Breast disease corpus construction and case text standardization method based on BioClinicalBERT model

Abstract

The invention discloses a breast disease corpus construction method based on a BioClinicalBERT model, through a breast MRI corpus construction method, in combination with man-machine collaborative labeling and an advanced self-supervised learning technology, discriminative features are extracted from a breast MRI image, and through calculation of multistage similarity, the accuracy and efficiency of breast MRI diagnosis are effectively improved. Powerful support is provided for early diagnosis and treatment of breast diseases, and meanwhile the wide application potential in other medical image analysis and diagnosis fields is shown due to the universality. And when the method is applied to the case text standardization method, the case texts not only contain accurate structured tags and medical image information, but also have high readability and understandability. A doctor can quickly browse the medical record text and accurately obtain key information, so that diagnosis and treatment decisions can be more efficiently made.

Description

Technical Field

[0001] The present invention relates to corpus construction, and specifically to a breast disease corpus construction and case text standardization method based on the BioClinicalBERT model. Background Art

[0002] Breast MRI (magnetic resonance imaging) plays a vital role in the diagnosis of breast diseases. It can provide detailed information about tissue structure and lesions to help doctors make accurate diagnoses. However, the breast MRI diagnostic process usually involves a large amount of medical image and medical record text analysis, and in most medical record reports, doctors describe the disease in the form of unstructured natural language text reports of hundreds of words, which is not only time-consuming and labor-intensive, but the diagnostic results may be affected by the doctor's personal experience and subjective judgment. With the continuous development of medical technology and the increase in the number of patients, traditional breast MRI diagnostic methods have been unable to meet the growing medical needs. Therefore, improving the efficiency and accuracy of breast MRI diagnosis has become an urgent problem to be solved. Summary of the invention

[0003] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a breast disease corpus construction and case text standardization method based on the BioClinicalBERT model, which can assist in improving the accuracy and efficiency of breast MRI diagnosis. And when applied to the case text standardization method, doctors can quickly browse the medical record text and accurately obtain key information, thereby making diagnosis and treatment decisions more efficiently.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for constructing a breast disease corpus based on the BioClinicalBERT model, comprising the following steps:

[0005] S101 collects free-text medical reports related to breast diseases to construct the first corpus;

[0006] S102 manually annotating the medical reports in the first corpus through structured label assignment;

[0007] S103 trains a BioClinicalBERT model using the first corpus and the corresponding structured labels until the model can accurately identify and understand key information in the medical report;

[0008] S104 collects medical images related to breast diseases, and obtains free text medical reports associated with these images to construct a second corpus;

[0009] S105 uses the pre-trained BioClinicalBERT model to process the medical reports in the second corpus, extract key information from them, and convert them into structured tags for machine annotation;

[0010] S106 compares the machine annotation with the manual formal annotation, compares the two annotation results, and evaluates the annotation quality of the model;

[0011] S107 If the annotation quality is within the preset threshold range, the annotation result of the model is used as the final corpus; otherwise, it is necessary to expand the basic corpus again and retrain the model to improve the accuracy and consistency of the annotation.

[0012] As a further improvement of the present invention, the report in S104 at least includes description of the image, diagnosis, and treatment suggestion information.

[0013] As a further improvement of the present invention, the structured label in S105 at least includes information on the type, size, location, and severity of the lesion.

[0014] As a further improvement of the present invention, S102 manually annotates the first corpus at least twice, and compares the consistency of the two manual annotations. If they are inconsistent, the annotation specification is updated and re-annotated until the annotations are highly consistent.

[0015] A method for standardizing breast disease case text based on the BioClinicalBERT model, including

[0016] S201 constructs a breast disease corpus using the method for constructing a breast disease corpus based on the BioClinicalBERT model as described in claims 1 to 3;

[0017] S202 associates the medical image with its corresponding free-text medical report, and processes the medical report using a pre-trained BioClinicalBERT model to extract structured labels;

[0018] S203 associates the generated structured label with the medical image to ensure that each image has a corresponding structured label;

[0019] S204 unifies the text format and description terms of the text corresponding to the structured tags.

[0020] S205 generates standardized medical record text.

[0021] As a further improvement of the present invention, S202 also pre-processes the free text medical report, including removing special characters, punctuation marks, and stop words.

[0022] As a further improvement of the present invention, S202 further rearranges the paragraphs of the free text, adds titles and subtitles, and presents them in a list or table.

