Pet report generation system based on segmented extraction and reasoning learning

Abstract

The application relates to a PET report generation system based on segmented extraction reasoning learning, relates to the fields of medical image analysis and natural language processing, and aims to solve the problem that language models in the prior art are difficult to efficiently generate high-quality reports.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 PET reports, and a SERI training data set is constructed based on the PET reports; 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 reasoning conclusions of the PET report; the PET is a positron emission tomography.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image analysis and natural language processing (NLP), and in particular to a small model based on segmented extraction-reasoning inference (SERI) learning for generating PET (positron emission tomography) reports. BACKGROUND

[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 nervous system diseases. PET scan reports are usually manually written by doctors based on image results, describing the metabolic activity and functional characteristics of multiple anatomical regions, and making preliminary diagnoses. 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 susceptible to human factors (such as fatigue or subjective judgment), leading to errors or omissions.

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

[0004] Large computational resource and data demand:

[0005] Large language models usually require a large amount of computational resources and massive training data. For the medical imaging field, especially for PET reports of specific diseases (such as cancer), the amount of data is often small, which makes it difficult for large models to be fully trained. At the same time, running these models requires high-performance hardware devices (such as GPUs or TPUs), which is costly for clinical applications.

[0006] Lack of efficient data utilization mechanism:

[0007] In existing technologies, many methods rely on generating more high-quality data or distilling large models to improve the quality of report generation. However, these methods fail to fully exploit the potential information in existing data, especially in medical image reports, where there is a large amount of structured image description and reasoning conclusions that are not efficiently utilized.

[0008] Difficult to capture multi-step reasoning and logical summary:

[0009] PET reports often need to describe not only individual image results but also metabolic activity in multiple anatomical regions, make inferences and summaries, and determine the nature or development trend of the disease. Although existing large models can generate text content, they do not perform well in complex multi-step reasoning and logical coherence, which can 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 perform well in performance, their deployment costs and computing requirements are not conducive to widespread application in actual medical environments. Therefore, the industry has also attempted to transplant chain-of-thought (CoT) reasoning capabilities into small models (such as T5-small or T5-base), but small models perform poorly on complex reasoning tasks and are difficult to effectively learn these complex reasoning steps, resulting in low-quality generated reports.

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

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

[0014] Difficult to achieve efficient reasoning: Generating medical image reports requires complex reasoning and diagnostic conclusions based on image results. Although existing large models have the ability to generate reports, they lack efficient reasoning capabilities, especially in small models where accurate multi-step reasoning is difficult to achieve.

[0015] High cost of computing resources: Large models require high computing resources, which are difficult to deploy on a large scale in medical scenarios. Existing small models have advantages in resource requirements, but their reasoning capabilities are limited, resulting in low-quality generated reports.

[0016] Low information utilization efficiency: Existing report generation methods do not fully utilize existing chain-of-thought data, lacking sufficient mining and efficient use of the inherent information in the data. SUMMARY

[0017] The present application proposes a PET report generation system based on segmented extraction reasoning learning, which efficiently segments chain-of-thought data and divides the content of the image report into extraction segments and reasoning segments. This method effectively improves the performance of small models in complex reasoning tasks, solving the problem of language models being unable to efficiently generate high-quality reports in existing technologies.

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

[0019] The training data set acquisition module, the model acquisition module, and the to-be-tested PET report generation module;

[0020] The training data set acquisition module is configured to acquire PET reports and construct a SERI training data set based on the PET reports;

[0021] SERI refers to segmented extraction and reasoning.

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

[0023] The to-be-tested PET report generation module is configured to input a to-be-tested PET report into the trained model, and the trained model outputs extraction information and reasoning conclusions of the PET report;

[0024] The PET refers to positron emission tomography.

[0025] The present application has the following advantages:

[0026] The present application aims to solve the deficiencies of small language models in PET report generation in the prior art and proposes 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 ability of the model to generate medical image reports.

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

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

[0029] Small language models usually perform poorly in handling complex reasoning tasks, especially in the generation of medical image reports, which requires combining image information from multiple anatomical regions and reasoning and summarizing. The present application divides the content in the PET report into extractive segments (ES) and reasoning segments (RS), introducing a step-by-step generation mechanism for small models. The specific advantages are as follows:

[0030] Staged generation: the model first generates extractive segments and then generates reasoning segments based on the extractive segments, making the reasoning process more logical and controllable and solving the problem of logical confusion that may occur when small models generate full reports at once.

[0031] Multi-step reasoning capability: Through segmented processing, small models can process image information of multiple anatomical regions and combine these information to generate overall report inferences. This step-by-step reasoning approach significantly improves the performance of small models in complex logical tasks, enabling them to approach the inference effect of large models.

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

[0033] Medical image report generation, especially PET report generation, often faces the problem of insufficient data. Large language models perform poorly in data scarcity, while the present invention significantly reduces the dependence 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 inherent 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] Efficiency improvement of iterative learning: By breaking down the complex reasoning process into multiple steps, the model can learn fine extraction and reasoning at each step. This fine-grained learning approach enhances the model's understanding and reasoning ability for limited data.

