Cancer diagnosis method and system based on large language model

Abstract

The invention discloses a cancer diagnosis method and a cancer diagnosis system based on a large language model, which are characterized in that a comprehensive medical knowledge database is constructed by integrating multi-source medical data, and clinical data and image examination results of patients are deeply analyzed and comprehensively analyzed in combination with the powerful natural language processing capability of the large language model. The method comprises the following steps: firstly, converting clinical data into standardized medical terms by using a large language model, and preliminarily screening possible cancer types in a knowledge base; the range is further narrowed in combination with image features, and the positions of cancer stages, primary lesions and metastatic lesions are accurately judged; finally, the system generates personalized treatment suggestions of the next stage, and records diagnosis results to continuously optimize the accuracy of the system. According to the invention, the accuracy and efficiency of cancer diagnosis are improved, and more reliable and personalized medical services are provided for clinicians and cancer patients.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a cancer diagnosis method and system based on a large language model. BACKGROUND

[0002] A large language model is a deep learning model trained by large-scale text data, which can generate human language text and perform various natural language processing tasks; however, directly using a large language model in the medical field poses a "hallucination" risk, which is unacceptable in the medical field. If a large language model generates incorrect drug recommendations, diagnostic conclusions, or treatment plans, it may not only affect the treatment effect of patients, but even endanger the lives of patients; in the medical field, cancer is a general term for a class of diseases. Under certain conditions, cells may undergo genetic mutations, causing cells to no longer follow normal life cycles but to proliferate indefinitely and possibly invade surrounding tissues or metastasize to other parts of the body through the blood and lymphatic system, forming new tumors and tumor metastasis.

[0003] A doctor's diagnosis of a patient, especially for complex cases such as multiple tumors or cases with metastasis, requires comprehensive analysis of clinical data and imaging examination results. Only by combining the two can a comprehensive understanding of the disease be obtained, and more accurate diagnosis and treatment decisions can be made. Relying solely on one data source may lead to misinterpretation or omission of important information, and joint use can be mutually verified to reduce the likelihood of misdiagnosis. However, patients with cancer are prone to have multiple conditions, and the symptoms of these conditions may mask or confuse the manifestations of cancer, leading to incorrect judgments of the disease, especially the primary and metastatic lesions of cancer. Doctors may have difficulty distinguishing which symptoms are caused by cancer and which are caused by other diseases. In this case, misjudgment is easy to occur, leading to incorrect treatment plans.

[0004] Therefore, it is necessary to improve the existing technology of a cancer diagnosis method and system based on a large language model to solve the above problems. SUMMARY

[0005] The present application overcomes the shortcomings of the prior art and provides a cancer diagnosis method and system based on a large language model, aiming to solve the problems of information fragmentation, high risk of misdiagnosis, and difficulty in accurately distinguishing between primary and metastatic lesions in the prior art cancer diagnosis method.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows: a cancer diagnosis method based on a large language model, comprising the steps of: S1, collecting and integrating cancer multi-source data to construct a comprehensive medical knowledge database and performing correlation analysis on the medical knowledge of cancer; S2, the large language model receives and analyzes the clinical data of the patient and converts it into medical terms, screens the relevant cancer types in the medical knowledge database, and sorts them according to the matching degree to preliminarily screen out possible cancer types; S3, combined with the imaging examination results of the patient, the organ position of the cancer is identified, and the imaging characteristics are used to further narrow down the range of cancer types; S4, according to the imaging examination results and the converted clinical data, the cancer staging information is matched and obtained from the database; S5, combining the disease information in the clinical data and the tumor characteristics of the imaging examination results, the positions of the primary tumor and the metastatic tumor are analyzed and judged; S6, all the above information is summarized to determine the specific type and stage of the cancer, and the current treatment plan is compared with the treatment plan in the database to generate and output a personalized next-stage treatment recommendation; S7, record the diagnosis results in the database as a data source for training the system, and continuously optimize the system to improve accuracy.

