Image text report quality control method and system based on LLM and structured report

Abstract

The invention discloses an image text report quality control method and system based on LLM and a structured report, relates to the technical field of medical text report quality control, and solves the problems that a current text report quality control method is difficult to accurately extract key fields and is difficult to judge complex diagnostic reasoning logic. According to the technical scheme, the method is characterized in that iconography feature description related to a diagnosis conclusion is extracted from a text report, cue words are constructed according to a matched structured report template to analyze the iconography feature description into structured data, and an interface is called to fill the structured data into the structured report template. And obtaining a template diagnosis conclusion according to the built-in judgment logic of the structured report template, and comparing the template diagnosis conclusion with the current diagnosis conclusion to realize quality control auditing of the text report.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical text report quality control, more particularly, it relates to an image text report quality control method and system based on LLM and structured report. BACKGROUND

[0002] Currently, when writing an image diagnosis report in a domestic imaging department, a text template is usually selected according to the diagnostic impression, and then personalized modification is performed, and the range of personalized modification covers image manifestations to diagnostic conclusions. This template-based text report writing is fast, but it is difficult to control the quality of the text report, and the diagnostic quality control is uneven.

[0003] There are currently two main text report quality control methods. One is to use a structured report system, which inputs image manifestations and diagnostic conclusions by selecting points in the structured report template. The structured report template controls the description method of the image manifestations, and has built-in reasoning logic for disease diagnosis. The advantage is that it can prospectively control the quality of the text report, but the disadvantage is that frequent point selection and filling will sacrifice work efficiency. Therefore, under the current huge business volume of imaging departments, structured reports are difficult to promote. The other is to extract key fields from text reports based on rules, and perform simple rule matching (such as the patient's gender is male, so the description of female disease image manifestations and diagnostic conclusions should not appear in the report). A correction system can be configured with hundreds of rules using similar logic. The advantage is that the implementation cost is low, but the disadvantage is that it can only perform strict matching of the literal, lacks semantic matching, and cannot match when the description is slightly different. Only simple, literal errors can be checked, and the diagnostic logic cannot be judged in depth.

[0004] The current text report quality control method mainly has two difficulties. First, it is difficult to extract key words of image manifestations and diagnostic conclusions from text reports based on NLP (natural language processing) technology, with low accuracy and difficulty in practical application. Second, the rules for judging image diagnosis reasoning logic are complex, and it is difficult to cover all sub-scenes in the diagnosis process using traditional programming technology.

[0005] Therefore, the present application provides an image text report quality control method and system based on LLM and structured report to solve the above problems. SUMMARY

[0006] The purpose of the present application is to provide an image text report quality control method and system based on LLM and structured report, to solve the problem that the current text report quality control method is difficult to accurately extract key fields and difficult to judge complex diagnostic reasoning logic. The present application extracts the imaging feature description related to the diagnostic conclusion from the text report, parses the imaging feature description into structured data according to the prompt words constructed according to the matched structured report template, fills the structured data into the structured report template through the interface, obtains the template diagnostic conclusion according to the judgment logic built in the structured report template, and compares the template diagnostic conclusion with the current diagnostic conclusion to realize the quality control audit of the text report.

[0007] The present application first provides an image text report quality control method based on LLM and structured report, which includes: obtaining an image text report to be audited, the image text report including: imaging performance and diagnostic impression; extracting the first N diagnostic conclusions indicating certainty or doubt from the diagnostic impression; querying the corresponding structured report template according to the scanning device type, scanning site and diagnostic conclusion; inputting the extracted diagnostic conclusion and imaging performance into the LLM model to extract the imaging feature description related to the diagnostic conclusion from the imaging performance; constructing prompt words according to the structured report template, inputting the imaging feature description into the LLM model, and obtaining the structured data of the imaging feature description according to the prompt words; calling the interface of the structured report template based on the structured data of the imaging feature description, filling the structured data into the structured report template to obtain the template diagnostic impression; comparing the template diagnostic impression with the diagnostic impression of the image text report to obtain an audit result.

