A large model-based medical image conclusion diagnosis method and device

Abstract

This invention relates to the field of medical information technology, and in particular to a method and apparatus for diagnosing medical images based on a large model. The method includes: acquiring the medical image to be diagnosed, image description, and preliminary diagnostic information; extracting image type, examination site, image findings, and preliminary report to generate examination task data; calling predefined analysis instructions and medical image report specifications based on the examination task data to generate structured prompt information; inputting the structured prompt information, image description, and preliminary diagnostic information into a large language model to generate a preliminary diagnostic conclusion; parsing the preliminary diagnostic conclusion to generate a unified feature list and performing consistency checks to generate consistent items, potential discrepancies, supplementary findings, and a difference analysis report; and generating modification suggestions, review prompts, and a final structured diagnostic conclusion based on the difference analysis report. This invention improves the standardization, traceability, and review efficiency of conclusion generation.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method and apparatus for diagnosing medical imaging conclusions based on a large model. Background Technology

[0002] In the field of medical information technology, existing solutions for medical image diagnosis typically acquire the medical image to be diagnosed, image description, and preliminary diagnostic information, and then generate a preliminary diagnostic conclusion or report using a large language model pre-trained on medical text. However, this approach suffers from limitations such as unclear correspondence between image findings and diagnostic conclusions, unstable connection between diagnostic evidence and differential diagnosis, and lack of source location records for preliminary diagnostic conclusions. Existing methods largely rely on the direct input of image descriptions or preliminary diagnostic information, with the large language model outputting the diagnostic conclusion.

[0003] Under the constraints of medical imaging report standards, preliminary diagnostic conclusions are prone to problems such as the lack of correlation between the diagnostic conclusion and the diagnostic basis, and the absence of direct corresponding content for supplementary findings, making it difficult to meet the requirement of consistently generating final structured diagnostic conclusions. If the preliminary diagnostic conclusion is directly written into the preliminary report, the differences between the image description, preliminary diagnostic information, and diagnostic conclusion are not easily separated, and it is also difficult to establish a correspondence between modification suggestions and review prompts and their specific source locations.

[0004] For the joint processing of unified feature lists, image descriptions, preliminary diagnostic information, predefined analysis instructions, and medical image reporting standards, existing technologies generally lack a classification process for consistent items, potential discrepancies, and supplementary findings. They also lack a continuous link for generating final structured diagnostic conclusions based on differential analysis reports. This makes it difficult to form a consistent process of acquisition, generation, parsing, verification, and output in medical image diagnosis scenarios, resulting in an unclear processing flow between preliminary diagnostic conclusions and final structured diagnostic conclusions. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a medical imaging diagnosis method based on a large model, comprising:

[0006] S100: Obtain the medical images to be diagnosed, image descriptions, and preliminary diagnostic information; extract image type, examination site, image findings, and preliminary report; and generate examination task data.

[0007] S200: Based on the examination task data, call predefined analysis instructions and medical image report specifications to generate structured prompt information;

[0008] The structured prompt information includes fields for image discovery, diagnosis conclusion, diagnosis basis, differential diagnosis, insufficient basis prompt, and report generation result format.

[0009] S300. Input the structured prompt information, image description and preliminary diagnosis information into a large language model pre-trained for medical text to generate a preliminary diagnosis conclusion;

[0010] S400: Analyze the preliminary diagnostic conclusions and generate a unified feature list;

[0011] S500 performs consistency checks based on a unified feature list, image description, preliminary diagnostic information, predefined analysis instructions, and medical image reporting standards, generating consistent items, potential discrepancies, supplementary findings, and a discrepancy analysis report.

[0012] S600 generates modification suggestions, review prompts, and final structured diagnostic conclusions based on the differential analysis report.

[0013] Furthermore, in S100, the image description and preliminary diagnostic information are correlated, extracted, and conflict resolved to obtain the image findings, examination sites, lesion areas, diagnostic basis, differential diagnosis, and recommendations, which are then written into the examination task data.

[0014] Furthermore, according to the method described in claim 1, in S200, based on the medical image report specifications matching the image type and the examination site, a correspondence is established between the image discovery field, the diagnostic conclusion field, the diagnostic basis field, and the differential diagnosis field, and the correspondence is written into the structured prompt information.

[0015] Furthermore, in S300, the large language model pre-trained on medical text outputs image findings, diagnostic conclusions, diagnostic basis, differential diagnosis, and recommendations in segments according to the report generation result format fields, forming a preliminary diagnostic conclusion.

[0016] Furthermore, in S400, the analysis of the preliminary diagnostic conclusion includes extracting imaging findings, examination sites, lesion areas, diagnostic conclusions, diagnostic basis, differential diagnosis, and recommendations, recording the source position of each content in the preliminary diagnostic conclusion, and forming a unified feature list.

[0017] Furthermore, in S500, when the content of the unified feature list matches the image description, preliminary diagnostic information, predefined analysis instructions, and medical image reporting specifications, the content is classified as a consistent item.

[0018] Furthermore, in S500, when the content of the unified feature list is inconsistent with the image description or preliminary diagnostic information in any field such as examination site, lesion area, image findings, diagnostic basis, differential diagnosis, or recommendation content, the content is classified as a potential discrepancy item.

[0019] Furthermore, in S500, when the content of the unified feature list conforms to the medical image reporting specifications but does not form a direct corresponding content in the image description and preliminary diagnostic information, the content is classified as a supplementary discovery item; when the diagnostic conclusion is not associated with the diagnostic basis, an insufficient basis item is generated.

[0020] Furthermore, in S600, diagnostic conclusions and diagnostic basis are retained based on consistent items, modification suggestions or review prompts are generated based on potential discrepancies, supplementary discovery prompts are generated based on supplementary discovery items, and in cases where the basis is insufficient, the diagnostic conclusion is adjusted to a tentative conclusion, a suspected conclusion, a conclusion that cannot be ruled out, or a conclusion that suggests combining clinical findings.

[0021] Furthermore, a medical imaging conclusion diagnostic device based on a large model, applied to any of the methods described above, includes: a data acquisition module, a prompt information generation module, a conclusion generation module, a list generation module, a consistency verification module, a difference analysis module, and a structured output module.

[0022] The key innovations of this invention include:

[0023] (1) The preliminary diagnostic conclusion is parsed into a unified feature list, and a consistency test is performed based on the unified feature list, image description, preliminary diagnostic information, predefined analysis instructions and medical image reporting standards to generate consistent items, potential difference items, supplementary findings items and difference analysis reports.

[0024] (2) Based on the difference analysis report, generate modification suggestions, review prompts and final structured diagnostic conclusions, so that consistent items, potential difference items and supplementary findings can enter the corresponding processing procedures.