[0023] The beneficial effects of the present invention are that, through the breast MRI corpus construction method, combined with human-machine collaborative annotation and advanced self-supervised learning technology, discriminative features are extracted from breast MRI images, and the accuracy and efficiency of breast MRI diagnosis are effectively improved by calculating multi-level similarities. This innovative solution not only provides strong support for the early diagnosis and treatment of breast diseases, but also shows its potential for wide application in other medical image analysis and diagnosis fields due to its versatility. And when applied to the case text standardization method, it can also make these medical record texts not only contain accurate structured labels and medical image information, but also have a high degree of readability and comprehensibility. Doctors can quickly browse medical record texts and accurately obtain key information, so as to make diagnosis and treatment decisions more efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flowchart of a breast MRI corpus construction framework of the present invention;

[0025] Figure 2 A schematic diagram of the data preparation and model training process of the present invention;

[0026] Figure 3 It is a schematic diagram of the model application and evaluation optimization of the present invention. DETAILED DESCRIPTION

[0027] The present invention will be further described below in detail with reference to the embodiments shown in the accompanying drawings.

[0028] Reference Figure 1-3 As shown,

[0029] A method for constructing a breast disease corpus based on the BioClinicalBERT model, which fully utilizes the advantages of artificial intelligence and manual annotation, aims to improve the efficiency and accuracy of breast MRI corpus construction. It mainly includes the following steps:

[0030] Step 1: Data preparation and model training

[0031] Free-text medical reports related to breast diseases were collected to construct the first corpus and preprocessed. Medical experts were invited to assign structured labels to the medical reports in the first corpus to ensure that key information was accurately labeled. The BioClinicalBERT model was trained using the first corpus and the corresponding structured labels until the model could accurately identify and understand the key information in the medical reports.

[0032] Step 2: Model application and evaluation optimization

[0033] First, we need to collect medical images related to breast diseases and obtain the free text medical reports associated with these images. These reports may contain doctors’ descriptions of the images, diagnoses, treatment recommendations, etc., so as to construct the second corpus.

[0034] Furthermore, the pre-trained BioClinicalBERT model is used to process the medical reports in the second corpus, extract key information from them, and convert them into structured labels. These labels may include information such as lesion type, size, location, severity, etc.

[0035] Furthermore, the machine annotations are compared with the formal manual annotations. A portion of the text data will be automatically annotated by the model, and the other portion will be formally annotated by humans. Then, the results of the two annotation methods are compared to evaluate the annotation quality of the model. The comparison may involve the use of some annotation consistency metrics, such as precision, recall, etc.

[0036] Furthermore, the annotation consistency is evaluated. The annotation quality of the model can be evaluated by calculating the consistency between the machine annotation and the manual formal annotation. This usually involves comparing the results of the two annotation methods and calculating the difference between them. If the difference is within the preset threshold, the annotation quality of the model can be considered to be high, and the annotation results of the model can be accepted as the final corpus. Otherwise, it is necessary to reconsider expanding the basic corpus and retrain the model to improve the accuracy and consistency of the annotation.

[0037] Step 3: Medical record text standardization

[0038] In this step, the first step is to associate the medical images with their corresponding free-text medical reports. This may involve building a system that can match the unique identifiers of the medical images (such as file names, IDs, etc.) with the identifiers of the medical reports. Once the medical images and medical reports are correctly associated, the next step is to process these medical reports using the pre-trained BioClinicalBERT model to extract structured labels.

[0039] Free-text medical reports were preprocessed, including removing special characters, punctuation marks, stop words, etc., to facilitate model processing.

[0040] Furthermore, the preprocessed text is input into the BioClinicalBERT model, which uses its ability to understand text semantics to extract key information and features.

[0041] Furthermore, based on the features of the model output and combined with knowledge in the medical field, corresponding structured labels are generated, such as lesion type, size, location, severity, etc.

[0042] Furthermore, the generated structured labels are associated with medical images to ensure that each image has a corresponding structured label. This can be achieved by establishing an association between images and labels in the database.

[0043] Furthermore, the medical record text is standardized. The main goal is to standardize the text extracted from the medical report to improve its readability and comprehensibility and ensure consistency in subsequent processing. This includes the following aspects.

[0044] Consistent terminology. Ensure consistent terminology and expressions are used to describe the same concepts, avoiding confusion with synonyms or near-synonyms.