[0036] 3. Reduce the demand for computing resources;

[0037] Although large language models perform well in many NLP tasks, their huge demand for computing resources limits 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 under limited computing resources:

[0038] Significant reduction in computing cost: By reasonably allocating the computing 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, using small models can reduce deployment costs.

[0039] High adaptability: This method can be deployed in lower configuration hardware environments, such as regular servers in hospitals or cloud computing platforms, making this technology have the potential for large-scale application.

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

[0041] Medical image reports are not just simple descriptions of image data, but often require reasoning and drawing medical conclusions based on multiple aspects of the image. The present invention processes image data in segments, making the generated report 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 descriptions and inference parts, thus avoiding the confusion of mixing image data with diagnostic inference. This helps to improve the accuracy of the report, making the generated report 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, with clear hierarchical structure throughout the process, avoiding the logical jumps or inconsistent reasoning that may occur in traditional one-time report generation.

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

[0045] Although the invention takes PET report generation as the core application scenario, the method has wide applicability. Since many medical image reports (such as CT reports, MRI reports) also need to analyze multiple anatomical regions and make inferences and summaries, the method of the invention can be applied to different types of medical image report generation tasks:

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

[0047] Support for task complexity: The invention can handle complex image report tasks from a single organ to multiple organs, especially suitable for diagnostic tasks that require comprehensive analysis of multiple image results, such as tumor metastasis analysis.

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

[0049] Traditional manual report generation methods may be affected by doctors' experience, fatigue, and other subjective factors, leading to inconsistent descriptions or insufficient reasoning in the report. The invention reduces human intervention by automatically generating reports, improving the reliability and consistency of the report:

[0050] Automatic generation reduces errors: The systematic report generation process ensures that each generated report is based on uniform reasoning rules, avoiding the bias caused by individual differences between different doctors.

[0051] Higher consistency: The report generated by the machine learning model is more consistent in language style and logical structure, especially when dealing with a large number of image reports, this consistency is particularly important.

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

[0053] The present application not only can generate accurate image reports, but also can generate suggestions for future diagnostic steps through the reasoning section. For example, in PET images, the system can give a reasoning section of "suggested to further perform CT or MRI examination" according to the abnormal degree of metabolic activity. This ability improves the practicability of the system and helps doctors quickly formulate the next step of treatment plan. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The system flowchart of the present application. DETAILED DESCRIPTION

[0055] Embodiment I: The PET report generation system based on segmented extraction reasoning learning in the embodiment includes:

[0056] a training data set acquisition module, a model acquisition module, and a to-be-tested PET report generation module;

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

[0058] SERI is segmented extraction reasoning;

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

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

[0061] The PET is a positron emission tomography.

[0062] Embodiment II: The difference between the embodiment and embodiment I is that the training data set acquisition module is used to acquire PET reports and construct a SERI training data set based on the PET reports;

[0063] The specific process is as follows:

[0064] I. Acquire PET reports, segment the PET reports according to anatomical regions, and obtain paragraphs of each anatomical region;

[0065] The purpose of this step is to logically divide the content of the whole body report according to organs or specific anatomical regions, so that the content of each region is independent, facilitating subsequent processing;

[0066] II. Decompose the paragraphs of each anatomical region obtained in I into finer clauses, and all the clauses form a clause list;

[0067] After identifying and segmenting the anatomical regions, the next step is to further decompose each paragraph of an anatomical region into finer sub-sentence units. The purpose of this step is to prepare for the subsequent classification of the extracted segment (ES) and the reasoning segment (RS);

[0068] III. Classify each sub-sentence in the sub-sentence list obtained in II to obtain the category of each sub-sentence in the sub-sentence list;

[0069] IV. Label each sub-sentence in the sub-sentence list obtained in III according to the category of each sub-sentence, and save the labeled sub-sentence list to obtain the SERI training data set.

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

[0071] Embodiment III: The difference between this embodiment and the first or second embodiment is that in the first step, the PET report is segmented according to anatomical regions to obtain paragraphs for each anatomical region.

[0072] The purpose of this step is to logically divide the content of the whole body report according to organs or specific anatomical regions, so that the content of each region is independent, facilitating subsequent processing.

[0073] The specific process is as follows:

[0074] III. 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, we first need to identify the paragraphs of each anatomical region from the report. For example, the report may have the following descriptions:

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

[0077] Pancreas: Abnormal increase in metabolic activity in the head of the pancreas.

[0078] Lungs: No abnormal metabolic foci in the lungs.

[0079] Through natural language processing technology, we can use keyword extraction or regular expression-based matching algorithms to automatically divide these regions. We can define a set of keywords or terms corresponding to common anatomical regions, such as "liver", "pancreas", "lungs", etc., to mark the division.

[0080] II. Segment the identified paragraphs of each anatomical region into independent paragraphs.

[0081] Each paragraph of the anatomical region contains two parts:

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

[0083] summarize the preliminary diagnostic impression for each anatomical region.

[0084] Split the report into separate paragraphs for each anatomical region description. Each paragraph should include all information for that anatomical region. For each anatomical region paragraph, there are usually two parts:

[0085] describe the imaging findings for that region (e.g., description of normal or abnormal metabolism);

[0086] summarize the preliminary diagnostic impression for that region (e.g., possible pathological inference).