[0025] (3) Based on the examination task data, call the predefined analysis instructions and medical image report specifications to generate structured prompt information, and input the structured prompt information, image description and preliminary diagnosis information into the medical text pre-trained big language model to generate preliminary diagnosis conclusions.

[0026] The following are its main beneficial effects:

[0027] (1) To address the problem of unclear correspondence between preliminary diagnostic conclusions and image descriptions and preliminary diagnostic information, a consistency test is conducted through a unified feature list, so that image findings, diagnostic conclusions, diagnostic basis and differential diagnosis enter the same test link, forming consistent items, potential difference items and supplementary findings.

[0028] (2) To address the problem that the differences are not easily separated, the consistency test results are adopted through the difference analysis report, so that potential differences can be included in the modification suggestions and review prompts, and supplementary findings can be included in the processing process before the final structured diagnostic conclusion.

[0029] (3) To address the problem of unclear field relationships when a large language model pre-trained in medical text directly generates diagnostic conclusions, structured prompts are used to limit the image discovery field, diagnostic conclusion field, diagnostic basis field, differential diagnosis field, insufficient basis prompt field, and report generation result format field, so that the preliminary diagnostic conclusions have a parsable field source.

[0030] (4) To address the issue that the diagnostic conclusions are not associated with the diagnostic basis, potential difference items and difference analysis reports are generated through consistency tests, and modification suggestions and review prompts are formed in the final structured diagnostic conclusions, so that such diagnostic conclusions enter the corresponding processing process before output.

[0031] (5) To address the issue of unclear processing between preliminary diagnostic conclusions and final structured diagnostic conclusions in existing schemes, a closed-loop processing process for medical imaging conclusion diagnosis is formed by checking the continuous link of task data, structured prompt information, preliminary diagnostic conclusions, unified feature list, differential analysis report and final structured diagnostic conclusions. Attached Figure Description

[0032] Figure 1 A flowchart illustrating a medical imaging diagnosis method based on a large model, provided for an embodiment of this application;

[0033] Figure 2 This is a structural block diagram of a medical imaging diagnostic device based on a large model, provided in an embodiment of this application. Detailed Implementation

[0034] Example 1: Refer to Figure 1 This is a flowchart illustrating a medical imaging diagnosis method based on a large model, provided by an embodiment of the present invention. The process may include at least steps S100-S600:

[0035] S100: Obtain the medical images to be diagnosed, image descriptions, and preliminary diagnostic information; extract image type, examination site, image findings, and preliminary report; and generate examination task data.

[0036] S200: Based on the examination task data, call the predefined analysis instructions and medical image report specifications to generate structured prompt information. The structured prompt information includes image discovery field, diagnosis conclusion field, diagnosis basis field, differential diagnosis field, insufficient basis prompt field, and report generation result format field.

[0037] S300: Input structured prompts, image descriptions, and preliminary diagnostic information into a large language model pre-trained with medical text to generate a preliminary diagnostic conclusion;

[0038] S400: Analyze the preliminary diagnostic conclusions and generate a unified feature list;

[0039] S500 performs consistency checks based on a unified feature list, image description, preliminary diagnostic information, predefined analysis instructions, and medical image reporting standards, generating consistent items, potential discrepancies, supplementary findings, and a discrepancy analysis report.

[0040] S600 generates modification suggestions, review prompts, and final structured diagnostic conclusions based on the differential analysis report.

[0041] S100: Acquire the medical images to be diagnosed, image descriptions, and preliminary diagnostic information; extract image type, examination site, image findings, and preliminary report; and generate examination task data.

[0042] The data acquisition module performs this step. Specifically, the data acquisition module initiates data acquisition when the image reporting system receives a new preliminary report, image description, or preliminary diagnostic information. The medical image to be diagnosed is the medical image data generated by the current examination, including the image type, examination site, and associated content of the lesion area. The image description is the image discovery text recorded in the preliminary report, including anatomical structures, lesion areas, image findings, and comparative descriptions of normal and abnormal image features. The preliminary diagnostic information is the diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations recorded in the preliminary report. The data acquisition module uses the medical image to be diagnosed, image description, and preliminary diagnostic information as input to the same examination task, establishes a correspondence between the three, and obtains the original input content for subsequent processing.

[0043] Specifically, the data acquisition module first reads the image type and examination site of the medical image to be diagnosed, and then reads the image findings in the image description. The image type represents the examination category of the current medical image, the examination site represents the anatomical structure range corresponding to the image description, and the image findings represent the lesion area, the comparative description of normal and abnormal image features, and related diagnostic evidence. When reading preliminary diagnostic information, the data acquisition module records the diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations separately. If the preliminary report already contains image description and preliminary diagnostic information, the data acquisition module extracts these contents from the preliminary report. If the image description originates from the image description generation module or the report interpretation result, the data acquisition module matches the image description with the medical image to be diagnosed and writes the matched image description into the associated location in the preliminary report.

[0044] Furthermore, the data acquisition module performs correlation extraction on image descriptions and preliminary diagnostic information. The correlation extraction process involves extracting the examination site, lesion area, and imaging findings from the image description; extracting the diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations from the preliminary diagnostic information; establishing a correspondence between imaging findings and diagnostic basis; and establishing a correspondence between the diagnostic conclusion and differential diagnosis. In practice, the data acquisition module reads the preliminary report according to the field order in the medical imaging report specifications: first, the examination site; then, the imaging findings; and finally, the diagnostic conclusion. For cases where multiple descriptions appear for the same lesion area, the data acquisition module merges the same examination site, the same lesion area, and the same imaging findings into a single imaging finding record. If a negative description exists in the image description, but a contradictory diagnostic conclusion exists in the preliminary diagnostic information, the data acquisition module writes this content into a conflict resolution record and retains the original text of the image description and preliminary diagnostic information.

[0045] Furthermore, the data acquisition module resolves conflicts between image descriptions and preliminary diagnostic information. This conflict resolution involves checking and marking inconsistencies between the examination site, lesion area, imaging findings, diagnostic basis, differential diagnosis, and recommendations. Specifically, when the examination site in the image description differs from that in the preliminary diagnostic information, the data acquisition module retains both texts and marks them as pending verification. When an imaging finding in the image description does not provide corresponding diagnostic basis in the preliminary diagnostic information, the data acquisition module marks that finding as insufficient evidence. When a diagnostic conclusion in the preliminary diagnostic information does not have a direct counterpart in the image description, the data acquisition module marks that conclusion as a supplementary finding. All these markings are written into the examination task data along with the original text, source location, and the associated preliminary report.

