Quality control system of medical image diagnosis report based on artificial intelligence

Abstract

The invention discloses a medical image diagnosis report quality control system based on artificial intelligence, and the system comprises a report obtaining module which is used for obtaining medical image diagnosis report data, and the medical image diagnosis report obtaining modes comprise on-line transmission and off-line import; the content identification module is used for identifying the content of the medical image diagnosis report to obtain the report content; the checking module is used for checking and correcting the content of the report to obtain a checked and corrected report; the report generation module is used for analyzing the check and correction completion report to generate a structured report; the feedback module is used for collecting feedback information of the user on the structured report and comprehensively processing the feedback information to generate optimization management information; and the intelligent evaluation module is used for comprehensively analyzing all the generated structured reports to generate overall quality evaluation information. According to the invention, more comprehensive quality control of the medical image diagnosis report can be carried out, so that the quality of the medical image diagnosis report is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of quality control systems, and more particularly to a quality control system for medical imaging diagnosis reports based on artificial intelligence. Background Art

[0002] Medical imaging diagnosis reports should follow certain quality control standards, which usually include the integrity, accuracy, standardization, and timeliness of the report content. Integrity requires that the report contains all necessary examination information, such as patient basic information, examination site, examination method, imaging findings, and diagnostic opinions, etc. Accuracy requires that the descriptions and diagnoses in the report match the actual condition of the patient. Standardization requires that the report writing conforms to professional norms, uses professional terms, and avoids ambiguity. Timeliness requires that the report be completed and issued within the specified time;

[0003] A quality control system is used in the process of quality control of medical imaging diagnosis reports.

[0004] The existing quality control systems have poor quality control effects and low overall intelligence, and cannot effectively improve the quality of medical imaging diagnosis reports, which has a certain impact on the use of quality control systems. Therefore, a quality control system for medical imaging diagnosis reports based on artificial intelligence is proposed. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention provides a quality control system for medical imaging diagnosis reports based on artificial intelligence.

[0006] The present invention solves the above technical problems through the following technical solutions. The present invention includes:

[0007] A report acquisition module for acquiring medical imaging diagnosis report data. The ways to acquire medical imaging diagnosis reports include online transmission and offline import;

[0008] A content recognition module for recognizing the content of medical imaging diagnosis reports to obtain the report content;

[0009] An inspection module for inspecting and correcting the report content to obtain a report after inspection and correction;

[0010] A report generation module for analyzing the report after inspection and correction to generate a structured report;

[0011] A feedback module for collecting feedback information from users on the structured report and comprehensively processing the feedback information to generate optimized management information;

[0012] An intelligent evaluation module for comprehensively analyzing all generated structured reports to generate overall quality evaluation information.

[0013] Furthermore, the specific process of content recognition by the content recognition module is as follows:

[0014] Step 1: Image acquisition and preprocessing, that is, obtaining a clear image, and then performing grayscale processing and noise reduction processing;

[0015] Step 2: Text localization based on layout analysis. After that, through edge detection and contour extraction, use the edge detection algorithm to find the edges of the text area, and then obtain the contours of the text area through the contour extraction algorithm; Screen and analyze the contours of the text area, and determine the text area according to the characteristics such as the area, aspect ratio, etc. of the contours;

[0016] Step 3: Perform text segmentation. When the text is handwritten, first perform character segmentation, and segment the characters through vertical and horizontal projection methods;

[0017] For multi-line text, perform line segmentation. Segment the lines through vertical projection, calculate the vertical projection of the image, and find the low valley area of the projection value to determine the boundaries of each line of text;

[0018] Step 4: Perform feature extraction. Extract the text structure features. For each text area or character, extract its structure features;

[0019] Statistical feature extraction. Extract the statistical features of the text;

[0020] The structure features and statistical features constitute the text features;

[0021] Step 5: Pre-train a text recognition model using medical text samples;

[0022] Input the text areas or characters with extracted text features into the trained text recognition model, and output the recognized text results;

[0023] Step 6: Perform post-processing and proofreading. First, perform grammar and semantic correction. Since medical examination reports have specific grammar and semantic rules, check and correct the grammar and semantics of the recognized text;

[0024] Then, determine whether manual proofreading is required according to the actual situation;

[0025] Step 7: Output the recognized text. Output the corrected text recognition results in text format, and integrate and associate the recognized text with other medical information of the patient to form a complete medical record.

[0026] Furthermore, the content of the inspection by the inspection module includes typo inspection, medical common sense and imaging professional knowledge inspection, and diagnostic logic inspection.

[0027] Furthermore, the specific process of the typo inspection is as follows:

[0028] Tokenize the text, perform part-of-speech tagging, and syntactic analysis on the text simultaneously through conditional random fields, LSTM, and BERT to obtain the recognition processing results;

[0029] Obtain three recognition processing results, compare the similarities between any two of the three recognition processing results. When the similarities of the three recognition processing results are all greater than the preset value, randomly select any one of the recognition processing results for the next step;

[0030] When the similarity between any two of the three recognition processing results is greater than the preset value, randomly select one of the two recognition processing results with similarities greater than the preset value for the next step;

[0031] When the similarities of the three recognition processing results are all less than the preset value, re-perform the recognition processing.

[0032] Furthermore, the specific process of the medical common sense and imaging professional knowledge inspection is as follows:

[0033] S1: Verify the medical common sense imaging professional knowledge, medical terms, and diagnostic descriptions in the report through a preset intelligent model combined with a medical knowledge base and an imaging professional database to obtain the intelligent inspection results;

[0034] S2: Before exporting the intelligent inspection results for the first f times, it is necessary to compare the intelligent inspection results with the manual inspection results to obtain the manual analysis comparison results;

[0035] S3: When the proportion of passing the comparison in the manual analysis comparison results exceeds the preset proportion, directly use the intelligent inspection results in the subsequent medical common sense and imaging professional knowledge inspection process;

[0036] S4: After directly using the intelligent inspection results reaches the preset number of times, perform manual analysis again, that is, repeat step S2.

