CT image report information extraction method, device, computer equipment and storage medium based on multi-round question and answer

Abstract

This invention discloses a method, apparatus, computer equipment, and storage medium for extracting information from CT image reports based on multi-turn question answering, including: (1) pre-setting head entity question-and-answer questions and tail entity question-and-answer question templates according to the information extraction task; (2) constructing and optimizing an information extraction model based on a reading comprehension framework; (3) concatenating sentences extracted from the CT image report with each head entity question-and-answer question to obtain the input text required for extracting the head entity, and inputting it into the trained information extraction model to obtain the head entity; then, concatenating the head entity with the tail entity question-and-answer question template to obtain the tail entity question-and-answer question for the head entity, and concatenating it with the sentence to obtain the input text required for extracting the tail entity, and inputting the input text into the trained information extraction model to obtain the tail entity; (5) pairing the obtained head entity and tail entity into triples, and outputting the extracted triple information after standardization.

Description

Technical Field

[0001] The present invention belongs to the field of medical data mining, and specifically relates to a method, device, computer equipment and storage medium for extracting CT image report information based on multi-round question and answer. Background Art

[0002] Computed tomography (CT) is the main means of clinical examination. The CT image report obtained records the physiological status information of the scanned part and a large amount of valuable information interpreted by the radiologist, which can be used for clinical staging analysis and prediction. However, CT image reports are stored in the computer in the form of unstructured text, which brings challenges to the direct use of CT image reports. If you want to use the unstructured text in the CT image report, you need to extract structured information from the CT image report. If you use manual extraction, it will cause problems such as long extraction time, error-prone and high cost.

[0003] Information Extraction (IE) refers to the process of extracting entities, relationships between entities and events that users are interested in from unstructured text, that is, converting unstructured information into structured or semi-structured information and storing it in the form of a database for users to query and further analyze and use.

[0004] In view of the problems existing in manual information extraction, the existing technology combines the idea of ​​information extraction and adopts the pipeline method to automatically extract information. They first apply the labeling model to identify entities, and then use the relationship extraction model to identify the relationship between each entity pair. Although the pipeline strategy has the flexibility to integrate different data sources and learning algorithms, the accuracy of the extracted information is often not high, mainly due to the following limitations: 1) errors in the previous named entity module will seriously affect the performance of the subsequent relationship extraction; 2) the identified entities are paired and then the relationship is extracted, so many entity pairs that originally have no relationship will generate redundant information, which brings difficulties to model recognition.

[0005] Machine Reading Comprehension (MRC) is a technology that uses algorithms to enable computers to understand the semantics of articles and answer related questions. Since both articles and questions are in the form of human language, machine reading comprehension belongs to the category of natural language processing (NLP). Machine reading comprehension mainly gives an article and a question based on the article, and lets the machine answer the question after reading the article.

[0006] Early MRC systems were rule-based and performed poorly in many scenarios. With the rise of deep learning and large-scale datasets, deep learning-based MRC significantly outperforms rule-based MRC and even surpasses human performance. Common MRC tasks can be divided into four types: cloze, multiple choice, fragment extraction, and free response. Summary of the invention

[0007] In view of the above, the purpose of the present invention is to provide a CT image report information extraction method, device, computer equipment and storage medium based on multi-round question and answer, so as to realize the automatic extraction of CT image report information and improve the accuracy of information extraction.

[0008] To achieve the above-mentioned purpose of the invention, the embodiment provides a method for extracting CT image report information based on multi-round question-answering, comprising:

[0009] Step 1: preset a question-answering template according to the information extraction task, including a head entity question-answering question template and a tail entity question-answering question template;

[0010] Step 2: Build an information extraction model based on the reading comprehension framework;

[0011] Step 3: In the information extraction model training phase, first, use the sentence segmentation algorithm to extract each sentence in the CT image report; then, concatenate the sentence with each head entity question and answer question to obtain the input text required for extracting the head entity; then, according to the manually annotated head entity gold standard data, concatenate it with each corresponding tail entity question and answer question template to form a tail entity question and answer question for the head entity, and then concatenate the tail entity question and answer question with the sentence to obtain the input text required for extracting the tail entity; finally, input the constructed input text for extracting the head entity and tail entity into the information extraction model for model training optimization;

