnostic opinion in the field of what is seen in the images, it is identified as an error of missing diagnostic sign.

[0013] Optionally, the structured quality control report output based on logical conflict points includes: encapsulating each logical conflict point into an independent JSON object, wherein the JSON object includes at least one of an error type field, an error source fragment field, an error target fragment field, and a correction suggestion field; and using an inference framework to merge and summarize multiple JSON objects to generate a structured JSON array containing all identified errors as the quality control report.

[0014] According to another aspect of the embodiments of this application, this application provides a multi-dimensional quality control device for medical image reports, including: an acquisition module for acquiring medical image reports to be quality controlled; a parsing module for inputting the medical image report into a large-scale medical quality control model to parse the medical image report into multiple logical domains using the large-scale medical quality control model, wherein the large-scale medical quality control model is a large model trained with training samples containing specific medical errors and fine-tuned with all parameters; a logic verification module for using the large-scale medical quality control model to perform reasoning based on internalized medical logic rules to perform multi-dimensional semantic logic verification on the medical image report and obtain verification results; an identification module for identifying logical conflict points in the medical image report based on the verification results; and an output module for outputting a structured quality control report based on the logical conflict points.

[0015] According to another aspect of the embodiments of this application, this application provides an electronic device, including a memory, a processor, a communication interface and a communication bus. The memory stores a computer program that can run on the processor. The memory and the processor communicate with each other through the communication bus and the communication interface. When the processor executes the computer program, it implements the steps of the above method.

[0016] According to another aspect of the embodiments of this application, this application also provides a computer-readable medium having processor-executable non-volatile program code that causes the processor to perform the above-described method.

[0017] Compared with related technologies, the technical solutions provided in this application have the following advantages: This application provides a multi-dimensional quality control method for medical image reports, including: acquiring the medical image report to be quality controlled; inputting the medical image report into a large-scale medical quality control model to parse the medical image report into multiple logical domains using the large-scale medical quality control model, wherein the large-scale medical quality control model is trained with training samples containing specific medical errors and fine-tuned with all parameters; using the large-scale medical quality control model to perform reasoning based on internalized medical logic rules to perform multi-dimensional semantic logic verification on the medical image report and obtain verification results; identifying logical conflict points in the medical image report based on the verification results; and outputting a structured quality control report based on the logical conflict points. This application applies a large-scale medical quality control model trained with training samples containing specific medical errors and fine-tuned with all parameters to image report quality control. Relying on the model's reasoning ability based on internalized medical logic rules, it first achieves accurate parsing of multiple logical domains in the report, then completes comprehensive verification of multi-dimensional semantic logic, accurately identifies logical conflict points in the report, and outputs a structured quality control report. This improves the automation level and accuracy of logical verification in image report quality control, effectively reduces the false negative rate, and solves the technical problem of high false negative rates in existing image report quality control. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the hardware environment for an optional multidimensional quality control method for medical image reports provided according to an embodiment of this application; Figure 2 This is a schematic diagram of an optional multidimensional quality control method for medical image reports provided according to an embodiment of this application; Figure 3 This is a block diagram of an optional multidimensional quality control device for medical image reports provided according to an embodiment of this application; Figure 4 This is a schematic diagram of an optional electronic device structure provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustration and has no specific meaning in itself. Therefore, "module" and "part" may be used interchangeably.

[0023] To address the problems mentioned in the background art, according to one aspect of the embodiments of this application, an embodiment of a multidimensional quality control method for medical image reports is provided.

[0024] Optionally, in the embodiments of this application, the above-described multidimensional quality control method for medical image reports can be applied to, for example... Figure 1 The hardware environment shown consists of terminal 101 and server 103. Figure 1 As shown, server 103 is connected to terminal 101 via a network and can be used to provide services to the terminal or clients installed on the terminal. Database 105 can be set up on the server or independently of the server to provide data storage services for server 103. The network mentioned above includes, but is not limited to, wide area network, metropolitan area network or local area network. Terminal 101 includes, but is not limited to, PC, mobile phone, tablet computer, etc.

[0025] A multidimensional quality control method for medical image reports in this application embodiment can be executed by server 103, or it can be jointly executed by server 103 and terminal 101, such as... Figure 2 As shown, the method may include the following steps: Step S202: Obtain the medical image report to be quality controlled; Step S204: Input the medical image report into the medical quality control big model, so as to use the medical quality control big model to parse the medical image report into multiple logical domains. The medical quality control big model is a big model that is trained with training samples containing specific medical errors and fine-tuned with all parameters. Step S206: Using the medical quality control big model, reasoning is performed based on internalized medical logic rules to perform multi-dimensional semantic logic verification on medical image reports and obtain verification results; Step S208: Based on the verification results, identify logical conflict points in the medical image report; Step S210: Output a structured quality control report based on logical conflict points.

[0026] In this embodiment, a medical imaging report refers to a text-based diagnostic report generated by equipment such as radiology, CT, MRI, and ultrasound, containing basic patient information, descriptions of imaging features, and diagnostic opinions. The medical quality control model refers to a lightweight, large-scale language model trained on specific medical error samples and fine-tuned with all parameters, possessing semantic understanding and logical verification capabilities for imaging reports. A logical domain refers to a structured area of ​​the imaging report divided according to information function, including a metadata domain, a clinical information domain, an imaging findings domain, and a diagnostic conclusion domain. The metadata domain includes information such as gender, age, and examination equipment; the clinical information domain includes information such as surgical history and chief complaint; the imaging findings domain includes descriptions of objective features; and the diagnostic conclusion domain includes diagnostic opinions. Multidimensional semantic logic verification refers to performing contextual logic verification on the report from multiple dimensions, such as metadata matching, anatomical common sense, and consistency between findings and conclusions. Internalized medical logic rules refer to anatomical norms, terminology norms, consistency verification rules, etc., solidified in the model parameters through model training. Logical conflict points refer to automatically identifiable errors in the report, such as inconsistencies, misidentification, missing evidence, and misuse of terminology. A structured quality control report is a machine-parseable quality control result that is output in a fixed format and includes error types, original text snippets, and correction instructions.

