Thyroid nodule result analysis method based on large model and series of ultrasonic reports

Abstract

The invention discloses a thyroid nodule result analysis method based on a large model and a series of ultrasonic reports, and relates to the technical field of medical image analysis. The thyroid nodule result analysis method based on the large model and the series of ultrasonic reports mainly comprises the steps of preprocessing thyroid ultrasonic report data to obtain preprocessed thyroid ultrasonic report data, and performing fine tuning on a large language model by utilizing a specific data set and a medical knowledge base to obtain a fine-tuned large language model, and designing a professional prompt template, performing nodule information extraction and unit unified processing on a to-be-analyzed ultrasonic report by using the fine-tuned large language model to obtain standardized structured data, and performing size comparison and analysis on the extracted nodule information by using the fine-tuned large language model to obtain a structured report. By implementing the thyroid nodule result analysis method based on the large model and the series of ultrasonic reports, the efficiency and accuracy of thyroid nodule diagnosis and follow-up visit management can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and more particularly to a method for analyzing thyroid nodules based on a large model and serial ultrasound reports. Background Art

[0002] Thyroid nodules are a common clinical endocrine disease. Ultrasound examination, due to its noninvasive, high-resolution, and real-time imaging capabilities, is widely used for screening, diagnosis, and follow-up management of thyroid nodules. However, nodule descriptions in existing ultrasound reports often rely on manual recording by physicians. Differences in language expression, diagnostic habits, and measurement units exist between physicians, leading to subjectivity and inconsistency in the recording and comparison of nodule size, thus hindering accurate tracking of nodule changes and early intervention during follow-up.

[0003] Traditional named entity recognition (NER) technology plays a vital role in the processing and analysis of medical text. Named entity recognition (NER) technology primarily uses predefined rules, statistical models, or deep learning models to automatically extract meaningful medical entities, such as disease names, symptoms, medications, and test indicators, from texts such as medical literature, electronic health records (EHRs), and ultrasound reports. In the early days of NER, rule-driven methods and dictionary-based matching techniques were prevalent. Rule-driven methods utilize predefined linguistic rules and rely on manually constructed rule sets for entity extraction. While these methods offer the advantage of high-precision extraction in specific domains, they also suffer from a lack of flexibility and scalability, making them difficult to address the challenges of linguistic diversity and complexity.

[0004] With the introduction of statistical methods, NER technologies based on probabilistic graphical models such as conditional random fields (CRFs) and hidden Markov models (HMMs) have gained widespread application. By learning linguistic features from large amounts of annotated data, these models can automatically identify medical entities and reduce reliance on manual rules. However, these statistical models still have limitations in understanding contextual relationships and extracting complex entities.

[0005] In recent years, the rapid development of deep learning technology, particularly Transformer-based models (such as BERT, BioBERT, and SciBERT), has enabled significant progress in named entity recognition (NER). NER methods based on large-scale pre-trained models can more accurately capture complex entity relationships in medical text through context-aware mechanisms, achieving particularly impressive results in areas such as medical imaging reports and clinical notes. By pre-training on large-scale corpora, models like BERT can learn linguistic patterns and the inherent connections between medical terminology from a vast corpus of medical literature, resulting in stronger generalization and higher recognition accuracy in practical applications.

[0006] Despite this, the application of existing NER technology to medical imaging reports still faces several challenges. First, the diverse named entities, inconsistent terminology, and diverse units (such as different units for nodule size) in medical reports complicate automated extraction. Second, existing NER technology has weak standardization capabilities, making it difficult to directly align the extracted results with the standards used in clinical practice. Summary of the Invention

[0007] The purpose of the present invention is to provide a thyroid nodule result analysis method based on a large model and serial ultrasound reports, which can improve the efficiency and accuracy of thyroid nodule diagnosis and follow-up management.

[0008] The present invention provides a thyroid nodule result analysis method based on a large model and a series of ultrasound reports, comprising the following steps: S1: preprocessing thyroid ultrasound report data to obtain preprocessed thyroid ultrasound report data; S2: fine-tuning a large language model based on the preprocessed thyroid ultrasound report data using a specific data set and a medical knowledge base to obtain a fine-tuned large language model; S3: designing a professional prompt template based on the fine-tuned large language model; S4: extracting nodule information and performing unit uniform processing on the ultrasound report to be analyzed using the fine-tuned large language model to obtain standardized structured data; S5: comparing and analyzing the extracted nodule information based on the standardized structured data using the fine-tuned large language model to obtain a structured report.

[0009] Furthermore, the above-mentioned thyroid nodule result analysis method based on a large model and a series of ultrasound reports also includes: performing accuracy verification based on the accuracy verification data set and the structured report to obtain an accuracy verification result; performing evaluation based on the structured report and the accuracy verification result, and reviewing and optimizing the structured report based on the evaluation result to obtain a final analysis report.