[0007] In a preferred embodiment of the present application, in step S1, the medical knowledge includes: cancer types, cancer symptoms, cancer signs, cancer clinical manifestations, TNM staging standards, AJCC cancer staging guidelines, cancer local treatment plans and systemic treatment plans; In a preferred embodiment of the present application, in step S1, the step of constructing the medical knowledge database includes: S11, collect and integrate multi-source data from public medical literature, clinical guidelines, research papers, authoritative medical websites, professional books and journals; S12, design a database mode, including: a cancer type table, a symptom table, a sign table, a TNM staging table, an AJCC staging guideline table, and a treatment plan table, each table containing key fields to support connection queries between different tables; S13, enter the medical knowledge into the database, associate the data between different tables through the key fields, and form a complete medical knowledge network; S14, perform correlation analysis on the associated medical knowledge to classify the cancer symptom correlations of different cancer types.

[0008] In a preferred embodiment of the present application, in step S14, the step of classifying the cancer symptom correlations of different cancer types includes: S141, extract the cancer type table and the symptom table from the medical knowledge database; S142, construct a symptom-cancer type matrix, with cancer types as rows and symptoms as columns, to construct a two-dimensional matrix, and the elements in the matrix represent the probability or frequency of a certain cancer type appearing a certain symptom; S143, calculate the maximum information coefficient between the cancer types and the symptoms.

[0009] In a preferred embodiment of the present application, the joint distribution data of the cancer types and the symptoms is grid divided in a two-dimensional space, and each grid has the same size; Under each grid division, the mutual information value between the cancer types and the symptoms is calculated, and the mutual information value , wherein the cancer types are , and the symptoms are , is the joint probability of the occurrence of a certain cancer type and a certain symptom , and is the marginal probability of the occurrence of a certain cancer type and a certain symptom alone; the mutual information value is standardized; Under all grid divisions, the maximum standardized mutual information value is selected as the MIC, which represents the maximum correlation degree between the variables under all possible scales.

[0010] In a preferred embodiment of the present application, in step S2, S21, receiving patient clinical data; S22, using a pre-trained large language model to analyze the clinical data, identifying and labeling medical concepts in the clinical data, including: symptoms, signs and treatment plans; mapping the identified medical concepts to the standard medical vocabulary ICD-10 for medical terminology standardization; S23, based on the standardized clinical data, matching the standardized symptoms in the clinical data with the symptoms in the database according to a matching degree calculation formula, sorting according to the matching degree, and combining the MIC values of the related cancer types with the symptoms in the database to sort the cancer types, and preliminarily screening the possible cancer types.

[0011] In a preferred embodiment of the present application, the standardized patient symptom vector is , and the symptom vector set corresponding to the cancer types in the medical knowledge database is For each cancer type , the similarity score between it and the patient symptoms is calculated ; Query the precomputed maximum information coefficient matrix to obtain the MIC value between each cancer type and the patient symptoms; Calculate the comprehensive score ; is a weight parameter for adjusting the importance of the similarity score and the MIC value; All cancer types are sorted in descending order of the comprehensive score.

[0012] In a preferred embodiment of the present application, in step S4, the steps of judging the cancer stage are: S41, integrate the imaging examination results of the patient and the transformed clinical data, and extract key information, including: tumor size, invasion range, lymph node metastasis; S42, according to the extracted key information, match with the TNM staging standard in the medical knowledge database, to obtain the T stage, N stage and M stage of the tumor; S43, according to the TNM staging result, combined with the AJCC staging guide table, determine the cancer stage of the patient.

[0013] In a preferred embodiment of the present application, in step S5, the steps of judging the primary focus and metastatic focus are: S51, according to the morphological characteristics, distribution mode and surrounding tissue influence of the tumor, judge whether there are multiple lesions, and whether there are local invasion or distant metastasis; When the imaging shows only one tumor lesion and no distant metastasis, it is judged as a primary focus; When multiple lesions are located in the same organ, it is a primary multiple center cancer or local spread; When the lesion appears in the opposite organ or distant organ, it needs to be further judged whether it is a metastatic focus; S52, combined with the clinical data and the position and morphological characteristics of the tumor, analyze the origin and diffusion path of the tumor, and further judge the primary focus and metastatic focus of the tumor; According to the primary focus and metastatic focus, determine and verify the specific cancer type and stage.