[0012] Step 4: In the information extraction model testing phase, first use the sentence segmentation algorithm to extract each sentence in the CT imaging report; then, concatenate the sentence with each head entity question to obtain the input text required for extracting the head entity, and input the text into the trained information extraction model to obtain the head entity as the first round of question-answering results; then, concatenate the extracted head entity with the tail entity question-answering question template to obtain the tail entity question-answering question for the head entity, and then concatenate the tail entity question-answering question with the sentence to obtain the input text required for extracting the tail entity, and input the text into the trained information extraction model again to obtain the tail entity as the second round of question-answering results;

[0013] Step 5: Pair the head entity and tail entity extracted from multiple rounds of question and answer into triplets, and standardize the extracted triplet information before outputting it.

[0014] Preferably, the reading comprehension framework includes BiLSTM, GRU, BERT, GPT, XLNET, etc., and sample data is used to optimize the reading comprehension framework parameters to build an information extraction model.

[0015] Preferably, in the training stage, each sentence in the CT imaging report is extracted, the head entity question and answer questions of each sentence are concatenated to form the head entity question and answer input text, the manually annotated head entity gold standard data is concatenated with the tail entity question and answer question template and the input text to form the tail entity question and answer input text, and a data set is constructed; the data set is divided into a training set, a validation set and a test set; the training set is used to train a reading comprehension framework to optimize the reading comprehension framework parameters, and the reading comprehension framework after parameter optimization is used as an information extraction model.

[0016] Preferably, in the training phase, in each iteration of model training, the information extraction model is evaluated using the validation set to obtain F1-marco. If the F1-marco value is better than the previous global best F1-marco value, the global best F1-marco value is updated and the information extraction model is saved. This process is continued until the end of the training, and the model with the global best F1-marco value is used as the final information extraction model.

[0017] Preferably, in the test phase, the head entity results obtained from the first round of question-answering are used in combination with the tail entity question-answering question template to generate the tail entity question-answering question, and the tail entity text is obtained by concatenating the text with the sentence; the tail entity text is input into the trained model to extract the tail entity. The accuracy of information extracted by the information extraction model is evaluated using the test set, and the gold standard data manually annotated in the test set is compared with the extraction results obtained by the information extraction model to obtain the precision, recall rate, and F1 value to evaluate the accuracy of information extracted by the information extraction model.

[0018] To achieve the above-mentioned purpose of the invention, the embodiment further provides a CT image report information extraction device based on multi-round question and answer, comprising:

[0019] The template construction module is used to preset question-answering templates according to the information extraction task, including the head entity question-answering question template and the tail entity question-answering question template;

[0020] The model building and optimization module is used to build an information extraction model based on the reading comprehension framework. When training the information extraction model, first, use the sentence segmentation algorithm to extract each sentence in the CT image report; then, splice the sentence with each head entity question and answer question to obtain the input text required for extracting the head entity; then, splice the manually annotated head entity gold standard data with each corresponding tail entity question and answer question template to form a tail entity question and answer question for the head entity, and then splice the tail entity question and answer question with the sentence to obtain the input text required for extracting the tail entity; finally, input the constructed input text for extracting the head entity and tail entity into the information extraction model for model training and optimization;

[0021] The multi-round question-answering extraction module is used to first extract each sentence in the CT imaging report using a sentence segmentation algorithm; then, the sentence is concatenated with each head entity question-answering question to obtain the input text required for extracting the head entity, and the text is input into the trained information extraction model to obtain the head entity as the result of the first round of question-answering; subsequently, the extracted head entity is concatenated with the tail entity question-answering question template to obtain the tail entity question-answering question for the head entity, and then the tail entity question-answering question is concatenated with the sentence to obtain the input text required for extracting the tail entity, and the text is input into the trained information extraction model again to obtain the tail entity as the result of the second round of question-answering;

[0022] The standardization module is used to pair the head entity and tail entity extracted from multiple rounds of question and answer into triplets, and output the extracted triplet information after standardization.