[0046] Understandably, the examination task data is the output of this step. The examination task data consists of the medical image to be diagnosed, image description, preliminary diagnostic information, image type, examination site, image findings, preliminary report, lesion area, diagnostic basis, differential diagnosis, suggested content, conflict resolution record, candidate content for insufficient evidence, and candidate content for supplementary findings. When generating the examination task data, the data acquisition module marks the source of each item, which includes the medical image to be diagnosed, image description, preliminary diagnostic information, and preliminary report. If the medical image to be diagnosed lacks an image type or examination site, the data acquisition module reads the corresponding content from the preliminary report; if the preliminary report lacks diagnostic basis, the data acquisition module sets the diagnostic basis field to an empty field and writes candidate content for insufficient evidence; if both the image description and preliminary diagnostic information lack an examination site, the data acquisition module generates a supplementary record and stops sending the examination task data to the next step.

[0047] In one engineering embodiment, after the image reporting system generates a preliminary report, the data acquisition module automatically reads the medical images to be diagnosed, image descriptions, and preliminary diagnostic information under the current examination task. The data acquisition module writes the examination site from the preliminary report into the examination site field, the lesion area and image findings from the image description into the image findings field, and the diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations from the preliminary diagnostic information into the corresponding fields. Subsequently, the data acquisition module performs correlation extraction and conflict resolution on the image description and preliminary diagnostic information to generate examination task data with source location. The examination task data is sent to S200, which uses predefined analysis instructions and medical image reporting specifications to generate structured prompt information.

[0048] The technical effect of this step can be summarized as follows: the examination task data incorporates the medical images to be diagnosed, image descriptions, and preliminary diagnostic information into the same examination task, and subsequent steps read content from the same source relationship. This step completes association extraction and conflict resolution before entering the large language model for processing, reducing field confusion before the preliminary diagnostic conclusion is generated. The examination task data output by this step follows the structured prompt information generation process of S200, forming the data foundation for subsequent consistency checks.

[0049] S200: Based on the examination task data, call the predefined analysis instructions and medical image report specifications to generate structured prompt information. The structured prompt information includes image discovery field, diagnosis conclusion field, diagnosis basis field, differential diagnosis field, insufficient basis prompt field, and report generation result format field.

[0050] The prompt message generation module performs this step. Specifically, the prompt message generation module receives the examination task data generated by S100 and reads the image type, examination site, image findings, preliminary report, lesion area, diagnostic basis, differential diagnosis, suggested content, conflict resolution record, insufficient basis candidate content, and supplementary finding candidate content from the examination task data. The predefined analysis instructions are pre-configured medical text processing instructions, including image finding reading rules, diagnostic conclusion generation rules, diagnostic basis association rules, differential diagnosis citation rules, insufficient basis record rules, and report generation result format rules. The medical image report specification is the report specification content stored in the image report system, including the correspondence between examination site, image findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and suggested content. After receiving the examination task data, the prompt message generation module calls the corresponding predefined analysis instructions according to the image type and examination site, and reads the medical image report specification that matches the image type and examination site.

[0051] Specifically, the prompt message generation module first reads fields from the examination task data. During field reading, the image discovery field receives descriptions of image findings, lesion areas, and comparisons of normal and abnormal image features from the examination task data; the diagnostic conclusion field receives the diagnostic conclusion from the preliminary diagnostic information; the diagnostic basis field receives the corresponding content between the image findings and the diagnostic basis; the differential diagnosis field receives the differential diagnosis and the corresponding image findings; the insufficient basis prompt field receives candidate content for insufficient basis and diagnostic conclusions that do not form a direct correspondence; and the report generation result format field receives the field order and segmented output requirements from the medical imaging report specifications. The prompt message generation module writes all the above fields into the same structured prompt message and retains the source position of each field in the examination task data.

[0052] Furthermore, the prompt message generation module matches the image type and examination site against medical imaging report standards. The matching process involves first comparing the image type with the examination category in the medical imaging report standards, and then comparing the examination site with the anatomical structure range in the medical imaging report standards. After the comparison is complete, the prompt message generation module reads the correspondence between the corresponding image discovery field, diagnostic conclusion field, diagnostic basis field, and differential diagnosis field. For image discoveries in the examination task data for which diagnostic basis already exists, the prompt message generation module writes the image discovery into the image discovery field and the corresponding diagnostic basis into the diagnostic basis field. For content in the examination task data that has a diagnostic conclusion but lacks diagnostic basis, the prompt message generation module writes the diagnostic conclusion into the diagnostic conclusion field and records the corresponding source location in the insufficient basis prompt field.

[0053] Furthermore, the prompt message generation module invokes predefined analysis instructions to process conflict resolution records in the examination task data. These conflict resolution records include inconsistencies in examination sites, lesion areas, imaging findings, diagnostic evidence, differential diagnosis, and recommendations. The prompt message generation module performs field mapping on these conflict resolution records. During field mapping, content related to imaging findings is written to the imaging findings field, content related to diagnostic conclusions is written to the diagnostic conclusion field, content related to diagnostic evidence is written to the diagnostic evidence field, and content related to differential diagnosis is written to the differential diagnosis field. For content that still lacks a corresponding relationship, the prompt message generation module writes an insufficient evidence prompt field and retains this content as input for consistency checks in S500.

[0054] Understandably, the structured prompt information is the output of this step. The structured prompt information includes fields for image discovery, diagnosis conclusion, diagnostic basis, differential diagnosis, insufficient basis prompt, and report generation result format. The image discovery field records the image findings, lesion area, and examination site in the image description; the diagnosis conclusion field records the diagnosis conclusion in the preliminary diagnosis information; the diagnostic basis field records the correspondence between the image findings and the diagnosis conclusion; the differential diagnosis field records the differential diagnoses related to the diagnosis conclusion; the insufficient basis prompt field records content lacking diagnostic basis or with incomplete source location; and the report generation result format field records the field order used by S300 when outputting the preliminary diagnosis conclusion in segments. The prompt information generation module sends the structured prompt information, image description, and preliminary diagnosis information to S300, allowing S300 to input the medical text pre-trained large language model and generate the preliminary diagnosis conclusion.

[0055] In one engineering embodiment, after the image reporting system completes S100, the prompt information generation module automatically reads the image type and examination site from the examination task data. If the image type is a chest computed tomography (CT) scan, and the CT scan first appears as "Computed Tomography," the prompt information generation module calls the medical image reporting standard corresponding to the chest CT scan. The prompt information generation module writes the lung image findings from the preliminary report into the image findings field, the diagnostic conclusion from the preliminary diagnosis information into the diagnosis conclusion field, the content in the image description that supports the diagnostic conclusion into the diagnostic basis field, and the content that needs to be excluded into the differential diagnosis field. If a diagnostic conclusion appears in the preliminary diagnosis information, but there is no corresponding diagnostic basis in the image description, the prompt information generation module writes the diagnostic conclusion into the insufficient basis prompt field. Subsequently, the prompt information generation module forms a structured prompt information according to the report generation result format field and sends the structured prompt information to the conclusion generation module to execute S300.