[0027] In step S202, the text of the image report that has been written and is awaiting review is collected through the hospital information system interface or by importing local files.

[0028] In step S204, the medical imaging report is input into the medical quality control model. The model performs semantic parsing on the report, automatically dividing it into multiple logical domains, providing a structural foundation for subsequent cross-domain logical verification. The medical quality control model is a lightweight large language model trained using training samples containing specific medical errors and fine-tuned with all parameters. The logical domains include metadata, clinical information, imaging findings, and diagnostic conclusions, used to distinguish different functional sections in the report.

[0029] In step S206, the medical quality control model, based on internalized medical logic rules after training, performs reasoning to conduct multi-dimensional semantic logic consistency verification on the report, obtaining verification results including error judgment, conflict location, and problem type. Internalized medical logic rules refer to anatomical specifications, equipment terminology specifications, medical history and imaging consistency specifications, and sight-to-conclusion correspondence specifications, all solidified in the model parameters through full parameter fine-tuning. Multi-dimensional semantic logic verification includes metadata consistency verification, anatomical and factual consistency verification, and sight-to-conclusion verification. The conclusion is verified by two-way implication, which is performed in parallel during a single inference.

[0030] In step S208, logical conflict points in the report are identified based on the verification results. Logical conflict points refer to quality control issues in the report text that can be identified by the model, such as inconsistencies, description errors, missing evidence, and misuse of terminology.

[0031] In step S210, a structured quality control report is output based on the identified logical conflict points. The structured quality control report is in a machine-parseable format and includes information such as error type, original text fragment, and correction instructions.

[0032] Through steps S202 to S210, this application applies a large-scale medical quality control model trained with training samples containing specific medical errors and fine-tuned with all parameters to the quality control of image reports. Relying on the model's reasoning ability for internalized medical logic rules, it first achieves accurate parsing of multiple logical domains in the report, and then completes comprehensive verification of multi-dimensional semantic logic. It accurately identifies logical conflict points in the report and outputs a structured quality control report, thereby improving the automation level of image report quality control and the accuracy of logical verification, effectively reducing the false negative rate, and solving the technical problem of high false negative rate in the quality control of existing image reports.

[0033] In an optional embodiment, before training the large medical quality control model using training samples containing specific medical errors, the method further includes generating training samples containing specific medical errors through reverse data construction, specifically including: Step 1: Collect correct reports and perform anonymization and structuring on the correct reports to obtain a correct report corpus. The structuring process includes dividing the correct reports into multiple logical domains, which include at least one of the following: metadata domain, clinical information domain, imaging findings domain, and diagnostic conclusion domain. Step 2: Using the reverse injection method, based on the medical knowledge of various error types preset in the correct report corpus, the corresponding logical error features are automatically injected to generate a text sample to be controlled containing specific medical errors. Step 3: Take the text sample to be controlled after the error is injected as the input instruction, and encapsulate the corresponding error type, original error text fragment and correction suggestion into structured data as the output target to construct the instruction data pair; Step 4: Use all instruction data pairs as training samples containing specific medical errors.

[0034] In this embodiment, reverse data construction refers to a dataset construction method that automatically injects logical errors to generate negative samples based on correct reports; correct reports refer to standard image reports that have been manually reviewed and confirmed to be error-free; desensitization refers to removing patient privacy information, such as removing patient names, ID numbers, and other personal privacy information to ensure data compliance; structuring refers to dividing the report into different logical domains; reverse injection refers to a method that automatically generates controllable errors according to type in correct text based on medical quality control rules; instruction data pairs refer to model training samples consisting of input instructions and standardized outputs. For example, the input content is: "You are a professional image report quality control physician. Please carefully review the following report and point out all the errors in it in JSON format. The following are the definitions and explanations of 14 error types. The full report text is as follows: [Insert the report text with injected errors here]", and the output is a list of errors in JSON format. For each injected error, generate a structured JSON object that must contain the following fields: "errorType" (error type, such as "inconsistent organ removal"), "source" (the original text fragment where the error occurred), "target" (the original text fragment where the error occurred), and "detail" (correction suggestion, such as "Medical history suggests gallbladder removal; imaging findings should no longer describe gallbladder morphology; please verify"). Combine all error objects into a single JSON array as the final output.

[0035] In this embodiment, the reverse data construction process involves collecting verified correct reports, performing privacy desensitization and structured segmentation on the reports to form a clean corpus of correct reports. Using a reverse injection method, based on preset medical knowledge corresponding to various quality control error types, error features consistent with clinical logic are automatically injected into the correct reports, generating a text sample to be quality controlled containing specific errors. The text with injected errors is used as input instructions, and the corresponding error type, original text fragment, and correction suggestions are encapsulated into a structured output target to form a complete instruction. Output data pairs. All data pairs are aggregated and used as training samples containing specific medical errors to train a large-scale medical quality control model.

[0036] This application addresses the problem of scarce labeled data in the medical quality control field by automatically generating large-scale, high-quality, and accurately labeled error samples through reverse data construction, thus providing reliable data support for model training.