[0087] Other steps and parameters are the same as in embodiment one or two.

[0088] Embodiment four: The difference between this embodiment and one of embodiments one to three is that the paragraph for each anatomical region obtained in one is further divided into more detailed clauses, and all clauses form a clause list.

[0089] After identifying and segmenting the anatomical regions, the next step is to further divide the paragraph for each anatomical region into more detailed clause units; the purpose of this step is to prepare for the subsequent classification of extraction sentences (ES) and reasoning sentences (RS);

[0090] The specific process is as follows:

[0091] One two one, based on punctuation segmentation tools (such as NLTK library or spaCy library in Python), divide the paragraph for each anatomical region obtained in one into clauses.

[0092] Punctuation segmentation:

[0093] Use punctuation segmentation tools (such as NLTK library or spaCy library in Python) to cut the paragraph into smaller clauses; common punctuation marks include period ("."), comma (","), semicolon (";"), etc. Punctuation marks can be used as natural segmentation points, making complex sentences simpler for subsequent processing;

[0094] For example, for "No obvious metabolic abnormalities were found in the liver, and the metabolic activity of the head of the pancreas was abnormally increased.", the sentence 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 was abnormally increased."

[0097] One two two, clean each clause to obtain a cleaned clause.

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

[0099] The unit and term standardization is performed on the cleaned clause to obtain a standardized processed clause;

[0100] After the clause is segmented, the text needs to be cleaned and pre-processed to remove noise and unnecessary symbols. Common processing steps include:

[0101] Removing spaces and punctuation: removing redundant spaces in the clause to maintain the semantic integrity of the clause;

[0102] Standardizing units and terms: for example, standardizing all radioactivity uptake values to a uniform unit, or unifying certain pathological terms (such as "increased activity" can be simplified to "metabolic abnormality");

[0103] The cleaned and pre-processed clause should be a format specification phrase or sentence to facilitate the subsequent classification;

[0104] One two three, based on the standardized processed clause to form a clause list.

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

[0106] For example, for a PET report containing liver and pancreas information, the segmented clause list may be:

[0107] Clause 1: "No obvious metabolic abnormality in the liver."

[0108] Clause 2: "Abnormal increase in metabolic activity in the head of the pancreas."

[0109] The other steps and parameters are the same as one of the first three specific embodiments.

[0110] Specific embodiment five: this embodiment is different from one of the first four specific embodiments in that: in one three, each clause in the clause list obtained in one two is preliminarily classified to obtain the category of each clause in the clause list; the specific process is:

[0111] One three one, define the extraction segment (ES) as the clause directly describing the observation result of the image;

[0112] Define the reasoning segment (RS) as the preliminary diagnosis or reasoning based on the image result;

[0113] One three two, classify each clause in the clause list obtained in one two based on a classification method to determine whether each clause is an extraction segment or a reasoning segment;

[0114] After the clause segmentation is completed, the next step is to preliminarily classify each clause as an extraction segment (ES) or a reasoning segment (RS). The specific classification criteria are as follows:

[0115] 1. Identification of extraction segments:

[0116] The definition of an extraction segment (ES) is a clause that directly describes the results of an image observation. These clauses usually contain quantitative or qualitative observation data and do not involve reasoning or analysis. For example:

[0117] "There is no obvious metabolic abnormality in the liver."

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

[0119] These clauses mainly provide physical or functional observations of the image and are direct explanations of the image data.

[0120] 2. Identification of reasoning segments:

[0121] The definition of a reasoning segment (RS) is a clause that makes a preliminary diagnosis or reasoning based on the image results. These clauses usually contain inferential, summary, or hypothetical content. For example:

[0122] "The metabolic activity of the head of the pancreas is abnormally high, and malignancy is considered possible."

[0123] "It is recommended to conduct further examinations to rule out tumor metastasis."

[0124] These clauses not only describe the image results but also further deduce possible pathological diagnosis results or propose next steps for handling.

[0125] 3. Selection of classification methods:

[0126] Rule-based classification: Use pre-defined rules (such as keyword matching or sentence analysis) to identify whether each clause in the clause list is an extraction segment or a reasoning segment. For example, clauses containing descriptions of physical characteristics such as "metabolic abnormalities" and "abnormal absorption" can be classified as extraction segments, while clauses containing inferential words such as "consider" and "suggest" can be classified as reasoning segments.

[0127] Classification by machine learning model: A trained classification model (such as SVM, logistic regression, etc.) can also be used to identify whether each clause in the clause list is an extraction segment or a reasoning segment,

[0128] Through manually annotated training data, the model is trained to distinguish between extraction segments and reasoning segments. 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 in one of the first four embodiments.

[0130] Specific implementation six: Different from one to five, the specific implementation is as follows: the category of each clause in the clause list obtained according to 13 is marked for each clause, and the marked clause is saved to obtain a SERI training data set; the specific process is as follows:

[0131] One four, the category of each clause in the clause list obtained according to 13 is marked for each clause, and the marked clause is saved to obtain a SERI training data set; the specific process is as follows:

[0132] One four, the category of each clause in the clause list obtained according to 13 is marked for each clause, and the marked clause is saved to obtain a SERI training data set; the specific process is as follows:

[0133] The saved format is as follows:

[0134] Region, clause text, category.