[0037] Furthermore, the process of obtaining the manual analysis comparison results is as follows: Extract the intelligent inspection results and the manual inspection results, compare them. When the contents of the intelligent inspection results and the manual inspection results are both the same, the manual analysis comparison result is that the comparison passes;

[0038] When there are deviations between the intelligent inspection results and the manual inspection results, the manual analysis comparison result is that the comparison fails.

[0039] Furthermore, the specific process of the diagnostic logic inspection is as follows: After performing diagnostic logic inspection through machine learning and deep learning algorithms, obtain the logic inspection results, and then export the logic inspection results;

[0040] Before the first f export logical check results, a verification of the diagnostic logic check needs to be carried out through a processing procedure similar to the intelligent check results.

[0041] The specific process of generating the structured report is as follows:

[0042] First, perform data parsing and segmentation. The text data of the inspection and correction completed report is initially segmented according to the logical structure of natural language to obtain the initially segmented data.

[0043] Extract preset category information. Using natural language processing technology and predefined medical term dictionaries and rule bases, calibration information is extracted from the initially segmented data.

[0044] Then, identify and extract the basic patient information in the report from the initially segmented data. The basic patient information includes name, age, gender, and medical record number.

[0045] After that, perform information classification and annotation. According to the preset database, the extracted calibration information and basic patient information are classified to obtain the classification information.

[0046] Add the corresponding metadata label Bz to each classification information, where z is the number of label types.

[0047] Pre-establish a filling benchmark template for the structured report. Filling boxes for each metadata label Bz are set in the filling benchmark template.

[0048] Fill each classification information into the filling box corresponding to the metadata label Bz.

[0049] Check the classification information filled into the filling box. When all classification information is filled correctly, export the filled benchmark template to obtain the structured report.

[0050] The specific process of checking the classification information filled into the filling box is as follows:

[0051] After the classification information is filled into the filling box, first perform a semantic content check to check whether the classification information belongs to the current filling box.

[0052] When the classification information belongs to the current filling box, proceed to the next check. When the classification information does not belong to the current filling box, fill it in again.

[0053] When the classification information belongs to the current filling box, use the document parsing method to determine the position information of each filling box on the page. For each filling box, record its top vertical coordinate and bottom vertical coordinate as well as the height h1 of the filling box.

[0054] Collect the starting top ordinate of the filled classification content through the same document parsing method and the ending ordinate value and calculate the height of the filled classification content

[0055] First, perform a top anomaly determination. Calculate the relative distance ratio r1 between the top of the filled classification content and the top of the filling box Set a top position threshold range. When r1 is within the top position threshold range, it indicates no anomaly. When r1 exceeds the top position threshold range, it indicates an anomaly;

[0056] Then, perform a bottom anomaly determination. Calculate the relative distance ratio r2 between the bottom of the filled classification content and the bottom of the filling box

[0057] When r2 is within the preset range, it indicates no anomaly. When r2 exceeds the preset range, it indicates an anomaly;

[0058] Perform an overall position anomaly comprehensive determination. Obtain the ratio r3 of the height of the filled content to the height of the filling box through r3 = h2 / h1;

[0059] Preset a ratio threshold for the height of the filled content to the height of the filling box. When r3 exceeds the ratio threshold of the height of the filled content to the height of the filling box, it indicates an anomaly. When r3 is within the ratio threshold of the height of the filled content to the height of the filling box, it indicates no anomaly;

[0060] When the top, bottom, and overall position of the filled classification content are all without anomalies, it indicates that all classification information is filled correctly.

[0061] Furthermore, the specific acquisition process of the optimized management information is as follows:

[0062] Step 1: Extract the feedback information of the structured reports obtained within a preset time period, that is, the total number of all feedback medical image diagnosis reports, denoted as A;

[0063] For each diagnostic report, record its corresponding image acquisition parameters, denoted as the set P = {p1, p2, p3... pn}, where Pn is the number of types of image acquisition parameters, and Pi represents the value of the i-th type of image acquisition parameter;

[0064] Compare the final diagnosis result of the diagnostic report with the clinical actual diagnosis result. If the comparison of the diagnosis results is exactly the same, denote it as R0;

[0065] If there are partial deviations in the comparison of the diagnosis results, that is, the similarity of the comparison of the diagnosis results is between 80% and 100%, denote it as R1;

[0066] If the similarity of the comparison of the diagnostic results is less than 80%, it is recorded as R2;

[0067] Count the number of various results, that is, R0, R1, and R2 are respectively recorded as C0, C1, and C2;

[0068] Then collect the qualification information of the diagnostic doctors, recorded as the set Q = {q1, q2, q3... qm}, where m is the number of types of qualification information, and qj represents the value of the jth type of qualification information;

[0069] Step 2: Calculate the abnormal ratio of the image acquisition parameters: For each image acquisition parameter pi, according to the preset range standard, count the ratio of the number of images exceeding the normal range to the total number of images, recorded as Ei;

[0070] After that, calculate the abnormal ratio E of the image acquisition parameters by the weighted average method. The weight Wp is set according to the influence degree of each parameter on the diagnostic accuracy. The weight set is Wp = {wp1, wp2, wp3... Wpn}, then

[0071] Calculate the deviation degree of the diagnostic results: Define the diagnostic deviation index Among them, 2*C2 is to increase the weight of the completely wrong diagnostic results;

[0072] Quantitatively score the qualifications of the diagnostic doctors:

[0073] According to the scoring criteria of factors such as professional title and working years within the hospital, assign the corresponding score sj to the qualification information qj of each doctor, and then calculate the comprehensive score S of the doctor's qualifications by the weighted average method. The weight set is Wq = {wq1, wq2, wq3... Wqj}, that is

[0074] Step 3: Set the weights of each factor. The weight of the abnormal ratio of the image acquisition parameters is α, the weight of the diagnostic deviation index is β, and the weight of the comprehensive score of the doctor's qualifications is γ, and α + β + γ = 1;

[0075] β > α > γ;

[0076] Calculate the optimization management index M = E×α + D×β + (1 - S)×γ;

[0077] According to the historical data and actual management requirements of the hospital, determine the threshold T. When M > T, generate optimization management information, indicating that it is necessary to optimize and improve the image acquisition process, diagnostic doctor training, or other related aspects.