[0014] The present application provides a cancer diagnosis system based on a large language model, comprising modules: A data acquisition and preprocessing module is used to collect medical literature, clinical guidelines, research paper information, and preprocess these data, such as cleaning, labeling and formatting; A medical knowledge base construction module is used to establish and maintain a comprehensive medical knowledge database; A clinical data analysis module is used to use a pre-trained large language model to analyze the clinical data of the patient in free text form, convert it into standardized medical terminology, and map it to ICD-10 code; A cancer type matching module is used to compare the standardized clinical data with the information in the medical knowledge database, screen out possible cancer types, and sort them according to the matching degree; An image data analysis module is used to analyze the imaging examination results, identify the location of the cancerous organ, and further narrow down the range of cancer types using image features; The cancer staging judgment module is used for determining the specific stage of the cancer in combination with the imaging examination result and the converted clinical data. The primary and metastatic lesion judgment module is used for comprehensively analyzing and judging the positions of the primary and metastatic lesions by comprehensively combining the clinical data and the imaging features. The treatment scheme generation module is used for determining the specific type and stage of the cancer and generating the individualized treatment suggestion.

[0015] The present application solves the defects in the background art and has the following beneficial effects: (1) The present application proposes a cancer diagnosis method and system based on a large language model, which integrates multiple sources of medical data to construct a comprehensive medical knowledge database, and combines the powerful natural language processing capability of the large language model to comprehensively analyze the clinical data and imaging examination result of the patient, uses the large language model to analyze the clinical data and convert it into standardized medical terminology, and screens possible cancer types in the knowledge base; then, the imaging features are combined to further narrow down the range; by comprehensively analyzing the imaging data and the clinical data, the stage of the cancer and the positions of the primary and metastatic lesions are accurately judged; finally, the system generates an individualized next-stage treatment suggestion and records the diagnosis result to continuously optimize the accuracy of the system. The present application significantly improves the accuracy and efficiency of cancer diagnosis, and provides more reliable and individualized medical services for clinicians and cancer patients.

[0016] (2) The present application realizes accurate judgment of the stage of the cancer by integrating the clinical data and imaging examination result of the patient, combining the TNM staging standard and AJCC staging guidelines in the medical knowledge database. In step S4, the system extracts key information including tumor size, invasion range and lymph node metastasis, and matches it with the staging standard in the database to determine the T stage, N stage and M stage of the tumor, and finally determines the stage of the cancer in combination with the AJCC staging guidelines. The comprehensive analysis method based on multiple data sources avoids the inaccuracy of the existing technology which only relies on a single data source, and further improves the accuracy of the cancer staging judgment, providing a more reliable basis for the selection of subsequent treatment schemes.

[0017] (3) The present application realizes accurate judgment of the primary and metastatic lesions by comprehensively analyzing the morphological features, distribution patterns, surrounding tissue influence and clinical data of the tumor; in step S5, the system judges whether there are multiple lesions and metastasis according to the imaging features, and analyzes the origin and diffusion path of the tumor in combination with the clinical data to distinguish the primary and metastatic lesions; compared with the existing technology, the present application overcomes the limitation that it is difficult to accurately distinguish the primary and metastatic lesions by only relying on imaging or clinical data, and further enhances the ability to distinguish cancer lesions, providing important support for the accurate treatment of cancer.

