Large-model intelligent TNM staging system construction method based on pathological report and TNM standard

Abstract

The invention discloses a method for establishing a large-model intelligent TNM staging system based on a pathological report and a TNM standard, and the method comprises the steps: firstly, generating a detailed TNM staging explanation which serves as an intermediate process to help refine staging judgment, and can also directly provide valuable auxiliary information for a doctor; and key information is further extracted and concentrated from the detailed explanations, so that an accurate TNM staging result is obtained. The method ensures that the model not only can provide detailed staging instructions, but also can output simple and clear final staging conclusions, meets the requirements in different scenes, provides real-time, rapid and high-accuracy TNM diagnosis for clinicians, and improves the actual application effect of clinical diagnosis.

Description

Technical Field

[0001] The invention relates to a method for establishing a large-model intelligent TNM staging system based on pathological reports and TNM standards, and belongs to the technical field of biomedicine. Background Art

[0002] The Chinese invention patent application with publication number CN119048360A, published on November 29, 2024, discloses a TNM staging management system for the respiratory system. By adding a TMN staging prediction management unit, it is possible to predict the conditions of primary tumors, lymph node metastasis and distant metastasis, and manage subsequent treatments based on the predicted conditions. By adding a regional frame selection and recognition unit, it is used to verify the status of lymph nodes through a lymph node model, identify the presence of lymph nodes and abnormal conditions of images, and by adding an abnormal lymph node recognition unit, it is possible to change and enhance the image color of the lymph node image, thereby accurately identifying abnormal conditions of lymph nodes and improving the accuracy of TNM staging management.

[0003] The Chinese invention patent application with publication number CN118840360A, published on October 25, 2024, discloses a method for predicting the TNM staging of liver cancer based on plain scan CT images, and discloses a technology for predicting the TNM staging of liver cancer that integrates deep learning, radiomics, and clinical diagnostics, aiming to accurately predict the TNM staging of liver cancer through plain scan CT images;

[0004] The above two methods are mainly based on images for TNM staging, and neither of them mentions whether the system is real-time, accurate, and fast. In particular, oral CT images cannot detect oral cancer, and it is impossible to achieve intelligent TNM staging for oral cancer. In addition, how to ensure the generalization ability of the model and how to apply the model in actual clinical environments are also key issues that need to be solved. Summary of the invention

[0005] The technical problem to be solved by the present invention is that in clinical practice, not all pathology reports contain pathological staging when they are issued, which causes great trouble to doctors' diagnosis, patient personal information acquisition and cognition, scientific research and other statistical analyses. At the same time, there is also a lack of a pathology classification system based on medical history and pathology reports that can perform batch processing and provide real-time feedback.

[0006] In order to solve the above technical problems, the technical solution of the present invention is to disclose a method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards, which is characterized by comprising the following steps:

[0007] Step 1: After obtaining the oral cancer pathology report, based on the established TNM staging industry standard, organize and initialize the data structure, label the oral cancer pathology report dataset according to the TNM staging industry standard, and generate a TNM pathology report dataset with real labels;

[0008] Step 2: Build a question-answering system based on natural language understanding, with oral cancer pathology report text as input data and oral cancer status of patients as output data. Use the voting classifier of machine learning method to train the input data, generate TNM staging explanation, and obtain TNM staging report, where:

[0009] The training data input into the question-answering system includes text information from oral cancer pathology reports, lesion type descriptions, clinical features of tumors, surgical findings, and other standardized medical records related to TNM staging. The labels are composed of detailed TNM staging explanation text or standardized labels to ensure that the question-answering system can accurately understand and generate professional content related to TNM staging.

[0010] Step 3: Establish a large TNM staging model, which further includes an embedded model module and a multi-model joint module, wherein:

[0011] The TNM staging report output by the question-answering system is used as the input data of the embedding model module, which converts high-dimensional data into low-dimensional dense vectors.