[0078] Furthermore, the specific process for obtaining the overall quality assessment information is as follows: Extract the information related to the structured reports obtained within a preset time period. The information related to the structured reports includes the number of structured reports and the number of abnormal structured reports.

[0079] Calculate the ratio of the number of abnormal structured reports to the total number of structured reports to obtain an evaluation parameter. When the evaluation parameter is greater than or equal to the preset value, the overall quality assessment information is generated. At this time, the overall quality assessment information indicates that the overall evaluation is poor.

[0080] When the evaluation parameter is less than the preset value, the overall quality assessment information is generated. At this time, the overall quality assessment information indicates that the overall evaluation is excellent.

[0081] Furthermore, the determination process for abnormal structured reports is as follows: Compare the obtained structured reports with a preset standard report template to obtain the real-time similarity. When the real-time similarity is less than the preset value, it indicates that the structured report is abnormal.

[0082] The beneficial effects of the present invention are as follows:

[0083] The quality control system for medical imaging diagnosis reports based on artificial intelligence can automatically obtain, identify, check, and correct medical imaging diagnosis reports, greatly reducing the time and cost of manual operations. The automated process improves the efficiency of report processing, enabling doctors to focus more on patient diagnosis and treatment. The content recognition module can accurately identify the text information in the report through steps such as image preprocessing, text localization and segmentation, and feature extraction, reducing recognition errors. The inspection module ensures the accuracy and professionalism of the report through spelling error checks, medical common sense and imaging professional knowledge checks, and diagnostic logic checks. The intelligent evaluation module can comprehensively analyze all generated structured reports to generate overall quality assessment information, helping the system to continuously optimize and improve. The feedback module collects the feedback information from users on the structured reports and comprehensively processes it to generate optimization management information, enabling the system to self-adjust and optimize according to user feedback. Comprehensiveness and thoroughness: The system not only focuses on the accuracy of the report but also comprehensively evaluates the report through multiple aspects such as spelling error checks, medical common sense and imaging professional knowledge checks, and diagnostic logic checks. The report generation module can generate structured reports for subsequent analysis and utilization. The system can add new inspection rules and algorithms according to actual needs to cope with the development of new medical knowledge and technologies. Through regular manual analysis and comparison and verification of the intelligent inspection results, the accuracy and reliability of the system are ensured. By extracting and analyzing the feedback information, the system can promptly discover and improve existing problems, enhance the overall quality, and achieve more intelligent and comprehensive quality control of medical imaging diagnosis reports, making the system more worthy of promotion and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts do not necessarily draw to actual scale.

[0085] Figure 1 This is the system block diagram of the present invention. Specific embodiments

[0086] The following will describe in detail the embodiments of the technical solutions of the present invention in conjunction with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, so they are only examples and cannot be used to limit the protection scope of the present invention.

[0087] It should be noted that unless otherwise specified, the technical terms or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art to which the present invention belongs.

[0088] As Figure 1 shown, this embodiment provides a technical solution: a quality control system for medical imaging diagnosis reports based on artificial intelligence, including:

[0089] A report acquisition module, used to acquire medical imaging diagnosis report data, and the ways to acquire medical imaging diagnosis reports include online transmission and offline import;

[0090] A content recognition module, used to recognize the content of the medical imaging diagnosis report and obtain the report content;

[0091] An inspection module, used to inspect and correct the report content and obtain the report after inspection and correction;

[0092] A report generation module, used to analyze the report after inspection and correction to generate a structured report;

[0093] A feedback module, used to collect the feedback information of users on the structured report and comprehensively process the feedback information to generate optimized management information;

[0094] An intelligent evaluation module, used to comprehensively analyze all the generated structured reports to generate overall quality evaluation information.

[0095] The specific process of the content recognition module for content recognition is as follows:

[0096] Step 1: Image acquisition and preprocessing, that is, acquire a clear image, and then perform grayscale processing and noise reduction processing;

[0097] Step 2: Based on layout analysis for text localization, then through edge detection and contour extraction, use the edge detection algorithm to find the edges of the text region, and then obtain the contour of the text region through the contour extraction algorithm; Screen and analyze the contour of the text region, and determine the text region according to features such as the area and aspect ratio of the contour;

[0098] Step 3: Perform text segmentation. When the text is handwritten, first perform character segmentation, and segment the characters through the vertical and horizontal projection methods;

[0099] For multi-line text, perform line segmentation. Segment the lines through vertical projection, calculate the vertical projection of the image, and find the low valley area of the projection value to determine the boundary of each line of text;

[0100] Step 4: Perform feature extraction. Extract the text structure features. For each text region or character, extract its structure features;

[0101] Statistical feature extraction. Extract the statistical features of the text;

[0102] The structure features and statistical features form the text features;

[0103] Step 5: Pre-train a text recognition model using medical text samples;

[0104] Input the text regions or characters with extracted text features into the trained text recognition model, and output the recognized text results;

[0105] Step 6: Perform post-processing and proofreading. First, perform grammar and semantic correction. Since medical examination reports have specific grammar and semantic rules, check and correct the grammar and semantics of the recognized text;

[0106] Then select whether manual proofreading is required according to the actual situation;

[0107] Step 7: Output the recognized text. Output the corrected text recognition results in text format, and integrate and associate the recognized text with other medical information of the patient to form a complete medical record;

[0108] Through image preprocessing (such as grayscale processing and noise reduction processing), the noise and interference in the image can be removed, and the accuracy of subsequent text recognition can be improved.