[0135] After the above steps are completed, the paragraphs of each anatomical region will be divided into multiple clauses, and each clause has been marked as an extraction segment or a reasoning segment.

[0136] The final output result should be a clause list that has been classified. For example:

[0137] Clause 1: "No obvious metabolic abnormalities in the liver." (Extraction segment ES)

[0138] Clause 2: "Abnormal increase in metabolic activity in the head of the pancreas." (Extraction segment ES)

[0139] Clause 3: "Consider the possibility of malignant lesions." (Reasoning segment RS)

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

[0141] Annotate the clauses:

[0142] After classification is completed, all clauses will be marked as extraction segments or reasoning segments. The output result includes the text of each clause, the anatomical region identifier, 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 subsequent construction of SERI data set. File format example:

[0145] Region, clause text, category

[0146] Liver, no obvious metabolic abnormalities in the liver, ES

[0147] Pancreas, metabolically active, ES

[0148] Pancreas, consider malignancy possible, RS

[0149] In this step, by using entropy, loss value or similarity and other methods, combined with pre-defined rules, accurately classify the clauses of the image report into extraction section (ES) or reasoning section (RS). This process lays a solid foundation for subsequent model training and iteration 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 section and reasoning section. This step provides a basic data set for the subsequent model training and iteration generation. In implementation, the rules of segmentation and classification 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 one of the first to fifth embodiments.

[0152] Specific embodiment seven: different from one of the first to sixth embodiments is that: in the first one three two, based on the classification method, each clause in the clause list obtained in the first one two is classified to determine whether each clause is an extraction section or a reasoning section; the specific process is:

[0153] I. Based on rules, each clause in the clause list obtained in the first one two is classified to determine whether each clause is an extraction section or a reasoning section; the specific process is:

[0154] I-1. Construct a keyword library; the specific process is:

[0155] According to the common words and expressions in the PET report, construct an extraction section keyword library and a reasoning section keyword library;

[0156] Extraction section keywords: such as "no abnormalities", "metabolically active", "abnormal increase in metabolism", "normal", "no abnormalities" and the like.

[0157] Reasoning section keywords: such as "consider", "possible", "speculate", "suggest" and the like.

[0158] II. Determine whether each clause in the clause list obtained in the first one two contains words in the extraction section keyword library, if it contains words in the extraction section keyword library, the clause is an extraction section; if not, the clause is not an extraction section;

[0159] Determine whether each clause in the clause list obtained in the first one two contains words in the reasoning section keyword library, if it contains words in the reasoning section keyword library, the clause is a reasoning section; if not, the clause is not a reasoning section;

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

[0161] "pancreatic metabolism is active" is marked as the extraction segment (ES).

[0162] "consider the possibility of a malignant tumor" is marked as the reasoning segment (RS).

[0163] If a clause contains both the extraction segment and the reasoning segment, the splitting algorithm is applied to further split the clause to ensure that each clause can only belong to one of the extraction segment or the reasoning segment, and the previous division has solved this problem;

[0164] For example, combined with punctuation marks and natural language tokenization tools (such as NLTK, spaCy), complex clauses are further split to ensure that each clause can only belong to one of the extraction segment or the reasoning segment;

[0165] II. Based on the entropy value, classify each clause in the clause list obtained in I to determine whether each clause is an extraction segment or a reasoning segment; the specific process is as follows:

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

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

[0168] Set an entropy value threshold;

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

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

[0171] The entropy value expression is:

[0172]

[0173] Where P(R i ) is the prediction 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 the entropy; i is the i-th; and V is the vocabulary;

[0175] The trained language model is a language model trained using the originally acquired PET report or a general pre-trained model.

[0176] For example, the entropy value of the clause "no obvious abnormalities in the liver" can be low because it directly describes the image results, and the model has high certainty in generating it. Therefore, this clause is marked as an extraction segment.

[0177] III. Based on the loss value, classify each clause in the clause list obtained in I and II to determine whether each clause is an extraction segment or a reasoning segment; the specific process is as follows:

[0178] III. I. Loss value calculation:

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

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

[0181] Use cross-entropy loss or other loss functions to evaluate the quality of the model's generation of the clause. III. II.

[0183] Set a loss value threshold:

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

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

[0186] IV. Based on the similarity, classify each clause in the clause list obtained in I and II to determine whether each clause is an extraction segment or a reasoning segment; the specific process is as follows:

[0187] IV. I. Similarity calculation:

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

[0189] For example, the clause "pancreatic metabolic activity" is directly related to the "pancreas" part in the image report, so the similarity is high, and it belongs to the extraction segment ES; while "considering the possibility of malignant tumor" has no direct correspondence with the context, and it belongs to the reasoning segment RS; IV. II.

[0191] Set a similarity threshold;

[0192] If the similarity of a clause with its adjacent upper and lower clauses is higher than the similarity threshold, mark the clause as an extraction segment ES;

[0193] Otherwise, the clause is labeled as a Reasoning Segment RS.

[0194] 2 Detailed explanation of classification into Extractive Segments ES and Reasoning Segments RS

[0195] The purpose of this step is to classify the segmented PET report clauses into Extractive Segments (ES) and Reasoning Segments (RS) based on specific algorithms and indicators, in order to subsequently build the SERI dataset and conduct model training.