[0010] Furthermore, the above preprocessing includes data cleaning, labeling of key information, structural processing, text segmentation and part-of-speech tagging.

[0011] Furthermore, the above-mentioned large language model is ChatGLM3; the specific dataset includes thyroid nodule descriptions, diagnostic criteria, and treatment plans; the medical knowledge base, built on LangChain, includes a knowledge graph, a terminology library, and a vector database.

[0012] Furthermore, the above-mentioned professional prompt template includes prompt instructions, which are used to identify key information, including nodule size, unit, location and morphology.

[0013] Furthermore, the standardized structured data includes the size, location and morphology of the nodules.

[0014] Furthermore, the structured report includes size differences between nodules, classification results, and corresponding risk assessments.

[0015] Furthermore, the above accuracy verification results include precision, recall rate and F1 score.

[0016] Furthermore, the above evaluation includes nodule information extraction accuracy evaluation, unit unification correctness evaluation, nodule size comparison and analysis effectiveness evaluation, and overall process execution efficiency evaluation.

[0017] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for analyzing thyroid nodules based on a large model and serial ultrasound reports.

[0018] The implementation of the thyroid nodule analysis method based on a large model and serial ultrasound reports provided by the present invention has the following beneficial effects: The present invention uses deep learning and natural language processing technology, introduces the ChatGLM3 large language model and its powerful semantic understanding ability, and can automatically extract relevant key information of thyroid nodules from ultrasound reports, including the size, location, morphology and other characteristics of the nodules; this automated extraction significantly improves the efficiency and accuracy of data processing, and reduces the probability of human error and processing time compared to traditional manual methods; in traditional methods, doctors need to manually record, organize, compare and analyze nodule information, which is not only time-consuming but also prone to information omissions or entry errors; through the present invention, a large number of ultrasound reports can be quickly processed and standardized outputs can be provided, saving doctors' time and improving diagnostic efficiency; for the processing of multi-site data, the model can not only process the nodule information of a single site, but also effectively utilize various relevant data points across sites, further improving the accuracy and consistency of data analysis, which is particularly important in multi-center or multi-location medical application scenarios; By leveraging the deep parsing capabilities of large models for medical text, the present invention can deeply understand and accurately identify semantic features in thyroid nodule descriptions, such as terms such as "size," "morphology," and "location," and their contextual relationships. By calculating vector similarity, the system can automatically match nodule information from different reports and automatically compare the sizes of nodules in the left and right thyroid lobes. This automatic matching not only reduces the interference of manual operations but also provides more objective and accurate comparative data, thereby providing high-precision data support for doctors' diagnosis and follow-up decisions. For example, the system can automatically identify and calculate subtle changes in nodule size, helping doctors determine whether the nodule is at risk of malignant transformation, thereby enabling them to take proactive measures to prevent the condition from worsening. In ultrasound reports, the units used for nodule size are often inconsistent. For example, sometimes "centimeter" is used, and sometimes "millimeter" or "inch" is used. If these different units are not handled properly, they may lead to inaccurate data comparison. To address this problem, the present invention combines a large model external medical knowledge base to automatically complete the conversion and unification of nodule size units, such as converting "1.5cm" to "15mm", ensuring the standardization and consistency of data formats. Through this automated unit conversion, the system effectively avoids human errors, ensures the scientific nature and accuracy of the analysis results, and improves the reliability and comparability of the overall data. In practical applications, this automatic conversion can also reduce the tedious work faced by doctors in report processing, thereby improving work efficiency and diagnostic quality. The present invention uses a large language model to analyze thyroid nodules in ultrasound reports, extract nodule size, units and left and right side information, unify nodule units, automatically calculate nodule size and compare them, and realize accurate automated evaluation of thyroid nodule detection and size results. It can overcome the subjectivity and inefficiency that may occur in traditional manual analysis, reduce human intervention, improve diagnostic efficiency and accuracy, and improve the system's processing capabilities for diverse nodule descriptions and report formats; it can perform personalized fine-tuning according to the format and style of different ultrasound reports, so that it can operate efficiently and accurately in various clinical scenarios; taking into full consideration the actual needs of clinical use, the output format of the report is concise and clear, which is convenient for doctors to quickly obtain the required information, thereby providing data support in clinical decision-making and helping doctors judge the nature of the nodules and the direction of treatment; overall, the present invention not only focuses on technological innovation, but also can be deeply integrated with actual application scenarios, promote the intelligent process of thyroid disease diagnosis, improve the efficiency and accuracy of thyroid nodule diagnosis and follow-up management, improve diagnostic efficiency, and improve patient treatment effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 This is a flow chart of a method for analyzing thyroid nodules based on a large model and serial ultrasound reports provided by the present invention; Figure 2 This is a flowchart of the method for analyzing thyroid nodules based on a large model and serial ultrasound reports provided by the present invention; Figure 3 This is a schematic diagram of the structure of the langchain external knowledge base provided by the present invention; Figure 4 The ChatGLM3 connection transformer structure provided by the present invention; Figure 5 This is the principle diagram of ChatGLM3 provided by the present invention; Figure 6 This is the medical knowledge base format provided by the present invention. DETAILED DESCRIPTION

