Image report process AI quality control system based on data fusion and dynamic rule engine

Abstract

The invention discloses an image report process AI quality control system based on data fusion and a dynamic rule engine, and the system comprises a data layer module which is used for integrating multi-source data from an image archiving and communication system, a radiation information system and a hospital information system, and carrying out the preprocessing of the data, and forming a high-quality training data pool; furthermore, the image archiving and communication system obtains historical images in a DICOM format and corresponding report data; the radiation information system collects examination application information and doctor diagnosis information. According to the image report process AI quality control system based on data fusion and the dynamic rule engine, full-dimension and high-efficiency accurate quality control is realized: the system can perform comprehensive automatic screening of grammar, text, logic consistency and image-text relevance on an image report by integrating natural language processing, computer vision and a cross-modal matching technology; error correction response time is shortened to a second level from several minutes of each case of traditional manual auditing, and quality control efficiency is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of medical imaging information technology, specifically to an AI-based quality control system for image reporting processes based on data fusion and a dynamic rule engine. Background Technology

[0002] In the context of high-quality hospital development, imaging reports serve as a crucial basis for clinical diagnosis and treatment decisions, and their quality directly impacts patient safety, treatment effectiveness, and hospital management efficiency. Traditional imaging report quality control heavily relies on manual review by senior physicians, which has several inherent drawbacks: First, it is inefficient. With thousands of reports processed daily in tertiary hospitals, manual review is time-consuming and labor-intensive, resulting in long review cycles and delayed feedback. Second, standards are inconsistent. Quality control is affected by the reviewing physician's personal experience, subjective judgment, and workload, making it difficult to guarantee consistency and objectivity in evaluation. Third, accuracy is poor. Manual review is prone to overlooking errors such as spelling mistakes, non-standard terminology, logical contradictions, and inconsistencies between imaging features and text descriptions due to fatigue or negligence; research data shows that the report failure rate can reach 5%-13%. Finally, management methods are outdated. Traditional models are mostly passive, sampling-based post-event checks, lacking full-process, standardized real-time monitoring and closed-loop management capabilities, making accurate data tracing, error analysis, and performance evaluation difficult, hindering continuous improvement in departmental management.

[0003] To address these challenges, existing research has explored the application of artificial intelligence (AI) technology to image report quality control. Current technologies largely focus on using specific natural language processing (NLP) or computer vision models to detect single types of errors in reports. However, these methods generally suffer from the following limitations: first, they are functionally limited, lacking the comprehensive capability for full-scale, multi-dimensional quality control of reports; second, they lack deep integration with hospital information systems, making it difficult to build a quality control system covering the entire process; and third, their model generalization ability is limited, with insufficient capacity to handle complex medical contexts, rare cases, and multimodal data correlations. Summary of the Invention

[0004] The purpose of this invention is to provide an AI quality control system for image reporting workflow based on data fusion and dynamic rule engine, in order to solve the problems mentioned in the background art, such as lack of comprehensive and multi-dimensional quality control of reports, lack of deep integration with hospital information systems, difficulty in building a quality control system covering the entire process, limited model generalization ability, and insufficient ability to handle complex medical contexts, rare cases, and multimodal data associations.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI quality control system for image reporting workflow based on data fusion and dynamic rule engine, comprising: The data layer module is used to integrate multi-source data from image archiving and communication systems, radiological information systems, and hospital information systems, and to preprocess the data to form a high-quality training data pool. The technology layer module, connected to the data layer module, includes: The natural language processing submodule uses a pre-trained language model that has been deeply trained on medical text and image reports to identify text errors, non-standard terminology, and logical contradictions in the reports. The image-text verification submodule extracts image features based on a convolutional neural network model and performs cross-modal consistency matching with the report description to identify inconsistencies between images and text and missed diagnoses. The structured template library submodule contains standardized report templates for various common diseases; The dynamic rule engine submodule generates and continuously iterates quality control rules based on clinical guidelines and historical error data. The application layer module, connected to the technology layer module, includes: The AI ​​quality control submodule is used to receive the output of the technical layer module, automatically mark the error type and risk level of the report, and generate a quality control report; The manual review submodule provides a human-computer interaction interface for reviewing physicians to review and correct the AI-marked results; The statistical analysis submodule is used to generate quality control reports at the department and individual levels; The model iteration and upgrade submodule is used to continuously optimize the model in the technical layer module by using the correction results of manual review as new training data.