[0012] The multi-model joint module is composed of multiple sub-models. The input data of each sub-model corresponds to different TNM staging-related factors, and the output data of each sub-model is the T-item prediction value, N-item prediction value, and M-item prediction value in the tumor TNM staging. The output data of each sub-model are integrated through a comprehensive module to obtain the final TNM staging result.

[0013] The output data of the TNM staging model include the final TNM staging results or the predicted values ​​of T items, N items, and M items;

[0014] Step 4: Use the trained TNM staging model to perform tumor staging;

[0015] Step 5: Regularly evaluate the TNM staging model to ensure its accuracy and effectiveness.

[0016] Preferably, in step 1, the oral cancer pathology report includes a professional diagnosis, clarifying whether the biopsy result is benign or malignant.

[0017] Preferably, if the biopsy result is benign, the impact of the biopsy result on future health, that is, whether the biopsied tumor has the risk of malignant transformation, is described in the oral cancer pathology report; if the biopsy result is malignant, the pathologist evaluates the severity of the malignant tumor in the oral cancer pathology report based on the observation of the slice.

[0018] Preferably, in step 2, the question-answering system selects the LLaMA Chinese large language model as a basis and adjusts its parameters to adapt to specific medical task requirements.

[0019] Preferably, in step 2, the TNM staging report includes the preliminary tumor staging or serves as an intermediate variable for prediction by the subsequently established TNM staging model.

[0020] Preferably, in step 2, the label includes a detailed TNM staging classification, a standardized staging label, and a detailed description text of the staging.

[0021] Preferably, after step 2 and before step 3, the method further includes: fine-tuning and optimizing the question-answering system established in step 2:

[0022] Fine-tuning was performed using the oral cancer pathology report data and labels collected in the early stage. During fine-tuning, the input data included a detailed description of the clinical report, basic patient information, and TNM staging standards. The question-and-answer system continuously optimized and adjusted the accuracy of the prediction.

[0023] Preferably, in step 3, the multi-model joint module is composed of a tumor feature sub-model, a lymph node metastasis sub-model and a distant metastasis sub-model, wherein:

[0024] The input data of the tumor feature sub-model include the tumor size, location, and differentiation degree in the TNM staging report, and the output data is the predicted value of the T item in the TNM staging;

[0025] The input data of the lymph node metastasis submodel are the size and metastasis of lymph nodes in the TNM staging report, and the output data are the predicted values ​​of N items in the TNM staging;

[0026] The input data of the distant metastasis submodel is the description of distant metastasis in the TNM staging report, and the output data is the predicted value of the M item in the TNM staging.

[0027] Preferably, in step 4, when the TNM staging large model is applied, the doctor can view the TNM staging results and detailed explanations provided by the TNM staging large model on the doctor's computer. If there is any question about the result, the doctor can further check the patient's pathology report and other related information by clicking on the system page.

[0028] Preferably, in step 5, the evaluation of the TNM staging model includes:

[0029] The evaluation criteria are set based on actual clinical cases, combined with expert opinions and TNM staging standards to ensure the comprehensiveness and authority of the evaluation;

[0030] Doctors participate in the evaluation: Doctors can view the TNM staging results and their detailed explanations generated by the TNM staging model and, as human evaluators, score the TNM staging results and their detailed explanations generated by the model according to established standards;

[0031] Conduct multiple rounds of evaluation at multiple time points, regularly collect feedback from doctors on the TNM staging model, and use the evaluation data collected from doctors as the basis for model evaluation;

[0032] Based on the obtained evaluation evidence, conclusions were drawn on the performance of the TNM staging model in practical applications, and the TNM staging model was fine-tuned;

[0033] After each fine-tuning, re-training is performed and a new evaluation cycle is entered. Through a continuous feedback loop, it is continuously optimized to adapt to more clinical scenarios and complex cases, ultimately achieving higher staging accuracy and clinical applicability.