[0020] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0021] Figure 1 A schematic diagram of a thyroid nodule analysis method based on a large model and serial ultrasound reports is shown. In this embodiment, the thyroid nodule analysis method based on a large model and serial ultrasound reports includes the following steps: S1: preprocessing the thyroid ultrasound report data to obtain preprocessed thyroid ultrasound report data; In an exemplary embodiment, preprocessing includes data cleaning, tagging key information, structural processing, text segmentation and part-of-speech tagging; As an exemplary embodiment, in step S1, thyroid ultrasound reports from different medical institutions and patients are collected and integrated, the text content in the reports is cleaned to remove irrelevant information, duplicate data, and non-nodule related parts; key information of the nodules is annotated and structured into a JSON data format; S2: Based on the pre-processed thyroid ultrasound report data, the large language model is fine-tuned using a specific dataset and medical knowledge base to obtain a fine-tuned large language model; In an exemplary embodiment, the large language model is ChatGLM3; the specific dataset includes thyroid nodule descriptions, diagnostic criteria, and treatment plans; the medical knowledge base is built on LangChain and includes a knowledge graph, a term base, and a vector database; As an exemplary embodiment, in step S2, ChatGLM3 is used as the basic model for fine-tuning in the medical field, so that it can effectively process clinical text information related to thyroid nodules; the large language model ChatGLM3 is suitable for medical text understanding, and through fine-tuning it can accurately identify and process key information of thyroid nodules; it is fine-tuned using specific data sets in the fields of thyroid nodule descriptions, diagnostic criteria, treatment plans, etc.; through fine-tuning, the ChatGLM3 model can better adapt to the use of medical professional terms, especially when processing ultrasound reports and clinical records related to thyroid diseases, accurately identify and process the size, shape, location and other features of nodules, thereby significantly improving the accuracy of medical text analysis. Accuracy and practicality; in addition, to further enhance the model's medical reasoning ability, an external medical knowledge base is integrated into the training process of ChatGLM3; this knowledge base contains the classification standards of thyroid nodules, relevant diagnostic criteria, clinical manifestations of common nodules, and the latest medical research results; through the knowledge graph, terminology library and vector database, external medical knowledge is combined with the semantic representation of ChatGLM3 to provide rich medical background knowledge support for the model; this knowledge injection mechanism enhances the model's reasoning ability, enabling it to enhance the model's medical reasoning ability when extracting nodule information, and combine background knowledge in the medical field for more accurate analysis, thereby ensuring that the medical significance of the nodule description is accurately parsed; S3: Design professional prompt templates based on the fine-tuned large language model; In an exemplary embodiment, the professional prompt template includes prompt instructions for identifying key information, the key information including nodule size, unit, location, and morphology; As an exemplary embodiment, in step S3, a professional prompt template for ultrasound reports is designed and applied, combined with the LangChain medical knowledge base to achieve interactive processing to improve the accuracy and scientificity of information extraction; the professional prompt template includes instructions for identifying key information such as nodule size, unit, location and morphology to improve the scientificity and accuracy of information extraction; specifically, the prompt template fully considers the diagnostic characteristics of thyroid nodules and common expression habits in ultrasound reports when designing, covering the entire process from text preprocessing to key information extraction; the template requires the model to automatically identify detailed descriptions of the location, size, morphology, etc. of the nodules in the report, and Automatic conversion is performed for different possible units of measurement to ensure that all extracted data conforms to a unified standard format. Furthermore, a medical verification mechanism is embedded in the template, which uses real-time access to standard data, clinical guidelines, and expert-defined rules in the LangChain medical knowledge base to calibrate and verify the extracted results, thereby correcting errors caused by linguistic ambiguity or inconsistent professional terminology. Through this interactive processing flow, the model can not only adaptively adjust information extraction strategies but also dynamically optimize output results, significantly improving the accuracy and reliability of nodule information analysis in ultrasound reports and providing a solid data foundation for subsequent automated comparison and quantitative analysis. S4: Use the fine-tuned large language model to extract nodule information and unify units from the ultrasound reports to obtain standardized structured data; In an exemplary embodiment, the standardized structured data includes the size, location, and morphology of the nodule; As an exemplary embodiment, in step S4, based on the fine-tuned ChatGLM3 model, nodule information is automatically extracted from the ultrasound report and unitized to ensure the