PET report generation system based on segmented extraction reasoning learning

Abstract

The invention discloses a PET report generation system based on segmented extraction reasoning learning, and relates to the field of medical image analysis and natural language processing. The objective of the invention is to solve the problem that a language model is difficult to generate a high-quality report efficiently in the prior art. The PET report generation system based on segmented extraction reasoning learning comprises a training data set acquisition module, a model acquisition module and a to-be-tested PET report generation module. The training data set acquisition module is used for acquiring a PET report and constructing an SERI training data set based on the PET report; sERI is segmented extraction reasoning; the model acquisition module is used for training a model based on the SERI training data set to obtain a trained model; the to-be-tested PET report generation module is used for inputting a to-be-tested PET report into the trained model, and the trained model outputs extraction information and a reasoning conclusion of the PET report; the PET is positron emission tomography.

Description

Technical Field

[0001] The present invention relates to the field of medical image analysis and natural language processing (NLP), and in particular to a small model based on segmented extraction and reasoning inference (SERI) learning, which is used for generating a PET (positron emission tomography) report system. Background Art

[0002] In modern medical diagnosis, PET (positron emission tomography) is a commonly used imaging technology, widely used in the detection and monitoring of tumors, cardiovascular diseases and neurological diseases. PET scan reports are usually written manually by doctors based on the imaging results, describing the metabolic activities and functional characteristics of multiple anatomical regions and making a preliminary diagnosis. This process requires doctors to have rich medical knowledge and high experience to ensure the accuracy, completeness and logic of the report content. However, manual report generation is not only time-consuming, but also easily affected by human factors (such as fatigue or subjective judgment), resulting in errors or omissions.

[0003] With the development of artificial intelligence technology, the demand for automated generation of medical imaging reports continues to increase. In recent years, large language models (such as GPT-3, T5, etc.) have been applied to automatic report generation tasks. These models can convert imaging data into text reports through natural language generation technology, reducing the burden on doctors and improving work efficiency. However, the large language models that are mainly relied on in existing technologies still face several key problems in PET report generation:

[0004] Large computing resources and data requirements:

[0005] Large language models usually require a lot of computing resources and huge training data. In the field of medical imaging, especially PET reports of specific diseases (such as cancer), the amount of data is often small, which makes it impossible to fully train large models. At the same time, running these models requires high-performance hardware devices (such as GPU or TPU), which is costly for clinical applications.

[0006] Lack of efficient data utilization mechanism:

[0007] In the existing technology, many methods rely on the ability to generate more high-quality data or distill large models to improve the quality of report generation. However, these methods fail to fully explore the potential information in the existing data, especially in medical imaging reports, where there are a large number of structured image descriptions and reasoning conclusions that are not efficiently utilized.

[0008] Difficulty capturing multi-step reasoning and logical summaries:

[0009] PET reports usually not only need to describe a single imaging result, but also need to reason and summarize the metabolic activities of multiple anatomical regions to determine the nature or development trend of the disease. Although existing large-scale models can generate text content, they are not ideal in complex multi-step reasoning and logical coherence, which can easily lead to a lack of medical logic in the report content and affect the accuracy of clinical diagnosis.

[0010] Limitations of small models:

[0011] Although large language models have excellent performance, their deployment cost and computing requirements are not conducive to their widespread application in actual medical environments. Therefore, the industry has also tried to transplant the chain-of-thought (CoT) capability into small models (such as T5-small or T5-base), but small models perform poorly on complex reasoning tasks and find it difficult to effectively learn these complex reasoning steps, resulting in low quality of generated reports.

[0012] In summary, the existing technology mainly has the following problems:

[0013] High dependence on data: Existing technologies usually require a large amount of data for model training when generating medical imaging reports. However, in certain fields (such as PET imaging), data resources are limited, which limits the performance of the model.

[0014] Difficulty in achieving efficient reasoning: Generating medical imaging reports requires complex reasoning and diagnostic conclusions based on imaging results. Although existing large models are capable of generating reports, they lack efficient reasoning capabilities, especially in small models where it is difficult to achieve accurate multi-step reasoning.

[0015] High computing resource cost: Large models require expensive computing resources, which makes it difficult to deploy them on a large scale in medical scenarios. Although existing small models have advantages in resource requirements, their reasoning capabilities are limited and the quality of the generated reports is low.

[0016] Low efficiency of information utilization: Existing report generation methods fail to make full use of existing chain reasoning data and lack full mining and efficient use of the intrinsic information of the data. Summary of the invention

[0017] The present invention proposes a PET report generation system based on segmented extraction reasoning learning. By efficiently segmenting the chain reasoning data, the content in the image report is divided into extraction segments and reasoning segments. This method can effectively improve the performance of small models in complex reasoning tasks and solve the problem in the prior art that language models are difficult to efficiently generate high-quality reports.

[0018] The PET report generation system based on segmented extraction inference learning includes:

[0019] Training data set acquisition module, model acquisition module, and PET report generation module;

[0020] The training data set acquisition module is used to obtain PET reports and build a SERI training data set based on the PET reports;

[0021] SERI stands for segment extraction reasoning;

[0022] The model acquisition module is used to train the model based on the SERI training data set to obtain the trained model;

[0023] The PET report generation module is used to input the PET report to be tested into the trained model, and the trained model outputs the extracted information and reasoning conclusion of the PET report;

[0024] The PET is positron emission tomography.

[0025] The beneficial effects of the present invention are:

[0026] The purpose of the present invention is to solve the shortcomings of the existing small language model in PET report generation, and propose a small model PET report generation method based on segmented extraction-reasoning (SERI) learning. The method decomposes the data in the PET report into extractive segments (ES) and reasoning segments (RS), and iteratively trains them to improve the model's ability to generate medical imaging reports.

[0027] The present invention proposes an effective solution to the problems in the prior art by introducing the segmented extraction-inference learning (SERI) method, especially the requirements of PET report generation and the limitations of small language models in multi-step reasoning tasks. The advantages of the present invention can be summarized as follows:

[0028] 1. Improve the reasoning ability of small models;

[0029] Small language models usually perform poorly when dealing with complex reasoning tasks, especially in the generation of medical imaging reports, which require combining imaging information from multiple anatomical regions while performing reasoning and summarization. This paper introduces a step-by-step generation mechanism for small models by dividing the content in PET reports into extraction segments (ES) and reasoning segments (RS). Specifically, it has the following advantages:

[0030] Phased generation: The model first generates the extraction segment, and then generates the reasoning segment based on the extraction segment, making the reasoning process more logical and controllable, and solving the logical confusion problem that may occur when a small model generates a full report at one time.