[0034] The present invention discloses a method for establishing a large-model intelligent TNM staging system based on pathology reports and TNM standards. First, a detailed TNM staging explanation is generated. This explanation not only helps to refine the staging judgment as an intermediate process, but also can directly provide valuable auxiliary information for doctors, and then further extracts and condenses key information from these detailed explanations to obtain accurate TNM staging results. This method ensures that the model can not only provide detailed staging descriptions, but also output concise and clear final staging conclusions, meet the needs of different scenarios, provide clinicians with real-time, fast and highly accurate TNM diagnosis, and improve the actual application effect of clinical diagnosis. Compared with the existing technical solutions, the present invention has the following specific beneficial effects:

[0035] 1. Improve diagnostic efficiency: By automatically processing pathology reports, the system established by the present invention can quickly perform TNM staging on a large amount of pathological data, greatly improving the efficiency of staging, reducing the workload of doctors, and allowing doctors to devote more time and energy to other clinical work.

[0036] 2. Enhanced diagnostic accuracy: Using large models and a large amount of pathology report data for training, the system established by the present invention can learn the relationship between complex pathological features and TNM staging, thereby improving the accuracy of staging. In addition, the system can update and learn the latest pathological data in real time to adapt to the ever-changing pathological features and diagnostic criteria.

[0037] 3. Ensure standardization and consistency: Follow industry standards for the integration of pathology reports and staging judgments. The system established by the present invention ensures that the staging results between different hospitals and different regions are highly standardized and consistent, which is of great significance for clinical research and statistical analysis.

[0038] 4. Improve the quality of clinical decision-making: Accurate TNM staging is crucial for the selection of treatment options for cancer patients. The highly accurate staging results provided by the system established by the present invention can help doctors formulate more accurate treatment plans, thereby improving treatment effects and the quality of life of patients.

[0039] 5. Reduce medical costs: Through automated and standardized processing, human errors and duplication of work are reduced, and medical costs are reduced. Especially in areas with limited pathology resources, the application of the system established by the present invention can significantly improve the accessibility and economy of pathology services.

[0040] 6. Promote scientific research and education: The system established by the present invention is not only valuable in clinical applications, but can also be used as a scientific research and educational tool to help medical students and researchers better understand the principles and practices of TNM staging and promote the development of medical education and scientific research.

[0041] 7. Strong adaptability: The design of the system established by the present invention takes into account the diversity of different pathology report formats and data sources, has strong adaptability and flexibility, can be seamlessly connected with the existing medical information system, and is easy to promote and apply in different medical institutions.

[0042] In summary, the intelligent TNM staging system established by the present invention not only solves the problems existing in the prior art through technological innovation, but also brings about the above-mentioned beneficial effects in many aspects, and has important clinical and social value. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Illustrate the model training and evaluation logic;

[0044] Figure 2 The overall operation logic of the TNM system is shown as follows: 1. Based on the vector database;

[0045] Figure 3 It illustrates the second overall operation logic of the TNM system - based on file assistant. DETAILED DESCRIPTION

[0046] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0047] The embodiment of the present invention discloses a method for constructing a large-model intelligent TNM staging system based on pathology reports and TNM standards, which specifically includes the following steps:

[0048] Step 1: Data preparation, including the following steps:

[0049] Step 101. Obtain an oral cancer pathology report (hereinafter referred to as the "TNM pathology report"), which includes a professional diagnosis provided by a doctor to clarify whether the biopsy result is benign or malignant. If the result is benign, the TNM pathology report should describe the impact of the biopsy result on future health, that is, whether the biopsied tumor has the risk of "malignant transformation"; if the report result is malignant, the pathologist should evaluate the severity of the malignant tumor in the TNM pathology report based on the observation of the slice.

[0050] Step 102, establish the TNM staging industry standard, organize and initialize the data structure. According to the TNM staging industry standard, the TNM pathology report dataset obtained in the previous step is accurately labeled to generate a TNM pathology report dataset with real labels. These real labels cover the primary tumor, regional lymph nodes, and distant metastasis. Subsequently, a preprocessing step is performed on these labeled TNM pathology report datasets to obtain a preprocessed TNM pathology report dataset, which serves as the basis for subsequent model training to ensure the quality of the data and the accuracy of the labels.