[0008] By using the above technical solution, the prompt words are constructed based on the matched structured report template, the structured data is extracted from the imaging feature description through the LLM model to realize the accurate extraction of key fields; the structured data is filled into the structured report template, the template diagnostic conclusion is obtained based on the judgment logic built in the structured report template, and the template diagnostic conclusion is compared with the current diagnostic conclusion to realize the judgment of complex diagnostic reasoning logic.

[0009] In a possible implementation, the prompt words are constructed according to the structured report template, the imaging feature description is input into the LLM model, and the structured data of the imaging feature description is obtained according to the prompt words; including: constructing prompt words according to the information required to be filled in the structured report template; extracting the information required to be filled in the structured report template from the imaging feature description through the prompt words as discrete information; based on the discrete information extracted from the imaging feature description, the structured data of the imaging feature description is composed.

[0010] In a possible implementation, the interface of the structured report template is invoked based on the structured data of the imaging feature description, the structured data is filled into the structured report template to obtain a template diagnosis impression; the interface of the structured report template is invoked, the structured report template includes a plurality of information to be filled; discrete information in the structured data is filled into a corresponding position of the structured report template; the structured report template executes built-in judgment rules to output the template diagnosis impression.

[0011] In a possible implementation, the template diagnosis impression is compared with a diagnosis impression of the imaging text report to obtain an audit result; when the template diagnosis impression is consistent with the diagnosis impression of the imaging text report, the imaging text report to be audited passes the quality inspection audit; when the template diagnosis impression is inconsistent with the diagnosis impression of the imaging text report, the imaging text report to be audited fails the quality inspection audit, and the template diagnosis impression is provided as a suggested diagnosis impression to the user; when the template diagnosis impression cannot be generated, information missing in the generation of the structured report template is analyzed and provided to the user.

[0012] The application also provides an imaging text report quality control system based on an LLM and a structured report, including: a report acquisition unit configured to acquire an imaging text report to be audited, the imaging text report including: imaging manifestations and a diagnosis impression; a diagnosis conclusion extraction unit configured to extract the first N diagnosis conclusions indicating a certain or suspected diagnosis from the diagnosis impression; a structured report template matching unit configured to query a corresponding structured report template according to a scanning device type, a scanning site, and a diagnosis conclusion; an imaging feature extraction unit configured to input the extracted diagnosis conclusion and the imaging manifestations into an LLM model to extract imaging feature descriptions related to the diagnosis conclusion from the imaging manifestations; a structured data analysis unit configured to construct prompt words according to the structured report template, input the imaging feature descriptions into the LLM model, and obtain structured data of the imaging feature descriptions according to the prompt words; a structured data filling unit configured to invoke an interface of the structured report template based on the structured data of the imaging feature descriptions, fill the structured data into the structured report template to obtain a template diagnosis impression; and an audit unit configured to compare the template diagnosis impression with a diagnosis impression of the imaging text report to obtain an audit result.

[0013] In a possible implementation, the structured data analysis unit includes: a prompt word construction unit configured to construct prompt words according to information to be filled in the structured report template; an information extraction unit configured to extract the information to be filled in the structured report template from the imaging feature descriptions as discrete information through the prompt words; and a structured data composition unit configured to compose the structured data of the imaging feature descriptions based on the discrete information extracted from the imaging feature descriptions.

[0014] In a possible implementation, the structured data filling unit comprises: a template calling unit configured to call an interface of a structured report template, the structured report template comprising a plurality of information to be filled; a template filling unit configured to fill discrete information in the structured data into corresponding positions of the structured report template; and a template diagnosis unit configured to execute built-in judgment rules of the structured report template and output a template diagnosis impression.

[0015] In a possible implementation, the auditing unit is specifically configured to: when the template diagnosis impression is consistent with a diagnosis impression of the image text report, the image text report to be audited passes the quality control audit; when the template diagnosis impression is inconsistent with the diagnosis impression of the image text report, the image text report to be audited fails the quality control audit, and the template diagnosis impression is provided as a suggested diagnosis impression to a user; and when the template diagnosis impression cannot be generated, information missing in the analysis of the structured report template is provided to the user.