[0031] Multi-step reasoning capability: Through segmented processing, the small model can process image information from multiple anatomical regions and combine this information to generate an overall report inference. This step-by-step reasoning method greatly improves the performance of the small model in complex logical tasks, allowing it to approach the reasoning effect of the large model.

[0032] 2. Efficient use of limited data resources;

[0033] Medical imaging report generation, especially PET report generation, often faces the problem of insufficient data. Large language models perform poorly when data is scarce, and the present invention significantly reduces the reliance on additional data generation by efficiently utilizing existing chain reasoning data. This advantage is reflected in the following aspects:

[0034] No need for additional data expansion: By effectively segmenting existing image data, the present invention can fully utilize the intrinsic information of existing data without relying on generating more high-quality CoT data. This reduces data acquisition costs and time and improves data utilization.

[0035] Improved efficiency of iterative learning: By breaking down complex reasoning processes into multiple steps, the model can learn detailed extraction and reasoning processes in each step. This fine-grained learning method enhances the model's ability to understand and reason about limited data.

[0036] 3. Reduce computing resource requirements;

[0037] Although large language models perform well in many NLP tasks, their huge computing resource requirements limit their application in actual medical scenarios. In contrast, the present invention optimizes the performance of small models through the SERI method, enabling them to generate high-quality medical reports with limited computing resources:

[0038] The computational cost is significantly reduced: By reasonably allocating the computational tasks of the model, the present invention does not require the high-performance hardware support required by large language models and is suitable for ordinary computing resource environments. Especially in medical scenarios, the use of small models can reduce deployment costs.

[0039] Strong adaptability: This method can be deployed in a low-configuration hardware environment, such as a regular server in a hospital or a cloud computing platform, which makes this technology have the potential for large-scale promotion and application.

[0040] 4. Enhance the accuracy and consistency of report generation;

[0041] Medical imaging reports are not just a simple description of imaging data, but often require reasoning and drawing medical conclusions based on multiple aspects of the imaging. The present invention processes imaging data in segments, making the generated reports more accurate and logically coherent:

[0042] Improved accuracy of information extraction: By introducing the division of extraction segments (ES) and reasoning segments (RS), the model can distinguish between image description and inference, thus avoiding the problem of confusing image data with diagnostic inferences. This helps to improve the accuracy of reports, making the generated reports highly consistent with the actual image results.

[0043] Enhanced logical coherence: The segmented generation mechanism ensures the logical order of report generation. From description to inference to diagnosis, the entire process is clearly structured, avoiding the logical jumps or incoherent reasoning that may occur when traditional one-time report generation is used.

[0044] 5. Adapt to various medical imaging report generation tasks;

[0045] Although the present invention takes PET report generation as the core application scenario, the method has wide adaptability. Since many medical imaging reports (such as CT reports and MRI reports) also need to analyze and reason about multiple anatomical regions, the method of the present invention can be extended to different types of medical imaging report generation tasks:

[0046] Strong scalability: The SERI learning method is not only applicable to PET reports, but can also be flexibly extended to other imaging report generation tasks. Without changing the core architecture of the method, it can quickly adapt to different report types.

[0047] Task complexity support: The present invention can handle complex imaging reporting tasks ranging from a single organ to multiple organs, and is particularly suitable for diagnostic tasks that require the integration of multiple imaging results, such as tumor metastasis analysis.

[0048] 6. Reduce human errors and subjective bias;

[0049] The traditional manual report generation method may be affected by subjective factors such as doctor's experience and fatigue, which may easily lead to inconsistent descriptions or insufficient reasoning in the report. This invention reduces human intervention and improves the reliability and consistency of the report by automatically generating reports:

[0050] Automated generation reduces errors: The systematic report generation process ensures that each report generated is based on unified reasoning rules, avoiding deviations caused by personal differences between different doctors.

[0051] Greater consistency: The reports generated by the machine learning model are more consistent in language style and logical structure, which is particularly important when processing a large number of imaging reports.

[0052] 7. Ability to provide further diagnostic suggestions;

[0053] The present invention can not only generate accurate imaging reports, but also generate suggestions for future diagnostic steps through reasoning segments. For example, in PET images, the system can give a reasoning segment of "suggesting further CT or MRI examination" based on the abnormal degree of metabolic activity. This capability improves the practicality of the system and helps doctors quickly formulate the next step of diagnosis and treatment plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of the system of the present invention. DETAILED DESCRIPTION

[0055] Specific implementation method 1: The PET report generation system based on segmented extraction inference learning in this implementation method includes:

[0056] Training data set acquisition module, model acquisition module, and PET report generation module;

[0057] The training data set acquisition module is used to obtain PET reports and build a SERI training data set based on the PET reports;

[0058] SERI stands for segment extraction reasoning;

[0059] The model acquisition module is used to train the model based on the SERI training data set to obtain the trained model;

[0060] The PET report generation module is used to input the PET report to be tested into the trained model, and the trained model outputs the extracted information and reasoning conclusion of the PET report;

[0061] The PET is positron emission tomography.

[0062] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that: the training data set acquisition module is used to acquire a PET report and construct a SERI training data set based on the PET report;

[0063] The specific process is:

[0064] 1. Obtain a PET report, segment the PET report according to anatomical regions, and obtain paragraphs for each anatomical region;

[0065] The purpose of this step is to logically divide the contents of the whole body report according to organs or specific anatomical regions, so that the contents of each region are independent and easy to process later;

[0066] 1.2. Break down the paragraphs of each anatomical region obtained one by one into finer clauses, and all clauses form a list of clauses;

[0067] After identifying and segmenting the anatomical regions, the next step is to further decompose the paragraphs of each anatomical region into finer clause units; the purpose of this step is to prepare for the subsequent classification of the extraction segment (ES) and reasoning segment (RS);

[0068] 13. Preliminarily classify each clause in the clause list obtained in 12 to obtain the category of each clause in the clause list;

[0069] 14. Label each clause according to the category of each clause in the clause list obtained in 13, save the labeled clauses, and obtain the SERI training data set.

[0070] The other steps and parameters are the same as those in the first embodiment.

[0071] Specific implementation method three: This implementation method is different from specific implementation method one or two in that: in the specific implementation method one or two, a PET report is obtained, and the PET report is segmented according to anatomical regions to obtain paragraphs of each anatomical region;

[0072] The purpose of this step is to logically divide the contents of the whole body report according to organs or specific anatomical regions, so that the contents of each region are independent and easy to process later;

[0073] The specific process is:

[0074] 1.1.1. Identify the paragraphs of each anatomical region from the PET report based on natural language processing technology;

[0075] Typically, a PET report will scan and describe multiple organs or regions in the body, such as the liver, pancreas, lungs, lymph nodes, etc. In this step, you first need to identify the paragraphs for each anatomical region from the report. For example, the report may have the following description:

[0076] Liver: No obvious metabolic abnormalities were found in the liver.