[0109] Edge detection and contour extraction algorithms can accurately locate the text region and reduce the situation of misrecognition.

[0110] The text segmentation step ensures that each character or text region is accurately segmented, which helps to improve the recognition accuracy.

[0111] The automated process reduces the need for manual intervention and improves the processing speed.

[0112] Through a pre-trained text recognition model, the system can quickly output recognition results, shortening the report processing time.

[0113] The recognized text is output in text format, facilitating reading and understanding by doctors and other medical staff.

[0114] Integrating and correlating the recognized text with other medical information of the patient to form a complete medical record improves the integrity and traceability of medical information.

[0115] The content checked by the said checking module includes typo checking, medical common sense and imaging professional knowledge checking, and diagnostic logic checking.

[0116] The specific process of the said typo checking is as follows:

[0117] Simultaneously perform word segmentation, part-of-speech tagging, and syntactic analysis on the text through Conditional Random Field (CRF), Long Short-Term Memory (LSTM), and Bidirectional Encoder Representations from Transformers (BERT) to obtain recognition processing results;

[0118] Obtain three recognition processing results, compare the similarities between every two of the three recognition processing results. When the similarities of all three recognition processing results are greater than a preset value, arbitrarily select any one of the recognition processing results for the next step;

[0119] When the similarity between any two of the three recognition processing results is greater than the preset value, arbitrarily select one of the two recognition processing results with similarities greater than the preset value for the next step;

[0120] When the similarities of all three recognition processing results are less than the preset value, re-perform recognition processing;

[0121] The processing process of simultaneously performing word segmentation, part-of-speech tagging, and syntactic analysis on the text through Conditional Random Field (CRF), Long Short-Term Memory (LSTM), and Bidirectional Encoder Representations from Transformers (BERT) can be roughly divided into the following steps:

[0122] I. Text preprocessing:

[0123] Preprocess the input text, including removing irrelevant characters, word segmentation, etc. This step is the basis for subsequent processing to ensure that the text data meets the requirements of model input.

[0124] II. Word segmentation processing:

[0125] In word segmentation processing, BERT and LSTM are usually not directly used for the word segmentation task, but can be used as feature extractors or combined with other models. The following is a possible word segmentation processing flow:

[0126] Feature extraction using BERT: Input the text into the BERT model to obtain the context representation of the text. The bidirectional training feature of BERT enables it to understand the context and capture semantic information, which is helpful for the word segmentation task.

[0127] Word segmentation combined with CRF: After obtaining the context representation of BERT, it can be input into the CRF model. CRF is a discriminative undirected graph model that can make full use of context information and the dependency relationships between features for accurate inference and classification. By training the CRF model, the boundary information between words in the text can be learned, thus achieving word segmentation.

[0128] III. Part-of-speech tagging:

[0129] The task of part-of-speech tagging is to determine the part of speech of each word in the text. BERT, LSTM, and CRF can all play a role in this task:

[0130] BERT feature extraction: Similarly, first use BERT to encode the text to obtain the context representation of the text.

[0131] LSTM sequence modeling: Use the output of BERT as the input of LSTM, and use the network structure of LSTM to model the sequence data. LSTM can handle the long-term dependencies in the sequence data, which is beneficial for the part-of-speech tagging task.

[0132] CRF layer optimization: Apply the CRF layer to the output of LSTM to utilize the constraints of CRF to ensure that the final part-of-speech tagging result is valid. These constraints can be automatically learned by the CRF layer during the training process.

[0133] IV. Syntactic analysis:

[0134] The task of syntactic analysis is to determine the structure of the sentences in the text, including the grammatical relationships and dependency relationships between words. This task is usually more complex than word segmentation and part-of-speech tagging and requires more advanced models to handle. Although BERT, LSTM, and CRF can all be used for natural language processing tasks, their direct application in syntactic analysis may be relatively limited. One possible method is:

[0135] BERT pre-trained model: First use BERT to encode the text to obtain the context representation of the text. The bidirectional training feature of BERT enables it to capture rich semantic information, which is helpful for the syntactic analysis task.

[0136] Combined with a specific syntactic analysis model: Then, the output of BERT can be input into a model specifically for syntactic analysis, such as a dependency parser. These models are usually specially trained to identify the grammatical relationships and dependency relationships in the sentence.

[0137] Post - processing: Finally, post - process the output of the syntactic analyzer to obtain the final syntactic analysis result. This may include adjusting the direction of dependency relations, merging similar structures, etc.

[0138] V. Obtaining the recognition and processing results:

[0139] After the processing of the above - mentioned steps, the word - segmentation result, part - of - speech tagging result, and syntactic analysis result of the text can be obtained. These results can be used for subsequent natural language processing tasks, such as text classification, information extraction, etc.

[0140] Through the above process, the accuracy of misspelling checking can be improved. By combining three different models, namely conditional random field, LSTM, and BERT, the text can be analyzed from different perspectives, so as to more accurately identify the misspellings in the text.

[0141] These models have high accuracy in word - segmentation, part - of - speech tagging, and syntactic analysis, can understand the text content more precisely, and reduce the possibility of misjudgment.

[0142] Due to the use of multiple models for parallel processing, when one model has an error, other models can be used as a supplement to improve the fault - tolerance ability of the overall system. By comparing the output results of different models, the most reliable recognition result can be selected for subsequent processing. Parallel processing of the recognition tasks of multiple models can make full use of computing resources and improve the processing speed.

[0143] The similarity comparison strategy can quickly screen out the most reliable recognition result without increasing additional computational complexity, avoiding unnecessary repeated processing. The automated misspelling checking process reduces the need for manual proofreading and lowers the labor cost. Accurate misspelling checking can reduce misunderstandings and errors caused by misspellings, and improve the accuracy and reliability of medical information. This process reflects the application potential of artificial intelligence technology in misspelling checking and provides strong support for the intelligent upgrade of the medical field.