consistency and accuracy of the nodule information. The fine-tuned ChatGLM3 model is responsible for extracting relevant information about thyroid nodules from the ultrasound report and automatically completing unitized processing. Specifically, the input ultrasound report is pre-processed and fed into the ChatGLM3 model. The model, through its deep semantic understanding capabilities, automatically identifies and extracts key information related to thyroid nodules in the report. Based on its pre-trained language model and knowledge fine-tuned in the medical field, ChatGLM3 can accurately identify nodule descriptions, including nodule size, location, morphology, and other characteristics. For example, the model can identify "nodule size is 1.5 cm × 1.2cm, located on the left thyroid, regular morphology"; in the information extraction process, ChatGLM3 uses the self-attention mechanism and context modeling capabilities to capture the semantic relationship between the nodule features in the report and other medical terms (such as "thyroid", "left", "morphology", etc.); for example, the model can recognize the association between "nodule" and terms such as "size" and "morphology", and ensure the consistency of these features in different contexts; through this mechanism, the model can correctly extract structured information from complex medical language, such as the size of the nodule (including numerical values and units), the location of the nodule (including numerical values and units), the morphology of the nodule (such as regular or irregular), etc.; at the same time, ChatGLM3 also applies named entity recognition (NER) technology to identify and annotate medical entities in the text; NER technology can help the model accurately identify medical terms such as "nodule", "thyroid", "left" or "right", and convert them into structured data; this step enables the model to effectively identify the relevant features of the nodule from the ultrasound report, and can further distinguish where the nodule is located. The model not only identifies the numerical value of the nodule, but also automatically detects and unifies the representation of different units. For example, if the report contains two units, "2 mm" and "2 cm", ChatGLM3 will automatically convert the units and convert the sizes of all nodules into standard units (such as centimeters). This unit conversion is based on the model's preset rules, ensuring the consistency of nodule size data and avoiding calculation errors caused by unit differences. In the unit unification process, ChatGLM3 first detects the units involved in the text, identifies possible unit inconsistencies, and unifies them according to preset rules. For example, converting millimeters (mm) to centimeters (cm), or inches (inches) to centimeters. This step not only ensures the consistency of the data in units, but also provides an accurate data basis for subsequent nodule size comparison and risk assessment. Finally, after extraction and unit unification, all nodule information will be formatted into standardized structured data. This data will include the size of the nodule (for example, "nodule 1: 1.5 cm × 1.2cm”), location (e.g., “right thyroid”), and morphology (e.g., “regular”), and this information is integrated into a standardized format for output. The formatted nodule data will serve as input for subsequent nodule comparative analysis and risk assessment models, providing accurate data support for clinical decision-making. S5: Based on the standardized structured data, the fine-tuned large language model is used to compare and analyze the extracted nodule information to obtain a structured report. In an exemplary embodiment, the structured report includes size differences between nodules, classification results, and corresponding risk assessments; As an exemplary embodiment, in step S5, the size of the extracted nodule information is compared and analyzed, and the trend analysis of the nodule size is performed in combination with the historical ultrasound data to assist the doctor in judging the risk of malignant transformation of the nodule; the nodule size trend analysis is combined with the historical ultrasound data to generate a nodule size change curve to assist the doctor in making further clinical decisions; that is, the nodule data after information extraction and unit unification will enter the nodule size comparison and analysis stage; in this process, the ChatGLM3 model first systematically compares the sizes of different nodules, evaluates the differences in their horizontal and vertical dimensions, and analyzes the relative sizes of the left and right thyroid nodules; this process can not only group the nodules, for example For example, nodules can be divided into small nodules, medium nodules and large nodules, and the potential risks of nodules can be automatically judged according to medical diagnostic standards, assisting doctors in identifying the pathological trends of nodules; in addition, ChatGLM3 can also combine historical ultrasound data to perform trend analysis on the size changes of nodules, detect whether there is abnormal growth of nodules, and thus help determine whether there is the possibility of malignant transformation; by integrating external medical knowledge bases, the model can comprehensively analyze the size, shape, location and other characteristics of nodules, conduct a comprehensive risk assessment, and provide doctors with specific diagnostic support suggestions; finally, the model will generate a structured report, presenting the size differences between nodules, classification results and corresponding risk assessments, and provide recommendations for further examinations.