[0051] In the embodiment of the present invention, the established TNM staging industry standard is shown in the following table.

[0052] 2024 CSCO Guidelines for the Diagnosis and Treatment of Head and Neck Tumors

[0053] UICC / AJCC TNM Analysis System (8th Edition)

[0054]

[0055] Tumor lesions and lymph nodes measurement standards

[0056]

[0057] Step 2. In order to improve the accuracy of the model's understanding of TNM staging, the industry standard text of TNM staging is added in the input stage. Specifically, a question-answering system based on natural language understanding (NLU) is constructed, and it is optimized specifically for the medical field. The question-answering system selects the LLaMA Chinese large language model as the basis, and adjusts its parameters to meet the specific medical task requirements. The question-answering system takes the TNM pathology report text as input and the patient's oral cancer condition as output, and uses the voting classifier of the machine learning method to train the input data. This not only provides more background information for the model, but also enhances its ability to understand the user's input intent. In this way, a detailed TNM staging explanation is first generated. This explanation not only serves as an intermediate process to help refine the staging judgment of the subsequent TNM staging model, but also can directly provide valuable auxiliary information for doctors.

[0058] In an embodiment of the present invention, the training data input into the question-answering system includes text information in the pathology report, description of lesion type, clinical characteristics of the tumor (e.g., size, distribution, metastasis), surgical findings, and other standardized medical records related to TNM staging. The label is composed of detailed TNM staging explanation text or standardized labels to ensure that the question-answering system can accurately understand and generate professional content related to TNM staging. The output data of the question-answering system is the TNM staging report, which may include preliminary tumor staging or intermediate variables (such as tumor size, lymph node status, etc.) used as predictions for the subsequent TNM staging model.

[0059] Step 3: Fine-tune and optimize the question-answering system established in step 2:

[0060] Fine-tuning is performed using the previously collected TNM staging pathology report data and labels, especially for specific disease types (such as oral cancer) and pathological stages. During this process, the Q&A system will be further adjusted based on different clinical data to adapt to specific application scenarios and disease types. During fine-tuning, the input data will include a detailed description of the clinical report, basic patient information, and TNM staging criteria. The Q&A system continuously optimizes and adjusts the accuracy of the prediction.

[0061] In the embodiment of the present invention, the label data includes detailed TNM staging classification, standardized staging labels and detailed description text of the staging. The goal of the subsequent TNM staging model is to generate accurate staging results by understanding these labels.

[0062] Step 4: Establish a TNM staging model. The TNM staging model further includes an embedding model module and a multi-model joint module, wherein:

[0063] Embedding Model module:

[0064] The embedding model module converts complex high-dimensional data into low-dimensional dense vectors. The TNM staging report output by the question-answering system is used as the input data of the embedding model module, which converts it into a text vector so that the data can be processed and analyzed efficiently. The embedding model module optimizes the performance of the model by adjusting hyperparameters such as dimension size, window size, and number of training iterations, further improving the ability to handle complex cases.

[0065] In the embodiment of the present invention, the embedded model module converts the key information (such as tumor size, location, metastasis, etc.) in the TNM staging report text into a numerical vector representation for easy calculation and further analysis. In this process, the input data includes the preprocessed TNM staging report text, and the output data is the vectorized pathological information and the corresponding intermediate results of staging prediction.

[0066] The multi-model joint module is composed of multiple existing sub-models. The input data of each sub-model corresponds to different TNM staging-related factors, including:

[0067] Tumor feature sub-model: Its input data includes information such as tumor size, location, and degree of differentiation in the TNM staging report, and its output data is the predicted value of the "T" item in the TNM staging of the tumor (such as T1, T2, etc.);

[0068] Lymph node metastasis submodel: Its input data is the size and metastasis of lymph nodes in the TNM staging report, and its output data is the predicted value of the "N" item in the TNM staging report;

[0069] Distant metastasis submodel: Its input data is the description of distant metastasis in the TNM staging report, and its output data is the predicted value of the "M" item in the TNM staging.