[0023] To achieve the above-mentioned purpose of the invention, an embodiment further provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores an information extraction model based on a reading comprehension framework, a head entity question and answer template, and a tail entity question and answer template, and the information extraction model is constructed by the above-mentioned CT image report information extraction method based on multi-round question and answer, and the processor implements the following steps when executing the computer program:

[0024] Step 1: First, use the sentence segmentation algorithm to extract each sentence in the CT imaging report; then, concatenate the sentence with each head entity question and answer question to obtain the input text required to extract the head entity, and input the text into the trained information extraction model to obtain the head entity as the first round of question and answer results;

[0025] Step 2: Use the extracted head entity and tail entity question answering template to obtain the tail entity question answering question for the head entity, then concatenate the tail entity question answering question with the sentence to obtain the input text required to extract the tail entity, and input the text into the trained information extraction model again to obtain the tail entity as the second round of question answering result;

[0026] Step 3: Pair the head entity and tail entity extracted from multiple rounds of question and answer into triplets, and output the extracted triplet information after standardization.

[0027] To achieve the above-mentioned purpose of the invention, an embodiment further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is processed and executed, the steps of the above-mentioned method for extracting CT image report information based on multi-round question and answer are implemented.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] According to the extraction task, the head entity question and answer question template and the tail entity question and answer question template are preset, and the information extraction model is constructed based on the reading comprehension framework. On this basis, the information extraction model is used in combination with the head entity question and answer question to extract the head entity of each sentence in the CT imaging report, and then the information extraction model is used in combination with the tail entity question and answer question template and the head entity to extract the tail entity of the sentence. The head entity and tail entity extracted by these two rounds form structured extracted information. Since the information is extracted under the guidance of the question and answer template, the extraction accuracy is high, and the information extraction model is used to automatically calculate the information, so the extraction efficiency is high. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 is a flow chart of a method for extracting CT image report information based on multi-round question and answer provided in an embodiment;

[0032] Figure 2 It is a structural schematic diagram of a CT image report information extraction device based on multi-round question and answer provided in an embodiment. DETAILED DESCRIPTION

[0033] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0034] In view of the problem that the accuracy of information extraction from existing CT image reports is not high, it has been found through research that, based on the principle of machine reading comprehension, it becomes very natural to convert the information extraction problem into a question-and-answer format, that is, to put forward the entity to be extracted in the form of a question, and let the reading comprehension model find the answer from the article. This makes it possible to extract structured information in one step and solve the error transmission problem existing in the pipeline-based method. Therefore, the embodiment provides a method, device, computer equipment and storage medium for extracting CT image report information based on multi-round question-and-answer.

[0035] Figure 1 FIG. 1 is a flow chart of a method for extracting CT image report information based on multi-round question-answering provided in an embodiment. Figure 1 As shown, the method for extracting CT image report information based on multi-round question-answering provided in the embodiment includes the following steps:

[0036] S1, preset question-answering templates according to the information extraction task, including head entity question-answering question templates and tail entity question-answering question templates.

[0037] In the embodiment, the information extraction task is generally to extract the subject and the related content of the subject, and these related contents may be the related attributes of the subject, or the related processing methods of the subject, etc. Based on such an information extraction task, the subject is generally considered as the head entity, and the head entity question and answer question is constructed around the head entity. At the same time, the related content of the subject is considered as the tail entity, and the tail entity question and answer question template is constructed around the tail entity.

[0038] S2, construct sample data and divide the sample data into training set, validation set, and test set.

[0039] In the embodiment, each sentence in the CT image report is extracted and spliced ​​with the pre-designed head entity question and tail entity question and answer template, and the two spliced ​​results are formed into a question and answer pair, which is used as sample data. The formed question and answer pair is suitable for the data format (question + answer) of the reading comprehension framework. The initial answer is a vector of all 0s. If the question has an answer, the corresponding subscript is changed to the serial number of the entity corresponding to the answer.