[0056] Summary of the technical effects of this step: This step transforms the examination task data generated by S100 into structured prompt information, ensuring that image findings, diagnostic conclusions, diagnostic evidence, and differential diagnoses are grouped into the same field relationship. The insufficient evidence prompt field retains content lacking diagnostic evidence, providing a source for consistency verification of S500. The report generation result format field provides fixed fields for the segmented output of the preliminary diagnostic conclusions from S300.

[0057] S300: Input structured prompts, image descriptions, and preliminary diagnostic information into a large language model pre-trained with medical text to generate a preliminary diagnostic conclusion;

[0058] The conclusion generation module performs this step. Specifically, the conclusion generation module receives structured prompt information sent by S200 and simultaneously receives image descriptions and preliminary diagnostic information generated by S100. The structured prompt information includes image discovery fields, diagnostic conclusion fields, diagnostic basis fields, differential diagnosis fields, insufficient basis prompt fields, and report generation result format fields. The large language model pre-trained by the medical text is a text generation model formed after training with medical text data, image sample reports, diagnostic conclusions, diagnostic basis, differential diagnoses, and suggestions. The large language model pre-trained by the medical text includes a text reading unit, a prompt information processing unit, a diagnostic reasoning unit, and a report generation unit. The text reading unit receives structured prompt information, image descriptions, and preliminary diagnostic information. The prompt information processing unit reads the content of each field. The diagnostic reasoning unit generates diagnostic conclusions, diagnostic basis, and differential diagnoses. The report generation unit outputs the preliminary diagnostic conclusions according to the report generation result format fields.

[0059] Specifically, the conclusion generation module first performs field validation on the structured prompt information. During field validation, the module reads the examination site, lesion area, and imaging findings from the image discovery field; the diagnostic conclusion from the diagnostic conclusion field; the diagnostic basis from the diagnostic basis field; the differential diagnosis from the differential diagnosis field; the insufficient basis prompt field; and the segmented output requirements from the report generation result format field. If the image discovery field, diagnostic conclusion field, and report generation result format field all contain content, the conclusion generation module concatenates the structured prompt information, image description, and preliminary diagnostic information into the model input content. If the diagnostic basis field is empty, the conclusion generation module retains the empty field state and adds the insufficient basis prompt field to the model input content. If the differential diagnosis field is empty, the conclusion generation module retains the empty field state of the differential diagnosis field and does not rewrite the missing content into a diagnostic conclusion.

[0060] Furthermore, the conclusion generation module feeds the model input content into a large language model pre-trained with medical text. This pre-trained large language model first has the text reading unit read the image findings and lesion regions from the image description, and then the prompt information processing unit reads the field relationships in the structured prompt information. Specifically, the prompt information processing unit associates the image findings field with the diagnostic basis field, the diagnostic conclusion field with the differential diagnosis field, and the insufficient basis prompt field with the diagnostic conclusion field. The diagnostic reasoning unit generates reasoning result information based on the above associations. The reasoning result information includes image findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations. The report generation unit, according to the report generation result format fields, segments the reasoning result information into a preliminary diagnostic conclusion.

[0061] Furthermore, the process of generating the preliminary diagnostic conclusion retains the source relationships. The conclusion generation module establishes a correspondence between each section of the preliminary diagnostic conclusion and the fields in the structured prompt information. The image discovery section corresponds to the image discovery field, the diagnostic conclusion section corresponds to the diagnostic conclusion field, the diagnostic basis section corresponds to the diagnostic basis field, the differential diagnosis section corresponds to the differential diagnosis field, and the suggestion content section corresponds to the report generation result format field. For content in the insufficient basis prompt field, the report generation unit writes it into the prompt position in the preliminary diagnostic conclusion and retains its source position. For content in the structured prompt information that contains conflict resolution records, the diagnostic reasoning unit does not directly delete the original text, but writes it as content to be parsed into the corresponding paragraph of the preliminary diagnostic conclusion, so that the S400 can extract the source position when parsing the preliminary diagnostic conclusion.

[0062] Understandably, the preliminary diagnostic conclusion is the output of this step. The preliminary diagnostic conclusion includes image findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations. It also includes prompts imported from the insufficient evidence prompt field. When outputting the preliminary diagnostic conclusion, the conclusion generation module records the model input content, the version of the structured prompt information, the report generation result format field, and the output time. If the large language model pre-trained in the medical text does not return complete image findings, diagnostic conclusion, diagnostic basis, and differential diagnosis, the conclusion generation module writes the missing fields into an anomaly record and sends the anomaly record along with the preliminary diagnostic conclusion to S400. S400 receives the preliminary diagnostic conclusion, parses it, and generates a unified feature list.

[0063] In one engineering embodiment, the conclusion generation module receives structured prompts, image descriptions, and preliminary diagnostic information corresponding to a chest computed tomography (CT) scan. The module reads the lung lesion region and image findings from the image findings field, the preliminary diagnostic conclusion from the diagnostic conclusion field, the corresponding content of the image findings from the diagnostic basis field, and the relevant content from the differential diagnosis field. A medical text pre-trained large language model outputs segmented text according to the report generation result format field. The first segment of the segmented text records the image findings, the second the diagnostic conclusion, the third the diagnostic basis, the fourth the differential diagnosis, and the fifth the suggested content. The conclusion generation module combines the above segmented text into a preliminary diagnostic conclusion and sends the preliminary diagnostic conclusion to the list generation module to execute S400.

[0064] Summary of the technical effects of this step: This step inputs structured prompts, image descriptions, and preliminary diagnostic information into a large language model pre-trained with medical text. The preliminary diagnostic conclusion is constrained by field relationships. This step preserves the source relationships of image discovery, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations, which can be directly parsed by S400. If insufficient prompts are used to enter the preliminary diagnostic conclusion, S500 can perform consistency checks on the relevant content.

[0065] S400: Analyze the preliminary diagnostic conclusions and generate a unified feature list;

[0066] The list generation module performs this step. Specifically, the list generation module receives the preliminary diagnostic conclusion generated by S300 and reads the structured prompt information, image description, and preliminary diagnostic information corresponding to the preliminary diagnostic conclusion. The preliminary diagnostic conclusion includes image findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations. The structured prompt information includes fields for image findings, diagnostic conclusion, diagnostic basis, differential diagnosis, insufficient basis prompt, and report generation result format. The list generation module consists of a text reading unit, a field processing unit, and a source location recording unit. The text reading unit reads the segmented text in the preliminary diagnostic conclusion. The field processing unit extracts the image findings, examination site, lesion area, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations from the segmented text. The source location recording unit records the source location of the above content in the preliminary diagnostic conclusion and associates it with the corresponding field in the structured prompt information.