[0018] (4) In step S14, the system calculates the correlation between cancer types and symptoms by the maximum information coefficient (MIC), which can capture nonlinear relationships, thereby more accurately associating symptoms with cancer types, overcoming the limitations of traditional linear correlation analysis, and more comprehensively reflecting the complex relationship between symptoms and cancer types; in combination with the matching degree sorting in step S2, the system not only considers the direct matching degree of patient symptoms and database cancer symptoms, but also combines the MIC value of symptoms and cancer types, further optimizing the sorting of cancer types, so that the system can more accurately identify cancer types and avoid misdiagnosis due to similar symptoms; in combination with step S3, the imaging examination result further narrows down the range of cancer types, and through two sorts, the system can not only quickly screen out possible cancer types, but also further verify and optimize the screening result through imaging characteristics.

[0019] (5) The present application records each diagnosis result into the medical knowledge database, and continuously trains the system by using machine learning algorithm, realizes the self-optimization of the system, and the system continuously learns through accumulated data, improves the accuracy of cancer type identification, staging judgment, primary and metastatic lesion distinction and treatment scheme generation. Compared with the prior art, the present application can continuously improve the diagnosis ability and efficiency of the system through continuous learning and optimization, better serve the clinicians and cancer patients, and promote the continuous development of cancer diagnosis technology. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings; Figure 1 is a flowchart of the preferred embodiment of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0023] SUMMARY In the medical field, cancer is a complex and diverse disease, characterized by abnormal cell proliferation and potential spread to other parts of the body. Even in different cancer types or different stages of the same type, patients may exhibit similar symptoms and tumor distribution patterns. Early-stage cancers may only present non-specific symptoms such as fatigue, weight loss, or mild pain. In the late stage, regardless of the type of cancer, it may present extensive metastasis, affecting multiple organ systems, increasing the difficulty of diagnosis, especially in distinguishing different types of cancer and their development stages, as well as distinguishing between primary and metastatic tumors.

[0024] Existing cancer diagnosis methods usually rely on a simple combination of clinical data and imaging examinations, failing to fully utilize the complementary advantages between the two, resulting in fragmented information and difficulty in forming a comprehensive and accurate diagnosis conclusion. In addition, cancer patients often have other coexisting diseases, and the symptoms of these diseases may mask or confuse the manifestations of cancer, making diagnosis more complex. Relying solely on any one data source not only easily leads to misinterpretation or omission of important information, but also increases the risk of misdiagnosis, thereby affecting the treatment effect and prognosis of patients.

[0025] To solve the above problems, the present application proposes a cancer diagnosis method and system based on a large language model, which utilizes existing medical knowledge bases and integrates the powerful natural language processing capabilities of a large language model to improve the accuracy and efficiency of cancer diagnosis.

[0026] Exemplary method: A cancer diagnosis method based on a large language model, comprising the steps of: S1, collecting and integrating multi-source cancer data to construct a comprehensive medical knowledge database for correlation analysis of cancer medical knowledge; S2, the large language model receives and analyzes the patient's clinical data and converts it into medical terminology, screens the relevant cancer types in the medical knowledge database, and sorts them according to the matching degree to preliminarily screen out possible cancer types; S3, combined with the patient's imaging examination results, identify the organ location of cancer, and further narrow down the range of cancer types using imaging features; S4, according to the imaging examination results and the converted clinical data, match and obtain the cancer staging information from the database; S5, combining the disease information in the clinical data and the tumor characteristics of the imaging examination results, analyze and determine the location of the primary tumor and the metastatic tumor; S6, summarize all the above information, determine the specific type and stage of cancer, compare the current treatment plan with the treatment plan in the database, and generate and output personalized next-stage treatment recommendations; S7, record the diagnosis result into the database as the data source of the training system, and continuously optimize the system to improve the accuracy.

[0027] In step S1, the step of constructing the medical knowledge database includes: S11, collect and integrate multi-source data from public medical literature, clinical guidelines, research papers, authoritative medical websites, professional books and journals; S12, design a database schema, including: a cancer type table, a symptom table, a sign table, a TNM staging table, an AJCC staging guideline table, and a treatment plan table, each table containing key fields to support connection queries between different tables; The key fields include: ID, name, description, and associated ID; S13, enter medical knowledge into the database, associate data between different tables through key fields, and form a complete medical knowledge network; S14, perform correlation analysis on the associated medical knowledge, and classify the cancer symptom correlations of different cancer types.