[0040] After obtaining the sample data, the sample data is divided into a training set, a validation set, and a test set. The specific division ratio may preferably be 7:1:2.

[0041] S3, builds an information extraction model based on the reading comprehension framework, and uses training sets, validation sets, and test sets to optimize and evaluate the information extraction model.

[0042] In the embodiment, the reading comprehension framework can adopt BiLSTM, GRU, BERT, GPT, XLNET, etc., preferably a BERT model. The parameters of the reading comprehension framework are optimized using the training set to construct an information extraction model, and the process includes: first, each sentence in the CT image report is extracted using a sentence segmentation algorithm; then, the sentence is spliced ​​with each head entity question and answer question to obtain the input text required for extracting the head entity; then, according to the manually annotated head entity gold standard data, it is spliced ​​with each corresponding tail entity question and answer question template to form a tail entity question and answer question for the head entity, and then the tail entity question and answer question is spliced ​​with the sentence to obtain the input text required for extracting the tail entity; finally, the input text for extracting the head entity and the tail entity is input into the information extraction model for model training optimization.

[0043] In each iteration of the training phase, the information extraction model is verified and evaluated using the validation set, and the F1-macro value is used as the evaluation result. If the F1-marco value exceeds the global optimal F1-marco value, the global optimal F1-marco value is updated, and the information extraction model after the validation set optimization parameters is saved. If the number of iterations is not reached, the training set is continued to be used to optimize the information extraction model parameters, and the validation set is continued to be used to verify and evaluate the information extraction model.

[0044] S4, using the information extraction model to perform two rounds of extraction tests on each sentence in the CT imaging report.

[0045] In the embodiment, each sentence in the CT imaging report is extracted, and the information extraction model is used in combination with the head entity question and answer question to perform the first round of extraction on the sentence to obtain the head entity. The specific process is: each sentence in the CT imaging report is extracted using a sentence segmentation algorithm; then, the sentence is spliced ​​with each head entity question and answer question to obtain the input text required to extract the head entity, and the text is input into the trained information extraction model to obtain the head entity as the result of the first round of question and answer. If the answer to the head entity question cannot be extracted from the sentence, the first round of extraction task ends.

[0046] In the embodiment, the head entity obtained by the first round of information extraction based on the head entity question is used to extract the tail entity information in the second round. The specific process is: the extracted head entity is spliced ​​with the tail entity question and answer question template to obtain the tail entity question and answer question for the head entity, and then the tail entity question and answer question is spliced ​​with the sentence to obtain the input text required to extract the tail entity, and the text is input into the trained information extraction model again to obtain the tail entity as the result of the second round of question and answer. If the answer to the tail entity question cannot be extracted from the sentence, the second extraction task ends, and the first and second rounds are linked by the head entity of the first round of the head entity question, so as to maximize the use of the information of the two rounds, so that the head entity and the tail entity have a strong correlation.

[0047] In the embodiment, the accuracy of information extracted by the information extraction model is evaluated using a test set, that is, the question-answer pairs contained in the test set are input into the information extraction model, and after two rounds of information extraction, the extracted information is obtained, and the extracted information is compared with the corresponding question-answer pair labels, and the accuracy of the information extracted by the information extraction model is evaluated by the evaluation indicators precision, recall, and F1 value. The information extraction model that has been evaluated and verified can be used for information extraction in clinical CT imaging reports.

[0048] S5, the head entity and tail entity extracted from multiple rounds of question and answer are paired into triples, and the extracted triple information is standardized and output.

[0049] After multiple rounds of extraction tasks are completed, the head entity and the tail entity are associated to form a completed triplet of extracted information. Since the same content may have different expressions in different CT imaging reports, the extracted information obtained is not standardized. To this end, in the embodiment, the extracted information is standardized, such as using the rules provided by the dictionary table to standardize the extracted information to obtain structured information and output it.