[0077] Pancreas: Abnormally increased metabolic activity in the head of the pancreas.

[0078] Lungs: No abnormal metabolic lesions were found in the lungs.

[0079] Through natural language processing technology, these regions can be automatically divided using keyword extraction or regular expression-based matching algorithms. A set of keywords or terms corresponding to common anatomical regions can be defined, such as "liver", "pancreas", "lungs", etc., and used as markers for segmentation.

[0080] 112. Segment each identified anatomical region into independent segments;

[0081] The paragraphs describing each anatomical region contain two parts:

[0082] Describe the imaging findings for each anatomical region;

[0083] Summarize the initial diagnostic impression for each anatomical region.

[0084] Divide the report into separate paragraphs describing each anatomical region. Each paragraph should include all the information about that anatomical region. For each anatomical region, the paragraph will usually contain two parts:

[0085] Description of imaging findings in the area (e.g., description of normal or abnormal metabolism);

[0086] Summarize the initial diagnostic impression (eg, possible pathology inferred) for the area.

[0087] The other steps and parameters are the same as those in the first or second embodiment.

[0088] Specific implementation method 4: This implementation method is different from the first one of the specific implementation methods 1 to 3 in that: in the first two, the paragraphs of each anatomical region obtained one by one are decomposed into finer clauses, and all the clauses form a clause list;

[0089] After identifying and segmenting the anatomical regions, the next step is to further decompose the paragraphs of each anatomical region into finer clause units; the purpose of this step is to prepare for the subsequent classification of the extraction segment (ES) and reasoning segment (RS);

[0090] The specific process is:

[0091] 121. Decompose the paragraphs of each anatomical region obtained one by one into clauses based on punctuation segmentation tools (such as NLTK library or spaCy library in Python);

[0092] Punctuation segmentation:

[0093] Use punctuation segmentation tools (such as the NLTK library or spaCy library in Python) to split paragraphs into smaller clauses; common punctuation marks include period ("."), comma (","), semicolon (";"), etc. Punctuation marks can serve as natural segmentation points, making complex sentences simpler and easier to process later;

[0094] For example, the sentence "No obvious metabolic abnormalities were found in the liver, but the metabolic activity of the pancreatic head was abnormally increased." can be divided into two clauses:

[0095] Clause 1: "No obvious metabolic abnormalities were found in the liver."

[0096] Clause 2: "The metabolic activity of the head of the pancreas is abnormally increased."

[0097] 122. Clean each clause to obtain a cleaned clause;

[0098] The cleaning is to remove redundant spaces in the clauses;

[0099] The cleaned clauses are standardized in terms of units and terms to obtain standardized clauses;

[0100] After the clause segmentation, the text needs to be cleaned and preprocessed to remove noise and unnecessary symbols. Common processing steps include:

[0101] Remove spaces and punctuation: Remove extra spaces in clauses to maintain the semantic integrity of the clauses;

[0102] Standardized units and terminology: for example, all radioactive uptake values ​​can be standardized into a unified unit, or certain pathological terms can be unified (e.g., "increased activity" can be simplified to "abnormal metabolism");

[0103] The cleaned and preprocessed clauses should be well-formatted phrases or sentences to facilitate subsequent classification;

[0104] 123. Form a clause list based on the standardized clauses.

[0105] After segmentation and cleaning, a list of clauses is finally formed, each clause corresponding to a specific description of an anatomical region.

[0106] For example, for a PET report containing information about the liver and pancreas, the list of clauses after segmentation might be:

[0107] Clause 1: "No obvious metabolic abnormalities were found in the liver."

[0108] Clause 2: "The metabolic activity of the head of the pancreas is abnormally increased."

[0109] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0110] Specific implementation method 5: This implementation method is different from the first one of the specific implementation methods 1 to 4 in that: in the first three, each clause in the clause list obtained in the first two is preliminarily classified to obtain the category of each clause in the clause list; the specific process is:

[0111] 131. Definition: An Extract Segment (ES) is a clause that directly describes the image observation results;

[0112] Definition The reasoning segment (RS) is the preliminary diagnosis or inference based on the imaging results;

[0113] 132. Classify each clause in the clause list obtained in 12 based on a classification method, and determine whether each clause is an extraction segment or an inference segment;

[0114] After the clause segmentation is completed, each clause is preliminarily classified and marked as an extraction segment (ES) or a reasoning segment (RS). The specific classification basis is as follows:

[0115] 1. Identification of extracted segments:

[0116] Extract segments (ES) are defined as clauses that directly describe the observations of an image. These clauses usually contain quantitative or qualitative observation data without involving reasoning or analysis. For example:

[0117] “No obvious metabolic abnormalities were found in the liver.”

[0118] "The metabolic activity of the head of the pancreas is abnormally increased."

[0119] These clauses primarily provide physical or functional observations of the images and are direct interpretations of the image data.

[0120] 2. Identification of reasoning segments:

[0121] A reasoning paragraph (RS) is defined as a preliminary diagnosis or inference based on the imaging results. This type of clause usually contains inferential, concluding or hypothetical content. For example:

[0122] "The metabolic activity of the pancreatic head is abnormally increased, and the possibility of malignancy is considered."

[0123] “Further investigation is recommended to rule out tumor metastasis.”

[0124] These clauses not only describe the imaging results, but also further deduce possible pathological diagnosis results or make suggestions for the next step of treatment.

[0125] 3. Selection of classification method:

[0126] Rule-based classification: Use predefined rules (such as keyword matching or sentence analysis) to identify whether each clause in the clause list is an extractive segment or an inference segment. For example, clauses containing "abnormal metabolism" and "abnormal absorption" that describe physical characteristics can be classified as extractive segments, while clauses containing inferential words such as "consider" and "suggest" can be classified as inference segments.

[0127] Classification of machine learning models: You can also use trained classification models (such as SVM, logistic regression, etc.) to identify whether each clause in the clause list is an extraction segment or an inference segment.

[0128] The model is trained to distinguish between extractive and inference segments using manually annotated training data. The model can classify based on lexical features, sentence structure, and contextual information in the clauses.