[0022] In an exemplary embodiment, the method for analyzing thyroid nodule results based on a large model and serial ultrasound reports further includes: S6: Perform accuracy verification based on the accuracy verification data set and structured report to obtain the accuracy verification results; In an exemplary embodiment, the accuracy verification results include precision, recall, and F1 score; As an exemplary embodiment, in step S6, the extracted nodule information, unit uniformity and nodule size analysis results are verified through the accuracy verification data set to ensure the accuracy and reliability of the model; the accuracy verification is carried out through indicators such as accuracy, recall rate and F1 score to ensure the accuracy and reliability of the model in practical applications; in the nodule comparison and analysis stage, the system will perform multi-dimensional comparison based on the extracted nodule information to analyze the size and morphological changes of the nodules; especially for ultrasound reports of the same patient at different time periods, the model will calculate the change trend of the nodule size and automatically detect whether the nodule has grown or shrunk. signs, thereby helping doctors determine whether the nodules are at risk of malignant transformation; the medical standards and clinical experience provided by the external knowledge base will also be applied to this step, helping the system to be more accurate in judging nodule changes and ensuring that the analysis results have clinical reference value; specifically, the model will first extract various features of the nodules from ultrasound reports of different time periods, including the size (value and unit), shape (such as regular or irregular), location, etc. of the nodules; for the comparison of nodule sizes, the system will first unify the units to ensure that the nodule sizes in different reports are consistent; if the units in the reports are different (such as "cm" and "mm"), the system will unify the units. The system will automatically convert units to ensure the accuracy of subsequent analysis; next, the system will calculate the trend of nodule size changes by comparing nodule characteristics in different time periods; in this process, the model will automatically calculate the percentage of nodule size change to determine the extent of nodule growth or shrinkage; if the nodule growth exceeds a certain threshold, the system will mark it as a sign of "growth"; if the nodule shrinks, it will be marked as a sign of "shrinkage"; in addition, the system will also analyze the morphological changes of the nodule. If the nodule changes from regular to irregular or the boundary is blurred, it may also indicate the potential risk of malignant changes; in order to improve the accuracy of the analysis results, the system will also refer to Medical standards and clinical experience provided by external medical knowledge bases; by comparing with the diagnostic standards and thyroid nodule classification standards in the knowledge base, the system can determine whether the changes in the nodules meet the clinical identification standards for malignant transformation; for example, the knowledge base may contain information about the growth pattern of malignant nodules, the morphological changes of malignant nodules, etc. The system will automatically call these standards during the comparative analysis to ensure that the analysis results have higher accuracy and clinical reference value; in the nodule comparative analysis stage, the system not only provides quantitative analysis results (such as the percentage of nodule size change), but also gives qualitative assessments, such as whether the nodule has a malignant risk. By integrating size, morphological changes and medical standards, the system provides doctors with comprehensive and scientific analysis results to help doctors make more accurate diagnostic decisions, detect potential malignant thyroid nodules in advance, and provide strong support for patient follow-up and treatment plans; S7: Evaluate the structured report and accuracy verification results, review and optimize the structured report based on the evaluation results, and obtain the final analysis report; In an exemplary embodiment, the evaluation includes nodule information extraction accuracy evaluation, unit unification correctness evaluation, nodule size comparison and analysis effectiveness evaluation, and overall process execution efficiency evaluation; As an exemplary embodiment, in step S7, the validated model is applied to an actual clinical setting to automatically generate nodule assessment reports and continuously monitor and optimize system performance. System performance monitoring includes regular review and optimization of the nodule assessment reports output by the model to improve the efficiency and quality of report generation. The accuracy of the proposed method is comprehensively evaluated to verify the accuracy and reliability of the model in actual application. The core content of the evaluation includes the evaluation of the accuracy of nodule information extraction, the correctness of unit unification, the effectiveness of nodule size comparison and analysis, and the overall process execution efficiency. First, to evaluate the accuracy of nodule information extraction, the system uses annotated standard datasets as a comparison benchmark. These standard datasets contain manually annotated thyroid nodule information, such as nodule size, location, and morphology. The model's extraction results are compared with the manually annotated data to calculate metrics such as accuracy, recall, and F1 score. Accuracy indicates the proportion of nodule information correctly identified by the model, recall measures the model's ability to identify all nodule information, and the F1 score comprehensively considers the balance between accuracy and recall. These metrics can quantify the model's performance in nodule information extraction and ensure that the system can accurately extract key nodule-related information. Secondly, the unit unification correctness assessment aims to ensure that the system can correctly handle the conversion of different units. In actual ultrasound reports, the size of nodules may be expressed in different units, such as "centimeter", "millimeters" or "inches". The system's processing capabilities of these different units will be compared to ensure that when extracting nodule information, the system can automatically perform unified unit conversion and that the converted data is consistent. For example, for a report that "the nodule size is 1.5cm × 1.2cm", the model should be able to unify it into "15mm × 12mm" in millimeters and ensure that there are no errors when comparing nodule sizes in different reports. The effectiveness evaluation of nodule size comparison and analysis focuses on whether the system can correctly analyze the changing trend of nodules when comparing ultrasound reports from different time periods. The system needs to be able to determine whether the nodules show signs of enlargement or reduction during the follow-up period and assess their malignancy risk based on this trend. To this end, the rate of change of nodule size will be calculated using ultrasound reports of patients with data at multiple time points and compared with the manually annotated nodule changes. The evaluation will be based on the accuracy of the trend of nodule size change, the accuracy of judging the malignancy risk, and the clinical relevance of the analysis results. Finally, the overall process execution efficiency assessment ensures that the system can continue to operate efficiently in a high-load clinical environment. The assessment will focus on the time required for the system to process a single ultrasound report, as well as its response speed and throughput when processing a large number of reports. These assessments not only ensure the smooth use of the system in daily clinical work, but also ensure that the system can maintain stability and efficiency when rapidly processing large amounts of data. By comprehensively considering multi-dimensional evaluation indicators such as the accuracy of nodule information extraction, the correctness of unit unification, the effectiveness of nodule size comparison and analysis, and the execution efficiency of the system, the accuracy and practicality of the proposed method can be fully verified, ensuring its reliability and effectiveness in practical applications.