[0037] In an optional embodiment, the injected logical error characteristics include at least one of the following: Identify descriptions of past surgical history or organ resection in correct reports and insert organ morphology descriptions that contradict the resection status into the imaging field to construct an organ resection inconsistency error; Identify the anatomical location descriptions of the same lesion in different paragraphs of the correct report, and create a conflict between the location of the imaging field and the diagnostic conclusion field by tampering with directional words or anatomical terms, so as to construct an anatomical location error. Identify the inspection equipment type corresponding to each inspection item in the correct report, and insert imaging terminology descriptions that do not match the imaging principle of the current inspection equipment type into the image seen field to construct an equipment type error.

[0038] In this application embodiment, the error of inconsistent organ resection refers to a contradiction where the surgical history mentions organ resection, but the imaging still describes the morphology of that organ; the error of anatomical localization refers to inconsistencies in the location, side, and anatomical structure of the same lesion in the findings and conclusions; and the error of equipment type refers to a mismatch between the examination equipment and the imaging terminology used.

[0039] In this embodiment, the implementation process of injecting logical error features is divided into three categories: constructing organ resection inconsistency errors, constructing anatomical location errors, and constructing device type errors. Specifically, constructing organ resection inconsistency errors involves identifying surgical history or organ resection descriptions in the report and inserting morphological descriptions that contradict the resection status into the imaging findings domain, creating cross-domain logical conflicts. For example, in a report containing "post-cholecystectomy," the description "thickened gallbladder wall, stones visible in the cavity" is forcibly inserted into the "imaging findings" paragraph, artificially creating a logical contradiction. Constructing anatomical location errors specifically refer to identifying the location descriptions of the same lesion in different paragraphs, altering locative terms or anatomical nouns, causing inconsistencies between the location information in the imaging findings domain and the diagnostic conclusion domain. For example, the description of "left lung" corresponds to the conclusion written as "right lung"; the description of "femur" corresponds to the conclusion written as "tibia"; the description of "bilateral pleural effusion" corresponds to the conclusion written as "right pleural effusion." Equipment type error refers to the insertion of technical terms that do not match the imaging principle into the image view field, resulting in incorrect terminology usage. For example, inserting the description "T1WI high signal" (MRI term) into a CT report. All of these errors are injected via automated scripts.

[0040] This application enables the model to learn and identify core error types by accurately generating high-frequency, high-risk clinical quality control error samples, thereby improving the model's ability to detect key errors.

[0041] In an optional embodiment, training a large medical quality control model through full parameter fine-tuning includes: Step 1: Select a lightweight pre-trained large model with parameters within a preset range as the base model, and unfreeze all weight parameters of the base model. Step 2: Set loss weight coefficients for locative words, negative words, and anatomical nouns to construct a logically weighted loss function; Step 3: Train the base model with the goal of optimizing the logical weighted loss function, so that all weight parameters of the base model participate in gradient updates, so that the base model learns long-distance, cross-domain medical logical dependencies, until the model converges to obtain a large medical quality control model.

[0042] In this application embodiment, the lightweight pre-trained large model refers to a Transformer architecture model with a moderate number of parameters that can be deployed locally, such as Qwen3-4B; full parameter fine-tuning refers to a training method that unfreezes all model weights and participates in gradient updates; the logistic weighted loss function refers to the optimization objective of assigning higher loss weights to key medical terms such as negation words, locative words, and anatomical terms; long-distance, cross-domain medical logical dependencies refer to the semantic associations across paragraphs between medical history and images, findings and conclusions, and metadata and main text.

[0043] In this embodiment of the application, the model training process is as follows: The first step is to select a lightweight pre-trained model with an appropriate number of parameters as the base model, and unfreeze all the weight parameters of the model so that it can participate in training and updates. For example, Qwen3-4B has a sufficient number of parameters to support complex medical logic, while having low requirements for hardware resources. It only requires a single consumer-grade graphics card (such as NVIDIA 3090 / 4090) to complete local deployment and inference, ensuring that the data does not leave the hospital.

[0044] The second step involves setting corresponding loss weight coefficients for error-prone keywords in medical quality control scenarios, including locative words, negative words, and anatomical terms, and constructing a logically weighted loss function: (1) (2) Formula (1) introduces a weight coefficient of Formula (2) for each word based on the standard cross-entropy loss. It is based on the current real word units The weight coefficients are defined by the category, where T is the sequence length. ) is the true word predicted by the model at time step t. The probability of.

[0045] This logistically weighted loss function imposes a stronger penalty on certain types of errors during training. For example, when the model predicts "right" as "left," the loss term is multiplied by [a certain factor]. When the model predicts "absence" as "existence", the loss term is multiplied by . This is equivalent to amplifying the gradients of these lexical terms during backpropagation, forcing the model to invest more capacity in learning these error-prone but crucial logical relationships.

[0046] Compared to the standard cross-entropy loss function, which treats all terms equally, the harm of miswriting "left kidney" as "right kidney" in medical reports is far greater than miswriting "kidney" as "kidney storage" (the latter may simply be a synonym or a misspelling). Weighted loss is precisely designed to address this imbalance in the cost of errors.

[0047] The third step, aiming to optimize the logistic weighted loss function, involves fully training the base model using training samples. This allows the model to learn long-distance, cross-logistic domain medical dependencies until convergence, resulting in a large-scale medical quality control model suitable for reasoning. Unlike efficient fine-tuning techniques like LoRA (Peripherally Optimized Elementary Theory) that only adjust partial parameters, the full-parameter fine-tuning method unfreezes the parameter weights of all layers (including self-attention modules and feedforward networks) and allows them to participate in gradient updates during training. This method more thoroughly reshapes the model's knowledge structure and reasoning ability, transforming it from a "general dialogue model" into a "domain expert model" proficient in image report quality control rules. It is particularly effective for learning long-distance, cross-domain logical dependencies (such as consistency between medical history and conclusions).