[0129] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0130] Specific implementation method 6: This implementation method is different from the first to fifth implementation methods in that: in the fourth implementation method, each clause in the clause list obtained in the third implementation method is labeled according to its category, and the labeled clauses are saved to obtain a SERI training data set; the specific process is as follows:

[0131] 141. label each clause according to the category of each clause in the clause list obtained in 13, labeling it as an extraction segment or an inference segment;

[0132] 142. Save each clause in the marked clause list;

[0133] Save format:

[0134] Region, Clause Text, Category.

[0135] Labeling completed: After completing the above steps, each anatomical region paragraph will be segmented into multiple clauses, and each clause has been labeled as an extraction segment or an inference segment.

[0136] The final output should be a list of classified clauses. For example:

[0137] Clause 1: "No obvious metabolic abnormalities were found in the liver." (Extracted ES)

[0138] Clause 2: "The metabolic activity of the pancreatic head is abnormally increased." (Extracted ES)

[0139] Clause 3: "Consider the possibility of malignancy." (Inference RS)

[0140] Save results: Save the annotation results in a dataset format that can be used for subsequent training. The annotation information of each clause should include the anatomical region identifier, clause text, and clause category (ES or RS).

[0141] Annotation clause:

[0142] After classification, all clauses are labeled as either extractable or inference segments. The output includes the text of each clause, the anatomical region identification, and the category of the clause (ES or RS).

[0143] Save the classification results:

[0144] Save the classification results as a structured file (such as CSV, JSON format) to facilitate the subsequent construction of the SERI dataset. File format example:

[0145] area, clause text, category

[0146] Liver: No obvious metabolic abnormalities were found in the liver.

[0147] Pancreas, pancreatic metabolic activity, ES

[0148] Pancreas, consider the possibility of malignancy, RS

[0149] In this step, by using various methods such as entropy, loss or similarity, combined with predefined rules, the clauses of the image report are accurately classified into extraction segments (ES) or reasoning segments (RS). This process lays a solid foundation for subsequent model training and iterative generation.

[0150] Through this step, the complete PET report has been segmented into multiple clauses divided by anatomical regions, and each clause has been classified into extraction segments and reasoning segments. This step provides a basic data set for subsequent model training and iterative generation. In implementation, the segmentation and classification rules can be flexibly adjusted according to the structure of the actual report to ensure accurate processing of the report.

[0151] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0152] Specific implementation method 7: This implementation method is different from specific implementation methods 1 to 6 in that: in the above-mentioned 132, each clause in the clause list obtained in 12 is classified based on the classification method to determine whether each clause is an extraction segment or an inference segment; the specific process is:

[0153] 1. Classify each clause in the clause list obtained in step 1 based on the rules to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows:

[0154] 1. Build a keyword library; the specific process is:

[0155] Based on the common words and expressions in the PET report, construct the keyword library of the extraction segment and the keyword library of the inference segment;

[0156] Extract segment keywords: for example, "no abnormality", "active metabolism", "increased metabolic abnormality", "normal", "no abnormality", etc.

[0157] Keywords for reasoning paragraphs: for example, "consider", "possibly", "speculate", "suggest", etc.

[0158] 12. Determine whether each clause in the clause list obtained in 12 contains a word in the keyword library of the extraction segment. If it contains a word in the keyword library of the extraction segment, the clause is an extraction segment; if not, the clause is not an extraction segment;

[0159] Determine whether each clause in the obtained clause list contains a word in the inference segment keyword library. If it contains a word in the inference segment keyword library, the clause is an inference segment; if not, the clause is not an inference segment.

[0160] For example, in the clause "The pancreas has active metabolism, and the possibility of malignant tumor is considered", "active metabolism" belongs to the extraction segment keyword, and "consider" belongs to the reasoning segment keyword. Therefore, it can be divided into two clauses and marked as the extraction segment and the reasoning segment respectively:

[0161] “Pancreatic metabolically active” was marked as the extracted segment (ES).

[0162] “Consider the possibility of malignancy” is marked as the reasoning segment (RS).

[0163] If a clause contains both an extraction segment and an inference segment, the segmentation algorithm is applied to further split the clause to ensure that each clause can only belong to one of the extraction segment and the inference segment. The previous clause segmentation has solved this problem.

[0164] For example, by combining punctuation and natural language segmentation tools (such as NLTK and spaCy), complex clauses can be further split to ensure that each clause can only belong to one of the extraction segments or the inference segment;

[0165] 2. Classify each clause in the clause list obtained in step 1 based on the entropy value to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows:

[0166] Input each clause in the obtained clause list into the trained language model, and the trained language model outputs the predicted probability of each clause;

[0167] Calculate the entropy value of each clause based on its predicted probability;

[0168] Set the entropy threshold;

[0169] If the entropy value of the clause is less than or equal to the entropy threshold, the clause is marked as an extraction segment ES;

[0170] If the entropy value of a clause is greater than the entropy threshold, the clause is marked as a reasoning segment RS;

[0171] The entropy value expression is:

[0172]

[0173] Among them, P(R i ) is the predicted probability of each clause output by the trained language model;

[0174] R i is the i-th word in each clause; Q is the question; P(R i |Q) is the conditional probability; H n is entropy; i is the i-th; V is the vocabulary;

[0175] The trained language model is a language model trained using the originally acquired PET report, or a common pre-trained model;

[0176] For example, the entropy value of the clause "The liver has no obvious abnormalities" may be low because it directly describes the imaging results and the model has a high degree of certainty in its generation. Therefore, this clause will be marked as an extracted segment.

[0177] 3. Classify each clause in the clause list obtained in step 1 based on the loss value to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows:

[0178] 3.1. Calculation of loss value:

[0179] Input each clause in the obtained clause list into the trained language model, and the trained language model outputs the predicted probability of each clause;

[0180] Calculate the loss value based on the predicted probability and the true probability;

[0181] Use cross entropy loss or other loss functions to evaluate the quality of the model's generation of this sentence. 32.

[0183] Set the loss value threshold:

[0184] If the loss value of a clause is higher than the loss value threshold, the clause is marked as reasoning segment RS;

[0185] If the loss value of a clause is less than or equal to the loss value threshold, the clause is marked as an extraction segment ES;

[0186] 4. Classify each clause in the clause list obtained in step 1 based on similarity to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows:

[0187] 4.1. Similarity calculation:

[0188] Calculate the similarity between each clause and its adjacent clauses (or input image descriptions), which can be measured using text similarity evaluation indicators such as BLEU or ROUGE;

[0189] For example, the clause “the pancreas has active metabolism” is directly related to the “pancreas” part in the imaging report, so it has a high similarity and belongs to the extraction segment ES; while “consider the possibility of malignant tumor” has no direct correspondence with the context and belongs to the inference segment RS; 42.