[0006] Preferably, the image archiving and communication system acquires historical images in DICOM format and corresponding report data; The radiology information system collects examination request information and physician diagnosis information; The hospital information system integrates basic patient information, laboratory test results, and pathological diagnosis reports. The data preprocessing unit cleans, standardizes, and transforms the multi-source data to a structured form, thereby constructing a multi-source training data pool for model training.

[0007] Preferably, the natural language processing submodule specifically adopts the BERT or RoBERTa model for: Detect and correct spelling errors in the report; Based on the pre-established standardized medical terminology system, verify the standardization of terminology usage; Based on semantic parsing, logical contradictions in the diagnostic description are identified.

[0008] Preferably, the image-text verification submodule specifically adopts the TResNet model for: Automatically extract the location, size, and morphological features of lesions from medical images; The extracted image features are associated and matched with descriptive statements in the report text; When there is a significant inconsistency between the image features and the text description, or when the text omits key image features, an inconsistency or missed diagnosis warning is generated.

[0009] Preferably, the dynamic rule engine submodule is used for: Based on authoritative clinical practice guidelines and the aforementioned historical error data, a quality control rule base is initialized and generated. During system operation, quality control rules are automatically generated or revised based on newly discovered error patterns and updated clinical guidelines; feedback from clinicians from the manual review submodule is received to optimize rule weights and judgment thresholds, thereby achieving adaptive iterative upgrades of the rule base.

[0010] Preferably, the model iteration and upgrade submodule in the application layer module specifically uses transfer learning technology to use the correction reports confirmed by doctors in the manual review submodule as new training samples to optimize the parameters of the natural language processing model and image-text verification model in the technology layer module, so as to achieve continuous improvement in model performance.

[0011] Preferably, the system combines the automatic marking of the AI ​​quality control submodule with the expert review of the manual review submodule to perform graded quality control on medical imaging reports. High-risk errors are marked first and submitted for review. The overall report quality control process takes more than 50% less time than traditional pure manual review.

[0012] Preferably, a medical image report quality control method is applied to the system as described in any one of claims 1 to 7, and the method is implemented in the following manner: The data layer module integrates multi-source medical data and constructs a training data pool. The natural language processing submodule and image-text verification submodule of the technical layer module are used to perform automated error detection on the quality control image report; The AI ​​quality control submodule of the application layer module generates a quality control report with error markers and risk levels based on the test results. Through the manual review submodule of the application layer module, the reviewing physician reviews and finally confirms the AI ​​quality control results; The statistical analysis submodule of the application layer module generates management reports based on the quality control results; The application layer module iterates and upgrades its sub-modules based on model iteration, and then optimizes the technology layer module's model using verified data. Compared with the prior art, the beneficial effects of the present invention are: 1. This AI-powered quality control system for image reports, based on data fusion and a dynamic rule engine, achieves comprehensive and efficient precision quality control. By integrating natural language processing, computer vision, and cross-modal matching technologies, the system can perform comprehensive automated screening of image reports for grammatical, textual, and logical consistency, as well as the correlation between images and text. This reduces the error correction response time from several minutes per case in traditional manual review to seconds, significantly improving quality control efficiency.

[0013] 2. This AI-based quality control system for image reporting workflows, built upon data fusion and a dynamic rule engine, establishes a standardized and iterative end-to-end management system. The system adopts a data layer-technology layer-application layer architecture, deeply integrating multi-source data from PACS / RIS / HIS, ​​enabling online management of the entire process from data access, AI preprocessing, manual review to statistical analysis. Through pre-set structured report templates and a dynamic rule engine built based on clinical guidelines and historical error data, the system ensures the uniformity and standardization of quality control standards.

[0014] 3. This AI quality control system for image reporting workflow based on data fusion and dynamic rule engine has formed a data-driven, closed-loop feedback continuous learning capability. The system not only has a powerful initial detection capability, but the correction results of manual review can also be automatically fed back to the training data pool for continuous training and optimization of the model.