[0048] This application enhances the learning depth of the model for key medical logic by using full-parameter fine-tuning and weighted loss, which significantly improves the recognition accuracy of cross-domain, long-distance, and high-risk logical conflicts.

[0049] In this embodiment, after model training, it can be deployed using the VLLM framework. The underlying PagedAttention algorithm is highly efficient with low GPU memory usage and offers significant advantages in continuous batch processing and ultra-long context optimization. Furthermore, it supports OpenAI format requests on the client side. The finely tuned model is deployed on a hospital's internal server to build an inference service, enabling automated, batch quality control of doctor-submitted reports using a large-scale medical quality control model.

[0050] In this embodiment, the medical quality control model, upon receiving the full report, is internally capable of automatically identifying and distinguishing the report's metadata domain, image findings domain, and diagnostic conclusion domain, laying the foundation for subsequent cross-domain logical verification. During a single inference calculation, the model simultaneously activates and executes three core logical verification chains in parallel: the metadata consistency chain, the anatomy and factual consistency chain, and the "seeing-conclusion" bidirectional implicit logical closed-loop chain. These three core logical verification chains are described below.

[0051] In an optional embodiment, performing multi-dimensional semantic and logical verification of medical image reports using a large-scale medical quality control model includes performing metadata consistency verification of medical image reports using the large-scale medical quality control model, specifically including: Step 1: Using the medical quality control big data model, analyze the gender metadata, age metadata, and examination equipment metadata in the medical imaging report, and extract the physiological characteristics and technical terms involved in the medical imaging report. Step 2: Calculate the matching degree between gender metadata, age metadata, and examination equipment metadata and physiological characteristics and various technical terms, respectively; Step 3: Based on the matching degree, identify errors in the description of specific organs due to gender mismatch and errors in the use of imaging terminology due to mismatch of examination equipment.

[0052] In this embodiment of the application, metadata refers to basic information such as gender, age, examination equipment, and examination site in the report; physiological characteristics refer to anatomical structure descriptions related to gender and age; and technical terms refer to imaging descriptive terms corresponding to equipment such as CT, MR, and ultrasound.

[0053] In this embodiment, the metadata consistency verification process involves parsing metadata such as gender, age, and examination equipment in the report using a large-scale medical quality control model, while simultaneously extracting physiological feature descriptions and imaging technology terms from the main text. The matching degree between the metadata and the physiological features and technical terms is calculated to determine if any conflicts exist. Errors are identified based on the matching results, including mismatches between gender and specific organs, and mismatches between examination equipment and terminology. For example, a contradiction is detected between "gender: male" and "no abnormalities were found in the uterine adnexa" in the main text; or between "examination equipment: CT" and the term "high signal on T2WI" in the main text.

[0054] This application enables automated consistency verification between metadata and report text, quickly identifies basic information errors, and improves the comprehensiveness of quality control coverage.

[0055] In an optional embodiment, performing multi-dimensional semantic and logical verification of medical imaging reports using a large-scale medical quality control model includes performing anatomical-fact consistency verification on the medical imaging reports using the large-scale medical quality control model, specifically including: Step 1: Analyze the anatomical location terms, anatomical terms, and measurement values ​​of lesions in the medical imaging report, and identify the corresponding anatomical structural units; Step 2: Verify whether the spatial description logic of anatomical locative terms and anatomical nouns in the context is contradictory, and compare the measured values ​​with the preset physiological constant range of the anatomical structural unit to identify anatomical positioning contradictions or physiological value exceeding the limit.

[0056] In this embodiment of the application, anatomical directional terms refer to spatial location terms such as left and right, up and down, front and back; anatomical nouns refer to standard names of organs, tissues, structures, etc.; physiological constant range refers to the reasonable range of normal human body structure dimensions and values; anatomical location contradiction refers to inconsistencies in spatial descriptions; and physiological value exceeding the limit error refers to measured values ​​exceeding the normal range.

[0057] In this embodiment, the process of anatomical-fact consistency verification is as follows: the model parses the anatomical locative terms, anatomical nouns, and measurement values ​​in the report to identify the corresponding anatomical structural units. The model automatically verifies whether the spatial logic of the anatomical locative terms and anatomical nouns is consistent in the context, and determines whether there is a location contradiction. The model compares the measurement values ​​with the preset physiological constant range to determine whether there are numerical values ​​exceeding the limit or unreasonable situations, and finally identifies anatomical location errors or physiological value errors. For example, after identifying the description "right upper lobe nodule," the following "lesion located in the left subpleural region" constitutes a contextual anatomical location contradiction; or identifying "kidney size approximately 25cm x 15cm" as a numerical error exceeding the normal physiological range.

[0058] This application achieves automated verification from the perspectives of anatomical knowledge and physical numerical values, effectively identifying positioning errors and numerical anomalies, and improving the rigor and reliability of quality control.

[0059] In an optional embodiment, using a large-scale medical quality control model to perform multi-dimensional semantic logic verification on medical imaging reports includes using the large-scale medical quality control model to perform two-way implication verification of the seen-to-conclusion profile of medical imaging reports, specifically including: Step 1: Scan the description of positive signs in the image field and search the diagnostic conclusion field to see if it contains the corresponding diagnostic opinion in order to perform positive implication check. If the diagnostic conclusion field does not contain the corresponding diagnostic opinion, it is identified as an error of missing diagnostic conclusion. Step 2: For each diagnostic opinion in the diagnostic conclusion domain, backtrack to the imaging findings domain to check if there are any imaging features that support the diagnostic opinion, in order to perform reverse implication verification. If there are no imaging features in the imaging findings domain that support the diagnostic opinion, it is identified as a diagnostic error due to missing features.