[0191] Set similarity threshold;

[0192] If the similarity between a clause and its adjacent clauses is higher than the similarity threshold, the clause is marked as an extracted segment ES;

[0193] Otherwise, the clause is marked as a reasoning segment RS.

[0194] 2. Detailed description of the classification into extraction segment ES and reasoning segment RS

[0195] The purpose of this step is to classify the PET report clauses segmented in step 1 into extractive segments (ES) and reasoning segments (RS) based on specific algorithms and indicators, so as to subsequently construct the SERI dataset and perform model training.

[0196] 2.1 Feature definition of extraction segment (ES) and reasoning segment (RS);

[0197] In the classification process, we must first clarify the definitions and characteristics of the extraction segment and the reasoning segment:

[0198] Extract Segments (ES): These segments directly describe the imaging observations, usually quantitative or qualitative statements about the state of an organ or tissue in a PET scan. These statements directly reflect the observed phenomena or data.

[0199] Characteristics: Usually contains clear observational data or image information, and the language is relatively direct and declarative.

[0200] Keywords: such as "no abnormality", "active metabolism", "no abnormal metabolic absorption", etc.

[0201] Reasoning paragraphs (RS): These paragraphs draw inferences or conclusions based on the observational data, often providing a preliminary diagnosis, hypothesis, or recommendation.

[0202] Characteristics: Involves logical reasoning, diagnostic conclusions, or recommendations for future action; language is speculative or concluding.

[0203] Keywords: such as "consider", "possible", "speculate", "recommend further examination", etc.

[0204] 2.2 Preliminary rule-based classification method

[0205] In order to quickly classify clauses, a rule-based classification method can be used first. This method divides clauses into extraction segments or inference segments through keyword matching and language pattern recognition.

[0206] Keyword library construction:

[0207] Based on common words and expressions in a large number of PET reports, a keyword library related to the extraction segment and the inference segment is constructed.

[0208] Extract segment keywords: for example, "no abnormality", "active metabolism", "increased metabolic abnormality", "normal", "no abnormality", etc.

[0209] Keywords for reasoning paragraphs: for example, "consider", "possibly", "speculate", "suggest", etc.

[0210] Matching rules apply:

[0211] Perform keyword matching on each clause to see if it contains words in the keyword library. If a clause contains an extraction segment keyword, the clause is marked as an extraction segment (ES); if it contains a reasoning segment keyword, it is marked as a reasoning segment (RS).

[0212] For example, in the clause "The pancreas has active metabolism, and the possibility of malignant tumor is considered", "active metabolism" belongs to the extraction segment keyword, and "consider" belongs to the reasoning segment keyword. Therefore, it can be divided into two clauses and marked as the extraction segment and the reasoning segment respectively:

[0213] “Pancreatic metabolically active” was marked as the extracted segment (ES).

[0214] “Consider the possibility of malignancy” is marked as the reasoning segment (RS).

[0215] Boundary condition processing:

[0216] For clauses containing multiple keywords, segmentation algorithms are applied to further segment them. For example, punctuation marks and natural language word segmentation tools (such as NLTK and spaCy) are combined to further split complex clauses to ensure that each clause can only belong to one of the extraction segment and the inference segment.

[0217] 2.3 Classification method based on entropy

[0218] In addition to rule-based classification, entropy can also be used to measure the model's certainty about the generated content and distinguish between extraction segments and inference segments based on certainty. The lower the entropy value, the more certain the model is about the generated result, which usually means that the clause is an extraction segment. The higher the entropy value, the more uncertain the model is about the result, which may belong to the inference segment.

[0219] Entropy calculation:

[0220] Use the trained language model to input each clause in the report and calculate the entropy value of each clause when it is generated. Specifically, the model generates a predicted distribution for each clause, and the entropy value reflects the uncertainty of the distribution;

[0221] Classification by:

[0222] By setting the entropy threshold β, if the entropy value of a clause is lower than or equal to β, the clause is marked as an extraction segment (ES); if the entropy value is higher than β, it is marked as a reasoning segment (RS).

[0223] For example, the entropy value of the clause "The liver has no obvious abnormalities" may be low because it directly describes the imaging results and the model has a high degree of certainty in its generation. Therefore, this clause will be marked as an extracted segment.

[0224] 2.4 Classification method based on loss value

[0225] Similar to entropy, loss is another criterion used for classification. In the model generation process, the loss reflects the error of the model in generating a certain sentence. The extraction segment is often directly related to the input context, so the model's loss value is lower; while the reasoning segment involves more reasoning, so the loss value is higher.

[0226] Loss value calculation:

[0227] During model training or inference, record the loss value of each clause. Use cross entropy loss or other loss functions to evaluate the model's generation quality for the clause.

[0228] Loss value threshold setting:

[0229] By setting a threshold for the loss value, if the loss value of a clause is higher than the threshold, it is marked as a reasoning segment (RS); a clause with a lower loss value is marked as an extraction segment (ES).

[0230] 2.5 Similarity-based classification method

[0231] In order to further improve the accuracy of classification, a similarity-based method can be introduced. The extraction segment is usually information directly extracted from the input context, so the similarity with the context is high; the inference segment is derived based on the context, so the similarity is low.

[0232] Similarity calculation:

[0233] Calculate the similarity between each clause and its context (or input image description), which can be measured using text similarity evaluation metrics such as BLEU or ROUGE.

[0234] For example, the clause “pancreas has active metabolism” is directly related to the “pancreas” part in the imaging report, so it has a high similarity and belongs to the extraction segment (ES); while “consider the possibility of malignant tumor” has no direct correspondence with the context and belongs to the reasoning segment (RS).

[0235] Classification by:

[0236] By setting a similarity threshold, if the similarity between a clause and the context is higher than the threshold, it is marked as an extracted segment (ES); otherwise, it is marked as a reasoning segment (RS).

[0237] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0238] Specific implementation eight: This implementation differs from specific implementations one to seven in that: in the second implementation, the model is trained based on the SERI training data set to obtain a trained model; the specific process is:

[0239] Step 1: Set language model hyperparameters; language model such as BART;

[0240] Step 2: Take the PET report in the SERI training data set as the input of the language model, and the language model outputs the extracted segment ES; the language model is such as BART;

[0241] Step 3: The extracted segment is used as the input of the language model, and the model outputs the reasoning segment RS;

[0242] Step 4: Repeat steps 2 to 4, use the Adam optimizer to minimize the loss function through gradient descent. When the output of the language model contains the preset stop sign, the iterative generation process is automatically terminated, the language model outputs a complete report, and a trained language model is obtained.