[0028] In step S1, the medical knowledge includes: cancer type, cancer symptom, cancer sign, cancer clinical manifestation, TNM staging standard, AJCC cancer staging guideline, cancer local treatment plan, and systemic treatment plan; In a specific embodiment, the TNM staging table of step S13 is as follows, taking lung cancer as an example: Table 1 TNM staging standard for lung cancer

[0029] The symptoms of cancer are important clues for doctors to preliminarily judge the type and severity of the disease. Since different patients may have similar symptoms when they have different types of cancer, by analyzing the correlation between different symptoms, similar symptoms can be more accurately associated with the correct cancer type, thereby improving the accuracy of diagnosis. Traditional correlation analysis methods, such as Pearson correlation coefficient, can measure the linear relationship between two variables, but in the medical field, the relationship between symptoms and cancer types is often more complex and may involve non-linear, multi-dimensional associations.

[0030] In step S14, the step of classifying the cancer symptom correlations of different cancer types includes: S141, extract the cancer type table and the symptom table from the medical knowledge database; S142, construct a symptom-cancer type matrix, with cancer types as rows and symptoms as columns, to construct a two-dimensional matrix, where the elements in the matrix represent the probability or frequency of a certain cancer type having a certain symptom; S143, calculate the maximum information coefficient between the cancer type and the symptom; Mutual information is a non-negative index that measures the degree of mutual dependence between two random variables. The maximum information coefficient MIC is an extension of mutual information, which can capture the nonlinear relationship between variables. It is obtained by dividing the data points in two-dimensional space into grids and calculating the maximum mutual information under different grid divisions. The joint distribution data of cancer types and symptoms are divided into grids in two-dimensional space, and each grid has the same size. Under each grid division, the mutual information value between the cancer type and the symptom is calculated, and the mutual information value , where the cancer type is , the symptom is , is the joint probability of the occurrence of a certain cancer type and a certain symptom , and is the marginal probability of the occurrence of a certain cancer type and a certain symptom alone; Since different grid divisions will result in different mutual information values, in order to eliminate this influence, the mutual information value needs to be standardized. Under all grid divisions, the maximum standardized mutual information value is selected as MIC, which represents the maximum correlation between variables under all possible scales.

[0031] In a specific embodiment, when MIC is less than 0.4, it is considered weakly correlated.

[0032] In step S2, the clinical data of the patient is often in the form of free text, containing a large amount of subjective description and personal expression habit. Through NLP technology, these descriptions can be converted into standard medical terminology, and then by matching the clinical data with the cancer types in the medical knowledge database, the possible cancer types can be preliminarily screened out, providing direction for the subsequent detailed diagnosis.

[0033] In step S2, the clinical data includes: patient basic information, subjective symptoms, current illness history, current treatment plan and objective signs; In step S2, S21, receive patient clinical data; S22, use a pre-trained large language model to analyze the clinical data, identify and label the medical concepts in the clinical data, including: symptoms, signs and treatment plans; Map the identified medical concepts to the standard medical vocabulary ICD-10 for medical terminology standardization; For example: The input raw clinical data text is: The patient reported that he had unexplained weight loss of about 5 kg in the past month, accompanied by intermittent chest pain, especially at night; The analysis is: Symptoms: weight loss, chest pain; Time information: one month; Map the identified medical concepts to the standard medical vocabulary ICD-10 to achieve term standardization: ICD-10 code: R63.4 (weight loss, not specified), R07.2 (chest pain).

[0034] S23, based on the standardized clinical data, match the standardized symptoms in the clinical data with the symptoms in the database according to the matching degree calculation formula, sort them according to the matching degree, and combine the MIC values of the related cancer types according to the symptoms in the database to sort the cancer types and preliminarily screen out possible cancer types.