[0060] In this embodiment, a positive sign refers to an abnormal appearance in the image; positive implication means that the abnormality seen should be reflected in the conclusion; negative implication means that the diagnosis in the conclusion should be based on the observation; omission of diagnosis conclusion means that there is a positive sign but the conclusion does not mention it; and missing diagnosis sign means that the conclusion has a diagnosis but there is no corresponding description in the observation.

[0061] In this embodiment, the implementation process of the bidirectional implication verification between the observed and the conclusion is as follows: Forward implication verification is performed, that is, for positive signs in the observed field of the scanned image, the diagnostic conclusion field is searched to see if a corresponding diagnosis is given; if not, it is determined that the diagnostic conclusion is incorrect due to omission. Reverse implication verification is performed, that is, for each positive diagnosis in the conclusion, the observed field of the image is searched to see if there are supporting signs; if no corresponding signs are found, it is determined that the diagnosis is incorrect due to missing signs. Simultaneously, direct semantic conflicts between the observed and the conclusion are detected to ensure consistency in expression. For example, the "observed" section may describe "no obvious space-occupying lesion in the liver," while the "conclusion" section may suggest "hepatic space-occupying lesion."

[0062] This application constructs a two-way logical closed loop between observations and conclusions, effectively identifying core errors such as missed diagnoses, insufficient evidence, and contradictions, and significantly reducing the rate of missed diagnostic errors.

[0063] In an optional embodiment, outputting a structured quality control report based on logical conflict points includes: Step 1: Encapsulate each logical conflict point into an independent JSON object. The JSON object includes at least one of the following: error type field, error source fragment field, error target fragment field, and correction suggestion field. Step 2: Use the inference framework to merge and summarize multiple JSON objects to generate a structured JSON array containing all identified errors as a quality control report.

[0064] In this embodiment, a JSON object refers to a single error structure containing fields such as error type, error source fragment, error target fragment, and correction suggestion; an inference framework refers to the runtime environment used for efficient model inference and result processing; and a structured JSON array refers to the final output format after multiple errors are integrated according to specifications.

[0065] In this embodiment, the process of outputting a structured quality control report is as follows: each identified logical conflict point is encapsulated as an independent JSON object; an inference framework is used to organize and merge all JSON objects to generate a structured JSON array containing all errors, which serves as the final quality control report output. The JSON format output can take the following forms: [ { "errorType": "Gender error", "source": "Gender: Female", "target": "The prostate gland is normal in shape, with no obvious abnormalities observed." "detail": "Basic information indicates female gender, but the description includes a male-specific structure, the prostate." }, { "errorType": "Normal and abnormal errors", "source": "Multiple small patchy low-density lesions were visible in both basal ganglia regions". "target": "The bilateral basal ganglia regions are of normal morphology, with no obvious abnormalities observed." "detail": "In the same area, bilateral basal ganglia, one area is described as having an abnormal lesion, while the other is described as normal, which is contradictory." } ] This application outputs standardized, machine-analyzable quality control results, facilitating integration with hospital systems and enhancing the automation, standardization, and traceability of the quality control process.

[0066] This application also provides a prompt word applicable to the multidimensional quality control method for medical image reports provided in this application, which may specifically be: Role positioning: You are a professional medical imaging report quality control expert, responsible for comprehensive structured identification and cross-domain consistency verification of imaging reports, and outputting standardized quality control results.

[0067] Input instructions: The image report consists of three parts: Basic information (metadata field): gender, age, examination method, examination items; What is seen (image visible domain): a structured / natural language description of the image representation; Conclusion (Diagnostic Conclusion Domain): Based on basic information and observed diagnosis / summary.

[0068] You need to perform internal structure recognition on the entire report in a single reasoning process, automatically distinguishing the three major domains mentioned above, laying the foundation for subsequent cross-domain logic verification and error identification.

[0069] Core capabilities and three-dimensional parallel logic chain verification: Internal text structure recognition: Automatically recognizes and segments “metadata domain (basic information) - image seen domain - diagnostic conclusion domain”, supporting subsequent domain-by-domain and cross-domain comparison, implication and contradiction detection.

[0070] Three-dimensional parallel logic chain (parallel execution within a single inference): 1. Metadata consistency chain: Verify the consistency between gender, age, examination method, etc. and the description in the text (e.g., "gender: male" should not include "uterus, ovary" etc.; "examination method: CT" should not include "T2 high signal" etc.).

[0071] 2. The chain of consistency between the dissection and the facts: Consistency between anatomical location and lateral orientation, reasonableness of measurement values ​​and units, and matching of equipment / physics terminology (e.g., "right upper lobe nodule" should not contradict "located in the left subpleural region"; "kidney size 25cm×15cm" exceeds the physiological range; misuse of "high signal" in CT scans).

[0072] 3. The "See-Conclusion" Two-Way Implication Closed-Loop Chain: Positive (Sights → Conclusion): Scan all positive signs (nodules, masses, effusions, etc.) in "Sights" and check whether "Conclusion" covers them one by one; if any are missed, it is judged as "Diagnosis conclusion omitted".

[0073] Reverse (Conclusion → Observations): For each positive diagnosis in the "Conclusion", trace back to the "Observations" to find supporting signs; if there is no corresponding description, it is judged as "omitted sign error".

[0074] Direct contradiction detection: such as "What was found: No space-occupying lesion was found in the liver" and "Conclusion: Hepatic space-occupying lesion" conflicting with each other.

[0075] Quality control items and judgment criteria: ...... ...... ...... ...... (Define the corresponding error type and judgment criteria) Execution process: 1. Structured parsing: Segment and label the three fields: "Basic Information (gender, age, examination method, examination items)", "Sights", and "Conclusion".