[0144] The specific process of the above - mentioned misspelling checking brings significant benefits in terms of improving accuracy, enhancing robustness, increasing processing efficiency, reducing the cost of manual intervention, and promoting intelligent upgrade. These benefits not only improve the efficiency and quality of medical information processing, but also provide strong support for the intelligent development of the medical field.

[0145] The specific process of the medical common sense and imaging professional knowledge checking is as follows:

[0146] S1: Through a preset intelligent model combined with a medical knowledge base and an imaging professional database, verify the medical common sense, imaging professional knowledge, medical terms, and diagnostic descriptions in the report to obtain the intelligent inspection result;

[0147] S2: Before the first f times of exporting the intelligent inspection results, it is necessary to compare the intelligent inspection results with the manual inspection results to obtain the manual analysis and comparison results.

[0148] S3: When the proportion of passing the comparison in the manual analysis and comparison results exceeds the preset proportion, that is, in the subsequent inspection process of medical common sense and imaging professional knowledge, directly use the intelligent inspection results.

[0149] S4: After directly using the intelligent inspection results reaches the preset number of times, conduct manual analysis again, that is, repeat step S2.

[0150] The process of obtaining the manual analysis and comparison results is as follows: Extract the intelligent inspection results and the manual inspection results, and compare them. When the contents of the intelligent inspection results and the manual inspection results are both consistent, that is, the manual analysis and comparison result is passing the comparison.

[0151] When there are deviations between the intelligent inspection results and the manual inspection results, that is, the manual analysis and comparison result is not passing the comparison.

[0152] The above process improves the accuracy and efficiency of the inspection. The intelligent model combines the medical knowledge base and the imaging professional database, and can quickly and accurately verify the medical common sense, imaging professional knowledge, medical terms and diagnostic descriptions in the report, reducing the cumbersome and time-consuming manual inspection.

[0153] Through the preset intelligent model, batch processing of a large number of reports can be realized, significantly improving the inspection efficiency.

[0154] The automated inspection process reduces the need for manual intervention and reduces the manpower investment of medical institutions.

[0155] Especially when facing a large number of reports, automated inspection can significantly reduce the work burden of medical staff.

[0156] Enhance the objectivity and consistency of the inspection: The intelligent model conducts inspections based on preset rules and algorithms, avoiding subjectivity and inconsistency caused by human factors.

[0157] This helps to ensure the objectivity and fairness of the inspection results and improve the medical quality.

[0158] Promote the development of medical informatization and intelligence: This process reflects the application value of artificial intelligence technology in the medical field and provides strong support for the development of medical informatization and intelligence.

[0159] With the continuous progress of technology and the in-depth application, intelligent inspection will play an increasingly important role in the medical field.

[0160] Improving the quality and safety of medical services: Accurate medical knowledge and imaging expertise examinations help ensure the accuracy and reliability of medical reports, thereby enhancing the quality of medical services.

[0161] Meanwhile, by reducing medical risks caused by incorrect information, it also helps improve the safety of medical services.

[0162] The specific process of the above-mentioned diagnostic logic check is as follows: After performing a diagnostic logic check based on machine learning and deep learning algorithms, the logic check results are obtained, and then the logic check results are exported.

[0163] Before the first f times of exporting the logic check results, a verification process similar to the intelligent check results needs to be carried out for the diagnostic logic check, where f ≥ 10.

[0164] Improving the accuracy and logic of diagnosis: The diagnostic logic check based on machine learning and deep learning algorithms can automatically analyze the logical relationships and consistencies in the diagnostic process to ensure the accuracy and logic of diagnosis.

[0165] This helps reduce misdiagnosis and missed diagnosis cases caused by logical errors.

[0166] Optimizing the diagnostic process: By automatically checking the diagnostic logic, problems in the diagnostic process can be discovered and corrected in a timely manner, thereby optimizing the diagnostic process.

[0167] This helps improve the diagnostic efficiency and reduce the waiting time of patients.

[0168] Enhancing the professional level of doctors: The diagnostic logic check process is not only a verification of the diagnostic results but also a training and improvement of doctors' diagnostic thinking.

[0169] By continuously learning and optimizing the algorithms, doctors can continuously improve their professional level and provide better medical services for patients.

[0170] Promoting medical research and innovation: The data and results generated by the automated check process can provide valuable references and bases for medical research.

[0171] This helps promote the research and innovation development in the medical field and provides support for the continuous improvement and upgrading of medical services.

[0172] The specific processes of medical knowledge and imaging expertise examinations and diagnostic logic checks bring significant benefits in terms of improving the accuracy and efficiency of examinations, reducing labor costs, enhancing the objectivity and consistency of examinations, promoting the development of medical informatization and intelligentization, and improving the quality and safety of medical services.

[0173] The specific process of the above-mentioned structured report generation is as follows:

[0174] First, perform data parsing and segmentation. Initially segment the text data of the inspection and correction completion report according to the logical structure of natural language to obtain the initially segmented data. For example, delimit the report into different information blocks using paragraphs, sentences, or specific punctuation marks. This may involve identifying the content of different parts such as titles, subtitles, body descriptions, conclusions, etc.

[0175] Extract preset category information. Use natural language processing techniques and predefined medical term dictionaries and rule bases to extract calibration information from the initially segmented data. Calibration information includes, for the disease diagnosis part, extracting disease names, locations, severity levels, etc.; for inspection items, extracting inspection types, inspection times, inspection result values (such as various indicator values in laboratory tests).

[0176] Then, identify and extract the basic patient information in the report from the initially segmented data. The basic patient information includes name, age, gender, and medical record number.

[0177] After that, perform information classification and annotation. Classify the extracted calibration information and basic patient information according to a preset database to obtain classified information. The preset database includes the knowledge system in the medical field and clinical practice norms. For example, classify symptom information as "clinical manifestations", inspection results as "auxiliary inspection results", and diagnosis conclusions as "diagnostic opinions".

[0178] Add the corresponding metadata label Bz to each classified information, where z is the number of label types.