[0243] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0244] Specific implementation method 9: This implementation method is different from the specific implementation methods 1 to 7 in that: in the second implementation method, the model is trained based on the SERI training data set to obtain a trained model; the specific process is:

[0245] Step 1: Set ES model hyperparameters;

[0246] Step 2: The PET reports in the SERI training dataset are used as the input of the ES model, and the ES model outputs the extracted segment (ES);

[0247] Step 3: Repeat steps 2 and 3, use the Adam optimizer to minimize the loss function through gradient descent. When the output of the ES model contains the preset stop sign, the iterative generation process automatically terminates, and the ES model outputs the complete extraction segment to obtain a trained ES model.

[0248] The ES model is a language model, and the language model is BART;

[0249] Step 4: Set RS model hyperparameters;

[0250] Step 5: The complete extracted segment output by the ES model is used as the input of the RS model, and the RS model outputs the reasoning segment (RS);

[0251] Step 6: Repeat steps 5 and 6, use the Adam optimizer to minimize the loss function through gradient descent. When the output of the RS model contains the preset stop sign, the iterative generation process automatically terminates, and the RS model outputs a complete inference segment to obtain a trained RS model.

[0252] Obtain a trained model based on the trained ES model and the trained RS model;

[0253] The RS model is a language model, and the language model is BART.

[0254] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0255] Step 3: Build the SERI dataset and perform iterative training

[0256] 3.1 Construction of SERI Dataset

[0257] The SERI dataset is constructed based on classified extracted segments (ES) and reasoning segments (RS). This dataset will be used for model training to help the model gradually generate extracted information and reasoning conclusions of PET reports.

[0258] Data Organization:

[0259] Data for each anatomical region should include:

[0260] Input: Image description or patient imaging examination information.

[0261] Output: corresponding extraction segment (ES) and reasoning segment (RS).

[0262] The data format should be organized according to the input-output structure of the model. For each anatomical region, the imaging description of the region is first input, and then the model generates the corresponding extraction segment and inference segment.

[0263] ES data sample:

[0264] Input: Image description (e.g. "pancreas image").

[0265] Output: Extracted segment (e.g. "Pancreas is metabolically active").

[0266] RS data sample:

[0267] Input: Image description and extracted segments (e.g. “Pancreas is metabolically active”).

[0268] Output: Reasoning paragraph (such as "Consider the possibility of malignant tumor").

[0269] Input and output formats:

[0270] The input can be the original descriptive information in the imaging examination or the patient data organized by the system (including imaging observation information), and the output is the joint generation of the extraction segment and the reasoning segment. The ES and RS in each data sample need to be clearly labeled and arranged in order for the model to learn.

[0271] Data sample:

[0272] Area, Input, Output

[0273] Liver, "Images showed no obvious metabolic abnormalities in the liver", "No obvious metabolic abnormalities in the liver"

[0274] Pancreas, "Images show active pancreatic metabolism", "Active pancreatic metabolism"

[0275] Pancreas, "Images show active pancreatic metabolism, considering the possibility of malignant tumors", "Considering the possibility of malignant tumors"

[0276] Labeling and segmentation:

[0277] The label information of each clause should be consistent with the classification result. According to the classification result of step 2, the extraction segment and inference segment of each clause should be marked with labels so as to distinguish the segments with different functions during the training process.

[0278] 3.2 Iterative training model

[0279] After the SERI dataset is built, the model will be trained through an iterative generation mechanism. The process includes training two independent tasks: generating extraction segments (ES) and generating reasoning segments (RS). The training process uses a recursive generation method to gradually generate a complete report through model iteration.

[0280] Single Path Learning (USM):

[0281] In the single-path learning framework, the model handles the generation tasks of the extraction segment and the reasoning segment at the same time. The model first generates the extraction segment (ES), and then generates the reasoning segment (RS) based on the generated ES segment. This process is recursive. After each ES is generated, the next RS is generated until the entire report is generated.

[0282] Generation process:

[0283] Step 1: The model inputs an image description and outputs an extracted segment (ES).

[0284] Step 2: Take the extracted segment as input and generate the reasoning segment (RS).

[0285] Step 3: Continue iterating and repeating the above process until a complete report is generated.

[0286] For example:

[0287] Input: "Imaging shows that the pancreas is metabolically active."

[0288] Output 1(ES): "Pancreas metabolically active."

[0289] Input 2: Add "Pancreas metabolically active" as a new input.

[0290] Output 2 (RS): "Consider the possibility of malignancy."

[0291] This single-path generation allows the model to recursively generate new content at each stage until the entire report is generated.

[0292] Dual Path Learning (DSM):

[0293] In the dual-path learning framework, two independent models are used to generate the extraction segment (ES) and the reasoning segment (RS). The two models are optimized independently to handle the extraction task of ES and the reasoning task of RS respectively.

[0294] Generation process:

[0295] Step 1: The ES model inputs the image description and outputs the extracted segment (ES).

[0296] Step 2: The RS model inputs the extraction segment and outputs the reasoning segment (RS).

[0297] The advantage of dual-path learning is that it can perform more specialized training for the extraction segment and the reasoning segment, and each model can be better optimized for its own task. The generation process is as follows:

[0298] ES model: input image description and generate extraction segments.

[0299] RS model: input extraction segment and generate reasoning segment.

[0300] Iteration details:

[0301] The generation process of each clause is iterative, that is, after generating a clause, the next clause is generated based on the clause until the entire report is completed.

[0302] 3.3 Model Optimization and Training Process

[0303] Loss function settings:

[0304] The error of each generation is calculated using the loss function of the standard sequence generation model (such as the cross entropy loss function). The generation loss of the extraction segment and the inference segment can be calculated separately or jointly optimized.

[0305] Paragraphs with lower loss values ​​mean higher generation accuracy, and the model should be optimized towards these directions.

[0306] Optimization strategy:

[0307] During the training process, use the Adam optimizer or other suitable optimization method to minimize the loss function by gradient descent.

[0308] Use early stopping mechanism to prevent the model from overfitting.

[0309] Hyperparameter tuning:

[0310] Learning rate: Adjust the learning rate of the model to ensure that the model converges during training but does not fall into a local optimum.

[0311] Training batch: Set a reasonable batch size so that the model can process enough training data in each iteration step.

[0312] 3.4 Stop sign during generation

[0313] Stop sign definition:

[0314] At each step of the generation process, a stop sign needs to be defined to indicate the end of the model generation task.

[0315] For example, in a PET report, a specific marker word (such as "end of report" or a specific symbol) can be used as a stop sign.

[0316] Termination conditions:

[0317] When the model detects that the generated text contains a predefined stop sign, the iterative generation process is automatically terminated and a complete report is output.