[0023] In some embodiments, the above-mentioned thyroid nodule result analysis method based on a large model and serial ultrasound reports can also be implemented in the following manner. Figure 2A flowchart is constructed for the thyroid nodule result analysis method based on a large model and a series of ultrasound reports; the thyroid nodule result analysis method based on a large model and a series of ultrasound reports includes the following steps: Step 1: Collect and integrate thyroid ultrasound reports from different medical institutions and patients to ensure the diversity and representativeness of the data, and clean the text content in the report, remove irrelevant information, duplicate data and non-nodule related parts, mark the key information of the nodule (such as the size, unit, location, morphology and other characteristics of the nodule), and structure this information into a Json data format; Step 2: Select ChatGLM3, a large language model suitable for medical text understanding, and fine-tune it through a data set containing thyroid nodule-related descriptions, diagnostic criteria, etc., so that it can accurately identify and process the key information of thyroid nodules; During the fine-tuning process, integrate an external medical knowledge base to provide background knowledge support and enhance the medical reasoning ability of the model; Step 3: Design and apply professional prompt templates for ultrasound reports, combined with the LangChain medical knowledge base , realize interactive processing to improve the accuracy and scientificity of information extraction; Step 4: Based on the fine-tuned ChatGLM3 model, nodule information is automatically extracted from the ultrasound report, and the units are unified to ensure the consistency and accuracy of the nodule information, including nodule size extraction, unit conversion and position comparison; Step 5: Compare and analyze the size of the extracted nodule information, combine historical ultrasound data, perform trend analysis of nodule size, assist doctors in judging the risk of malignant transformation of nodules, and generate structured reports to provide data support for clinical decision-making; Step 6: Use the precision verification data set to verify the extracted nodule information, unit unification and nodule size analysis results, and use evaluation indicators such as accuracy, recall rate and F1 score to ensure the accuracy and reliability of the model in practical applications; Step 7: Apply the verified model to the actual clinical environment, automatically generate nodule assessment reports, and continuously monitor and optimize the system performance to ensure efficiency and accuracy under different report formats and diverse data sources.