[0050] Figure 2 Schematic diagram of the structure of the CT image report information extraction device based on multi-round question-answering provided by the embodiment. Figure 2 As shown, the CT image report information extraction device provided in the embodiment includes:

[0051] The template construction module is used to preset question-answering templates according to the information extraction task, including the head entity question-answering question template and the tail entity question-answering question template;

[0052] The sample data construction module is used to divide the sample data into training sets, validation sets, and test sets;

[0053] Model building and optimization module, which is used to build an information extraction model based on the reading comprehension framework, and optimize and evaluate the information extraction model using training sets, validation sets, and test sets;

[0054] The multi-round question-answering extraction module uses the information extraction model to perform two rounds of extraction tests on each sentence in the CT imaging report.

[0055] The standardization module is used to perform two rounds of extraction tests on each sentence in the CT imaging report using the information extraction model.

[0056] It should be noted that the CT image report information extraction device provided in the above embodiment should be illustrated by the division of the above functional modules when performing CT image report information extraction. The above functions can be assigned to different functional modules as needed, that is, the internal structure of the terminal or server is divided into different functional modules to complete all or part of the functions described above. In addition, the CT image report information extraction device provided in the above embodiment and the CT image report information extraction method embodiment belong to the same concept. The specific implementation process is detailed in the CT image report information extraction method embodiment, which will not be repeated here.

[0057] The embodiment further provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores an information extraction model based on a reading comprehension framework, a head entity question and answer question template, and a tail entity question and answer question template, and the information extraction model is constructed by the above-mentioned CT image report information extraction method based on multi-round question and answer. When the processor executes the computer program, the following steps are implemented:

[0058] Step 1: First, use the sentence segmentation algorithm to extract each sentence in the CT imaging report; then, concatenate the sentence with each head entity question and answer question to obtain the input text required to extract the head entity, and input the text into the trained information extraction model to obtain the head entity as the first round of question and answer results;

[0059] Step 2: Use the extracted head entity and tail entity question answering template to obtain the tail entity question answering question for the head entity, then concatenate the tail entity question answering question with the sentence to obtain the input text required to extract the tail entity, and input the text into the trained information extraction model again to obtain the tail entity as the second round of question answering result;

[0060] Step 3: Pair the head entity and tail entity extracted from multiple rounds of question and answer into triplets, and output the extracted triplet information after standardization.

[0061] It should be noted that the CT image report information extraction method implemented by the computer device through steps 1-3 is the same as the CT image report information extraction method based on multi-round question and answer provided in the above embodiment, and will not be repeated here.

[0062] In practical applications, the computer memory can be a volatile memory at the near end, such as RAM, or a non-volatile memory, such as ROM, FLASH, floppy disk, mechanical hard disk, etc., or a remote storage cloud. The computer processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), that is, the extraction step of CT image report information can be implemented through these processors.

[0063] The embodiment further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is processed and executed, the steps of the above-mentioned method for extracting CT image report information based on multiple rounds of question and answer are implemented.

[0064] The CT image report information extraction method and device provided in the embodiment presets a question template based on the actual information extraction task, so that the CT image report information extraction method has strong versatility. An information extraction model is constructed based on a reading comprehension framework, and the information extraction model is used in combination with the head entity question and answer question to extract the head entity of each sentence in the CT image report, and then the information extraction model is used in combination with the tail entity question and answer question template and the head entity to extract the tail entity of the sentence. The head entity and tail entity extracted by these two rounds form structured extracted information. Since it is extracted under the guidance of the question and answer template, the extraction accuracy is high, and the information extraction model is automatically calculated, so the extraction efficiency is high.