[0179] Pre-establish a filling benchmark template for the structured report. The filling benchmark template sets filling boxes for each metadata label Bz.

[0180] Fill each classified information into the filling box corresponding to the metadata label Bz.

[0181] Check the classified information filled into the filling box. When all classified information is filled correctly, export the filled benchmark template, that is, obtain the structured report.

[0182] The specific process of checking the classified information filled into the filling box is as follows:

[0183] After the classified information is filled into the filling box, first perform a semantic content check to check whether the classified information belongs to the current filling box.

[0184] When the classified information belongs to the current filling box, proceed to the next step of the check. When the classified information does not belong to the current filling box, re-fill it.

[0185] When the classification information belongs to the current input box, use the document parsing method to determine the position information of each input box on the page. For each input box, record its top vertical coordinate and bottom vertical coordinate as well as the height h1 of the input box,

[0186] Through the same document parsing method, collect the starting top vertical coordinate and the ending vertical coordinate value of the filled classification content and calculate the height

[0187] First, perform a top anomaly determination. Calculate the relative distance ratio r1 between the top of the filled classification content and the top of the input box. Set a top position threshold range. When r1 is within the top position threshold range, it indicates no anomaly. When r1 exceeds the top position threshold range, it indicates an anomaly;

[0188] When 0 ≤ r1 ≤ 0.2, it means that the distance from the top of the filled content to the top of the input box should not exceed 20% of the height of the input box);

[0189] If r1 < 0, it means that the filled content exceeds the top boundary of the input box, and it is determined as a top position anomaly; if r1 > 0.2, it is considered that there is an abnormal deviation in the top position of the filled content, and it may be that the filling position is too high;

[0190] Then, perform a bottom anomaly determination. Calculate the relative distance ratio r2 between the bottom of the filled classification content and the bottom of the input box.

[0191] When r2 is within the preset range, it indicates no anomaly. When r2 exceeds the preset range, it indicates an anomaly;

[0192] When 0 ≤ r2 ≤ 0.2, it means that the distance from the top of the filled content to the top of the input box should not exceed 20% of the height of the input box);

[0193] If r2 < 0, it means that the filled content exceeds the top boundary of the input box, and it is determined as a top position anomaly; if r2 > 0.2, it is considered that there is an abnormal deviation in the top position of the filled content, and it may be that the filling position is too high;

[0194] Perform an overall position anomaly comprehensive determination. Through r3 = h2 / h1, obtain the ratio r3 of the height of the filled content to the height of the input box;

[0195] A ratio threshold of the height of the filled content to the height of the filling box is preset. When r3 exceeds the ratio threshold of the height of the filled content to the height of the filling box, it indicates an abnormality. When r3 is within the ratio threshold of the height of the filled content to the height of the filling box, it indicates no abnormality;

[0196] For example, r3 < 0.5 indicates that the height of the filled content is less than 50% of the height of the filling box, which may mean that the filled content is not completely filled or the position is severely deviated, or r3 > 1.5, which means that the height of the filled content exceeds 1.5 times the height of the filling box, and there may be an abnormal height caused by duplicate filling of content or abnormal format;

[0197] When the top, bottom, and overall positions of the filled classification content are all normal, it means that all classification information is filled correctly.

[0198] The specific process of obtaining the optimized management information is as follows:

[0199] Step 1: Extract the feedback information of the structured reports obtained within a preset time period, that is, the total number of all feedback medical imaging diagnosis reports, denoted as A;

[0200] For each diagnosis report, record its corresponding image acquisition parameters, denoted as the set P = {p1, p2, p3... pn}, where Pn is the number of types of image acquisition parameters, and Pi represents the value of the i-th type of image acquisition parameter;

[0201] Compare the final diagnosis result of the diagnosis report with the clinical actual diagnosis result. If the comparison of the diagnosis results is exactly the same, it is denoted as R0;

[0202] If there are some deviations in the comparison of the diagnosis results, that is, the similarity of the comparison of the diagnosis results is between 80% and 100%, it is denoted as R1;

[0203] If the similarity of the comparison of the diagnosis results is less than 80%, it is denoted as R2;

[0204] Count the number of various results, that is, R0, R1, and R2 are respectively denoted as C0, C1, and C2;

[0205] Then collect the qualification information of the diagnosing doctor, denoted as the set Q = {q1, q2, q3... qm}, where m is the number of types of qualification information, and qj represents the value of the j-th type of qualification information;

[0206] Step 2: Calculate the abnormal ratio of the image acquisition parameters: For each type of image acquisition parameter pi, according to the preset range standard, count the ratio of the number of images exceeding the normal range to the total number of images, denoted as Ei;

[0207] After that, the abnormal ratio E of the image acquisition parameters is calculated by the method of weighted average. The weight Wp is set according to the influence degree of each parameter on the diagnostic accuracy. The weight set is Wp = {wp1, wp2, wp3... Wpn}, then

[0208] Calculate the deviation degree of the diagnostic result: Define the diagnostic deviation index where 2*C2 is to increase the weight of the completely wrong diagnostic result. The above situation is only used when C1 + 2×C2 ≤ A. When C1 + 2×C2 > A, use to obtain the defined diagnostic deviation index D;

[0209] Quantitatively score the qualifications of the diagnostic doctors:

[0210] According to the scoring criteria of factors such as professional title and working years within the hospital, assign the corresponding score sj to the qualification information qj of each doctor, and then calculate the comprehensive score S of the doctor's qualifications by the method of weighted average. The weight set is Wq = {wq1, wq2, wq3... Wqj}, that is

[0211] Step 3: Set the weights of each factor. The weight of the abnormal ratio of the image acquisition parameters is α, the weight of the diagnostic deviation index is β, and the weight of the comprehensive score of the doctor's qualifications is γ, and α + β + γ = 1;

[0212] β > α > γ;

[0213] Calculate the optimization management index M = E×α + D×β + (1 - S)×γ. 1 - S is because a higher comprehensive score of the doctor's qualifications should correspond to a lower optimization requirement, that is, take the inverse of S to reflect the correlation with the optimization management;

[0214] According to the historical data and actual management requirements of the hospital, determine the threshold T;

[0215] When M > T, generate optimization management information, indicating that it is necessary to optimize and improve the image acquisition process, diagnostic doctor training or other related aspects. For example, if M is greater than 0.6, issue a detailed optimization suggestion report, including the calibration of the image acquisition equipment, the special training plan for doctors, etc., to improve the quality and accuracy of the medical image diagnostic report.