[0196] 2.1 Feature definition of Extractive Segments (ES) and Reasoning Segments (RS);

[0197] In the classification process, the definitions and features of extractive segments and reasoning segments need to be clarified first:

[0198] Extractive Segments (ES): These paragraphs directly describe the imaging observation results, usually quantitative or qualitative statements about the state of organs or tissues in PET scans. These statements directly reflect the observed phenomena or data.

[0199] Features: Usually contain explicit observation data or image information, language is more direct and declarative.

[0200] Keywords: such as "no abnormalities", "metabolic activity", "no abnormal metabolic absorption", etc.

[0201] Reasoning Segments (RS): These paragraphs make inferences or summaries based on observation data, usually providing preliminary diagnoses, hypotheses or suggestions.

[0202] Features: Involve logical reasoning, diagnostic conclusions or future action suggestions, language is speculative or summary.

[0203] Keywords: such as "consider", "may", "speculate", "suggest further examination", etc.

[0204] 2.2 Rule-based preliminary classification method

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

[0206] Key word library construction:

[0207] According to the common words and expressions in a large number of PET reports, a keyword library related to extractive segments and reasoning segments is constructed.

[0208] Extractive segment keywords: such as "no abnormalities", "metabolic activity", "abnormal increase in metabolism", "normal", "no abnormalities", etc.

[0209] Reasoning segment keywords: e.g. "consider", "may", "speculate", "suggest" etc.

[0210] Matching rule application:

[0211] For each clause, keyword matching is performed to see if it contains any words in the keyword library. If a clause contains an extraction segment keyword, it 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 "pancreatic metabolism is active, consider the possibility of a malignant tumor", "metabolism is active" 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 extraction segment and reasoning segment respectively:

[0213] "Pancreatic metabolism is active" is marked as an extraction segment (ES).

[0214] "Consider the possibility of a malignant tumor" is marked as a reasoning segment (RS).

[0215] Boundary condition handling:

[0216] For clauses containing multiple keywords, a segmentation algorithm is applied for further subdivision. For example, by combining punctuation marks and natural language tokenization tools (such as NLTK, spaCy), complex clauses are further split to ensure that each clause can only belong to one of the extraction segment or reasoning segment.

[0217] 2.3 Entropy-based classification method

[0218] In addition to rule-based classification, entropy can also be used to measure the certainty of the model's generated content, and to distinguish between extraction segments and reasoning segments based on certainty. The lower the entropy value, the higher the certainty of the model's generated results, usually indicating that the clause is an extraction segment. The higher the entropy value, the greater the uncertainty of the model's results, which may belong to the reasoning segment.

[0219] Entropy calculation:

[0220] Using a trained language model, input each clause in the report and calculate the entropy value of each clause when it is generated. Specifically, the model will generate a prediction distribution for each clause, and the entropy value reflects the uncertainty of the distribution;

[0221] Classification basis:

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

[0223] For example, the entropy value of the clause "No obvious abnormalities in the liver" may be low because it directly describes the image results, and the model generates it with high certainty. Therefore, this clause will be marked as an extraction segment.

[0224] 2.4 Classification method based on loss value

[0225] Similar to entropy value, loss value is another criterion for classification. During model generation, loss value reflects the error of the model when generating a certain clause. Extraction segments are often directly related to the input context, so the model has a low loss value; reasoning segments involve more reasoning, so the loss value is higher.

[0226] Loss value calculation:

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

[0228] Loss value threshold setting:

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

[0230] 2.5 Classification method based on similarity

[0231] To further improve the accuracy of classification, a similarity-based method can be introduced. Extraction segments are usually information directly extracted from the input context, so they have a high similarity to the context; reasoning segments are based on context and therefore have a lower similarity.

[0232] Similarity calculation:

[0233] Calculate the similarity of each clause to its context (or input image description) using text similarity evaluation indicators such as BLEU or ROUGE.

[0234] For example, the clause "Pancreatic metabolic activity" is directly related to the "pancreas" part of the image report, so it has a high similarity and belongs to the extraction segment (ES); "Consider the possibility of malignant tumor" has no direct correspondence with the context and belongs to the reasoning segment (RS).

[0235] Classification basis:

[0236] By setting a similarity threshold, if the similarity of a clause to the context is higher than the threshold, it will be marked as an extraction segment (ES); otherwise, it will be marked as a reasoning segment (RS).

[0237] Other steps and parameters are the same as one of the first six embodiments.

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

[0239] First step: set the language model hyperparameters; the language model is BART;

[0240] Second step: take the PET report in the SERI training data set as the input of the language model, and the language model outputs the extracted section ES; the language model is BART;

[0241] Third step: take the extracted section as the input of the language model, and the model outputs the reasoning section RS;

[0242] Fourth step: repeat the second to fourth steps, use the Adam optimizer, minimize the loss function through gradient descent, when the output of the language model contains the preset stop flag, the iterative generation process is automatically terminated, the language model outputs the complete report, and a trained language model is obtained.

[0243] Other steps and parameters are the same as one of the specific implementations one to seven.