[0076] 2. Metadata extraction: Parse key metadata such as gender, age, examination method, and examination items.

[0077] 3. Parallel verification: Initiate three core logical chains (metadata consistency chain, dissection and fact consistency chain, and see-conclusion bidirectional implication closed-loop chain). At the same time, check all error types against the "Quality Control Items and Judgment Criteria".

[0078] 4. Result merging: Generate an independent error object for each problem found to avoid merging errors of different categories.

[0079] 5. Output standardized JSON.

[0080] Output requirements: Output format: Must be a standard JSON array.

[0081] For each error, output object fields: errorType: Error category name (e.g., "anatomical location error", "typo error", "normal and abnormal error", etc.); Source: Error text 1, strictly quoting the original report, maximum 20 characters; if the limit is exceeded, use "..." to truncate in the middle, retaining the first and last characters; target: Error original text 2 or null (must strictly come from the original report, or null if there is no corresponding text); also follows the 20-character truncation rule and "..."; Detail: Briefly explain the reason for the error (concise and clear).

[0082] If multiple errors are found, output multiple objects; if no errors are found, output an empty array.

[0083] The output must be enclosed in triple quotes and include a type label: If there are no errors, output: json [] ``` If there is an error, output: json [ { ...}, { ...} ] ``` Special attention The source / target must strictly use the original report text, and no characters may be modified; if the length exceeds 20 characters, it should be centered and truncated to "......".

[0084] If there is no corresponding second paragraph of the original text, set target to null.

[0085] Negative / normal descriptions in the findings do not need to be repeated in the conclusion; only when the corresponding conclusion for positive signs is missing is it judged as "omitted diagnosis conclusion".

[0086] The "error of missing signs" only applies when the corresponding site for a positive diagnosis is completely absent from the findings; if a neutral / negative / vague description exists, this error is not considered.

[0087] The determination of errors beyond the scope of the examination site should take into account the routine clinical imaging and the reasonable range of accompanying findings (such as chest CT scans which may include upper abdominal organs).

[0088] The omission of a special MR sequence is only determined when the examination item explicitly includes the sequence.

[0089] Numerical units and equipment terminology must be consistent with the examination method: CT (density shadow, HU), MR (signal, T1 / T2 / DWI / ADC, etc.), US (echo, blood flow, cm / s), DR / X-ray (luminous / dense shadow).

[0090] Age and gender information must be consistent with the metadata; resected organs must not include routine descriptions of normal or diseased organs, nor follow-up recommendations.

[0091] When encountering semantic ambiguity that does not constitute a clear contradiction, remain cautious and avoid making overly extended judgments.

[0092] This application provides a multi-dimensional quality control method for medical image reports, including: acquiring the medical image report to be quality controlled; inputting the medical image report into a large-scale medical quality control model to parse the medical image report into multiple logical domains using the large-scale medical quality control model, wherein the large-scale medical quality control model is trained with training samples containing specific medical errors and fine-tuned with all parameters; using the large-scale medical quality control model to perform reasoning based on internalized medical logic rules to perform multi-dimensional semantic logic verification on the medical image report and obtain verification results; identifying logical conflict points in the medical image report based on the verification results; and outputting a structured quality control report based on the logical conflict points. This application applies a large-scale medical quality control model trained with training samples containing specific medical errors and fine-tuned with all parameters to image report quality control. Relying on the model's reasoning ability based on internalized medical logic rules, it first achieves accurate parsing of multiple logical domains in the report, then completes comprehensive verification of multi-dimensional semantic logic, accurately identifies logical conflict points in the report, and outputs a structured quality control report. This improves the automation level and accuracy of logical verification in image report quality control, effectively reduces the false negative rate, and solves the technical problem of high false negative rates in existing image report quality control.

[0093] According to another aspect of the embodiments of this application, such as Figure 3 As shown, a multi-dimensional quality control device for medical image reports is provided, comprising: The acquisition module 301 is used to acquire medical image reports to be quality controlled; The parsing module 303 is used to input the medical image report into the medical quality control big model, so as to use the medical quality control big model to parse the medical image report into multiple logical domains. The medical quality control big model is a big model that is trained with training samples containing specific medical errors and fine-tuned with all parameters. The logic verification module 305 is used to perform reasoning based on internalized medical logic rules using the medical quality control big model to perform multi-dimensional semantic logic verification on medical image reports and obtain verification results. The identification module 307 is used to identify logical conflict points in the medical image report based on the verification results. Output module 309 is used to output structured quality control reports based on logical conflict points.

[0094] It should be noted that the acquisition module 301 in this embodiment can be used to execute step S202 in this application embodiment, the parsing module 303 in this embodiment can be used to execute step S204 in this application embodiment, the logic verification module 305 in this embodiment can be used to execute step S206 in this application embodiment, the identification module 307 in this embodiment can be used to execute step S208 in this application embodiment, and the output module 309 in this embodiment can be used to execute step S210 in this application embodiment.

[0095] It should be noted that the examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should also be noted that the above modules, as part of the device, can be implemented in software or hardware.

[0096] Optionally, the multi-dimensional quality control device for medical image reports also includes a reverse data construction module, specifically used for: collecting correct reports, and performing desensitization and structuring processing on the correct reports to obtain a correct report corpus. The structuring processing includes dividing the correct reports into multiple logical domains, including at least one of a metadata domain, a clinical information domain, an image findings domain, and a diagnostic conclusion domain; using a reverse injection method, automatically injecting corresponding logical error features into the correct report corpus based on preset medical knowledge of various error types to generate a text sample to be quality controlled containing specific medical errors; using the injected error-laden text sample as input instructions, encapsulating the corresponding error type, original error fragment, and correction suggestions into structured data as output targets to construct instruction data pairs; and using all instruction data pairs as training samples containing specific medical errors.