[0016] The application further provides a computer storage medium comprising computer instructions, which, when executed on a mobile terminal, cause the mobile terminal to perform the image text report quality control method based on LLM and structured report as described above.

[0017] The application further provides a computer program product, which, when executed on a mobile terminal, causes the mobile terminal to perform the image text report quality control method based on LLM and structured report as described above.

[0018] Compared with the prior art, the application has the following beneficial effects: first, the LLM model is used to extract the imaging feature description related to the diagnosis conclusion from the image performance of the text report, and the LLM model can more accurately extract the key field compared with the traditional NLP technology; second, the diagnosis logic of the structured report is used to fill the extracted imaging feature description into the structured report template after the imaging feature description is disassembled into structured data to generate a template diagnosis conclusion, and the text report is quality controlled by comparing the template diagnosis conclusion with the current diagnosis conclusion, so that the judgment on the diagnosis logic of the text report can be realized. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and constitute a part of this application, do not constitute limitations to the embodiments of the application. In the drawings: Figure 1 A flowchart of the image text report quality control method based on LLM and structured report provided by the embodiments of the application; Figure 2 A structural diagram of the image text report quality control system based on LLM and structured report provided by the embodiments of the application. DETAILED DESCRIPTION

[0020] Hereinafter, the term "include" or "may include" used in various embodiments of the present application indicates existence of the applied function, operation, or element, and does not limit one or more functions, operations, or elements to be added. Also, as used in various embodiments of the present application, the term "include" or "have" and variations thereof are merely intended to denote presence of the stated features, numbers, steps, operations, elements, components, or combinations thereof, and are not intended to exclude presence or addition of one or more other features, numbers, steps, operations, elements, components, or combinations thereof.

[0021] In various embodiments of the present application, the expression "or" or "at least one of A, B, and C" includes any combination of the listed terms or all the terms. For example, the expression "A or B" or "at least one of A or B" can include A, B, or both A and B.

[0022] The terms used in the various embodiments of the present application are used only to describe specific embodiments and are not intended to limit the various embodiments of the present application. As used herein, the singular forms are intended to include the plural forms as well, unless the context clearly indicates otherwise. All the terms used herein, including technical terms and scientific terms, have the same meanings as those generally understood by those skilled in the art to which various embodiments of the present application belong. The terms, such as those defined in a generally used dictionary, are to be interpreted as having the same meanings as those in the context of relevant related technology and are not to be interpreted as having ideal or excessively formal meanings, unless clearly defined in various embodiments of the present application.

[0023] In order to make the purposes, technical solutions, and advantages of the present application more clear, further specific description will be made to the present application with reference to the embodiments and the accompanying drawings, and the schematic embodiments of the present application and the description thereof are only used to explain the present application, and do not limit the present application.

[0024] Please refer to Figure 1 as shown, Figure 1The flowchart of the image text report quality control method based on LLM and structured report provided by the embodiments of the present application. The method comprises: S1, obtaining an image text report to be audited, the image text report comprising: image performance and diagnostic impression; S2, extracting the first N diagnostic conclusions indicating certainty or doubt from the diagnostic impression; S3, querying the corresponding structured report template according to the scanning device type, scanning site and diagnostic conclusion; S4, inputting the extracted diagnostic conclusion and image performance into the LLM model to extract the imaging feature description related to the diagnostic conclusion from the image performance; S5, constructing a prompt word according to the structured report template, inputting the imaging feature description into the LLM model, and obtaining the structured data of the imaging feature description according to the prompt word; S6, calling the interface of the structured report template based on the structured data of the imaging feature description, and filling the structured data into the structured report template to obtain a template diagnostic impression; and S7, comparing the template diagnostic impression with the diagnostic impression of the image text report to obtain an audit result.

[0025] Specifically, a doctor directly writes an image text report using a text editor. During quality inspection and auditing, the key diagnostic conclusions are first extracted from the diagnostic impression of the text report, the imaging feature description related to the diagnostic conclusions is extracted from the text report through the LLM model, and the related imaging feature description is converted into acceptable structured data of the structured report template. The structured data is automatically filled into the structured report template by calling the template interface, and the template diagnostic conclusion can be obtained by the structured report template through the built-in judgment logic. By comparing the template diagnostic conclusion with the diagnostic conclusion of the image text report, the report quality control auditing can be realized.