[0065] Experimental example:

[0066] The information extraction task is to extract lung cancer staging information from lung CT image reports. Based on this extraction task, the constructed head entity question answering template and tail entity question answering template are shown in Table 1:

[0067] Table 1 Question template

[0068]

[0069] For multiple chest CT imaging reports, the labeling results of lung cancer staging related information by engineers with medical background are used as sample labels. At the same time, each sentence in multiple chest CT imaging reports is extracted, and sample data is constructed in combination with the question and answer template in Table 1, and then the information extraction model is constructed in accordance with step 2 of the above-mentioned multi-round question and answer CT imaging report information extraction method. The information extraction model, Pipeline-BERT and Pipeline-BiLSTM are verified using test samples. The verification results are shown in Tables 2 and 3. Analysis of Tables 2 and 3 shows that the information extraction model based on multi-round question and answer provided in the embodiment is superior to the models based on Pipeline-BERT and Pipeline-BiLSTM in terms of precision, recall and F1 value. The information extraction model achieves the best performance with an F1 score of 93.66%, and is significantly better than Pipeline-BERT and Pipeline-BiLSTM, with an increase of 11.35% and 15.10% in F1 scores, respectively.

[0070] Table 2

[0071]

[0072] Table 3 F1 values ​​of various features extracted by different methods of structured information

[0073]

[0074]

[0075] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for extracting CT image report information based on multi-round question-answering, characterized in that: The following steps are involved: Step 1: preset a question-and-answer template according to the information extraction task, including a head entity question-and-answer question template and a tail entity question-and-answer question template, wherein the information extraction task is to extract the subject and the relevant content of the subject, the relevant content is the relevant attribute of the subject or the relevant processing method of the subject, and the subject is considered as the head entity, and a head entity question-and-answer question is constructed around the head entity, and the relevant content of the subject is considered as the tail entity, and a tail entity question-and-answer question template is constructed around the tail entity. The information extraction task is specifically to extract lung cancer staging information in the lung CT imaging report; Step 2: Build an information extraction model based on a reading comprehension framework, where the reading comprehension framework includes BiLSTM, GRU, BERT, GPT, or XLNET; Step 3: In the information extraction model training phase, first, use the sentence segmentation algorithm to extract each sentence in the CT image report; then, concatenate the sentence with each head entity question answering question to obtain the input text required to extract the head entity; Subsequently, the manually annotated gold standard data of the head entity is spliced ​​with each corresponding question-answering template of the tail entity to form a question-answering question for the head entity. The question-answering question of the tail entity is then spliced ​​with the sentence to obtain the input text required to extract the tail entity, construct a data set, and divide the data set into a training set, a validation set, and a test set. Under supervision, the reading comprehension framework is trained using the training set to optimize the parameters of the reading comprehension framework. The reading comprehension framework with optimized parameters is used as the information extraction model. In each iteration of model training, the information extraction model is evaluated using the validation set to obtain F1-marco. If the F1-marco value is better than the previous global best F1-marco value, the global best F1-marco value is updated and the information extraction model is saved until the end of training. The model with the global best F1-marco value is used as the final information extraction model. Step 4: In the information extraction model testing phase, first use the sentence segmentation algorithm to extract each sentence in the CT image report; then, concatenate the sentence with each head entity question and answer question to obtain the input text required to extract the head entity, and input the text into the trained information extraction model to obtain the head entity as the first round of question and answer results; Subsequently, the extracted head entity is spliced ​​with the tail entity question-answering template to obtain the tail entity question-answering question for the head entity, and the tail entity question-answering question is spliced ​​with the sentence to obtain the input text required to extract the tail entity. The text is input into the trained information extraction model again to obtain the tail entity as the second round of question-answering result. If the answer to the tail entity question cannot be extracted from the sentence, the second extraction task is terminated. The first and second rounds are associated through the head entity of the head entity question of the first round, and the information of the two rounds is used to the maximum extent, so that the head entity and the tail entity have a strong correlation. Step 5: Pair the head entity and tail entity extracted from multiple rounds of question and answer into triplets, and standardize the extracted triplet information before outputting it.

2. The method for extracting CT image report information based on multi-round question-answering according to claim 1, characterized in that: During the testing phase, the head entity results obtained from the first round of question-answering are used in combination with the tail entity question-answering question template to generate a tail entity question-answering question, which is then concatenated with the sentence to obtain the text of the extracted tail entity. The text of the extracted tail entity is input into the trained model to extract the tail entity, and the test set is used to evaluate the accuracy of the information extracted by the information extraction model. The gold standard data manually annotated in the test set is compared with the extraction results obtained by the information extraction model to obtain the precision, recall rate, and F1 value, and to evaluate the accuracy of the information extracted by the information extraction model.