[0015] 4. This AI-powered quality control system for image reporting workflows, based on data fusion and a dynamic rule engine, empowers departments with refined and precise management decisions. The system can collect quality control data across the entire process in real time and automatically, enabling traceability management of all operations and assigning responsibility to individuals. Through the generated multi-dimensional quality control reports at both the departmental and individual levels, it provides managers with objective and quantifiable decision-making support. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the AI ​​quality control system for image reporting based on the data fusion and dynamic rule engine of this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides a technical solution: an AI quality control system for image reporting workflow based on data fusion and a dynamic rule engine, comprising: The data layer module is used to integrate multi-source data from image archiving and communication systems, radiological information systems, and hospital information systems, and to preprocess the data to form a high-quality training data pool. Furthermore, the image archiving and communication system acquires historical images and corresponding report data in DICOM format; The radiology information system collects examination request information and physician diagnosis information; The hospital information system integrates basic patient information, laboratory test results, and pathology diagnostic reports. The data preprocessing unit cleans, standardizes, and transforms multi-source data into a structured form, thereby constructing a multi-source training data pool for model training.

[0019] The technology layer module, connected to the data layer module, includes: The natural language processing submodule uses a pre-trained language model that has been deeply trained on medical text and image reports to identify text errors, non-standard terminology, and logical contradictions in the reports. The natural language processing submodule specifically uses either the BERT or RoBERTa model for: Detect and correct spelling errors in the report; Based on the pre-established standardized medical terminology system, verify the standardization of terminology usage; Based on semantic parsing, logical contradictions in diagnostic descriptions are identified. The image-text verification submodule extracts image features based on a convolutional neural network model and performs cross-modal consistency matching with the report description to identify inconsistencies between images and text and missed diagnoses. A further implementation of the image-text validation submodule specifically employs the TResNet model for: Automatically extract the location, size, and morphological features of lesions from medical images; The extracted image features are associated and matched with descriptive statements in the report text; When there is a significant inconsistency between the image features and the text description, or when the text omits key image features, an inconsistency or missed diagnosis warning is generated. The structured template library submodule contains standardized report templates for various common diseases; The dynamic rule engine submodule generates and continuously iterates quality control rules based on clinical guidelines and historical error data. This submodule is used for: Based on authoritative clinical practice guidelines and historical error data, a quality control rule base is initialized and generated. During system operation, quality control rules are automatically generated or revised based on newly discovered error patterns and updated clinical guidelines. Receive feedback from clinicians in the manual review submodule, optimize rule weights and judgment thresholds, and achieve adaptive iterative upgrades of the rule base; The application layer module, including the model iteration and upgrade sub-module, specifically utilizes transfer learning techniques to optimize the parameters of the natural language processing model and image-text verification model in the technology layer module. This optimization employs physician-confirmed correction reports from the manual review sub-module as new training samples to continuously improve model performance. This module connects with the technology layer module and includes: The AI ​​quality control submodule receives the output from the technical layer module, automatically marks the error type and risk level in the report, and generates a quality control report. The manual review submodule provides a human-computer interaction interface for reviewing physicians to review and correct the AI-marked results; The statistical analysis submodule is used to generate quality control reports at the department and individual levels; The model iteration and upgrade submodule is used to continuously optimize the model in the technical layer module by using the correction results of manual review as new training data.

[0020] The system combines the automatic labeling of the AI ​​quality control submodule with the expert review of the manual review submodule to conduct graded quality control of medical imaging reports. High-risk errors are marked first and submitted for review. The overall report quality control process takes more than 50% less time than the traditional pure manual review.

[0021] The quality control method for medical image reports in this plan is implemented through the following means: The data layer module integrates multi-source medical data and constructs a training data pool. The natural language processing submodule and image-text verification submodule of the technical layer module are used to perform automated error detection on the quality control image report; The AI ​​quality control submodule of the application layer module generates a quality control report with error markers and risk levels based on the test results. Through the manual review submodule of the application layer module, the reviewing physician reviews and finally confirms the AI ​​quality control results; The statistical analysis submodule of the application layer module generates management reports based on the quality control results; The application layer module upgrades its sub-modules through model iteration, and optimizes the technology layer module's model using verified data.