[0026] The improvement of the present application is that the LLM model trained based on the disease corpus first extracts the imaging feature description related to the diagnostic conclusion from the image performance of the text report. Compared with the traditional NLP technology, the LLM model can more accurately extract the key fields. Secondly, based on the diagnostic logic of the structured report, the extracted imaging feature description is disassembled into structured data and filled into the structured report template to generate a template diagnostic conclusion. By comparing the template diagnostic conclusion with the current diagnostic conclusion, the text report can be quality controlled, and the judgment of the diagnostic logic of the text report can be realized.

[0027] It should be noted that the method of the present application depends on the structured report template and the LLM model. As for the structured report template, the standard description of the image performance and the reasoning logic of the diagnostic conclusion are set therein, which is the basis for the diagnostic logic judgment of the present application. The structured report template has been used in the existing structured report system, and only needs to input the image performance and diagnostic conclusion by clicking. Here, no further description is given.

[0028] As for the LLM model, it matches the diagnostic conclusion and the impact, which is the basis for the application to accurately extract the image manifestations related to the diagnostic conclusion. As for the selection of the LLM model, in actual testing, the online GPT4 extraction effect is the best, and the domestically deployed domestic GLM4-6B can also fully meet the demand. As for the training of the LLM model, for different anatomical sites and different diseases, the LLM model can be trained through professional term corpus to realize accurate matching of image manifestations and diagnostic conclusions. There are many related papers and tests, which will not be repeated here.

[0029] Step S1 is to obtain an image text report to be audited, and the image text report includes: image manifestations and diagnostic impressions. For example: “Image manifestations: Prostate diameter: 5.9 cm x 2.9 cm x 3.0 cm (left-right x front-back x up-down) x 0.52 = 26.69 ml; Prostate transitional zone hyperplasia, peripheral zone shows mild abnormal signal, signal distribution is uneven; Prostate capsule is complete; bilateral nerve and vascular bundles show no abnormalities; Bilateral seminal vesicles show no obvious abnormalities; No abnormalities in the bladder; No abnormalities in the rectum; No enlarged lymph nodes in the scanning range; No bone destruction in the scanning range; Right seminal vesicle shows bleeding; Diagnostic impression: Inflammatory changes in the peripheral zone of the prostate (PIRADS 2); Mild prostate hyperplasia; Possible bleeding in the right seminal vesicle”.

[0030] Step S2 is to extract the first N diagnostic conclusions indicating a certain or suspected diagnosis from the diagnostic impression. Specifically, N diagnostic conclusions can be extracted from the diagnostic impression by setting judgment rules. Or according to the type of scanning equipment, scanning site, select the LLM model in this scene to extract N diagnostic conclusions from the diagnostic impression. The N diagnostic conclusions mentioned here refer to the information that the doctor will give when writing the diagnostic impression, such as certain diagnostic conclusions, suspected diagnostic conclusions, negative diagnostic conclusions, etc. The diagnostic process is a reasoning and guessing process, and often cannot simply rely on image manifestations to give a completely certain diagnostic conclusion. According to the writing habit of the image text report, the importance of these diagnostic conclusions is arranged according to the order of appearance in the diagnostic impression, and the earlier the more important.

[0031] Therefore, when extracting the diagnostic conclusion from the diagnostic impression, the diagnostic conclusion indicating negation is excluded first, and only the diagnostic conclusion indicating certainty and doubt is left. This step can be completed by an LLM model or by setting a judgment rule. Certain imaging examinations are performed to confirm or exclude certain doubts, so the diagnostic conclusion indicating negation is also common, and it is not necessary to perform quality control analysis on these excluded diagnoses.