[0035] The standardized patient symptom vector is , and the symptom vector set corresponding to the cancer types in the medical knowledge database is For each cancer type Calculate the similarity score between it and the patient's symptoms ; Query the precomputed maximum information coefficient matrix to obtain the MIC value between each cancer type and the patient's symptoms; Calculate the comprehensive score ; is a weight parameter used to adjust the importance of the similarity score and the MIC value.

[0036] Sort all cancer types according to the comprehensive score from high to low, and do not consider those with a comprehensive score below 0.4.

[0037] In step S3, the imaging examination results include: tumor size, location, morphological characteristics, distribution pattern and surrounding tissue impact; In step S3, according to the imaging examination results, determine the organ of canceration, exclude some cancer types that are not possible to occur in this organ, and further filter the corresponding cancer types.

[0038] In step S4, the steps of cancer staging are: S41, integrate the patient's imaging examination results and the transformed clinical data to extract key information, including: tumor size, invasion range, and lymph node metastasis; S42, according to the extracted key information, match with the TNM staging standard in the medical knowledge database to obtain the T stage, N stage and M stage of the tumor; S43, according to the TNM staging results, combined with the AJCC staging guide table, determine the cancer stage of the patient; In a specific embodiment, the AJCC staging guide table of step S42 is as follows, taking lung cancer as an example: Table 1 AJCC Staging Guide Table for Lung Cancer

[0039] The primary site is the starting position of the cancer, usually located at the site where the tumor is first discovered; the metastatic site is the lesion formed by the spread of cancer cells from the primary site to other sites; for multiple lesions, it is difficult to determine which is the primary site and the metastatic site, which misleads the judgment and staging of cancer, and is not conducive to the diagnosis of the disease.

[0040] In step S5, the steps of judging the primary site and the metastatic site are: S51, according to the morphological characteristics, distribution pattern and surrounding tissue influence of the tumor, judge whether there are multiple lesions, and whether there are local invasion or distant metastasis; When the imaging shows only one tumor lesion and no distant metastasis, it is usually judged as the primary site; When multiple lesions are located in the same organ, it is a primary multiple center cancer or local spread; When the lesion appears in the opposite organ or distant organ, it needs to be further judged whether it is a metastatic lesion; S52, combined with clinical data and the location and morphological characteristics of the tumor, analyze the origin and spread path of the tumor, and judge the primary site and metastatic site of the tumor; The primary site is located at the site where the tumor is first discovered, with local lymph node metastasis, irregular morphological characteristics, unclear boundaries, and may invade surrounding tissues; The metastatic site is multiple, with regular morphological characteristics, clear boundaries, and located in distant organs, accompanied by distant lymph node metastasis.

[0041] According to the primary site and the metastatic site, further determine and verify the specific type and stage of the cancer.

[0042] In step S6, S61, integrate all relevant information such as clinical data, imaging results, cancer type matching degree, TNM staging, AJCC staging guide, and judgment results of the primary site and metastatic site, to determine the specific type and stage of the cancer; S62, compare the current treatment plan with the treatment plan in the medical knowledge database, consider the patient's basic information, medical history and current treatment plan, generate and output personalized next-stage treatment recommendations.

[0043] In a specific embodiment, the treatment plan table of step S62 is shown below, taking lung cancer as an example:

[0044] In step S7, the diagnosis results of each time, including the clinical data of the patient, the imaging examination results, the diagnosis conclusion, the treatment plan and the treatment effect, etc. are recorded in detail into the medical knowledge database. These records will serve as the data source for the subsequent learning and optimization of the system, and through machine learning algorithms, the system is continuously trained to improve its accuracy and efficiency in cancer diagnosis.

[0045] With the accumulation of data and the continuous optimization of algorithms, the system will be able to more accurately identify cancer types, judge cancer stages, distinguish primary and metastatic lesions, and generate more personalized treatment plans, thereby better serving clinicians and cancer patients.