[0022] The specific implementation steps of this AI-assisted image report quality control system are as follows: Construction and implementation of the data layer of this invention: First, multi-source medical data is integrated. This involves connecting to the hospital's internal information system interfaces, specifically the image archiving and communication systems, to obtain DICOM format image data and corresponding historical report texts; connecting to the radiology information system to collect examination request forms, patient clinical histories, and physician preliminary diagnostic information; and connecting to the hospital information system and laboratory information systems to integrate basic patient information, laboratory test results, and pathology reports. This multi-source data is then aggregated to construct a unified training data pool.

[0023] Next, the data imported into the data pool is preprocessed. For text data, it is cleaned to remove irrelevant characters and formatting errors. For image data, it undergoes standardization processing, including resolution adjustment and algorithmic removal of common artifacts. Finally, using natural language processing technology, the free report text is mapped to standard medical terminology systems such as RadLe's AI-powered image report workflow quality control system based on data fusion and dynamic rule engines, achieving preliminary structured transformation and providing a standardized data foundation for subsequent AI analysis.

[0024] The core module implementation of this invention at the technical layer: (1) Natural Language Processing Module: The BERT model is used to perform word segmentation and deep semantic analysis on the report text to identify and correct spelling errors and non-standard terminology. At the same time, the RoBERTa model is used to detect logical contradictions within the report, such as the contradictory descriptions of "no fractures found" and "discontinuous cortical bone" in the same report.

[0025] (2) Image-text verification module: Based on convolutional neural network models such as TResNet, it automatically extracts key features (such as the location, size, and density of lesions) in medical images. It performs cross-modal matching and consistency verification between these features and the description in the report text, thereby identifying missed diagnoses such as "abnormal image display but not mentioned in the report" or errors such as "the report description does not match what is seen in the image".

[0026] (3) Structured Template Library: Based on clinical guidelines and departmental standards, structured report templates are created for various common diseases such as rectal cancer MRI, lung nodule CT, and breast cancer MRI. These templates guide or mandate physicians to fill out reports according to the standard format, achieving standardized management of report content.

[0027] (4) Dynamic Rule Engine: This engine contains two core mechanisms. First, rule generation and updating: The system automatically parses the latest clinical guideline text and combines it with the error data pattern library accumulated in historical quality control to automatically generate or revise quality control rules through algorithms. Second, real-time verification and iterative closed loop: The rule set is called in real time for verification during the report generation process. For newly discovered error patterns, potential rule vulnerabilities are identified through anomaly detection algorithms, and the rule library is automatically optimized in combination with feedback from clinicians to achieve continuous iteration and self-evolution of rules.

[0028] The workflow implementation of this invention at the application layer: (1) AI-automated quality control: After the physician completes the initial draft of the report, the system automatically starts the quality control process and runs the above-mentioned technical modules in a comprehensive manner. The system completes the detection within seconds and marks suspected errors with different colors or icons on the report interface, while generating a detailed quality control suggestion report.

[0029] (2) Manual review intervention: Second-line review physicians log in to the system via a web browser to view the reports marked by AI. The physicians review each AI prompt, confirm valid errors and make corrections, or reject the AI's misjudgments. All manual review operations and results are recorded by the system.

[0030] (3) Statistical analysis and management: The system backend automatically aggregates data from the entire process and generates multi-dimensional statistical analysis reports. Managers can view indicators such as the error rate trend of departments or individuals, the distribution of common error types, and the average time spent on report review, which can be used for performance management and quality improvement decisions.

[0031] (4) Model Iteration and Upgrade: The corrected results, which have been manually reviewed and confirmed, are automatically fed back into the training data pool as new high-quality labeled data. The AI ​​model is retrained regularly using these new data through techniques such as transfer learning, thereby continuously optimizing the model's accuracy and generalization ability, forming a closed loop of "use-feedback-optimization".