[0244] Specific implementation nine: different from one of the specific implementations one to seven is that: the second, based on the SERI training data set, the model is trained to obtain a trained model; the specific process is:

[0245] First step: set the ES model hyperparameters;

[0246] Second step: take the PET report in the SERI training data set as the input of the ES model, and the ES model outputs the extracted section (ES);

[0247] Third step: repeat the second and third steps, use the Adam optimizer, minimize the loss function through gradient descent, when the output of the ES model contains the preset stop flag, the iterative generation process is automatically terminated, the ES model outputs the complete extracted section, and a trained ES model is obtained;

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

[0249] Fourth step: set the RS model hyperparameters;

[0250] Fifth step: take the complete extracted section output by the ES model as the input of the RS model, and the RS model outputs the reasoning section (RS);

[0251] Step 6: Repeat step 5 and step 6 using Adam optimizer to minimize the loss function by gradient descent, and the iteration generation process is automatically terminated when the output of the RS model contains a preset stop flag, the RS model outputs a complete reasoning section, and a trained RS model is obtained;

[0252] Based on the trained ES model and the trained RS model, a trained model is obtained;

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

[0254] The other steps and parameters are the same as one of the first to seventh embodiments.

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

[0256] 3.1 Construct SERI dataset

[0257] The construction of the SERI dataset is based on the classified extraction section (ES) and reasoning section (RS). This dataset will be used for model training to help the model gradually generate the extraction information and reasoning conclusion of the PET report.

[0258] Data organization:

[0259] The data for each anatomical region should include:

[0260] Input: image description or patient image examination information.

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

[0262] The data format should be organized according to the input-output structure of the model. For each anatomical region, first input the image examination description of the region, and then the model generates the corresponding extraction section and reasoning section.

[0263] ES data sample:

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

[0265] Output: extraction section (e.g. "pancreatic metabolic activity").

[0266] RS data sample:

[0267] Input: image description and extraction section (e.g. "pancreatic metabolic activity").

[0268] Output: reasoning section (e.g. "considering the possibility of malignant tumor").

[0269] Input-output format:

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

[0271] Data examples:

[0272] Region, input, output

[0273] Liver, "imaging shows no obvious metabolic abnormalities in the liver", "no obvious metabolic abnormalities in the liver"

[0274] Pancreas, "imaging shows that the pancreas is metabolically active", "pancreas is metabolically active"

[0275] Pancreas, "imaging shows that the pancreas is metabolically active, and the possibility of malignant tumor should be considered", "the possibility of malignant tumor should be considered"

[0276] Label 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 section and the reasoning section of each clause should be marked with labels so that the training process can distinguish the different functions of the paragraphs.

[0278] 3.2 Iterative training model

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

[0280] Single-path learning (USM):

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

[0282] Generation process:

[0283] Step 1: The model inputs the image description and outputs the extraction section (ES).

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

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

[0286] For example:

[0287] Input: "The image shows a pancreas with high metabolic activity."

[0288] Output 1 (ES): "The pancreas is metabolically active."

[0289] Input 2: "The pancreas is metabolically active" as a new input.

[0290] Output 2 (RS): "Consider the possibility of a malignant tumor."

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

[0292] Dual-path learning (DSM):

[0293] Under the dual-path learning framework, two separate models are used to generate the extraction segment (ES) and the reasoning segment (RS) respectively. These two models are optimized independently and handle the extraction task for ES and the reasoning task for RS respectively.

[0294] Generation process:

[0295] Step 1: The ES model inputs the image description and outputs the extraction 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 train more specifically for the extraction segment and the reasoning segment, and each model can be better optimized on its respective task. The generation process is as follows:

[0298] ES model: Input image description, generate extraction segment.

[0299] RS model: Input extraction segment, 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 previous clause, until the entire report is completed.

[0302] 3.3 Model optimization and training process

[0303] Loss function setting:

[0304] Use the standard loss function of sequence generation models (such as cross-entropy loss function) to calculate the error of each generation. The generation loss of the extraction segment and the reasoning segment can be calculated separately or jointly optimized.

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

[0306] Optimization strategy:

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

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

[0309] Hyperparameter tuning:

[0310] Learning rate: Adjust the learning rate of the model to ensure convergence during training but not fall into 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 flag in the generation process

[0313] Stop flag definition:

[0314] In each step of the generation process, a stop flag 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 "report end" or a specific symbol) can be used as a stop flag.

[0316] Termination condition:

[0317] When the model detects the presence of a predefined stop flag in the generated text, the iterative generation process is automatically terminated, and the complete report is output.

[0318] 3.5 Post-processing of report generation

[0319] Report formatting:

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

[0321] Result verification:

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

[0323] 3.6 Model evaluation and verification

[0324] Evaluation indicators:

[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] Additionally, based on the specific tasks of image reporting (such as diagnostic accuracy), evaluate whether the generated report meets medical standards.