[0045] Standardize the format of the text, including font, font size, paragraph spacing, etc., to improve the overall aesthetics and readability of the document.

[0046] Optimize for readability. Adjust the structure of the text to make it easier to read and understand. This may involve rearranging paragraphs, adding headings and subheadings, and presenting information using lists or tables.

[0047] The standardized medical record text is finally generated, which makes it convenient for doctors to quickly and accurately obtain key information in the medical record and provide support for the diagnosis and treatment of breast diseases.

[0048] The present invention designs a breast disease corpus construction framework based on the BioClinicalBERT model, which combines the advantages of manual annotation and automatic annotation. An initial data set is formed by collecting a large amount of text data such as case reports, teaching materials, and medical literature related to breast MRI. Subsequently, meticulous data preprocessing is performed, which includes unifying the text format and removing noise information to ensure the quality and standardization of the data set.

[0049] During the annotation stage, an iterative approach of two manual pre-annotations and formal annotation was adopted.

[0050] First, detailed annotation specifications are developed based on the knowledge dimension and classification method in the medical field of breast MRI.

[0051] Furthermore, the processed breast MRI medical record text dataset was subjected to lexical and syntactic analysis to construct event triples, and a portion of the text was manually pre-annotated twice based on these triples. By comparing the results of the two pre-annotations, the consistency of the annotations was evaluated, and the annotation specifications were updated as needed. This iterative process ensured the accuracy and practicality of the annotation specifications.

[0052] In the formal annotation stage, another part of the text is manually formally annotated twice based on the updated annotation specifications, and a basic corpus is constructed.

[0053] Similar to the pre-annotation stage, the annotation consistency of the two formal annotations is calculated, and the results are used to determine whether further iteration or adjustment is needed. When the annotation consistency reaches the preset threshold, the manual annotation process is considered to be completed, and the manual formal annotation results are output as the first corpus. The first corpus is mainly used to train the BioClinicalBERT model to enable it to identify and understand the key information in medical reports on breast diseases. The collection of selected free-text medical reports on breast diseases contains detailed information on breast diseases, such as diagnosis results, description of the condition, treatment recommendations, etc. These reports have been assigned detailed structured labels by medical experts, which means that the key information in the reports (such as disease type, symptoms, test results, etc.) has been labeled or classified to facilitate model learning.

[0054] In order to improve the fine-grained retrieval accuracy of breast MRI diagnostic medical record text, the present invention designs a multi-level similarity calculation method based on deep learning and natural language processing technology.

[0055] First, we used the pre-trained BioClinicalBERT model, a deep learning model based on the Transformers architecture, which has been trained on a large-scale biomedical text corpus and can accurately capture the semantic information in biomedical text. We used this model to perform deep feature extraction on sentences in the breast MRI corpus and convert each sentence into a high-dimensional vector representation that can fully reflect the semantic content of the sentence.

[0056] Furthermore, the cosine similarity between these sentence vectors is calculated, which is a method to measure the cosine value of the angle between vectors, which can effectively reflect the similarity between vectors.

[0057] Furthermore, considering only cosine similarity may not be sufficient to accurately reflect the semantic co-occurrence relationship between sentences, so a normalization strategy is introduced to appropriately adjust the similarity to ensure its accuracy in different contexts.

[0058] Furthermore, based on the similarity calculation between sentences, a sentence similarity matrix is ​​constructed. Each row and column of this matrix represents a sentence, and the values ​​in the matrix are the similarity scores between the corresponding sentences. The similarity matrix can clearly show the semantic relationship between sentences, providing an important basis for subsequent retrieval and diagnosis.

[0059] At the same time, we also consider that a breast MRI report may contain multiple sentences, and there are certain semantic relationships between these sentences. Therefore, we further construct a report similarity matrix, and obtain the overall similarity between reports by comprehensively considering the similarities between all sentences in the report. This helps to more accurately judge the correlation between different reports, thereby achieving accurate retrieval of breast MRI medical record texts.

[0060] Finally, by jointly analyzing the multi-level similarity matrix, combining the information of the sentence similarity matrix and the report similarity matrix, accurate retrieval and diagnosis of breast MRI images are achieved. This method not only takes into account the semantic similarity between texts, but also fully considers the structural relationship and contextual information between texts, making the retrieval results more accurate and reliable. Through the multi-level similarity calculation method of the present invention, a more accurate and efficient retrieval tool can be provided for breast MRI diagnosis, providing strong support for doctors.