[0318] 3.5 Post-processing of report generation

[0319] Report formatting:

[0320] After the generation is completed, the generated PET report is formatted to ensure that the paragraph structure, grammar and medical terminology meet the standards.

[0321] Verification results:

[0322] After the report is generated, the model output should be verified manually or automatically to ensure the accuracy and completeness of the report content.

[0323] 3.6 Model Evaluation and Verification

[0324] Evaluation Metrics:

[0325] Use evaluation indicators in standard NLP generation tasks (such as BLEU, ROUGE, etc.) to evaluate the similarity between the text generated by the model and the real report.

[0326] In addition, based on specific tasks of imaging reports (such as diagnostic accuracy), the generated reports are evaluated to see whether they meet medical standards.

[0327] Validation set and test set:

[0328] During the training process, the validation set is used to monitor the model's real-time performance to prevent overfitting.

[0329] After training is completed, the test set is used to evaluate the final generation effect of the model and record various evaluation indicators.

[0330] This step enables the model to gradually generate a complete PET report by constructing a SERI dataset and performing iterative training. Single-path or dual-path generation methods can help the model extract key information from the image description and infer summary conclusions. This process ensures the consistency and accuracy of the report, and controls the termination of generation through a stop sign. During the training and optimization process, it is necessary to set loss functions, tune hyperparameters, define termination conditions and other techniques to ensure that the model can efficiently generate PET reports that meet medical standards.

[0331] The following examples are used to verify the beneficial effects of the present invention:

[0332] Embodiment 1:

[0333] In order to verify the effectiveness of the present invention, especially the performance improvement of the small model based on the segmented extraction-inference learning (SERI) method in the PET report generation task, the present invention implemented multiple experiments, compared and analyzed the differences between the present method and the prior art, and demonstrated the advantages of the method in medical imaging report generation. The following are the detailed steps of the implementation and the simulation experiment results.

[0334] Example 1: PET report generation task

[0335] This embodiment demonstrates the application of a small model based on the SERI learning method in the task of generating PET reports, especially generating automated reports for PET images of patients undergoing tumor detection.

[0336] Experimental setup:

[0337] Dataset: A real PET imaging report dataset from a large hospital was selected, which contains complete imaging reports of multiple patients. Each report describes the metabolic information of multiple anatomical regions (such as liver, lung, pancreas, etc.) and the corresponding doctor's diagnosis.

[0338] Task goal: Automatically generate image descriptions and reasoning conclusions that are consistent with real doctor reports, including extraction segments and reasoning segments for each anatomical region.

[0339] Model: Optimized based on T5-small model and applied SERI learning method for report generation.

[0340] Comparison method: Compare with the traditional large language model based on GPT-3 and the T5-small model that directly generates the entire report.

[0341] Data processing:

[0342] Data segmentation: The original PET report is segmented into multiple independent paragraphs according to anatomical regions, each of which contains the anatomical region description and diagnostic opinion. Then, natural language processing technology is used to further segment each paragraph into clauses.

[0343] Clause classification: According to the SERI method, each clause is classified into an extraction segment (ES) and a reasoning segment (RS). The extraction segment describes the specific information in the image, while the reasoning segment gives the doctor's diagnosis speculation and next step recommendation.

[0344] Model training:

[0345] Training method: The small T5 model is iteratively trained using the SERI method. The model is first trained to generate extraction segments (ES), and then based on the generated ES segments, the model is further trained to generate reasoning segments (RS).

[0346] Iterative generation: Use single path learning (USM) for recursive generation. The model first inputs the image data to generate the ES segment, and then uses the ES segment as input to generate the RS segment. The entire process is completed step by step on each anatomical area of ​​the report until the entire report is generated.

[0347] Simulation experiment process:

[0348] Training set: 80% of the data is used for model training, and the image description and doctor's reasoning conclusion of each report are used as label data.

[0349] Validation set and test set: Use the remaining 20% ​​of the data as validation set and test set to test the report generation ability of the model.

[0350] Evaluation indicators:

[0351] BLEU: used to evaluate the text similarity between the generated report and the doctor's report.

[0352] ROUGE: Used to evaluate the overlap between the generated text and the reference text, including vocabulary and phrase matching.

[0353] Medical consistency score: Professional doctors will manually evaluate the generated report to determine the medical logic and accuracy of the generated report.

[0354] Experimental results:

[0355] Method BLEU ROUGE-L medical consistency score (full score 10)

[0356] GPT-3 0.52 0.61 8.5

[0357] T5-small (without SERI) 0.35 0.47 6.8

[0358] T5-small(SERI)0.65 0.72 9.2

[0359] Improved BLEU and ROUGE-L scores: The T5-small model using the SERI method significantly outperforms the T5-small model without SERI in BLEU and ROUGE-L scores, and is close to the performance of the large model GPT-3.

[0360] Improved medical consistency score: Multiple doctors evaluated the medical logic of the generated reports, and the reports generated by the SERI method received higher scores. This shows that the reports generated by the SERI method through step-by-step reasoning are more consistent with medical logic and avoid inconsistent reasoning or omissions.

[0361] Qualitative analysis:

[0362] Report example analysis:

[0363] Real doctor report:

[0364] "There is no obvious metabolic abnormality in the liver, but the pancreas is metabolically active. We consider the possibility of pancreatic malignancy and recommend further CT examination."

[0365] GPT-3 generates a report:

[0366] "The liver has active metabolism, with no obvious metabolic abnormalities. The pancreas has active metabolism, so pancreatic malignancy is considered. Further examination is recommended."

[0367] T5-small (without SERI) generates a report:

[0368] "The liver has no obvious metabolic abnormalities, but the pancreas has active metabolism, which may be malignant."

[0369] T5-small(SERI) generates report:

[0370] "There is no obvious metabolic abnormality in the liver, but the pancreas is metabolically active. We consider the possibility of pancreatic malignancy and recommend further CT examination."

[0371] From the comparison, it can be seen that the reports generated by the SERI method are not only more similar to the real doctor's reports, but also logically coherent and the reasoning is more complete.

[0372] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A PET report generation system based on segmented extraction inference learning, characterized by: Said include: Training data set acquisition module, model acquisition module, and PET report generation module; The training data set acquisition module is used to obtain PET reports and build a SERI training data set based on the PET reports; SERI stands for segment extraction reasoning; The model acquisition module is used to train the model based on the SERI training data set to obtain the trained model; The PET report generation module is used to input the PET report to be tested into the trained model, and the trained model outputs the extracted information and reasoning conclusion of the PET report; The PET is positron emission tomography.