[0070] After each sub-model generates its corresponding staging information, the output data of each sub-model is finally fused through a comprehensive module to obtain the final TNM staging result. This fusion process may include rule-based judgment, or weighting or sorting the output of each sub-model through other algorithms to generate the most accurate clinical staging.

[0071] The output data of the TNM staging model includes the final TNM staging results, or intermediate variables (such as the predicted values ​​of T, N, and M items of the tumor). The TNM staging model needs to be fine-tuned according to the actual application scenario to improve its accuracy in specific cases. For example, the TNM staging model can be localized according to the reporting style of different hospitals and different equipment to ensure its application effect in different environments. During fine-tuning, parameter optimization is performed based on feedback from actual clinical data to further improve the accuracy of the staging results.

[0072] Step 5: Model application:

[0073] The trained TNM staging model is used for tumor staging.

[0074] On the doctor's computer, the doctor can view the TNM staging results and detailed explanations provided by the TNM staging model. If there are any questions about the results, the doctor can click on the system page to further view the patient's pathology report and other related information. This process is designed to assist doctors in making quick and accurate diagnoses and ensure that decisions are supported by sufficient information.

[0075] Step 6: Model evaluation:

[0076] Regularly evaluate the TNM staging model to ensure its accuracy and effectiveness. The evaluation process includes the following steps:

[0077] Evaluation criteria setting: The evaluation criteria should be based on actual clinical cases, combined with expert opinions and TNM staging standards to ensure the comprehensiveness and authority of the evaluation.

[0078] Doctors participate in the evaluation: Doctors can view the TNM staging results and their detailed explanations generated by the model. As human evaluators, they score the TNM staging results and their detailed explanations generated by the model according to established standards. The scoring content includes model staging accuracy, clarity of explanation, decision support capabilities, etc. This evaluation process ensures the accuracy and reliability of the model output, and provides valuable feedback for the continuous improvement of model performance.

[0079] Multiple rounds of evaluation: Evaluation should be conducted at multiple time points, and physician feedback on the model should be collected regularly to monitor the changing trend of model performance and ensure that it adapts to the continuous development of clinical practice.

[0080] Evaluation data collection: Evaluation data collected from doctors, including scoring results, the actual application effect of the model in clinical practice, specific problems and improvement suggestions reported by doctors, serve as the basis for model evaluation.

[0081] Obtaining and fine-tuning evaluation conclusions:

[0082] Evaluation conclusion: Through statistical analysis of the evaluation scores output by the TNM staging model, conclusions were drawn on the performance of the TNM staging model in actual applications, including indicators such as model staging accuracy, missed diagnosis rate, misdiagnosis rate, applicability and satisfaction of doctors' feedback.

[0083] Fine-tuning plan: After obtaining the evaluation conclusion, fine-tune the TNM staging model according to the evaluation results. For the parts with low accuracy, the model performance can be improved by increasing training data, adjusting model parameters, introducing new features (such as clinical information such as patient age and gender), or improving the model structure.

[0084] Feedback loop mechanism: After each fine-tuning, retraining is performed and a new evaluation cycle is entered. The TNM staging model is continuously optimized through a continuous feedback loop to adapt to more clinical scenarios and complex cases, ultimately achieving higher staging accuracy and clinical applicability.