[0327] Validation set and test set:

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

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

[0330] This step enables the model to gradually generate a complete PET report by building a SERI dataset and iteratively training. Single-path or double-path generation methods can help the model extract key information from image descriptions and infer summary conclusions. This process ensures the coherence and accuracy of the report, and controls the termination of generation through stop signs. During training and optimization, techniques such as setting loss functions, tuning hyperparameters, and defining termination conditions are required 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] Example 1:

[0333] To verify the effectiveness of the present invention, especially the performance improvement of small models based on the segmented extraction-inference learning (SERI) method in the PET report generation task, the present invention implements multiple experiments, compares and analyzes the differences between the present method and the prior art, and demonstrates the advantages of the method in medical image report generation. The following are the detailed steps and simulation experiment results.

[0334] Example 1: PET report generation task

[0335] This example demonstrates the application of small models based on the SERI learning method in the PET report generation task, especially the automatic generation of reports for patients with tumor detection PET images.

[0336] Experimental setup:

[0337] Dataset: A real PET image report dataset from a large hospital was selected, which contains complete image 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 inference conclusions consistent with real doctor reports, including extraction sections and inference sections for each anatomical region.

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

[0340] Comparison method: Compared with traditional large language models based on GPT-3 and T5-small model directly generating a complete report.

[0341] Data processing:

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

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

[0344] Model training:

[0345] Training method: Use SERI method to iteratively train small T5 model. First, train the model to generate extraction section (ES), then based on the generated ES segment, further train the model to generate reasoning section (RS).

[0346] Iterative generation: Use single-path learning (USM) for recursive generation. The model first inputs image data to generate ES section, and then generates RS section based on ES section. The whole process is completed step by step on each anatomical region of the report until the complete report is generated.

[0347] Simulation experiment process:

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

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

[0350] Evaluation index:

[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 generated text and reference text, including matching of words and phrases.

[0353] Medical consistency score: Artificially evaluate the generated report by professional doctors to judge the medical logic and accuracy of the generated report.

[0354] Experimental Results:

[0355] Method BLEU ROUGE-L Medical Consistency Score (Max 10)

[0356] GPT-3 0.52 0.61 8.5

[0357] T5-small (no SERI) 0.35 0.47 6.8

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

[0359] BLEU and ROUGE-L Score Improvement: 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] Medical Consistency Score Improvement: The generated reports by the SERI method received higher scores in the evaluation of medical logic by multiple doctors. This indicates that the reports generated by the SERI method through step-by-step reasoning are more consistent with medical logic, avoiding reasoning inconsistencies or omissions.

[0361] Qualitative Analysis:

[0362] Report Example Analysis:

[0363] Real Doctor Report:

[0364] "No obvious metabolic abnormalities in the liver, pancreatic metabolism is active, consider the possibility of pancreatic malignant tumor, suggest further CT examination."

[0365] GPT-3 Generated Report:

[0366] "No obvious metabolic abnormalities in the liver, pancreatic metabolism is active, consider the possibility of pancreatic malignant tumor, suggest further examination."

[0367] T5-small (no SERI) Generated Report:

[0368] "No obvious metabolic abnormalities in the liver, pancreatic metabolism is active, may be malignant."

[0369] T5-small (SERI) Generated Report:

[0370] "No obvious metabolic abnormalities in the liver, pancreatic metabolism is active, consider the possibility of pancreatic malignant tumor, suggest further CT examination."

[0371] From the comparison, it can be seen that the report generated by the SERI method not only has higher similarity to the real doctor's report, but also has logical coherence and more complete reasoning.

[0372] The application also can have other various embodiments, and those skilled in the art can make various corresponding changes and modifications according to the application without departing from the spirit and essence of the application, and these corresponding changes and modifications should all belong to the protection scope of the claims attached to the application.

Claims

1. A PET report generation system based on piecewise extraction and reasoning learning, characterized in that: The application relates to a PET report SERI (Segmented Extraction and Reasoning) method and device. The method comprises the following steps: A training data set acquisition module, a model acquisition module and a to-be-tested PET report generation module are provided. The training data set acquisition module is used for acquiring a PET report and constructing a SERI training data set based on the PET report. SERI is segmented extraction and 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 reasoning conclusions of the PET report. The PET is a positron emission tomography. The training data set acquisition module is used for acquiring a PET report and constructing a SERI training data set based on the PET report. The specific process is as follows: I. A PET report is acquired, and the PET report is segmented according to an anatomical region to obtain a paragraph of each anatomical region. II. Each paragraph of the anatomical region obtained in I is decomposed into a clause, and all the clauses form a clause list. III. Each clause in the clause list obtained in II is classified to obtain the category of each clause in the clause list. IV. Each clause is labeled according to the category of each clause in the clause list obtained in III, and the labeled clause is saved to obtain a SERI training data set. In III, each clause in the clause list obtained in II is classified to obtain the category of each clause in the clause list. III.

1. Define an extraction segment ES as a clause directly describing an image observation result. Define a reasoning segment RS as a preliminary diagnosis or reasoning based on an image result.

2. The PET report generation system based on segmental extraction and reasoning learning of claim 1, wherein: III.

2. Classify each clause in the clause list obtained in II based on a classification method to determine whether each clause is an extraction segment or a reasoning segment. In I, the PET report is segmented according to an anatomical region to obtain a paragraph of each anatomical region. I.

1. An image result of each anatomical region is recognized from the PET report based on a natural language processing technology. I.

2. The recognized paragraph of each anatomical region is segmented into an independent paragraph. Each paragraph of each anatomical region comprises two parts: An image result of each anatomical region is described.