3. A CT image report information extraction device based on multi-round question-answering, characterized in that: include: A template construction module is used to preset question-and-answer templates according to information extraction tasks, including head entity question-and-answer questions and tail entity question-and-answer question templates, wherein the information extraction task is to extract the subject and the relevant content of the subject, the relevant content is the relevant attributes of the subject or the relevant processing method of the subject, and the subject is considered as the head entity, and a head entity question-and-answer question is constructed around the head entity, and the relevant content of the subject is considered as the tail entity, and a tail entity question-and-answer question template is constructed around the tail entity. The information extraction task is specifically to extract lung cancer staging information in a lung CT imaging report; Model building and optimization module, which is used to build information extraction models based on reading comprehension frameworks, including BiLSTM, GRU, BERT, GPT, or XLNET; When training the information extraction model, first, use the sentence segmentation algorithm to extract each sentence in the CT image report; then, concatenate the sentence with each head entity question and answer question to obtain the input text required to extract the head entity; Subsequently, the manually annotated gold standard data of the head entity is spliced ​​with each corresponding question-answering template of the tail entity to form a question-answering question for the head entity. The question-answering question of the tail entity is then spliced ​​with the sentence to obtain the input text required to extract the tail entity, construct a data set, and divide the data set into a training set, a validation set, and a test set. Under supervision, the reading comprehension framework is trained using the training set to optimize the parameters of the reading comprehension framework. The reading comprehension framework with optimized parameters is used as the information extraction model. In each iteration of model training, the information extraction model is evaluated using the validation set to obtain F1-marco. If the F1-marco value is better than the previous global best F1-marco value, the global best F1-marco value is updated and the information extraction model is saved until the end of training. The model with the global best F1-marco value is used as the final information extraction model. The multi-round question-answering extraction module is used to first extract each sentence in the CT image report using a sentence segmentation algorithm; then, the sentence is concatenated with each head entity question-answering question to obtain the input text required for extracting the head entity, and the text is input into the trained information extraction model to obtain the head entity as the first round of question-answering results; Subsequently, the extracted head entity is spliced ​​with the tail entity question-answering template to obtain the tail entity question-answering question for the head entity, and the tail entity question-answering question is spliced ​​with the sentence to obtain the input text required to extract the tail entity. The text is input into the trained information extraction model again to obtain the tail entity as the second round of question-answering result. If the answer to the tail entity question cannot be extracted from the sentence, the second extraction task is terminated. The first and second rounds are associated through the head entity of the head entity question of the first round, and the information of the two rounds is used to the maximum extent, so that the head entity and the tail entity have a strong correlation. The standardization module is used to pair the head entity and tail entity extracted from multiple rounds of question and answer into triplets, and output the extracted triplet information after standardization.

4. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The memory stores an information extraction model based on a reading comprehension framework, a head entity question and answer question, and a tail entity question and answer template, the information extraction model is constructed by the CT image report information extraction method based on multi-round question and answer according to claim 1 or 2, and the processor implements the following steps when executing the computer program: Step 1: First, use the sentence segmentation algorithm to extract each sentence in the CT imaging report; then, concatenate the sentence with each head entity question and answer question to obtain the input text required to extract the head entity, and input the text into the trained information extraction model to obtain the head entity as the first round of question and answer results; Step 2: Use the extracted head entity and tail entity question answering template to obtain the tail entity question answering question for the head entity, then concatenate the tail entity question answering question with the sentence to obtain the input text required to extract the tail entity, and input the text into the trained information extraction model again to obtain the tail entity as the second round of question answering result; Step 3: Pair the head entity and tail entity extracted from multiple rounds of question and answer into triplets, and standardize the extracted triplet information before outputting it.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is processed and executed, the steps of the method for extracting CT image report information based on multi-round question and answer as described in claim 1 or 2 are implemented.

Images