[0024] In some embodiments, the above-mentioned thyroid nodule analysis method based on a large model and serial ultrasound reports can also be implemented in the following manner. The thyroid nodule analysis method based on a large model and serial ultrasound reports includes: Step 1: Ultrasound report data preprocessing; Ultrasound report data preprocessing is the basic link of the entire process, which aims to provide standardized and high-quality input for subsequent nodule information extraction; first, collect and integrate thyroid ultrasound reports from different medical institutions and patients to ensure the diversity and representativeness of the data; then, perform data cleaning on the text content in the ultrasound report, including removing irrelevant content (such as noise characters, non-medical terms, non-nodule-related parts, repeated data) and correcting text expressions with irregular formats; then, annotate key information and extract key information about nodules in the report through annotation, including features such as the size, unit, location and morphology of the nodules; secondly, by structuring the report, important information such as the size, location, and morphology of the nodules is extracted and formatted from the free text to ensure that this information can be efficiently recognized and parsed by the model. In this embodiment, this information is structured into JSON data format; finally, professional word segmentation tools in the field are used. The tool and medical terminology library are used to segment and tag the text, especially for core medical terms such as "thyroid nodule", "hypoechoic", and "size", and perform accurate vocabulary tagging and semantic encoding. It should be noted that text segmentation and part-of-speech tagging are important steps in data preprocessing. Especially when processing medical text, it is necessary to focus on nodule-related professional terms and medical terms (such as "nodule", "size", "morphology", etc.). By segmenting and tagging the text, it is ensured that the model can accurately identify and parse the medical terms and their contextual relationships in clinical reports, thereby improving the accuracy and effectiveness of nodule feature extraction. At the same time, this embodiment constructs a medical knowledge base based on LangChain, including medical standards, diagnostic specifications, professional term interpretations, etc., to enhance the text parsing ability in the preprocessing stage. This deep preprocessing in the medical field greatly improves the model's parsing ability in complex semantic environments, laying a solid foundation for subsequent links. Step 2: Nodule information extraction using the ChatGLM3 model. When using the ChatGLM3 model for nodule information extraction, the preprocessed ultrasound report data is input into a specially fine-tuned ChatGLM3 model. This model combines its advanced natural language processing capabilities and named entity recognition (NER) technology to accurately extract information related to thyroid nodules from the report. This information includes the specific location (e.g., "left thyroid"), size (e.g., "1.5 cm × 1.2 cm"), morphology (e.g., "clear boundaries"), and quantity (e.g., "single nodule" or "multiple nodules"). To further improve the accuracy of extraction, this example introduces a medical knowledge base built on LangChain. By comparing the descriptions in the report with the knowledge base content in real time, it provides contextual verification and semantic correction. This knowledge base covers a wealth of medical data, including standard term definitions, clinical diagnostic rules, and multiple expressions of common terms. This helps improve the model's understanding of medical context and ensures the accuracy and comprehensiveness of information extraction. Figure 3 This is a schematic diagram of the structure of the langchain external knowledge base provided by the present invention; Figure 4 The ChatGLM3 connection transformer structure provided by the present invention; Figure 5 This is the principle diagram of ChatGLM3 provided by the present invention; Step 3: Design prompt templates and interact with the knowledge base. Based on the ChatGLM3 model, professional prompt templates for ultrasound reports are designed and integrated with the LangChain medical knowledge base for interactive processing. These prompt templates are designed based on the diagnostic characteristics of thyroid nodules and the expression habits of medical reports, covering functions such as information extraction, unit conversion, and medical verification. As an example, the prompt template is: "Please extract relevant information about thyroid nodules based on the following ultrasound report content, including the size (length × width), location (left or right), and morphology (such as regular or irregular), and convert all size units to centimeters (cm). At the same time, verify the accuracy of the extracted information according to medical standards, and refer to the terms and standard definitions in the following medical knowledge base when necessary: {Medical Knowledge Base Content Example}. The output should be in a structured format, for example: Nodule location: left, Nodule size: 1.5cm × 1.2cm, nodule morphology: rules. The prompt template guides the ChatGLM3 model to generate structured output through clear instructions and examples. During this process, the LangChain medical knowledge base provides authoritative medical terminology and diagnostic standards as reference, such as standard nodule size units, nodule morphology definitions and classifications. Figure 6 This is the medical knowledge base format provided by the present invention; Step 4: Nodule size extraction and unit unification; Step 4 mainly involves nodule size extraction and unit unification. After extracting key information related to the nodule, the ChatGLM3 model is used to complete the unification of nodule size units. Ultrasound reports usually contain multiple size units (such as millimeters (mm) and centimeters (cm)). Such unit inconsistencies may lead to errors in subsequent analysis results. The ChatGLM3 model first identifies all nodule sizes and their corresponding unit expressions through contextual semantic analysis and the support of the external LangChain medical knowledge base. Then, based on the unified standards in the medical field, the model converts all nodule sizes into standard units (such as centimeters). During the unit conversion process, the model combines the standard unit conversion rules recorded in the knowledge base to further verify the accuracy and logic of the conversion results, thereby ensuring the consistency and reliability of the data. This method not only solves the problem of unit inconsistency, but also significantly improves the efficiency and accuracy of multi-source data integration and subsequent analysis. Step 5: Nodule size comparison and analysis: After nodule size extraction and unit unification, the model will conduct a comparative analysis of nodule size. The goal of this step is to quantify and evaluate the size differences between different nodules to help doctors make clinical decisions. The ChatGLM3 model will compare the size differences between different nodules and classify the nodule sizes according to standards (such as small nodules, medium nodules, and large nodules). The model will also automatically detect the relative size differences between left and right thyroid nodules, as well as the changing trends of nodules in different ultrasound reports. If a nodule increases significantly at different time points, the model will automatically label it and provide relevant risk assessments. In addition, the model will also combine the nodule morphology (regular or irregular) and other characteristics for a comprehensive analysis to determine the potential malignancy risk of the nodule. Step 6: Nodule risk assessment and decision support. During this stage, the ChatGLM3 model will assess nodule risk based on characteristics such as nodule size, morphology, and location. By combining the medical knowledge base with pre-set diagnostic criteria, the model can determine whether the nodule has malignant characteristics and provide doctors with further examination recommendations (such as whether a puncture biopsy is necessary). Nodules larger than a certain size (such as 2 cm) are marked as high risk, indicating that surgical intervention or further imaging examinations may be required. The results of this step will provide doctors with detailed decision support to help them choose the appropriate diagnosis and treatment plan based on the nature and risk level of the nodule. Step 7: Accuracy evaluation of the method; When evaluating the accuracy of the method, in order to ensure the effectiveness and reliability of the model, a manually annotated standard dataset will be used to compare with the model output results to evaluate the accuracy of nodule information extraction, unit unification, nodule size comparison and analysis; the main evaluation indicators include: accuracy (the proportion of nodule features correctly extracted by the model), recall rate (the proportion of all nodule information that the model can identify) and F1 score (the harmonic mean of accuracy and recall rate); in addition, the performance of the model on different cases and different datasets will be tested to ensure that the model has good generalization ability; finally, the accuracy, stability and efficiency of the model in actual clinical applications will be comprehensively evaluated to ensure its practicality and reliability in thyroid nodule assessment. Table 1 is a comparison of the results of different models in thyroid nodule information extraction, unit unification, and nodule size comparison (test set).