[0097] Optionally, the reverse data construction module is also used to: identify descriptions of past surgical history or organ resection in the correct report, and insert organ morphology descriptions that contradict the resection status into the imaging findings field to construct an organ resection inconsistency error; identify anatomical location descriptions of the same lesion in different paragraphs of the correct report, and create a conflict between the location of the imaging findings field and the diagnostic conclusion field by tampering with directional words or anatomical terms to construct an anatomical location error; identify the examination equipment type corresponding to each examination item in the correct report, and insert imaging terminology descriptions that do not match the imaging principle of the current examination equipment type into the imaging findings field to construct an equipment type error.

[0098] Optionally, the multidimensional quality control device for medical image reports also includes a model training module, specifically used for: selecting a lightweight pre-trained large model with parameter magnitudes within a preset range as a base model, and unfreezing all weight parameters of the base model; setting loss weight coefficients for directional words, negative words, and anatomical terms to construct a logically weighted loss function; training the base model with the goal of optimizing the logically weighted loss function, so that all weight parameters of the base model participate in gradient updates, so that the base model learns long-distance, cross-domain medical logical dependencies, until the model converges to obtain a large medical quality control model.

[0099] Optionally, the logic verification module is used to perform metadata consistency verification on medical imaging reports using a large-scale medical quality control model. Specifically, this includes: parsing gender metadata, age metadata, and examination equipment metadata in the medical imaging report using the large-scale medical quality control model, and extracting the physiological characteristics and technical terms involved in the medical imaging report; calculating the matching degree between gender metadata, age metadata, and examination equipment metadata and physiological characteristics and various technical terms, respectively; and based on the matching degree, identifying errors in the description of specific organs due to gender mismatch and errors in the use of imaging terms due to examination equipment mismatch.

[0100] Optionally, the logic verification module is also used to perform anatomical-fact consistency verification on medical imaging reports using a large medical quality control model. Specifically, this includes: parsing the anatomical location terms, anatomical terms, and measurement values ​​of lesions in the medical imaging report, and identifying the corresponding anatomical structural units; verifying whether the spatial description logic of the anatomical location terms and anatomical terms in the context is contradictory, and comparing the measurement values ​​with the preset physiological constant range of the anatomical structural units to identify anatomical location contradictions or physiological value exceedance errors.

[0101] Optionally, the logic verification module is also used to perform bidirectional implication verification of medical image reports using a large medical quality control model. Specifically, this includes: describing positive signs in the scanned image's seen field and searching whether the diagnostic conclusion field contains the corresponding diagnostic opinion to perform a forward implication verification. If the diagnostic conclusion field does not contain the corresponding diagnostic opinion, it is identified as an error of missing diagnostic conclusion. For each diagnostic opinion in the diagnostic conclusion field, the module backtracks to the image seen field to search for whether there are imaging signs that support the diagnostic opinion to perform a reverse implication verification. If there are no imaging signs that support the diagnostic opinion in the image seen field, it is identified as an error of missing diagnostic signs.

[0102] Optionally, the output module is specifically used to: encapsulate each logical conflict point into an independent JSON object, wherein the JSON object includes at least one of an error type field, an error source fragment field, an error target fragment field, and a correction suggestion field; and use an inference framework to merge and summarize multiple JSON objects to generate a structured JSON array containing all identified errors as a quality control report.

[0103] According to another aspect of the embodiments of this application, this application provides an electronic device, such as... Figure 4 As shown, the system includes a memory 401, a processor 403, a communication interface 405, and a communication bus 407. The memory 401 stores a computer program that can run on the processor 403. The memory 401 and the processor 403 communicate through the communication interface 405 and the communication bus 407. When the processor 403 executes the computer program, it implements the steps of the above method.

[0104] The memory and processor in the aforementioned electronic devices communicate with each other via a communication bus and a communication interface. The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0105] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0106] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0107] According to another aspect of the embodiments of this application, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of any of the above embodiments.

[0108] Optionally, in embodiments of this application, the computer-readable medium is configured to store program code for the processor to perform the following steps: Obtain medical imaging reports that require quality control; The medical imaging report is input into the medical quality control model, which is used to parse the medical imaging report into multiple logical domains. The medical quality control model is a large model that is trained with training samples containing specific medical errors and fine-tuned with all parameters. By using a large medical quality control model based on internalized medical logic rules to perform multi-dimensional semantic logic verification on medical image reports, the verification results are obtained. Based on the verification results, identify logical conflict points in the medical imaging report; Output structured quality control reports based on logical conflict points.

[0109] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments, and will not be repeated here.

[0110] In specific implementation, the embodiments of this application can be referred to the above embodiments and have corresponding technical effects.

[0111] It is understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described herein, or combinations thereof.

[0112] For software implementation, the techniques described herein can be implemented by units that perform the functions described herein. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or external to the processor.

[0113] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0114] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0115] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0116] 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.

[0117] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0118] If the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0119] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A multidimensional quality control method for medical image reports, characterized in that, include: Obtain medical imaging reports that are subject to quality control; The medical image report is input into the medical quality control model, which is used to parse the medical image report into multiple logical domains. The medical quality control model is a large model that is trained with training samples containing specific medical errors and fine-tuned with all parameters. The medical quality control big model is used to perform reasoning based on internalized medical logic rules to perform multi-dimensional semantic logic verification on the medical image report and obtain the verification results. Based on the verification results, identify the logical conflict points in the medical image report; A structured quality control report is output based on the logical conflict points.