[0032] Preferably, N = 2. When the diagnostic conclusion involved in the text report is greater than 2, only the first 2 key diagnostic conclusions can be extracted for quality control. Because other secondary diagnostic conclusions often lack sufficient description in the imaging findings, lack the necessary elements required for quality control in this application, not only the clinical significance is limited, but also a large number of prompts requiring additional information can be generated, reducing work efficiency.

[0033] For example, setting a judgment rule to extract the first 2 diagnostic conclusions indicating certainty or doubt from the diagnostic impression of the above image text report, we get: “Peripheral zone inflammatory changes of the prostate (PIRADS 2). Mild prostatic hyperplasia”.

[0034] Step S3 is to query the corresponding structured report template according to the scan device type, scan site and diagnostic conclusion. Specifically, according to the 2 diagnostic conclusions extracted above and the scan device type and scan site, query Table 1 to determine whether there is a corresponding structured report template to support quality control. If there is, go to the next step, if not, give up the quality control of this diagnostic conclusion.

[0035] Table 1. Structured report templates corresponding to scan device type, scan site and diagnostic conclusion

[0036] For the 2 diagnostic conclusions extracted above: “Peripheral zone inflammatory changes of the prostate (PIRADS 2)”, “Mild prostatic hyperplasia”, there are corresponding structured report templates, which can enter the next step of processing.

[0037] Step S4 is to input the extracted diagnostic conclusion and imaging findings into the LLM model to extract the imaging features related to the diagnostic conclusion from the imaging findings.

[0038] For example, according to the 2 diagnostic conclusions extracted above, the imaging features related to the diagnostic conclusion are extracted from the imaging findings: “Prompt: Please extract the relevant description from the following imaging findings according to the diagnostic impression: Diagnostic impression: Peripheral zone inflammatory changes of the prostate (PIRADS 2); Influencing performance: Prostate diameter: 5.9 cm x 2.9 cm x 3.0 cm (left-right x anteroposterior x superior-inferior) x 0.52 = 26.69 ml; Prostatic transition zone hyperplasia, peripheral zone shows mild abnormal signal, signal distribution is uneven; Prostate capsule is complete; bilateral neurovascular bundles show no abnormalities; Bilateral seminal vesicles show no obvious abnormalities; No abnormalities in the bladder; No abnormalities in the rectum; No enlarged lymph nodes in the scanning range; No bone destruction in the scanning range; Hemorrhage is visible in the right seminal vesicle; Model answer: Peripheral zone shows mild abnormal signal, signal distribution is uneven; Impression of diagnosis: Mild prostatic hyperplasia; Influencing performance: Prostate diameter: 5.9 cm x 2.9 cm x 3.0 cm (left-right x anteroposterior x superior-inferior) x 0.52 = 26.69 ml; Prostatic transition zone hyperplasia, peripheral zone shows mild abnormal signal, signal distribution is uneven; Prostate capsule is complete. Bilateral neurovascular bundles show no abnormalities; Bilateral seminal vesicles show no obvious abnormalities; No abnormalities in the bladder; No abnormalities in the rectum; No enlarged lymph nodes in the scanning range; No bone destruction in the scanning range; Hemorrhage is visible in the right seminal vesicle; Model answer: Prostatic transition zone hyperplasia”.

[0039] Step S5 is to construct a prompt word according to the structured report template, input the imaging feature description into the LLM model, and obtain the structured data of the imaging feature description according to the prompt word. In one possible implementation, step S5 includes: constructing a prompt word according to the information required to be filled in the structured report template; extracting the information required to be filled in the structured report template from the imaging feature description as discrete information through the prompt word; and composing the structured data of the imaging feature description based on the discrete information extracted from the imaging feature description.

[0040] Specifically, the prompt word of the LLM model is designed using the corresponding structured report template, and the number, location, size, shape, edge, content and other imaging manifestations of the lesion are extracted according to the filling requirements of the structured report.