[0046] Exemplary system: A cancer diagnosis system based on a large language model, comprising modules: A data collection and preprocessing module for collecting medical literature, clinical guidelines, research paper information, and preprocessing these data such as cleaning, labeling and formatting; A medical knowledge base construction module for establishing and maintaining a comprehensive medical knowledge database; A clinical data analysis module for using a pre-trained large language model to analyze the patient's free text form of clinical data, converting it into standardized medical terminology, and mapping it to ICD-10 coding; A cancer type matching module for comparing standardized clinical data with information in the medical knowledge database, screening possible cancer types, and ranking according to matching degree; An image data analysis module for analyzing imaging results, identifying cancerous organ locations, and further narrowing down the range of cancer types using image features; A cancer staging module for determining the specific stage of cancer in combination with imaging results and transformed clinical data; A primary and metastatic lesion judgment module for analyzing and determining the location of primary and metastatic lesions based on clinical data and imaging features; A treatment plan generation module for determining the specific type and stage of cancer and generating personalized treatment recommendations.

[0047] Based on the ideal embodiments of the present application, the above description, relevant personnel can make various changes and modifications without deviating from the scope of the present application. The technical scope of the present application is not limited to the contents of the specification, and must be determined by the scope of the claims.

Claims

1. A cancer diagnosis method based on a large language model, characterized in that, Including the following steps: S1. Collect and integrate multi-source cancer data to build a comprehensive medical knowledge database and conduct correlation analysis on medical knowledge related to cancer. S2: The large language model receives and parses the patient's clinical data and converts it into medical terms. It then filters relevant cancer types in the medical knowledge database and sorts them according to the degree of matching in order to initially screen out possible cancer types. S3. Based on the patient's imaging examination results, identify the location of the cancerous organ and further narrow down the range of cancer types using imaging features; S4. Based on the imaging examination results and the converted clinical data, match and obtain cancer staging information from the database; S5. Analyze and determine the location of the primary tumor and metastatic tumor by combining the symptom information in the comprehensive clinical data and the tumor characteristics in the imaging examination results. S6. Summarize all the above information, determine the specific type and stage of cancer, compare the current treatment plan with the treatment plans in the database, and generate and output personalized treatment suggestions for the next stage. S7. Record the diagnostic results in the database as a data source for the training system, and continuously optimize the system to improve accuracy.

2. The cancer diagnosis method based on a large language model according to claim 1, characterized in that: In step S1, the medical knowledge includes: cancer types, cancer symptoms, cancer signs, cancer clinical manifestations, TNM staging criteria, AJCC cancer staging guidelines, local cancer treatment plans, and systemic cancer treatment plans.

3. The cancer diagnosis method based on a large language model according to claim 2, characterized in that: In step S1, the steps for constructing the medical knowledge database include: S11. Collect and integrate multi-source data from public medical literature, clinical guidelines, research papers, authoritative medical websites, professional books and journals; S12. Design the database schema, including: cancer type table, symptom table, sign table, TNM staging table, AJCC staging guideline table, and treatment plan table. Each table contains key fields and supports join queries between different tables. S13. Enter medical knowledge into the database, and use key fields to link data between different tables to form a complete medical knowledge network; S14. Conduct correlation analysis on the associated medical knowledge and classify it according to the correlation of cancer symptoms of different cancer types.

4. The cancer diagnosis method based on a large language model according to claim 3, characterized in that: Step S14, the step of classifying cancer symptoms based on their relevance to different cancer types, includes: S141. Extract cancer type and symptom tables from the medical knowledge database; S142. Construct a symptom-cancer type matrix. With cancer type as the row and symptoms as the column, construct a two-dimensional matrix. The elements in the matrix represent the probability or frequency of a certain symptom occurring in a certain cancer type. S143. Calculate the maximum information coefficient between cancer type and symptoms.