[0025] Table 1 Comparison of different models in thyroid nodule information extraction, unit unification, and nodule size comparison (test set)

[0026] Evaluation results showed that the ChatGLM3 model achieved 94.1% accuracy, 92.8% recall, and 93.4% F1 score in the thyroid nodule information extraction task, significantly outperforming traditional machine learning models (85.4%, 83.2%, and 84.3%). In the nodule unit unification task, the ChatGLM3 model achieved both accuracy and F1 score of 99.6%, comparable to the performance of manual annotation (99.2%), but with higher efficiency. In the nodule size comparison and analysis task, the ChatGLM3 model achieved an F1 score of 90.3%, exceeding the 84.0% of the traditional algorithm. In the nodule risk assessment task, the ChatGLM3 model achieved 92.7% accuracy and 91.5% F1 score, significantly improving over the 86.8% of manual annotation. Overall, the ChatGLM3 model demonstrated superior accuracy and reliability compared to traditional methods across all tasks, particularly in automated information extraction and analysis.

[0027] This embodiment provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for analyzing thyroid nodules based on a large model and serial ultrasound reports.

[0028] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A method for analyzing thyroid nodules based on a large model and serial ultrasound reports, characterized in that: The following steps are involved: S1: preprocessing the thyroid ultrasound report data to obtain preprocessed thyroid ultrasound report data; S2: fine-tuning the large language model based on the preprocessed thyroid ultrasound report data using a specific data set and a medical knowledge base to obtain a fine-tuned large language model; S3: Designing a professional prompt template based on the fine-tuned large language model; S4: Use the fine-tuned large language model to extract nodule information and unify units from the ultrasound reports to obtain standardized structured data; S5: Based on the standardized structured data, the extracted nodule information is compared and analyzed using the fine-tuned large language model to obtain a structured report.

2. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 1, characterized in that: Also includes: Performing accuracy verification based on the accuracy verification data set and the structured report to obtain an accuracy verification result; An evaluation is performed based on the structured report and the accuracy verification result, and the structured report is reviewed and optimized based on the evaluation result to obtain a final analysis report.

3. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 1, characterized in that: The preprocessing includes data cleaning, key information labeling, structural processing, text segmentation and part-of-speech tagging.

4. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 1, wherein: The large language model is ChatGLM3; the specific data set includes thyroid nodule descriptions, diagnostic criteria, and treatment plans; the medical knowledge base is built based on LangChain and includes a knowledge graph, a term base, and a vector database.

5. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 1, characterized in that: The professional prompt template includes prompt instructions, and the prompt instructions are used to identify key information, and the key information includes nodule size, unit, location and morphology.

6. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 1, characterized in that: The standardized structured data includes the size, location and morphology of the nodules.

7. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 1, characterized in that: The structured report includes size differences between nodules, classification results, and corresponding risk assessment.

8. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 2, wherein: The accuracy verification results include precision, recall and F1 score.

9. The method for analyzing thyroid nodules based on a large model and serial ultrasound reports according to claim 2, wherein: The evaluation includes nodule information extraction accuracy evaluation, unit unification correctness evaluation, nodule size comparison and analysis effectiveness evaluation, and overall process execution efficiency evaluation.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method for analyzing thyroid nodules based on a large model and serial ultrasound reports are implemented.

Images