[0041] For example, construct the prompt words according to the structured report template, and parse the aforementioned extracted 2 imaging feature descriptions into structured data: "Prompt word 1: ```json { "Peripheral zone abnormal signal": ["Yes", "No"], "Degree": ["Mild", "Moderate", "Obvious"], "Signal distribution": ["Uniform", "Non-uniform"] } ``` Please refer to the above structured mode to structure the text For example Text that needs to be structured ```text There is an obvious abnormal signal in the prostate, and the distribution is non-uniform. ``` Structured information ```json { "Peripheral zone abnormal signal": "Yes", "Degree": "Obvious", "Signal distribution": "Non-uniform" } ``` The following is the text that needs to be structured ```text The peripheral zone shows mild abnormal signals, and the signal distribution is non-uniform. ``` Structured information Model answer information ```json { "Prostate abnormal signal": "Yes", "Degree": "Mild", "Signal distribution": "Non-uniform" } ``` Prompt word 2: ```json { "Prostate hyperplasia": ["Yes", "No"], "Degree": ["Mild", "Moderate", "Obvious"] } ``` Please refer to the above structured mode to structure the text For example Text that needs to be structured ```text Significantly abnormal signal of prostate, uneven distribution; ``` Structured information ```json { "Abnormal signal of prostate": "Yes", "Degree": "Significant", "Signal distribution": "Uneven" } ``` The following text needs to be structured ```text Mild prostatic hyperplasia; ``` Structured information Information answered by the model ```json { "Prostatic hyperplasia": "Yes", "Degree": "Mild" } ```".

[0042] Step S6 is an interface that calls a structured report template based on the structured data described by the imaging features, fills the structured data into the structured report template to obtain a template diagnosis impression. In one possible implementation, step S6 includes: calling an interface of a structured report template, the structured report template containing multiple information to be filled; filling discrete information in the structured data into the corresponding position of the structured report template; the structured report template executes built-in judgment rules to output a template diagnosis impression.

[0043] For example, calling the interface of the structured report template, filling the two structured data obtained above into the structured report template, and obtaining a template diagnosis impression.

[0044] "Input 1: ```json { "Abnormal signal of peripheral zone": "Yes", "Degree": "Significant", "Signal distribution": "Uneven" } ``` Output 1: ```text Inflammatory changes in the peripheral zone of the prostate (PIRADS 2 points); ``` Input 2: ```json { "prostatic hyperplasia": "yes", "degree": "mild" } ``` Output 2: ```text mild prostatic hyperplasia.

[0045] Step S7 is to compare the template diagnosis impression with the diagnosis impression of the image text report to obtain an audit result. In one possible implementation, step S7 includes: when the template diagnosis impression is consistent with the diagnosis impression of the image text report, the image text report to be audited passes the quality control audit; when the template diagnosis impression is inconsistent with the diagnosis impression of the image text report, the image text report to be audited fails the quality control audit, and the template diagnosis impression is provided as a suggested diagnosis impression and provided to the user; and when the template diagnosis impression cannot be generated, the missing information for generating the structured report template is analyzed and provided to the user.

[0046] It can be understood that the image text report quality control method based on LLM and structured report provided by the present application automatically generates a diagnosis conclusion by using an LLM model and the built-in logic of a structured report, and controls the quality of the text report by comparing whether the automatically generated diagnosis conclusion and the manually written diagnosis conclusion are the same. This quality control method relies on the personalized prompt word engineering of the LLM and the underlying diagnosis logic of the structured report, and can reasonably control the quality of the image diagnosis logic, improve the diagnosis efficiency, effectively track the quality and data mining, and be easily promoted. Since the text report is still the mainstream form of the radiology diagnosis report at present and in the future for a period of time, the promotion of this quality control mode has a positive significance for improving the diagnosis quality.

[0047] Please refer to Figure 2 as shown, Figure 2A structural diagram of an image text report quality control system based on LLM and structured report provided by an embodiment of the present application. The system comprises: a report acquisition unit configured to acquire an image text report to be audited, the image text report comprising: image manifestations and diagnostic impressions; a diagnostic conclusion extraction unit configured to extract the first N diagnostic conclusions indicating certain or suspected diagnoses from the diagnostic impressions; a structured report template matching unit configured to query a corresponding structured report template according to a scanning device type, a scanning site and the diagnostic conclusions; an imaging feature extraction unit configured to input the extracted diagnostic conclusions and the image manifestations into an LLM model, and extract imaging feature descriptions related to the diagnostic conclusions from the image manifestations; a structured data analysis unit configured to construct prompt words according to the structured report template, input the imaging feature descriptions into the LLM model, and obtain structured data of the imaging feature descriptions according to the prompt words; a structured data filling unit configured to call an interface of the structured report template based on the structured data of the imaging feature descriptions, and fill the structured data into the structured report template to obtain a template diagnostic impression; and an auditing unit configured to compare the template diagnostic impression with the diagnostic impression of the image text report, and obtain an auditing result.