[0067] Specifically, the list generation module first reads the preliminary diagnostic conclusion according to the report generation result format fields. If the preliminary diagnostic conclusion has been output in segments according to imaging findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations, the text reading unit directly reads each segment of text. If there is unsegmented content in the preliminary diagnostic conclusion, the text reading unit segments it according to the field order in the structured prompt information. During segmentation, content containing examination site, lesion area, and imaging findings is written as "Imaging Findings"; content containing diagnostic conclusion is written as "Diagnostic Conclusion"; descriptions supporting the diagnostic conclusion are written as "Diagnostic Basis"; text containing content requiring exclusion or differentiation is written as "Differential Diagnosis"; and text containing suggestions for re-examination, clinical consideration, or treatment is written as "Recommendations". After completing the above extraction, the field processing unit merges content with the same examination site, the same lesion area, and the same imaging findings into the same list record.

[0068] Furthermore, the list generation module performs corresponding processing on the diagnostic conclusions and diagnostic evidence. The field processing unit reads each conclusion text from the diagnostic conclusion and searches for the corresponding image discovery in the diagnostic evidence. If a diagnostic conclusion has a corresponding diagnostic evidence, the field processing unit writes the diagnostic conclusion, the diagnostic evidence, and the image discovery into the same list record. If a diagnostic conclusion is not associated with a diagnostic evidence, the field processing unit writes the diagnostic conclusion into the list record and writes the corresponding content from the insufficient evidence prompt field into the same list record. If an image discovery exists in the diagnostic evidence, but the diagnostic conclusion does not reference the image discovery, the field processing unit writes the image discovery into the content to be examined and retains its source location.

[0069] Furthermore, the list generation module processes the differential diagnoses and recommendations accordingly. The field processing unit reads the differential diagnosis from the differential diagnosis field and searches for the corresponding imaging findings, examination sites, and lesion areas. If the differential diagnosis and the diagnostic conclusion belong to the same examination site, the field processing unit writes the differential diagnosis into the same list record. If the examination sites of the differential diagnosis and the diagnostic conclusion are inconsistent, the field processing unit writes the differential diagnosis into the conflict resolution record and retains the corresponding source location in the unified feature list. For recommendations, the field processing unit reads the corresponding diagnostic conclusion and diagnostic basis. If the recommendation is not associated with a diagnostic conclusion, the field processing unit writes it into the content to be tested for consistency verification by S500.

[0070] Understandably, the unified feature list is the output of this step. The unified feature list includes image findings, examination sites, lesion areas, diagnostic conclusions, diagnostic basis, differential diagnosis, suggested content, source location, insufficient basis prompt field content, and conflict resolution records. The source location recording unit associates each item in the unified feature list with the paragraph position in the preliminary diagnostic conclusion and with the corresponding field in the structured prompt information. If the list generation module finds an empty field in the image findings field, diagnostic conclusion field, or diagnostic basis field, the list generation module retains the empty field status and writes it into the unified feature list. If the preliminary diagnostic conclusion does not return a diagnostic conclusion, the list generation module generates an anomaly record and sends the anomaly record along with the unified feature list to S500.

[0071] In one engineering embodiment, the list generation module receives the preliminary diagnostic conclusion corresponding to a chest computed tomography scan. The text reading unit reads the image discovery segment, diagnostic conclusion segment, diagnostic basis segment, differential diagnosis segment, and suggested content segment from the preliminary diagnostic conclusion. The field processing unit extracts the lung examination site, lesion area, and image findings from the image discovery segment; extracts the diagnostic conclusion from the diagnostic conclusion segment; extracts the image findings supporting the diagnostic conclusion from the diagnostic basis segment; extracts relevant differential diagnoses from the differential diagnosis segment; and extracts follow-up or clinically relevant content from the suggested content segment. The source location recording unit records the position of the above content in the preliminary diagnostic conclusion and generates a unified feature list. The unified feature list is sent to the consistency verification module and the differential analysis module, allowing S500 to perform consistency verification based on the unified feature list, image description, preliminary diagnostic information, predefined analysis instructions, and medical image reporting specifications.

[0072] Summary of the technical effects of this step: This step breaks down the preliminary diagnostic conclusion into a unified feature list, establishing a source-location relationship between image findings, diagnostic conclusions, diagnostic basis, differential diagnosis, and recommendations. The unified feature list retains fields indicating insufficient evidence and conflict resolution records, facilitating S500's identification of consistent items, potential discrepancies, and supplementary findings. The list generation module converts the text output of the large language model into readable field content for subsequent consistency checks.

[0073] S500 performs consistency checks based on a unified feature list, image description, preliminary diagnostic information, predefined analysis instructions, and medical image reporting standards, generating consistent items, potential discrepancies, supplementary findings, and a discrepancy analysis report.

[0074] The consistency verification module and the differentiation analysis module perform this step. Specifically, the consistency verification module receives the unified feature list generated by S400 and simultaneously reads the image description and preliminary diagnostic information generated by S100, as well as the predefined analysis instructions and medical image report specifications called by S200. The unified feature list includes image findings, examination site, lesion area, diagnostic conclusion, diagnostic basis, differential diagnosis, suggested content, source location, insufficient basis prompt field content, and conflict resolution records. The image description is the image findings text recorded in the preliminary report. The preliminary diagnostic information includes the diagnostic conclusion, diagnostic basis, differential diagnosis, and suggested content. The predefined analysis instructions include image findings reading rules, diagnostic conclusion generation rules, diagnostic basis association rules, differential diagnosis citation rules, insufficient basis recording rules, and report generation result format rules. The medical image report specifications include the correspondence between examination site, image findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and suggested content. The consistency verification module loads the above content into the same examination task to form homogeneous verification data.

[0075] Specifically, the consistency verification module first reads the unified feature list item by item. For each item, the module reads its examination site, lesion area, imaging findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations, and then checks the corresponding text in the preliminary diagnostic conclusion based on its source location. Subsequently, the module compares this item with the image description, preliminary diagnostic information, predefined analysis instructions, and medical imaging report specifications. During the comparison, the examination site is matched with the anatomical structure range in the medical imaging report specifications, the lesion area is matched with the lesion area in the image description, the imaging findings are matched with the diagnostic basis in the preliminary diagnostic information, the diagnostic conclusion is matched with the differential diagnosis, and the recommendations are matched with the diagnostic conclusion. If the content in the unified feature list has corresponding content in both the aforementioned texts and specifications, the consistency verification module classifies that content as a consistent item.

