Structured information processing method for imaging report text, lung disease monitoring method and system

Abstract

The present application relates to a structured information processing method for imaging report text, a lung disease monitoring method and system. The structured information processing method includes: S11, further dividing the part and morphological features in the imaging medical professional entity into two categories, negative and positive, to obtain eight entity labels, namely, vacancy filler, sentence starter, sentence terminator, part-negative, part-positive, morphology-negative, morphology-positive and disease name, performing named entity recognition on the imaging report text based on the eight entity labels, and extracting the named entity output in BIO format; S12, filtering out redundant information marked as O and entity labels as part-negative and morphology-negative based on the extracted named entities, calculating the sentence vector and storing it in the database. The present application realizes the efficient extraction of symptom information and obtains a more expressive sentence vector, so that a more accurate cluster of similar cases can be obtained for spatiotemporal distribution feature analysis.

Description

Technical Field

[0001] The present application relates to medical text processing and analysis technology, and more specifically, to a structured information processing method and system for imaging report text, a lung disease monitoring method and system based on imaging report text, and an electronic device. Background Art

[0002] The processing and analysis of medical texts has always been a focus of research both domestically and internationally. Medical texts contain a wealth of information, including patient admission records, pathology reports, and imaging reports, and play an important guiding role in clinical diagnosis. Unlike the analysis and application of English medical texts, which are the focus of much foreign research, the unique nature of the Chinese language, with its lack of clear delimiters, word roots, and prefixes, makes text analysis in Chinese more difficult. Furthermore, data encoding varies from hospital to hospital, and doctors' methods of writing electronic medical records vary widely. This results in electronic medical records containing a large amount of meaningless punctuation and stop words, as well as problems such as inconsistent formatting, ambiguous word usage, abbreviations, and spelling errors. These issues pose significant challenges to subsequent text mining and analysis.

[0003] Most medical texts are stored in a semi-structured format. Unstructured text offers greater expressiveness and presentation, but is not conducive to subsequent in-depth analysis. Therefore, extracting unstructured text information and implementing structured representations are crucial for text analysis. Currently, rule-based methods and deep learning-based methods are commonly used. Rule-based methods require extensive and in-depth domain expertise and rely on expert rule construction for development, resulting in high labor costs. Deep learning-based methods rely on labeled data and designed network structures for end-to-end training. Their advantage is that they do not require manual feature screening, which reduces the labor cost of expert feature extraction. They also possess strong generalization capabilities and are currently receiving widespread attention.

[0004] The most important step in achieving structured representation of medical text is nomenclature recognition. Nomenclature recognition involves extracting meaningful words from arbitrary text. These words serve as structured information for subsequent in-depth analysis. Generally, nomenclature recognition approaches fall into three categories: rule-based and dictionary-based methods, traditional statistical analysis, and neural network training.

[0005] Dictionary-based approaches rely on a dictionary of specialized terminology and then utilize matching algorithms for named entity recognition. Therefore, for highly specialized texts, such as medical texts, the size and quality of the annotated corpus significantly impacts model performance. In the practice of medical text analysis, annotation standards and corpus construction are under ongoing exploration, including the development of annotation standards for Chinese electronic medical records in 2015, which referenced the i2b2 2010 annotation standards. Rule-based approaches analyze and process the entire text by constructing rule templates, with named entity recognition implemented through matching. This approach is more intuitive and easier to maintain, but it also requires domain experts to spend time constructing rules, and template construction becomes very difficult without clear rules.

[0006] In response to the difficulties of dictionary- and rule-based approaches, statistical machine learning methods have gradually attracted considerable attention. Traditional machine learning methods for named face recognition require large labeled datasets, transforming the task into a classification problem. Commonly used sequence labeling models include the Hidden Markov Model (HMM), Maximum Entropy Model (EM), Conditional Random Field (CRF), and Support Vector Machine (SVM).

[0007] Neural networks are large, complex models composed of a large number of interconnected nonlinear units that can characterize complex nonlinear dynamic characteristics. They also rely on a large amount of data to drive model training, have self-organization, self-adaptation, and self-learning capabilities, and are particularly suitable for handling imprecise information processing problems that require simultaneous consideration of multiple factors. A neural network model based on CNN-BiLSTM-CRF has been proposed in the prior art to reduce the workload of manual feature extraction. This involves using a convolutional neural network to train character vectors with morphological features and word vectors with semantic feature information, combining the two and inputting them into a BiLSTM-CRF model. For details, see the following literature: Kupiec J, Robust part-of-speech tagging using a hidden Markov model [J]. Computer Speech & Language, 1992, 6(3): 2250-242. In order to extract the general features of latent semantic information and syntax in sentences, some people have proposed integrating the language model and the sentence-level reading control gate into the BiLSTM-CRF model. The reading control gate is used to integrate the implicit information of the sentence, and the language model learning can represent richer potential features. For details, please refer to the following literature: Li L, Jiang Y. Integrating language model and reading control gate in BiLSTM-CRF for biomedical named entity recognition [A]. In IEEE International Conference on Bioinformatics and Biomedicine (BIBM) [C]. Kansas: IEEE computer society, 2017: 380-385. To solve the problem of difficulty in utilizing the information of the entire text at the sentence level, there is a solution in the existing technology that introduces the attention mechanism into the BiLSTM-CRF model, and uses the attention mechanism to obtain the contextual representation of the current word in the entire text. For details, please refer to the following literature:

[12] Yang Pei, Yang Zhihao, Luo Ling, et al. Named entity recognition of chemical drugs based on attention mechanism [J]. Computer Research and Development, 2018, 55(7): 1548-1556.