[0048] In a possible implementation, the structured data analysis unit comprises: a prompt word construction unit configured to construct prompt words according to information required to be filled in the structured report template; an information extraction unit configured to extract information required to be filled in the structured report template from the imaging feature descriptions as discrete information through the prompt words; and a structured data composition unit configured to compose structured data of the imaging feature descriptions based on the discrete information extracted from the imaging feature descriptions.

[0049] In a possible implementation, the structured data filling unit comprises: a template calling unit configured to call an interface of the structured report template, the structured report template comprising a plurality of information required to be filled in; a template filling unit configured to fill the discrete information in the structured data into positions corresponding to the structured report template; and a template diagnosis unit configured to execute built-in judgment rules of the structured report template, and output a template diagnostic impression.

[0050] In a possible implementation, the auditing unit is specifically configured to: when the template diagnostic impression is consistent with the diagnostic impression of the image text report, the image text report to be audited passes the quality control audit; when the template diagnostic impression is inconsistent with the diagnostic impression of the image text report, the image text report to be audited fails the quality control audit, and the template diagnostic impression is provided as a suggested diagnostic impression to a user; and when the template diagnostic impression cannot be generated, information missing in the generation of the structured report template is analyzed and provided to the user.

[0051] Specifically, the doctor uses a text editor to write an image text report, clicks a quality control button when the image text report is edited, and the image text report quality control system performs the above quality control method on the image text report to obtain an audit result. According to the diagnostic impression, the image performance and the matched structured report template in the image text report, it is judged whether the quality inspection audit is passed. For the image text report that does not pass, a suggested diagnostic impression is given, or the image performance that is missing for generating the diagnostic impression is prompted.

[0052] It can be understood that the image text report quality control system based on LLM and structured report provided by the present application is used to implement the above-mentioned image text report quality control method based on LLM and structured report, and corresponds to the method, has the corresponding technical effect, so more details are not added.

[0053] The embodiment of the present application further provides a computer storage medium, including computer instructions, when the computer instructions run on the mobile terminal, make the mobile terminal execute the image text report quality control method based on LLM and structured report as described above.

[0054] The embodiment of the present application further provides a computer program product, when the computer program product runs on the mobile terminal, makes the mobile terminal execute the image text report quality control method based on LLM and structured report as described above.

[0055] The above specific embodiments, the purpose, technical scheme and beneficial effects of the present application are further described in detail, it should be understood that the above is only the specific embodiment of the present application, and is not used to limit the protection scope of the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. An image text report quality control method based on LLM and structured report, characterized in that, The method comprises the following steps: obtaining an image text report to be audited, wherein the image text report comprises an image manifestation and a diagnostic impression; extracting the first N diagnostic conclusions representing certain or suspected diagnostic conclusions from the diagnostic impression; querying a corresponding structured report template according to a scanning device type, a scanning site and the diagnostic conclusions; inputting the extracted diagnostic conclusions and the image manifestation into an LLM model to extract imaging feature descriptions related to the diagnostic conclusions from the image manifestation; constructing prompt words according to the structured report template, inputting the imaging feature descriptions into the LLM model, and obtaining structured data of the imaging feature descriptions according to the prompt words; calling an interface of the structured report template based on the structured data of the imaging feature descriptions, filling the structured data into the structured report template to obtain a template diagnostic impression; comparing the template diagnostic impression with the diagnostic impression of the image text report to obtain an audit result.