[0076] Furthermore, the consistency verification module identifies potential discrepancies. These potential discrepancies are those where the content in the unified feature list differs from the image description or preliminary diagnostic information in any of the fields: examination site, lesion area, imaging findings, diagnostic basis, differential diagnosis, or suggested content. Specifically, when the examination site in the unified feature list differs from the examination site in the image description, the consistency verification module classifies that list content as a potential discrepancy. When the diagnostic conclusion in the unified feature list does not have a corresponding imaging finding in the image description, the consistency verification module classifies that diagnostic conclusion as a potential discrepancy. When the differential diagnosis in the unified feature list does not match the correspondence in the medical imaging report specifications, the consistency verification module classifies that differential diagnosis as a potential discrepancy. When the suggested content is not associated with a diagnostic conclusion or diagnostic basis, the consistency verification module classifies that suggested content as a potential discrepancy. For conflict resolution records written by S400, the consistency verification module directly reads the source location of the conflict resolution record and includes the corresponding content in the potential discrepancies.

[0077] Furthermore, the consistency verification module identifies supplementary findings and items with insufficient evidence. Supplementary findings are those in the unified feature list that conform to medical image reporting standards but do not have a direct corresponding content in the image description and preliminary diagnostic information. Specifically, when a new image finding appears in the preliminary diagnostic conclusion, and this finding conforms to medical image reporting standards, but there is no direct corresponding text in the image description and preliminary diagnostic information, the consistency verification module classifies this content as a supplementary finding. When a diagnostic conclusion in the unified feature list is not associated with diagnostic evidence, the consistency verification module generates an item with insufficient evidence and records the source position of this diagnostic conclusion in the preliminary diagnostic conclusion. For the content of the insufficient evidence prompt field, the consistency verification module reads its corresponding diagnostic conclusion and writes it into the item with insufficient evidence.

[0078] Understandably, the differential analysis module receives consistent items, potential discrepancies, supplementary findings, and insufficient evidence items from the consistency verification module and generates a differential analysis report. This report includes records of consistent items, potential discrepancies, supplementary findings, insufficient evidence items, source location, and verification process. When generating the report, the differential analysis module associates and saves consistent items with their corresponding image findings, diagnostic conclusions, and diagnostic evidence; associates and saves potential discrepancies with their corresponding inconsistencies and source locations; associates and saves supplementary findings with their correspondence in medical image reporting standards; and associates and saves insufficient evidence items with diagnostic conclusions that are not associated with diagnostic evidence. If an empty field exists in the unified feature list, the differential analysis module writes it to the verification process record and retains the original source of the field.

[0079] In one engineering embodiment, the consistency check module receives a unified feature list corresponding to a chest computed tomography scan. The unified feature list records the lung examination site, lesion area, imaging findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations. The consistency check module compares the lung examination site with the examination site in the image description, the lesion area with the lesion area in the preliminary report, and the diagnostic conclusion with the diagnostic basis. If a diagnostic conclusion has corresponding diagnostic basis in both the image description and the preliminary diagnostic information, the consistency check module classifies it as a consistent item. If a diagnostic conclusion lacks corresponding diagnostic basis, the consistency check module generates an insufficient basis item. If an imaging finding originates from the preliminary diagnostic conclusion but does not appear in the image description, the consistency check module classifies it as a supplementary finding item. The differential analysis module generates a differential analysis report based on the above classification results and sends the differential analysis report to the structured output module to execute S600.

[0080] This step's technical effects can be summarized as follows: This step integrates the unified feature list with image descriptions, preliminary diagnostic information, predefined analysis instructions, and medical image reporting standards into a single examination task for consistency verification. Consistent items, potential discrepancies, supplementary findings, and items with insufficient evidence form the field basis for the differential analysis report. The differential analysis report incorporates S600's modification suggestions, review prompts, and the final structured diagnostic conclusion generation process.

[0081] S600: Generate modification suggestions, review prompts, and final structured diagnostic conclusions based on the differential analysis report.

[0082] The structured output module performs this step. Specifically, the structured output module receives the differential analysis report generated by S500 and reads the consistent items, potential differences, supplementary findings, insufficient evidence items, source location records, and testing process records from the differential analysis report. The modification suggestions are text processing content generated for potential differences and insufficient evidence items, including deletion, replacement, downgrading, and retention for review. The review prompts are prompts read during the preliminary report review, including examination site, lesion area, imaging findings, diagnostic conclusion, diagnostic basis, differential diagnosis, suggested content, and source location. The final structured diagnostic conclusion is the conclusion text output in this step, including the diagnostic conclusion, diagnostic basis, differential diagnosis, modification suggestions, and review prompts. The structured output module automatically starts after the differential analysis report is generated and uses the differential analysis report as input data for this step.

[0083] Specifically, the structured output module first reads the consistency items and extracts the diagnostic conclusion, diagnostic basis, and corresponding imaging findings from them. For the content in the consistency items, the structured output module retains the original diagnostic conclusion and original diagnostic basis, and writes them into the diagnostic conclusion field and diagnostic basis field of the final structured diagnostic conclusion. The structured output module then reads the potential discrepancies and generates modification suggestions based on the inconsistency fields in the potential discrepancies. When the potential discrepancy originates from inconsistency in the examination site, the structured output module generates a replacement suggestion for the examination site and records the examination site in the image description and the examination site in the preliminary diagnostic information. When the potential discrepancy originates from inconsistency in the lesion area, the structured output module generates a lesion area review prompt and records the source location. When the potential discrepancy originates from missing diagnostic basis, the structured output module generates a downgrade expression suggestion and writes the corresponding diagnostic conclusion into the review prompt.

[0084] Furthermore, the structured output module reads supplementary findings. These supplementary findings are a list of items that conform to medical imaging report standards but do not have a direct corresponding content in the image description and preliminary diagnostic information. The structured output module does not directly write supplementary findings into the definitive diagnostic conclusion; instead, it writes them into supplementary findings prompts, which are then included as part of the review prompts. For image findings in the supplementary findings, the structured output module records their source location, examination site, and lesion area. For differential diagnoses in the supplementary findings, the structured output module records their corresponding diagnostic conclusions and their correspondence with the medical imaging report standards. For suggestions in the supplementary findings, the structured output module records their associated diagnostic conclusions and diagnostic basis.

[0085] Furthermore, the structured output module reads items indicating insufficient evidence. These insufficient evidence items refer to content where the diagnostic conclusion is not associated with relevant diagnostic evidence. The structured output module downgrades the expression of the diagnostic conclusion based on these insufficient evidence items. Specifically, if a diagnostic conclusion lacks diagnostic evidence, but corresponding content exists in imaging findings and the examined site, the structured output module adjusts the diagnostic conclusion to a tentative conclusion. If a diagnostic conclusion lacks diagnostic evidence and the corresponding content in imaging findings is incomplete, the structured output module adjusts the diagnostic conclusion to a suspected conclusion. If there is overlap between the diagnostic conclusion and the differential diagnosis, the structured output module adjusts the diagnostic conclusion to a conclusion that cannot be ruled out. If the diagnostic conclusion requires consideration of recommendations, the structured output module adjusts the diagnostic conclusion to a conclusion that recommends consideration of clinical findings. These adjustments are written into modification suggestions and saved in conjunction with the review prompt.