2. The method according to claim 1, characterized in that, Before training the medical quality control model using the training samples containing specific medical errors, the method further includes generating the training samples containing specific medical errors through reverse data construction, specifically including: Collect correct reports, and perform anonymization and structuring processing on the correct reports to obtain a correct report corpus. The structuring processing includes dividing the correct reports into multiple logical domains, and the logical domains include at least one of the following: metadata domain, clinical information domain, imaging findings domain, and diagnostic conclusion domain. Using the reverse injection method, based on preset medical knowledge of various error types, corresponding logical error features are automatically injected into the correct report corpus to generate a text sample to be quality controlled containing specific medical errors; The text sample to be controlled after the injection of errors is used as the input instruction, and the corresponding error type, original error text fragment and correction suggestion are encapsulated as structured data as the output target to construct the instruction data pair; All the instruction data pairs are used as training samples containing specific medical errors.

3. The method according to claim 2, characterized in that, The logical error characteristics corresponding to the injection include at least one of the following: Identify descriptions of past surgical history or organ removal in the correct report and insert organ morphology descriptions that contradict the removal status into the image view area to construct an organ removal inconsistency error; Identify the anatomical location descriptions of the same lesion in different paragraphs of the correct report, and construct anatomical location errors by tampering with directional words or anatomical terms to create a conflict between the location of the image-seen domain and the diagnostic conclusion domain. Identify the inspection equipment type corresponding to each inspection item in the correct report, and insert an imaging term description that does not match the imaging principle of the current inspection equipment type into the image visible field to construct an equipment type error.

4. The method according to claim 1, characterized in that, The medical quality control model was trained through full parameter fine-tuning, including: Select a lightweight pre-trained large model with parameter magnitudes within a preset range as the base model, and unfreeze all weight parameters of the base model; Set loss weight coefficients for locative words, negative words, and anatomical nouns to construct a logically weighted loss function; The base model is trained with the goal of optimizing the logical weighted loss function, so that all weight parameters of the base model participate in gradient updates, so that the base model learns long-distance, cross-domain medical logical dependencies until the model converges to obtain the large medical quality control model.

5. The method according to claim 1, characterized in that, The multi-dimensional semantic logic verification of the medical image report using the medical quality control model includes metadata consistency verification of the medical image report using the medical quality control model, specifically including: The medical quality control model is used to analyze the gender metadata, age metadata, and examination equipment metadata in the medical imaging report, and to extract the physiological characteristics and technical terms involved in the medical imaging report. Calculate the matching degree between the gender metadata, the age metadata, and the examination device metadata and the physiological characteristics and various technical terms, respectively; Based on the matching degree, errors in the description of specific organs due to gender mismatch and errors in the use of imaging terminology due to mismatch of examination equipment are identified.

6. The method according to claim 1, characterized in that, The multi-dimensional semantic logic verification of the medical imaging report using the medical quality control model includes performing anatomy-fact consistency verification on the medical imaging report using the medical quality control model, specifically including: The anatomical location terms, anatomical terms, and measurement values ​​of lesions in the medical imaging report are analyzed, and the corresponding anatomical structural units are identified. Verify whether the spatial description logic of the anatomical locative terms and anatomical nouns in the context is contradictory, and compare the measured values ​​with the preset physiological constant range of the anatomical structural unit to identify anatomical positioning contradictions or physiological value exceeding limits.

7. The method according to claim 1, characterized in that, The multi-dimensional semantic logic verification of the medical image report using the medical quality control model includes performing a two-way implication verification of the medical image report based on the seen-conclusion approach using the medical quality control model. Specifically, this includes: The positive signs in the scanned image field are described, and the diagnostic conclusion field is searched to see if the corresponding diagnostic opinion is included, in order to perform positive implication verification. If the diagnostic conclusion field does not include the corresponding diagnostic opinion, it is identified as an error of missing diagnostic conclusion. For each diagnostic opinion in the diagnostic conclusion domain, backtracking to the image observation domain to check whether there are radiological signs supporting the diagnostic opinion is performed to perform reverse implication verification. If there are no radiological signs supporting the diagnostic opinion in the image observation domain, it is identified as a diagnostic missing sign error.

8. The method according to any one of claims 1 to 7, characterized in that, The structured quality control report output based on the logical conflict points includes: Each of the aforementioned logical conflict points is encapsulated into an independent JSON object, wherein the JSON object includes at least one of an error type field, an error source fragment field, an error target fragment field, and a correction suggestion field; The multiple JSON objects are merged and summarized using an inference framework to generate a structured JSON array containing all identified errors as the quality control report.

9. A multidimensional quality control device for medical image reports, characterized in that, include: The acquisition module is used to acquire medical image reports that need to be quality controlled. The parsing module is used to input the medical image report into the medical quality control model, so as to use the medical quality control model to parse the medical image report into multiple logical domains. The medical quality control model is a large model that is trained with training samples containing specific medical errors and fine-tuned with all parameters. The logic verification module is used to perform reasoning based on internalized medical logic rules using the medical quality control big model to perform multi-dimensional semantic logic verification on the medical image report and obtain the verification result. The identification module is used to identify logical conflict points in the medical image report based on the verification results. The output module is used to output a structured quality control report based on the logical conflict points.

10. An electronic device comprising a memory, a processor, a communication interface, and a communication bus, wherein the memory stores a computer program executable on the processor, and the memory and the processor communicate via the communication bus and the communication interface, characterized in that... When the processor executes the computer program, it implements the multidimensional quality control method for medical image reports as described in any one of claims 1 to 8.

11. A computer-readable medium having processor-executable non-volatile program code, characterized in that, The program code causes the processor to execute the multidimensional quality control method for medical image reports as described in any one of claims 1 to 8.

Images