2. The PET report generation system based on segmented extraction inference learning according to claim 1 is characterized in that: The training data set acquisition module is used to acquire a PET report and construct a SERI training data set based on the PET report; The specific process is:

1. Obtain a PET report, segment the PET report according to anatomical regions, and obtain paragraphs for each anatomical region; 1.

2. Decompose the paragraphs of each anatomical region obtained one by one into clauses, and all clauses form a clause list; 13. Preliminarily classify each clause in the clause list obtained in 12 to obtain the category of each clause in the clause list; 14. Label each clause according to the category of each clause in the clause list obtained in 13, save the labeled clauses, and obtain the SERI training data set.

3. The PET report generation system based on segmented extraction inference learning according to claim 2 is characterized in that: In the above, the PET report is obtained, and the PET report is segmented according to the anatomical regions to obtain the paragraphs of each anatomical region; the specific process is as follows:

1. Identify the paragraphs of each anatomical region from the PET report based on natural language processing technology; 112. Segment each identified anatomical region into independent segments; The paragraphs describing each anatomical region contain two parts: Describe the imaging findings for each anatomical region; Summarize the initial diagnostic impression for each anatomical region.

4. The PET report generation system based on segmented extraction inference learning according to claim 3 is characterized in that: In the first two, the paragraphs of each anatomical region obtained one by one are decomposed into clauses, and all clauses form a clause list; the specific process is:

121. The paragraphs of each anatomical region obtained one by one were decomposed into clauses based on punctuation segmentation tools; 122. Clean each clause to obtain a cleaned clause; The cleaning is to remove redundant spaces in the clauses; The cleaned clauses are standardized in terms of units and terms to obtain standardized clauses; 123. Form a clause list based on the standardized clauses.

5. The PET report generation system based on segmented extraction inference learning according to claim 4 is characterized in that: In the step 13, each clause in the clause list obtained in step 12 is preliminarily classified to obtain the category of each clause in the clause list; the specific process is as follows:

131. Definition: The extraction segment ES is a clause that directly describes the image observation results; Definition: The inference segment RS is a preliminary diagnosis or inference based on the imaging results; 132. Classify each clause in the clause list obtained in 12 based on a classification method, and determine whether each clause is an extraction segment or an inference segment.

6. The PET report generation system based on segmented extraction inference learning according to claim 5 is characterized in that: In the step 14, each clause is labeled according to the category of each clause in the clause list obtained in step 3, and the labeled clauses are saved to obtain a SERI training data set; The specific process is:

141. label each clause according to the category of each clause in the clause list obtained in 13, labeling it as an extraction segment or an inference segment; 142. Save each clause in the marked clause list; Save format: Region, Clause Text, Category.

7. The PET report generation system based on segmented extraction inference learning according to claim 6 is characterized in that: In the step 132, each clause in the clause list obtained in step 12 is classified based on a classification method to determine whether each clause is an extraction segment or an inference segment; the specific process is as follows:

1. Classify each clause in the clause list obtained in step 1 based on the rules to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows:

1. Build a keyword library; the specific process is: According to the vocabulary and expressions in the PET report, construct the keyword library of the extraction segment and the keyword library of the inference segment; 12. Determine whether each clause in the clause list obtained in 12 contains a word in the keyword library of the extraction segment. If it contains a word in the keyword library of the extraction segment, the clause is an extraction segment; if not, the clause is not an extraction segment; Determine whether each clause in the obtained clause list contains a word in the inference segment keyword library. If it contains a word in the inference segment keyword library, the clause is an inference segment; if not, the clause is not an inference segment.

2. Classify each clause in the clause list obtained in step 1 based on the entropy value to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows: Input each clause in the obtained clause list into the trained language model, and the trained language model outputs the predicted probability of each clause; Calculate the entropy value of each clause based on its predicted probability; Set the entropy threshold; If the entropy value of the clause is less than or equal to the entropy threshold, the clause is marked as an extraction segment ES; If the entropy value of a clause is greater than the entropy threshold, the clause is marked as a reasoning segment RS; 3. Classify each clause in the clause list obtained in step 1 based on the loss value to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows: 3.

1. Calculation of loss value: Input each clause in the obtained clause list into the trained language model, and the trained language model outputs the predicted probability of each clause; Calculate the loss value based on the predicted probability and the true probability; 32. Set the loss value threshold: If the loss value of a clause is higher than the loss value threshold, the clause is marked as reasoning segment RS; If the loss value of a clause is less than or equal to the loss value threshold, the clause is marked as an extraction segment ES; 4. Classify each clause in the clause list obtained in step 1 based on similarity to determine whether each clause is an extraction segment or an inference segment. The specific process is as follows: 4.

1. Similarity calculation: Calculate the similarity between each clause and its adjacent clauses; 42. Set similarity threshold; If the similarity between a clause and its adjacent clauses is higher than the similarity threshold, the clause is marked as an extracted segment ES; Otherwise, the clause is marked as a reasoning segment RS.

8. The PET report generation system based on segmented extraction inference learning according to claim 7 is characterized in that: In the second method, the model is trained based on the SERI training data set to obtain a trained model; the specific process is: Step 1: Set language model hyperparameters; Step 2: The PET report in the SERI training dataset is used as the input of the language model, and the language model outputs the extracted segment ES; Step 3: The extracted segment is used as the input of the language model, and the model outputs the reasoning segment RS; Step 4: Repeat steps 2 to 4, use the Adam optimizer to minimize the loss function through gradient descent. When the output of the language model contains the preset stop sign, the iterative generation process is automatically terminated, the language model outputs a complete report, and a trained language model is obtained.

9. The PET report generation system based on segmented extraction inference learning according to claim 7 is characterized in that: In the second method, the model is trained based on the SERI training data set to obtain a trained model; the specific process is: Step 1: Set ES model hyperparameters; Step 2: The PET report in the SERI training dataset is used as the input of the ES model, and the ES model outputs the extracted segment ES; Step 3: Repeat steps 2 and 3, use the Adam optimizer to minimize the loss function through gradient descent. When the output of the ES model contains the preset stop sign, the iterative generation process automatically terminates, and the ES model outputs the complete extraction segment to obtain a trained ES model. The ES model is a language model; Step 4: Set RS model hyperparameters; Step 5: The complete extraction segment output by the ES model is used as the input of the RS model, and the RS model outputs the reasoning segment RS; Step 6: Repeat steps 5 and 6, use the Adam optimizer to minimize the loss function through gradient descent. When the output of the RS model contains the preset stop sign, the iterative generation process automatically terminates, and the RS model outputs a complete inference segment to obtain a trained RS model. Obtain a trained model based on the trained ES model and the trained RS model; The RS model is a language model.

Images