[0086] Understandably, when generating the final structured diagnostic conclusion, the structured output module follows the field writing order specified in medical imaging reports. First, it writes the retained consistent diagnostic conclusions and diagnostic basis; then, it writes modification suggestions after processing potential discrepancies; next, it writes supplementary findings and review prompts; finally, it writes the downgraded expression content after processing insufficient evidence. The structured output module records the source location, the version of the differential analysis report, and the output time in each item. If there are empty fields in the differential analysis report, the structured output module writes the empty field status into the review prompt and retains the source location of the preliminary diagnostic conclusion corresponding to the empty field. If the differential analysis report does not contain consistent items, the structured output module writes all potential discrepancies, supplementary findings, and insufficient evidence items into the review prompt and generates a final structured diagnostic conclusion awaiting review.

[0087] In one engineering embodiment, the structured output module receives a differential analysis report corresponding to a chest computed tomography (CT) scan. The differential analysis report includes the lung examination site, lesion area, imaging findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations. The structured output module reads the diagnostic conclusion and diagnostic basis from the consistency section and writes them into the final structured diagnostic conclusion. If a potential difference section shows that the diagnostic conclusion is inconsistent with the lesion area in the image description, the structured output module generates a lesion area replacement suggestion and a review prompt. If a supplementary discovery section shows that there are imaging findings in the preliminary diagnostic conclusion that are not directly recorded in the image description, the structured output module writes this content into the supplementary discovery prompt. If an insufficient basis section shows that the diagnostic conclusion is not associated with diagnostic basis, the structured output module adjusts the diagnostic conclusion to a tentative conclusion, a suspected conclusion, a conclusion that cannot be ruled out, or a conclusion that recommends clinical correlation. The final structured diagnostic conclusion is output to the imaging reporting system and saved in conjunction with the preliminary report.

[0088] Summary of the technical effects of this step: This step generates modification suggestions, review prompts, and a final structured diagnostic conclusion based on the differential analysis report. Potential discrepancies, supplementary findings, and items with insufficient evidence are processed through different pathways in this step. The final structured diagnostic conclusion retains the source location and testing process records, forming a closed-loop output for medical imaging diagnosis.

[0089] Example 2: Figure 2 A structural block diagram of a medical imaging diagnostic device based on a large model according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include:

[0090] The data acquisition module 01 is used to acquire the medical image to be diagnosed, image description, and preliminary diagnostic information, extract image type, examination site, image findings, and preliminary report, and generate examination task data. Specifically, the data acquisition module receives the medical image to be diagnosed, image description, and preliminary diagnostic information from the image reporting system. The data acquisition module reads the image type and examination site from the medical image to be diagnosed, the image findings, lesion area, and diagnostic basis from the image description, and the diagnostic conclusion, differential diagnosis, and recommendations from the preliminary diagnostic information. The data acquisition module performs correlation extraction and conflict resolution on the image description and preliminary diagnostic information, writing the image findings, examination site, lesion area, diagnostic basis, differential diagnosis, and recommendations into the same examination task data. The examination task data retains the source location of the medical image to be diagnosed, image description, preliminary diagnostic information, and preliminary report, and passes it to the prompt information generation module for invocation.

[0091] The prompt information generation module 02, connected to the data acquisition module, is used to generate structured prompt information based on the examination task data by calling predefined analysis instructions and medical imaging report specifications. The structured prompt information includes fields for image discovery, diagnosis conclusion, diagnosis basis, differential diagnosis, insufficient basis prompt, and report generation result format. Specifically, the prompt information generation module receives the examination task data from the data acquisition module and reads the image type, examination site, image discovery, preliminary report, lesion area, diagnosis basis, differential diagnosis, and recommendations. The prompt information generation module calls predefined analysis instructions according to the image type and examination site, and matches the correspondence between the examination site, image discovery, diagnosis conclusion, diagnosis basis, differential diagnosis, and recommendations in the medical imaging report specifications. The prompt information generation module writes the image discovery into the image discovery field, the diagnosis conclusion into the diagnosis conclusion field, the diagnosis basis into the diagnosis basis field, the differential diagnosis into the differential diagnosis field, content not associated with diagnosis basis into the insufficient basis prompt field, and writes the fields sequentially into the report generation result format field. The structured prompt information, along with the image description and the preliminary diagnostic information, is transmitted to the conclusion generation module for invocation.

[0092] The conclusion generation module 03, connected to the prompt information generation module, is used to input the structured prompt information, the image description, and the preliminary diagnosis information into a medical text pre-trained large language model to generate a preliminary diagnosis conclusion. Specifically, the conclusion generation module receives the structured prompt information from the prompt information generation module and simultaneously reads the image description and the preliminary diagnosis information. The conclusion generation module sends the image discovery field, diagnosis conclusion field, diagnosis basis field, differential diagnosis field, insufficient basis prompt field, and report generation result format field into the medical text pre-trained large language model. The medical text pre-trained large language model reads the corresponding content between image discovery and diagnosis basis, reads the corresponding content between diagnosis conclusion and differential diagnosis, and segments the content according to the report generation result format field to form image discovery, diagnosis conclusion, diagnosis basis, differential diagnosis, and recommendation content. The conclusion generation module synthesizes the above segmented content into a preliminary diagnosis conclusion and passes the preliminary diagnosis conclusion to the list generation module for invocation.

[0093] The list generation module 04, connected to the conclusion generation module, is used to parse the preliminary diagnostic conclusion and generate a unified feature list. Specifically, the list generation module receives the preliminary diagnostic conclusion from the conclusion generation module and reads the structured prompt information corresponding to the preliminary diagnostic conclusion. The list generation module parses the preliminary diagnostic conclusion according to the report generation result format fields, extracting imaging findings, examination sites, lesion areas, diagnostic conclusions, diagnostic basis, differential diagnoses, and recommendations from the preliminary diagnostic conclusions. The list generation module records the source position of the above content in the preliminary diagnostic conclusions and establishes a correspondence between each item and the imaging findings field, diagnostic conclusion field, diagnostic basis field, differential diagnoses field, and insufficient basis prompt field. The unified feature list includes imaging findings, examination sites, lesion areas, diagnostic conclusions, diagnostic basis, differential diagnoses, recommendations, and source positions, and is passed to the consistency verification module for invocation.