Claims

1. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards, characterized in that: The following steps are involved: Step 1: After obtaining the oral cancer pathology report, based on the established TNM staging industry standard, organize and initialize the data structure, label the oral cancer pathology report dataset according to the TNM staging industry standard, and generate a TNM pathology report dataset with real labels; Step 2: Build a question-answering system based on natural language understanding, with oral cancer pathology report text as input data and oral cancer status of patients as output data. Use the voting classifier of machine learning method to train the input data, generate TNM staging explanation, and obtain TNM staging report, where: The training data input into the question-answering system includes text information from oral cancer pathology reports, lesion type descriptions, clinical features of tumors, surgical findings, and other standardized medical records related to TNM staging. The labels are composed of detailed TNM staging explanation text or standardized labels to ensure that the question-answering system can accurately understand and generate professional content related to TNM staging. Step 3: Establish a large TNM staging model, which further includes an embedded model module and a multi-model joint module, wherein: The TNM staging report output by the question-answering system is used as the input data of the embedding model module, which converts high-dimensional data into low-dimensional dense vectors. The multi-model joint module is composed of multiple sub-models. The input data of each sub-model corresponds to different TNM staging-related factors, and the output data of each sub-model is the T-item prediction value, N-item prediction value, and M-item prediction value in the tumor TNM staging. The output data of each sub-model are integrated through a comprehensive module to obtain the final TNM staging result. The output data of the TNM staging model include the final TNM staging results or the predicted values ​​of T items, N items, and M items; Step 4: Use the trained TNM staging model to perform tumor staging; Step 5: Regularly evaluate the TNM staging model to ensure its accuracy and effectiveness.

2. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: In step 1, the oral cancer pathology report includes a professional diagnosis, clarifying whether the biopsy result is benign or malignant.

3. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 2, characterized in that: If the biopsy result is benign, the oral cancer pathology report will describe the impact of the biopsy result on future health, that is, whether the biopsied tumor has the risk of malignant transformation; if the biopsy result is malignant, the pathologist will evaluate the severity of the malignant tumor in the oral cancer pathology report based on the observation of the slice.

4. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: In step 2, the question-answering system selects the LLaMA Chinese large language model as the basis and adjusts its parameters to adapt to specific medical task requirements.

5. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: In step 2, the TNM staging report includes the preliminary tumor staging or serves as an intermediate variable for prediction by the subsequently established TNM staging model.

6. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: In step 2, the label includes a detailed TNM staging classification, a standardized staging label, and a detailed description text of the staging.

7. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: After step 2 and before step 3, the method further includes: fine-tuning and optimizing the question-answering system established in step 2: Fine-tuning was performed using the oral cancer pathology report data and labels collected in the early stage. During fine-tuning, the input data included a detailed description of the clinical report, basic patient information, and TNM staging standards. The question-and-answer system continuously optimized and adjusted the accuracy of the prediction.

8. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: In step 3, the multi-model joint module is composed of a tumor feature sub-model, a lymph node metastasis sub-model, and a distant metastasis sub-model, wherein: The input data of the tumor feature sub-model include the tumor size, location, and differentiation degree in the TNM staging report, and the output data is the predicted value of the T item in the TNM staging; The input data of the lymph node metastasis submodel are the size and metastasis of lymph nodes in the TNM staging report, and the output data are the predicted values ​​of N items in the TNM staging; The input data of the distant metastasis submodel is the description of distant metastasis in the TNM staging report, and the output data is the predicted value of the M item in the TNM staging.

9. A method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: In step 4, when the TNM staging model is applied, the doctor can view the TNM staging results and detailed explanations provided by the TNM staging model on his / her computer. If the doctor has any questions about the results, the doctor can further check the patient's pathology report and other related information by clicking on the system page.

10. The method for constructing a large-model intelligent TNM staging system based on pathological reports and TNM standards as claimed in claim 1, characterized in that: In step 5, the evaluation of the TNM staging model includes: The evaluation criteria are set based on actual clinical cases, combined with expert opinions and TNM staging standards to ensure the comprehensiveness and authority of the evaluation; Doctors participate in the evaluation: Doctors can view the TNM staging results and their detailed explanations generated by the TNM staging model and, as human evaluators, score the TNM staging results and their detailed explanations generated by the model according to established standards; Conduct multiple rounds of evaluation at multiple time points, regularly collect feedback from doctors on the TNM staging model, and use the evaluation data collected from doctors as the basis for model evaluation; Based on the obtained evaluation evidence, conclusions were drawn on the performance of the TNM staging model in practical applications, and the TNM staging model was fine-tuned; After each fine-tuning, re-training is carried out and a new evaluation cycle is entered. Through a continuous feedback loop, it is continuously optimized to adapt to more clinical scenarios and complex cases, ultimately achieving higher staging accuracy and clinical applicability.

Images