3. The PET report generation system based on segmental extraction and reasoning learning of claim 2, wherein: A preliminary diagnosis impression of each anatomical region is summarized. In II, each paragraph of the anatomical region obtained in I is decomposed into a clause, and all the clauses form a clause list. II.

1. Each paragraph of the anatomical region obtained in I is decomposed into a clause based on a punctuation symbol segmentation tool. II.

2. Each clause is cleaned to obtain a cleaned clause. The cleaning is to remove redundant spaces in the clause. The cleaned clause is standardized in units and terms to obtain a standardized clause.

4. The PET report generation system based on segment-wise extraction and reasoning learning of claim 3, wherein: II.

3. A clause list is formed based on the standardized clause. In IV, each clause is labeled according to the category of each clause in the clause list obtained in III, and the labeled clause is saved to obtain a SERI training data set. IV.

1. Each clause is labeled according to the category of each clause in the clause list obtained in III, and the labeled clause is saved to obtain a SERI training data set. Ia2, save each clause in the labeled clause list; The saving format is: Region, clause text, category.

5. The PET report generation system based on segmental extraction and reasoning learning of claim 4, wherein: The classification method in Ia2 classifies each clause in the clause list obtained in Ia1 to determine whether each clause is an extraction section or a reasoning section; the specific process is: I, classify each clause in the clause list obtained in Ia1 based on rules to determine whether each clause is an extraction section or a reasoning section; the specific process is: Ia1, construct a keyword library; the specific process is: According to the vocabulary and expressions in the PET report, construct an extraction section keyword library and a reasoning section keyword library; Ia2, determine whether each clause in the clause list obtained in Ia1 contains vocabulary in the extraction section keyword library; if it contains vocabulary in the extraction section keyword library, the clause is an extraction section; if it does not contain, the clause is not an extraction section; Determine whether each clause in the clause list obtained in Ia1 contains vocabulary in the reasoning section keyword library; if it contains vocabulary in the reasoning section keyword library, the clause is a reasoning section; if it does not contain, the clause is not a reasoning section; II, classify each clause in the clause list obtained in Ia1 based on entropy value to determine whether each clause is an extraction section or a reasoning section; the specific process is: Input each clause in the clause list obtained in Ia1 into the trained language model, and the trained language model outputs the prediction probability of each clause; Calculate the entropy value of each clause based on the prediction probability; Set the entropy value threshold; If the entropy value of the clause is less than or equal to the entropy value threshold, the clause is marked as an extraction section ES; If the entropy value of the clause is greater than the entropy value threshold, the clause is marked as a reasoning section RS; III, classify each clause in the clause list obtained in Ia1 based on loss value to determine whether each clause is an extraction section or a reasoning section; the specific process is: IIIa, loss value calculation: Input each clause in the clause list obtained in Ia1 into the trained language model, and the trained language model outputs the prediction probability of each clause; Calculate the loss value based on the prediction probability and the true probability; IIIb, Set the loss value threshold: If the loss value of a certain clause is higher than the loss value threshold, mark the clause as a reasoning section RS; If the loss value of a certain clause is less than or equal to the loss value threshold, mark the clause as an extraction section ES; IV, classify each clause in the clause list obtained in Ia1 based on similarity to determine whether each clause is an extraction section or a reasoning section; the specific process is: IVa, similarity calculation: Calculate the similarity between each clause and its adjacent upper and lower clauses; IVb, Set the similarity threshold; If the similarity between the clause and its adjacent upper and lower clauses is higher than the similarity threshold, mark the clause as an extraction section ES; Otherwise, mark the clause as a reasoning section RS.

6. The PET report generation system based on segmental extraction and reasoning learning of claim 5, wherein: The model is trained based on the SERI training data set in II to obtain a trained model; the specific process is: Step 1: Set the language model hyperparameters; 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 extraction section ES; Step 3: Take the extraction section as the input of the language model, and the model outputs the reasoning section RS; Fourth step: repeat the second to fourth steps, use Adam optimizer, minimize the loss function by 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 the complete report, and the trained language model is obtained.

7. The PET report generation system based on segmental extraction and reasoning learning of claim 6, wherein: The second is to train the model based on the SERI training data set to obtain a trained model; the specific process is: First step: set the ES model hyperparameters; Second step: take the PET report in the SERI training data set as the input of the ES model, and the ES model outputs the extracted section ES; Third step: repeat the second and third steps, use Adam optimizer, minimize the loss function by gradient descent, when the output of the ES model contains the preset stop sign, the iterative generation process is automatically terminated, the ES model outputs the complete extracted section, and the trained ES model is obtained; The ES model is a language model; Fourth step: set the RS model hyperparameters; Fifth step: take the complete extracted section output by the ES model as the input of the RS model, and the RS model outputs the reasoning section RS; Sixth step: repeat the fifth and sixth steps, use Adam optimizer, minimize the loss function by gradient descent, when the output of the RS model contains the preset stop sign, the iterative generation process is automatically terminated, the RS model outputs the complete reasoning section, and the trained RS model is obtained; Based on the trained ES model and the trained RS model, a trained model is obtained; The RS model is a language model.

Images