5. The cancer diagnosis method based on a large language model according to claim 4, characterized in that: The combined distribution data of cancer types and symptoms are divided into grids in a two-dimensional space, with each grid being the same size; Under each grid division, calculate the mutual information value between cancer type and symptoms. Among them, the cancer type is Symptoms are , It is a certain type of cancer and a certain symptom The joint probability of simultaneous occurrence It is a certain type of cancer and a certain symptom Marginal probability of a single occurrence; standardization of mutual information values; Under all grid divisions, the largest standardized mutual information value is selected as the MIC, which represents the maximum degree of correlation between variables at all possible scales.

6. The cancer diagnosis method based on a large language model according to claim 5, characterized in that: In step S2, S21. Receive patient clinical data; S22. Use a pre-trained large language model to parse clinical data, identify and label medical concepts in the clinical data, including symptoms, signs and treatment plans; map the identified medical concepts to the standard medical vocabulary ICD-10 to standardize medical terminology; S23. Based on the standardized clinical data, according to the matching degree calculation formula, the standardized symptoms in the clinical data are matched with the symptoms in the database, and sorted according to the matching degree. Combined with the MIC value of the relevant cancer types in the database, the cancer types are sorted and possible cancer types are preliminarily screened.

7. The cancer diagnosis method based on a large language model according to claim 6, characterized in that: The standardized patient symptom vector is The symptom vector set corresponding to cancer types in the medical knowledge database is For each type of cancer Calculate its similarity score to the patient's symptoms. ; Query the pre-computed maximum information coefficient matrix to obtain the MIC value between each cancer type and patient symptoms; Calculate the overall score ; These are weighting parameters used to adjust the importance of similarity scores and MIC values; All cancer types are sorted from highest to lowest based on their overall scores.

8. The cancer diagnosis method based on a large language model according to claim 1, characterized in that: Step S4, the steps for determining cancer staging: S41. Based on the patient's imaging examination results and the converted clinical data, extract key information, including: tumor size, extent of invasion, and lymph node metastasis. S42. Based on the extracted key information, match it with the TNM staging standard in the medical knowledge database to obtain and determine the T stage, N stage and M stage of the tumor; S43. Based on the TNM staging results and in conjunction with the AJCC staging guidelines, determine the patient's cancer stage.

9. A cancer diagnosis method based on a large language model according to claim 8, characterized in that: Step S5 involves identifying the primary lesion and metastatic lesions: S51. Based on the morphological characteristics, distribution pattern, and impact on surrounding tissues of the tumor, determine whether there are multiple lesions, as well as whether there is local invasion or distant metastasis. When imaging shows only one tumor lesion with no distant metastasis, it is determined to be the primary lesion; When multiple lesions are located in organs on the same side, it indicates primary multicentric cancer or local spread. When lesions appear in contralateral or distant organs, further assessment is needed to determine whether they are metastatic lesions. S52. Combining clinical data with the location and morphological characteristics of the tumor, analyze the origin and spread pathway of the tumor to further determine the primary and metastatic lesions. Based on the primary and metastatic lesions, the specific cancer type and stage are determined and verified.

10. A cancer diagnosis system based on a large language model, comprising the cancer diagnosis method based on a large language model as described in any one of claims 1-9, characterized in that, Includes modules: The data acquisition and preprocessing module is used to collect information from medical literature, clinical guidelines, and research papers, and to preprocess this data, such as cleaning, labeling, and formatting. The medical knowledge base construction module is used to build and maintain a comprehensive medical knowledge database; The clinical data parsing module is used to parse patients' free text-based clinical data using a pre-trained large language model, convert it into standardized medical terms, and map it to ICD-10 encoding. The cancer type matching module is used to compare standardized clinical data with information in a medical knowledge database, filter out possible cancer types, and sort them according to the degree of matching. The image data analysis module is used to analyze image examination results, identify the location of cancerous organs, and further narrow down the range of cancer types using image features; The cancer staging module is used to determine the specific stage of cancer by combining imaging examination results and converted clinical data. The primary lesion and metastatic lesion determination module is used to analyze and determine the location of primary lesions and metastatic lesions by integrating clinical data and imaging features; The treatment plan generation module is used to determine the specific type and stage of cancer and generate personalized treatment recommendations.

Images