[0061] In another aspect of the present invention, a second corpus is constructed, i.e., a collection of free text medical reports associated with medical images of breast diseases. These reports may be similar to those in the first corpus, but are more focused on descriptions and diagnostic information related to specific medical images (such as mammary X-rays, MRI scans, etc.). The second corpus is used to apply the trained BioClinicalBERT model to actual scenarios to extract key information in the reports and generate corresponding structured tags for medical images. These structured tags can help doctors quickly understand the key information in medical records, thereby more accurately diagnosing and treating breast diseases. The second corpus trains the corpus automatic annotation model. After the training is completed, the model is used to machine-annotate new breast disease medical record texts, and again compared with the manual formal annotation results to calculate the annotation consistency. If the consistency of the machine annotation and the manual annotation reaches or exceeds a preset threshold, the trained corpus automatic annotation model is output and used as the final corpus tool.

[0062] However, if the annotation consistency fails to reach the preset threshold, further optimization measures will be taken. This includes expanding the basic corpus to introduce more diverse and complex text data; optimizing the annotation specifications to further improve the accuracy and consistency of annotations; and retraining the corpus automatic annotation model to improve its performance on new data. The entire process embodies the idea of ​​human-computer collaboration and iterative optimization, ensuring the efficiency, accuracy and reliability of the corpus automatic annotation model, and providing strong support for research and application in fields such as natural language processing and text mining.

[0063] Structured label generation based on deep learning models can identify and extract key information from medical record text, such as disease type, lesion severity, symptom description, etc. These labels not only provide doctors with a clear and concise overview of information, but also ensure the accuracy of the information. By associating structured labels with medical images (such as mammograms, MRI scans, etc.), a complete and consistent information system is constructed, allowing doctors to view medical record text and medical images at the same time, thereby gaining a more comprehensive understanding of the patient's condition.

[0064] On this basis, the medical record text was subjected to in-depth standardization to ensure its readability and comprehensibility. First, the terminology was unified to ensure that the use of medical terms in the medical record text was consistent, avoiding misunderstandings caused by terminology confusion. Secondly, the format of the medical record text was standardized to give it a clear and easy-to-read layout and typesetting. Finally, the readability of the medical record text was optimized by simplifying the language and adjusting the sentence structure to make it easier for doctors to understand and use.

[0065] More specific additional explanations are as follows:

[0066] A method for constructing a breast disease corpus based on the BioClinicalBERT model is used to solve the problem that existing free medical text sentences are lengthy and the descriptions are not standardized. The method for constructing the breast MRI corpus framework mainly includes three stages: data preparation and model training, model application and evaluation optimization, and medical record text standardization. The data preparation and model training stage includes data collection and processing, assigning structured labels, and training the BioClinicalBERT model; the model application and evaluation optimization stage includes the application of the BioClinicalBERT model for annotation, and comparing the results of manual annotation and machine annotation; the case text standardization stage includes the standardization of existing free medical texts and the standardization of future medical text formats.

[0067] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] Figure 1 A flowchart of a breast MRI corpus construction framework provided by the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0069] Step 101: construct a first breast MRI corpus and a corpus automatic annotation model BioClinicalBERT. The text in the corpus includes clinical case records and medical reports containing detailed disease descriptions and diagnosis results.

[0070] When constructing the first breast MRI corpus, we can use clinical medical records containing detailed disease descriptions and diagnosis results as raw data, and invite domain experts to assign structured labels to the medical reports in the first corpus. Here, the symptoms of breast diseases are divided into "111", "222", "333", and "444". These data can provide valuable information for BioClinicalBERT model training. However, in the process of using these data, we need to pay attention to privacy protection issues to ensure that patient information is fully protected. In order to complete the types of diseases included in the database, we can also obtain descriptions and diagnostic methods of related diseases from authoritative medical books, research reports, journals and other channels to improve the accuracy and reliability of the data.

[0071] The schematic diagram of data preparation and model training process is as follows Figure 2 As shown, the details are as follows:

[0072] Step 1011: pre-process the database to obtain a processed database.

[0073] As an optional implementation manner, preprocessing the database to obtain a processed database specifically includes:

[0074] The text is formatted to obtain a text in a preset format.