2. The method of claim 1, wherein, constructing prompt words according to the structured report template, inputting the imaging feature descriptions into the LLM model, and obtaining structured data of the imaging feature descriptions according to the prompt words; comprising: constructing prompt words according to information required to be filled in the structured report template; extracting information required to be filled in the structured report template from the imaging feature descriptions as discrete information through the prompt words; composing structured data of the imaging feature descriptions based on the discrete information extracted from the imaging feature descriptions.

3. The method of claim 1, wherein the structured report is a radiology report. calling an interface of the structured report template based on the structured data of the imaging feature descriptions, filling the structured data into the structured report template to obtain a template diagnostic impression; comprising: calling the interface of the structured report template, wherein the structured report template comprises a plurality of information required to be filled in; filling the discrete information in the structured data into a corresponding position of the structured report template; the structured report template executing built-in judgment rules to output the template diagnostic impression.

4. The method of claim 1, wherein, comparing the template diagnostic impression with the diagnostic impression of the image text report to obtain an audit result; comprising: when the template diagnostic impression is consistent with the diagnostic impression of the image text report, the image text report to be audited passes the quality inspection and audit; when the template diagnostic impression is inconsistent with the diagnostic impression of the image text report, the image text report to be audited fails the quality inspection and audit, and the template diagnostic impression is provided as a suggested diagnostic impression to a user; when the template diagnostic impression cannot be generated, information missing in the generation of the structured report template is analyzed and provided to the user.

5. An image text report quality control system based on LLM and structured report, characterized in that, The method comprises the following steps: a report obtaining unit is configured to obtain an image text report to be audited, wherein the image text report comprises an image manifestation and a diagnostic impression; a diagnostic conclusion extracting unit is configured to extract the first N diagnostic conclusions representing certain or suspected diagnostic conclusions from the diagnostic impression; a structured report template matching unit is configured to query a corresponding structured report template according to a scanning device type, a scanning site and the diagnostic conclusions; an imaging feature extracting unit is configured to input the extracted diagnostic conclusions and the image manifestation into an LLM model to extract imaging feature descriptions related to the diagnostic conclusions from the image manifestation; a structured data analyzing unit is configured to construct prompt words according to the structured report template, input the imaging feature descriptions into the LLM model, and obtain structured data of the imaging feature descriptions according to the prompt words; The structured data filling unit fills the structured data into the structured report template to obtain a template diagnosis impression. The auditing unit compares the template diagnosis impression with a diagnosis impression of the image text report to obtain an auditing result.

6. The LLM and structured reporting based image text report quality control system of claim 5, wherein, The structured data analysis unit comprises: A prompt word construction unit constructs prompt words according to information required to be filled in the structured report template; An information extraction unit extracts the information required to be filled in the structured report template from the imaging feature description as discrete information through the prompt words; A structured data composition unit composes the structured data of the imaging feature description based on the discrete information extracted from the imaging feature description.

7. The LLM and structured reporting based image text reporting quality control system of claim 5, wherein, The structured data filling unit comprises: A template calling unit calls an interface of the structured report template, and the structured report template comprises a plurality of information required to be filled in; A template filling unit fills the discrete information in the structured data into a corresponding position of the structured report template; A template diagnosis unit executes built-in judgment rules of the structured report template to output a template diagnosis impression.

8. The LLM and structured reporting based image text reporting quality control system of claim 5, wherein, The auditing unit is specifically configured to: When the template diagnosis impression is consistent with the diagnosis impression of the image text report, the image text report to be audited passes the quality inspection and auditing; When the template diagnosis impression is inconsistent with the diagnosis impression of the image text report, the image text report to be audited fails the quality inspection and auditing, and the template diagnosis impression is provided as a suggested diagnosis impression to the user; When the template diagnosis impression cannot be generated, information missing in the generation of the structured report template is analyzed and provided to the user.

9. A computer storage medium, characterized in that The computer program product comprises computer instructions, and when the computer instructions run on the mobile terminal, the mobile terminal executes the method according to any one of claims 1 to 4.

10. A computer program product, characterised in that, The computer program product makes the mobile terminal execute the method according to any one of claims 1 to 4 when the computer program product runs on the mobile terminal.

Images