[0216] The specific process of obtaining the overall quality assessment information is as follows: Extract the structured report-related information obtained within the preset time period. The structured report-related information includes the number of structured reports and the number of abnormal structured reports;

[0217] Calculate the ratio of the number of abnormal structured reports to the total number of structured reports to obtain an evaluation parameter. When the evaluation parameter is greater than or equal to the preset value, generate the overall quality evaluation information, and at this time, the overall quality evaluation information is that the overall evaluation is poor.

[0218] When the evaluation parameter is less than the preset value, generate the overall quality evaluation information, and at this time, the overall quality evaluation information is that the overall evaluation is excellent.

[0219] The determination process of the abnormal structured report is as follows: Compare the obtained structured report with the preset standard report template to obtain the real-time similarity. When the real-time similarity is less than the preset value, it means that the structured report is abnormal.

[0220] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. A quality control system for medical imaging diagnosis reports based on artificial intelligence, characterized in that, Including: A report acquisition module, which is used to acquire medical imaging diagnosis report data. The ways to acquire medical imaging diagnosis reports include online transmission and offline import; A content recognition module, which is used to recognize the content of medical imaging diagnosis reports and obtain the report content; An inspection module, which is used to inspect and correct the report content and obtain the inspected and corrected report; A report generation module, which is used to analyze the inspected and corrected report to generate a structured report; A feedback module, which is used to collect the feedback information of users on the structured report and comprehensively process the feedback information to generate optimized management information; An intelligent evaluation module, which is used to comprehensively analyze all generated structured reports to generate overall quality evaluation information.

2. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 1, characterized in that: The specific process of content recognition by the content recognition module is as follows: Step 1: Image acquisition and preprocessing, that is, acquiring clear images, and then performing grayscale processing and noise reduction processing; Step 2: Text localization based on layout analysis. After that, through edge detection and contour extraction, use the edge detection algorithm to find the edges of the text area, and then use the contour extraction algorithm to obtain the contours of the text area; Screen and analyze the contours of the text area, and determine the text area according to the characteristics such as the area and aspect ratio of the contours; Step 3: Perform text segmentation. When the text is handwritten, first perform character segmentation, and segment the characters by vertical and horizontal projection methods; For multi-line text, perform line segmentation. Segment the lines by vertical projection, calculate the vertical projection of the image, and find the low valley area of the projection value to determine the boundary of each line of text; Step 4: Perform feature extraction. Extract the text structure features. For each text area or character, extract its structure features; Statistical feature extraction, extract the statistical features of the text; The structural features and statistical features form the text features; Step 5: Pre-train a text recognition model with medical text samples; Input the text areas or characters with extracted text features into the trained text recognition model, and output the recognized text results; Step 6: Perform post-processing and proofreading. First, perform grammar and semantic correction. Since medical examination reports have specific grammar and semantic rules, perform grammar and semantic inspection and correction on the recognized text; Then determine whether manual proofreading is required according to the actual situation; Step 7: Output the recognized text. Output the corrected text recognition results in text format, and integrate and associate the recognized text with other medical information of the patient to form a complete medical record.

3. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 1, characterized in that: The content inspected by the inspection module includes misspelling inspection, medical common sense and imaging professional knowledge inspection, and diagnostic logic inspection.

4. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 3, characterized in that: The specific process of the misspelling inspection is as follows: Simultaneously perform word segmentation, part-of-speech tagging, and syntactic analysis on the text through conditional random fields, LSTM, and BERT to obtain the recognition processing results; Obtain three recognition processing results, compare the similarities between the three recognition processing results pairwise. When the similarities of the three recognition processing results are all greater than the preset value, select any one of the recognition processing results for the next step; When the similarity between any two of the three recognition processing results is greater than a preset value, any one of the two recognition processing results with similarity greater than the preset value is selected for the next step; When the similarity of the three recognition processing results is less than the preset value, the recognition processing is restarted.

5. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 3, characterized in that: The specific process of the medical common sense and imaging professional knowledge inspection is as follows: S1: Verify the medical common sense imaging professional knowledge, medical terms, and diagnosis descriptions in the report through a preset intelligent model combined with a medical knowledge base and an imaging professional database to obtain an intelligent inspection result; S2: Before the intelligent inspection results are exported for the first f times, it is necessary to compare the intelligent inspection results with the manual inspection results to obtain a manual analysis comparison result; S3: When the proportion of passing the comparison in the manual analysis comparison result exceeds the preset proportion, during the subsequent medical common sense and imaging professional knowledge inspection process, the intelligent inspection results are directly used; S4: After the intelligent inspection results are directly used for the preset number of times, a manual analysis is performed again, that is, step S2 is repeated.

6. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 5, characterized in that: The process of obtaining the manual analysis comparison result is as follows: Extract the intelligent inspection result and the manual inspection result, and compare them. When the contents of the intelligent inspection result and the manual inspection result are the same, the manual analysis comparison result is considered to pass the comparison; When there is a deviation between the intelligent inspection result and the manual inspection result, the manual analysis comparison result is considered to fail the comparison.

7. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 3, wherein: The specific process of the diagnostic logic inspection is as follows: After performing the diagnostic logic inspection through machine learning and deep learning algorithms, a logic inspection result is obtained, and then the logic inspection result is exported; Before the logic inspection results are exported for the first f times, a verification process similar to the intelligent inspection results needs to be performed for the diagnostic logic inspection.

8. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 1, characterized in that, The specific process of generating the structured report is as follows: First, perform data parsing and segmentation, and preliminarily segment the text data of the inspection and correction completed report according to the logical structure of natural language to obtain preliminary segmentation data; Extract preset category information, and use natural language processing technology and a predefined medical term dictionary and rule library to extract calibration information from the preliminary segmentation data; Then, identify and extract the patient's basic information in the report from the preliminary segmentation data. The patient's basic information includes name, age, gender, and medical record number; After that, perform information classification and annotation, classify the extracted calibration information and patient's basic information according to a preset database to obtain classification information; Add the corresponding metadata label Bz to each classification information, where z is the number of label types; Pre-establish a filling benchmark template for the structured report, and set a filling box for each metadata label Bz in the filling benchmark template; Fill each classification information into the filling box corresponding to the metadata label Bz; Check the classification information filled into the filling box. When all the classification information is filled correctly, export the filled benchmark template, that is, obtain the structured report; The specific process of checking the classification information filled into the filling box is as follows: After the classification information is filled into the filling box, first perform a semantic content check to check whether the classification information belongs to the current filling box; When the classification information belongs to the current filling box, the next step of inspection is carried out. When the classification information does not belong to the current filling box, filling is carried out again; When the classification information belongs to the current input box, use the document parsing method to determine the position information of each input box on the page. For each input box, record its top vertical coordinate and bottom vertical coordinate as well as the height h1 of the input box Collect the starting top ordinate of the filled classification content by the same document parsing method and the ending ordinate value and calculate the height of the filled classification content First, perform the top anomaly determination. Calculate the relative distance ratio r1 between the top of the filled classification content and the top of the filling box. Set the top position threshold range. When r1 is within the top position threshold range, it indicates no anomaly. When r1 exceeds the top position threshold range, it indicates an anomaly. Then, perform bottom-end anomaly determination, and calculate the relative distance ratio r2 between the bottom end of the classified content filled in and the bottom end of the filling box. When r2 is within the preset range, it indicates no abnormality. When r2 exceeds the preset range, it indicates an abnormality; Perform an overall position abnormality comprehensive determination. By r3 = h2 / h1, obtain the ratio r3 of the height of the filled content to the height of the filling box; A ratio threshold of the height of the filled content to the height of the filling box is preset. When r3 exceeds the ratio threshold of the height of the filled content to the height of the filling box, it indicates an abnormality. When r3 is within the ratio threshold of the height of the filled content to the height of the filling box, it indicates no abnormality; When the top, bottom, and overall position of the filled classification content are all without abnormality, it indicates that all classification information is filled correctly.

9. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 1, characterized in that: The specific acquisition process of the optimized management information is as follows: Step 1: Extract the feedback information of the structured reports obtained within a preset time period, that is, the total number of all feedback medical image diagnosis reports, denoted as A; For each diagnosis report, record its corresponding image acquisition parameters, denoted as the set P = {p1, p2, p3... pn}, where Pn is the type number of the image acquisition parameters, and Pi represents the value of the i-th type of image acquisition parameter; Compare the final diagnosis result of the diagnosis report with the clinical actual diagnosis result. If the comparison of the diagnosis results is exactly the same, denote it as R0; If there are some deviations in the comparison of the diagnosis results, that is, the similarity of the comparison of the diagnosis results is between 80% and 100%, denote it as R1; If the similarity of the comparison of the diagnosis results is less than 80%, denote it as R2; Count the number of various results, that is, R0, R1, and R2 are respectively denoted as C0, C1, and C2; Then collect the qualification information of the diagnosing doctor, denoted as the set Q = {q1, q2, q3... qm}, where m is the type number of the qualification information, and qj represents the value of the j-th type of qualification information; Step 2: Calculate the abnormal ratio of the image acquisition parameters: For each type of image acquisition parameter pi, according to the preset range standard, count the ratio of the number of images exceeding the normal range to the total number of images, denoted as Ei; After that, the abnormal ratio E of the image acquisition parameters is calculated by weighted average. The weight Wp is set according to the influence degree of each parameter on the diagnostic accuracy, and the weight set is Wp = {wp1, wp2, wp3... Wpn}, then Calculate the deviation degree of the diagnostic result: Define the diagnostic deviation index Quantitatively evaluate the qualification of the diagnosing doctor: According to the scoring criteria based on factors such as professional titles and working years within the hospital, assign corresponding scores sj to the qualification information qj of each doctor, and then calculate the comprehensive doctor qualification score S through the method of weighted average. The weight set is Wq = {wq1, wq2, wq3... Wqj}, that is Step 3: Set the weights of each factor. Let the weight of the abnormal ratio of the image acquisition parameters be α, the weight of the diagnosis deviation index be β, and the weight of the comprehensive score of the doctor's qualification be γ, and α + β + γ = 1; β > α > γ; Calculate the optimization management index M = E×α + D×β + (1 - S)×γ; According to the historical data and actual management requirements of the hospital, determine the threshold T. When M > T, generate the optimized management information.

10. The quality control system for medical imaging diagnosis reports based on artificial intelligence according to claim 1, characterized in that: The specific acquisition process of the overall quality assessment information is as follows: Extract the information related to the structured reports obtained within a preset time period. The information related to the structured reports includes the number of structured reports and the number of abnormal structured reports; Calculate the ratio of the number of abnormal structured reports to the total number of structured reports to obtain the evaluation parameter. When the evaluation parameter is greater than or equal to the preset value, generate the overall quality assessment information. At this time, the overall quality assessment information is that the overall assessment is poor; When the evaluation parameter is less than the preset value, the overall quality evaluation information is generated. At this time, the overall quality evaluation information indicates that the overall evaluation is relatively good. The determination process of the structured report anomaly is as follows: Compare the obtained structured report with the preset standard report template to obtain the real-time similarity. When the real-time similarity is less than the preset value, it indicates that the structured report is abnormal.

Images