[0075] Step 1012: The text in the preset format is assigned structured labels to obtain assigned breast disease symptom labels; the breast disease symptom labels include 111, 222, 333, and 444.

[0076] Step 1013: Obtain the assigned breast disease symptom label results and train the BioClinicalBERT model.

[0077] In practical applications, based on the model framework construction of the BioClinicalBERT model, the present invention takes breast disease medical report corpus data as an example, combines multi-source breast disease medical report texts such as breast MRI-related case reports, textbooks, medical literature, etc., and combines the structured labels assigned by field experts to the medical reports in the first corpus to prepare the training corpus of the BioClinicalBERT model and conduct model training, which is specifically divided into the following steps: data collection and processing, annotation specification formulation, lexical and syntactic analysis, and model training.

[0078] The existing medical report texts mainly include electronic and handwritten forms. There is some private information such as the doctor's name and the patient's name in the case report. The formats of electronic medical records are different, and there are some images in the electronic medical records, which is not conducive to the subsequent text label assignment work, so the original data needs to be processed. First, the collected data is converted into a format, and the texts of different formats are uniformly encoded in UTF-8 and stored in txt format. Secondly, the doctor's name, patient's name, and pictures in the text are removed to facilitate subsequent work.

[0079] In most medical diagnoses, doctors describe diseases in the form of unstructured natural language text reports of hundreds of words. Therefore, domain experts were invited to assign structured labels to the medical reports in the first corpus and then train the BioClinicalBERT model.

[0080] The first corpus text in the preset format is decoupled, split into sentences, and lexical and syntactic analysis is performed to construct event triples. In practical applications, lexical and syntactic analysis is performed, firstly Chinese word segmentation is performed on the text and the lexicon is introduced; then syntactic analysis is performed, and event triples are constructed in combination with semantic role analysis.

[0081] In order to obtain sentence similarity, the present invention intends to use the pre-trained BioClinicalBERT model as a feature extractor, and proposes a normalized similarity to measure the similarity between different sentences. This similarity measure eliminates the influence of common semantics between sentences and can better reflect the true similarity. The main forms are as follows:

[0082]

[0083] in is the i-th sentence, s j represents the encoded vector representation, K represents the number of sentences in the corpus, <.> represents the cosine similarity function, p is the average representation of all sentences, o * is the average similarity of each sentence with the average representation.

[0084] Next, calculate the similarity matrix:

[0085]

[0086] in represents the similarity matrix, represents the sentence representation matrix.

[0087] Similarly, the report similarity matrix C can be obtained r .

[0088] Step 102: Use the trained BioClinicalBERT model to machine-annotate a portion of the text, and compare it with the results of manual formal annotation to evaluate the annotation quality of the model and perform corresponding optimization.

[0089] Model application and evaluation optimization diagram Figure 3 As shown, the details are as follows:

[0090] Step 1021: construct a second corpus based on relevant medical report text data.

[0091] Step 1022: Apply the trained BioClinicalBERT model to the medical reports in the second corpus to automatically extract key information and generate corresponding structured tags.

[0092] Step 1023: Compare the generated corresponding structured tags with the results of manual annotation. The annotation quality of the model is evaluated by calculating the annotation consistency of the two annotation methods.

[0093] As an optional implementation, step 1023 specifically includes:

[0094] The annotation consistency between the generated corresponding structured label and the manual annotation result is calculated.

[0095] It is determined whether the labeling consistency reaches a preset threshold, and a determination result is obtained.

[0096] If the judgment result is that the annotation consistency reaches the preset threshold, it means that the annotation quality of the trained BioClinicalBERT model is high, and the trained BioClinicalBERT model is output.

[0097] The trained BioClinicalBERT model was used to construct a breast MRI corpus.

[0098] If the judgment result is that the annotation consistency does not reach the preset threshold, the basic corpus is expanded and the process returns to "training the BioClinicalBERT model" (step 1013).

[0099] In actual applications, 50 reports randomly selected during manual annotation are also used for machine annotation in order to calculate the annotation consistency. The 50 manually annotated reports are used as a benchmark, and the machine annotations are compared with the manual annotations for the part of the reports to determine whether the annotation consistency with the 50 machine annotated documents reaches 70%. If it does not reach 70%, the basic corpus needs to be expanded to continue training the model until the breast MRI corpus is finally formed. If it reaches 70%, the machine annotation is completed to form the final breast MRI corpus.