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An image report process AI quality control system based on data fusion and dynamic rule engine, characterized in that, The application relates to an artificial intelligence (AI) quality control system for medical imaging reports, which comprises: a data layer module for integrating multi-source data from a picture archiving and communication system (PACS), a radiology information system (RIS) and a hospital information system (HIS), and preprocessing the data to form a high-quality training data pool; a technology layer module connected with the data layer module, which comprises: a natural language processing (NLP) sub-module using a pre-trained language model deeply trained by medical texts and imaging reports to identify text errors, non-standard terms and logical contradictions in the reports; an image-text verification sub-module based on a convolutional neural network (CNN) model to extract image features and match them with report descriptions for cross-modal consistency to identify inconsistencies between images and texts and missed diagnoses; a structured template library sub-module containing standardized report templates for common diseases; a dynamic rule engine sub-module generating and continuously iterating quality control rules based on clinical guidelines and historical error data; an application layer module connected with the technology layer module, which comprises: an AI quality control sub-module receiving the output of the technology layer module, automatically marking report error types and risk levels, and generating a quality control report; a manual review sub-module providing a human-machine interaction interface for auditors to review and correct the AI marking results; a statistical analysis sub-module generating quality control reports at department and individual levels; a model iteration and upgrading sub-module using the manual review correction results as new training data to continuously optimize the models in the technology layer module.

2. The AI quality control system for image reporting workflow based on data fusion and dynamic rule engine according to claim 1, characterized in that, The PACS acquires historical images and corresponding report data in DICOM format; The RIS collects examination application information and physician diagnosis information; The HIS integrates patient basic information, laboratory examination results and pathological diagnosis reports; A data preprocessing unit is used to clean, standardize and structure the multi-source data to build a multi-source training data pool for model training. 3.The AI quality control system based on data fusion and dynamic rule engine for image reporting workflow according to claim 1, wherein, The NLP sub-module specifically uses a BERT or RoBERTa model to: detect and correct spelling errors in the reports; check the standardization of term usage according to a preset standardized medical terminology system; and identify logical contradictions in the diagnosis descriptions based on semantic analysis.

4. The AI quality control system for image reporting workflow based on data fusion and dynamic rule engine of claim 1, wherein, The image-text verification sub-module specifically uses a TResNet model to: automatically extract the location, size and morphological features of lesions from medical images; associate and match the extracted image features with descriptive sentences in the report text; and generate inconsistency or missed diagnosis warnings when there is significant inconsistency between the image features and the text description or the text misses key image features.

5. The AI quality control system for image reporting workflow based on data fusion and dynamic rule engine of claim 1, wherein, The dynamic rule engine sub-module is used to: initialize a quality control rule library based on authoritative clinical diagnosis and treatment guidelines and the historical error data; automatically generate or revise quality control rules during system operation according to newly discovered error patterns and updated clinical guidelines; receive feedback from the manual review sub-module, optimize rule weights and decision thresholds, and realize adaptive iteration and upgrading of the rule library.

6. The AI quality control system for image reporting workflow based on data fusion and dynamic rule engine of claim 1, wherein, The model iterative upgrading sub-module in the application layer module, specifically through the migration learning technology, takes the correction report confirmed by the physician in the artificial review sub-module as a new training sample, optimizes the parameters of the natural language processing model and the image-text verification model in the technical layer module, so as to continuously improve the performance of the model.

7. The AI quality control system for image reporting workflow based on data fusion and dynamic rule engine of claim 1, wherein, The system combines the automatic marking of the AI quality control sub-module with the expert review of the artificial review sub-module to grade and control the medical image report, wherein the high-risk errors are preferentially marked and submitted for review, and the overall report quality control process reduces the time consumption by more than 50% compared with the traditional pure manual review.

8. The AI quality control system for image reporting workflow based on data fusion and dynamic rule engine of claim 1, wherein, The system is applied as claimed in any one of claims 1 to 7, and is realized by the following ways: Integrating multi-source medical data through the data layer module and constructing a training data pool; Through the natural language processing sub-module and the image-text verification sub-module of the technical layer module, automatically detecting errors of the image report to be controlled; Through the AI quality control sub-module of the application layer module, generating a quality control report with error marking and risk level according to the detection result; Through the artificial review sub-module of the application layer module, reviewing and finally confirming the AI quality control result by the reviewer physician; Through the statistical analysis sub-module of the application layer module, generating a management report based on the quality control result; Through the model iterative upgrading sub-module of the application layer module, optimizing the model of the technical layer module by using the reviewed and confirmed data.

Images