[0094] The consistency verification module 05, connected to the list generation module, is used to perform consistency verification based on the unified feature list, the image description, the preliminary diagnostic information, the predefined analysis instructions, and the medical image reporting specifications, generating consistent items, potential discrepancies, and supplementary findings. Specifically, the consistency verification module receives the unified feature list from the list generation module and simultaneously reads the image description, the preliminary diagnostic information, the predefined analysis instructions, and the medical image reporting specifications. The consistency verification module reads the examination site, lesion area, image findings, diagnostic conclusion, diagnostic basis, differential diagnosis, and recommendations from the unified feature list item by item. The consistency verification module compares the above content with the image description, the preliminary diagnostic information, and the medical image reporting specifications, respectively. When the content in the unified feature list matches all of the image description, the preliminary diagnostic information, the predefined analysis instructions, and the medical image reporting specifications, it is classified as a consistent item; when any field is inconsistent, it is classified as a potential discrepancy item; when it conforms to the medical image reporting specifications but does not form a direct corresponding content in the image description and the preliminary diagnostic information, it is classified as a supplementary finding item. The consistency item, the potential difference item, and the supplementary discovery item are passed to the difference analysis module for invocation.

[0095] The differentiation analysis module 06, connected to the consistency verification module, is used to generate a differentiation analysis report based on the consistent items, potential difference items, and supplementary findings. Specifically, the differentiation analysis module receives the consistent items, potential difference items, and supplementary findings from the consistency verification module and reads the source location of each item. The differentiation analysis module associates and saves the consistent items with their corresponding image findings, diagnostic conclusions, and diagnostic criteria; associates and saves the potential difference items with their corresponding inconsistency fields and source locations; and associates and saves the supplementary findings with their corresponding relationships in the medical image reporting specifications. When a diagnostic conclusion is not associated with diagnostic criteria, the differentiation analysis module generates an insufficient criteria item and writes the insufficient criteria item into the differentiation analysis report. The differentiation analysis report contains records of consistent items, potential difference items, supplementary findings, insufficient criteria items, and source locations, and is passed to the structured output module for invocation.

[0096] The structured output module 07, connected to the differential analysis module, is used to generate modification suggestions, review prompts, and a final structured diagnostic conclusion based on the differential analysis report. Specifically, the structured output module receives the differential analysis report from the differential analysis module and reads the consistent items, potential difference items, supplementary findings, insufficient evidence items, and source location records. The structured output module retains the diagnostic conclusions and diagnostic evidence in the consistent items, writes the inconsistency fields corresponding to the potential difference items into modification suggestions, writes the supplementary findings into review prompts, and adjusts the diagnostic conclusions corresponding to the insufficient evidence items to a propensity conclusion, a suspected conclusion, a conclusion that cannot be ruled out, or a conclusion that suggests clinical consideration. The structured output module writes the diagnostic conclusion, diagnostic evidence, differential diagnosis, modification suggestions, and review prompts according to medical imaging report specifications to form a final structured diagnostic conclusion, and saves the final structured diagnostic conclusion along with the preliminary report.

Claims

1. A method for diagnosing a conclusion of a medical image based on a large model, characterized by, include: S100: Obtain the medical images to be diagnosed, image descriptions, and preliminary diagnostic information; extract image type, examination site, image findings, and preliminary report; and generate examination task data. S200: Based on the examination task data, call predefined analysis instructions and medical image report specifications to generate structured prompt information; The structured prompt information includes fields for image discovery, diagnosis conclusion, diagnosis basis, differential diagnosis, insufficient basis prompt, and report generation result format. S300. Input the structured prompt information, image description and preliminary diagnosis information into a large language model pre-trained for medical text to generate a preliminary diagnosis conclusion; S400: Analyze the preliminary diagnostic conclusions and generate a unified feature list; S500 performs consistency checks based on a unified feature list, image description, preliminary diagnostic information, predefined analysis instructions, and medical image reporting standards, generating consistent items, potential discrepancies, supplementary findings, and a discrepancy analysis report. S600 generates modification suggestions, review prompts, and final structured diagnostic conclusions based on the differential analysis report.

2. The method of claim 1, wherein, In S100, the image description and preliminary diagnostic information are correlated, extracted, and conflict resolved to obtain the image findings, examination sites, lesion areas, diagnostic basis, differential diagnosis, and recommendations, which are then written into the examination task data.

3. The method of claim 1, wherein, In S200, based on the medical imaging report standards matching the image type and examination site, a correspondence is established between the image discovery field, diagnostic conclusion field, diagnostic basis field, and differential diagnosis field, and the correspondence is written into the structured prompt information.

4. The method of claim 1, wherein, In S300, the large language model pre-trained in medical text outputs image findings, diagnostic conclusions, diagnostic basis, differential diagnosis, and recommendations in segments according to the report generation result format fields, forming a preliminary diagnostic conclusion.

5. The method of claim 1, wherein, In S400, the analysis of the preliminary diagnostic conclusion includes extracting imaging findings, examination sites, lesion areas, diagnostic conclusions, diagnostic basis, differential diagnoses, and recommendations. The source location of each item in the preliminary diagnostic conclusion is recorded to form a unified feature list.

6. The method of claim 1, wherein, In S500, when the content of the unified feature list matches the image description, preliminary diagnostic information, predefined analysis instructions, and medical image reporting specifications, the content is classified as a consistent item.

7. The method of claim 1, wherein, In S500, when the content of the Uniform Feature List is inconsistent with the image description or preliminary diagnostic information in any field such as examination site, lesion area, image findings, diagnostic basis, differential diagnosis, or recommendation content, the content is classified as a potential discrepancy.

8. The method of claim 1, wherein, In S500, when the content of the unified feature list conforms to the medical image reporting specifications but does not form a direct corresponding content in the image description and preliminary diagnostic information, the content is classified as a supplementary discovery item; when the diagnostic conclusion is not associated with the diagnostic basis, an insufficient basis item is generated.

9. The method of claim 1, wherein, In S600, diagnostic conclusions and diagnostic basis are retained based on consistent items, modification suggestions or review prompts are generated based on potential discrepancies, supplementary discovery prompts are generated based on supplementary discovery items, and in cases where the basis is insufficient, the diagnostic conclusion is adjusted to a probabilistic conclusion, a suspected conclusion, a conclusion that cannot be ruled out, or a conclusion that suggests combining clinical findings.

10. A medical imaging diagnostic device based on a large model, applied to the method of any one of claims 1-9, characterized in that, include: Data acquisition module, prompt information generation module, conclusion generation module, list generation module, consistency verification module, differential analysis module, structured output module. Data acquisition module, prompt information generation module, conclusion generation module, list generation module, consistency verification module, differential analysis module, structured output module.

Images