[0100] Specifically, after manual and machine annotation is completed, the annotation consistency of the corpus is analyzed and compared. The present invention uses F value or Kappa value to evaluate the annotation consistency of the corpus. Among them, Kappa value is usually used to evaluate the positive and negative examples of annotation. For the corpus of the present invention, only unannotated corpus can be used as negative examples, which is difficult to count. For this situation, the F value is generally used for evaluation. Let the two annotation parties (manual and machine) be A and B respectively, and A be used as the reference group (the selection of the reference group is random) to calculate the accuracy P of B. The calculation formula is:

[0101]

[0102] Then the recall rate R of B is:

[0103]

[0104] Based on the accuracy P and recall R of B, the F value is calculated:

[0105]

[0106] The higher the F value, the more accurate the B annotation is, that is, the higher the annotation consistency is. When calculating entity annotation consistency, the annotation is considered consistent when the entity string and type are the same. The above indicators are also used when calculating relationship consistency. The annotation is consistent when the entity pairs and the relationship between entities are the same.

[0107] If the consistency of the machine-annotated corpus does not reach 70%, the basic corpus needs to be expanded, and 1,000 pieces of data need to be manually annotated to expand the basic corpus. After the expansion is completed, iterate from the first step until the annotation consistency is greater than 70%.

[0108] Step 103: Standardize the medical record text to improve the readability and comprehensibility of the medical record text. The standardization process includes terminology unification, format specification, readability optimization, etc.

[0109] The construction of the breast MRI corpus framework in the present invention has reference value for the research on data-driven corpus construction methods, and the formed breast MRI corpus framework has a practical supporting role in the text standardization of subsequent breast disease medical reports.

[0110] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0111] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for constructing a breast disease corpus based on the BioClinicalBERT model, characterized in that: The steps include: S101 collects free-text medical reports related to breast diseases to construct the first corpus; S102 manually annotating the medical reports in the first corpus through structured label assignment; S103 trains a BioClinicalBERT model using the first corpus and the corresponding structured labels until the model can accurately identify and understand key information in the medical report; S104 collects medical images related to breast diseases, and obtains free text medical reports associated with these images to construct a second corpus; S105 uses the pre-trained BioClinicalBERT model to process the medical reports in the second corpus, extract key information from them, and convert them into structured tags for machine annotation; S106 compares the machine annotation with the manual formal annotation, compares the two annotation results, and evaluates the annotation quality of the model; S107 If the annotation quality is within the preset threshold range, the annotation result of the model is used as the final corpus; Otherwise, the base corpus needs to be expanded and the model retrained to improve the accuracy and consistency of annotations.

2. The method for constructing a breast disease corpus based on the BioClinicalBERT model according to claim 1, characterized in that: The report in S104 includes at least description of the image, diagnosis, and treatment recommendation information.

3. The method for constructing a breast disease corpus based on the BioClinicalBERT model according to claim 1 or 2, characterized in that: The structured label in S105 includes at least lesion type, size, location, severity, symptoms, and examination result information.

4. The method for constructing a breast disease corpus based on the BioClinicalBERT model according to claim 3, characterized in that: S102 manually annotates the first corpus at least twice, and compares the consistency of the two manual annotations. If they are inconsistent, the annotation specification is updated and re-annotated until the annotations are highly consistent.

5. A method for standardizing case text of breast diseases based on the BioClinicalBERT model, characterized in that: include S201 constructs a breast disease corpus using the method for constructing a breast disease corpus based on the BioClinicalBERT model as described in claims 1 to 4; S202 associates the medical image with its corresponding free-text medical report, and processes the medical report using a pre-trained BioClinicalBERT model to extract structured labels; S203 associates the generated structured label with the medical image to ensure that each image has a corresponding structured label; S204 unifies the text format and description terms of the text corresponding to the structured tags. S205 generates standardized medical record text.

6. The method for standardizing case text of breast diseases based on the BioClinicalBERT model according to claim 5, characterized in that: S202 also preprocesses free-text medical reports, including removing special characters, punctuation marks, and stop words.

7. The method for standardizing case text of breast diseases based on the BioClinicalBERT model according to claim 5, characterized in that: S202 also rearranges paragraphs of free text, adds headings and subheadings, and presents them through lists or tables.

Images