[0008] However, existing technologies lack methods and systems for extracting information from imaging report texts to analyze the characteristics of lung diseases and thus monitor lung diseases. Summary of the Invention

[0009] The technical problem to be solved by this application is to provide a structured information processing method and system for imaging report text, a lung disease monitoring method and system based on imaging report text, and an electronic device in response to the above-mentioned defects of the prior art.

[0010] In order to solve the technical problem, the present application proposes a method for processing structured information of an imaging report text, comprising the following steps:

[0011] S11. Further classify the part and morphological features in the medical imaging professional entity into two categories, negative and positive, to obtain eight entity labels, namely, a blank filler, a sentence starter, a sentence terminator, part-negative, part-positive, morphology-negative, morphology-positive, and a disease name; perform named entity recognition on the imaging report text based on the eight entity labels, and extract and obtain named entity output in BIO format;

[0012] S12. Based on the named entities extracted in step S11, redundant information marked as O and entity labels as part-negative and morphology-negative is filtered out, a sentence vector of the imaging report text is calculated, and the sentence vector is stored in a database.

[0013] According to an embodiment of the first aspect of the present application, in step S11, a BERT-BiLSTM-CRF model is used to perform named entity recognition on the radiology report text.

[0014] According to an embodiment of the first aspect of the present application, calculating the sentence vector in step S12 further includes: using the word vectors of the words corresponding to the three entity labels of part-positive, morphology-positive and disease name retained after removing redundant information to calculate the sentence vector, and setting respective weights for the word vectors corresponding to different entity labels when calculating the sentence vector, wherein the word vector is the output of the BERT model in step S11.

[0015] To solve the technical problem, the present application proposes a lung disease monitoring method based on imaging report text, comprising the following steps:

[0016] S21. Processing the lung imaging report text using the structured information processing method for imaging report text as described above to construct a database;

[0017] S22. Search the database, calculate the similarity between the sentence vectors of each imaging report text, and classify all imaging report texts with similarity greater than a threshold as similar cases;

[0018] S23. Analyze the spatiotemporal distribution characteristics of all imaging report texts classified as similar cases.

[0019] According to an embodiment of the second aspect of the present application, step S22 further includes:

[0020] Get n imaging report texts within a period of time;

[0021] Calculating the cosine similarity between the n radiology report texts to obtain an nxn cosine similarity matrix;

[0022] Taking each radiology report text as a graph node, constructing a weighted undirected graph representation between the radiology report texts according to the cosine similarity matrix and the cosine similarity threshold, and obtaining a weighted adjacency matrix A of the radiology report texts;

[0023] Calculate the Karplast matrix L of the weighted undirected graph representation between the texts of the radiology reports, L = D – A, where D is the degree matrix and D is a diagonal matrix with dimension nxn;

[0024] Calculate the normalized Laplace matrix L', L' = D -1 / 2 LD -1 / 2 ,The spectral clustering method is used to find the imaging report texts with similar symptoms, and obtain several case clusters of similar texts.

[0025] According to an embodiment of the second aspect of the present application, step S23 further includes:

[0026] Extract the word sets corresponding to the location-positive and morphology-positive entity labels in the imaging report texts that are grouped in the same case cluster as the imaging feature representation of this type of disease;

[0027] According to the consultation time and hospital location of the cases, the spatial distribution and temporal distribution of the cases corresponding to the imaging report texts of the same case cluster are drawn respectively.

[0028] To solve the technical problem, the present application proposes a structured information processing system for imaging report texts in a third aspect, comprising:

[0029] The named entity recognition module is used to further classify the location and morphological features of medical imaging entities into negative and positive categories, obtaining eight entity labels, namely, blank filler, sentence starter, sentence terminator, location-negative, location-positive, morphology-negative, morphology-positive, and disease name. Based on these eight entity labels, the named entity recognition is performed on the imaging report text to extract the named entity output in BIO format;

[0030] A sentence vector construction module is used to filter out redundant information marked as O and entity labels as part-negative and morphology-negative based on the named entities extracted by the named entity recognition module, calculate the sentence vector of the imaging report text and store the sentence vector in the database.

[0031] To solve the technical problem, the present application proposes a lung disease monitoring system based on imaging report text in a fourth aspect, comprising:

[0032] The structured information processing system for imaging report texts as described above is used to perform structured information processing on lung imaging report texts and build a database;

[0033] The similar case classification module is used to search the database, calculate the similarity between the sentence vectors of each imaging report text, and classify all imaging report texts with similarity greater than a threshold as similar cases;

[0034] The spatiotemporal feature analysis module is used to analyze the spatiotemporal distribution characteristics of all imaging report texts classified as similar cases.

[0035] In order to solve its technical problems, the present application proposes an electronic device in the fifth aspect, including a processor and a memory, wherein the memory stores a computer program, and is characterized in that when the computer program is executed by the processor, the steps of the structured information processing method of the imaging report text as described above are implemented.

[0036] In order to solve its technical problems, the present application proposes an electronic device in the sixth aspect, including a processor and a memory, wherein the memory stores a computer program, and is characterized in that when the computer program is executed by the processor, the steps of the lung disease monitoring method based on the imaging report text as described above are implemented.

[0037] The implementation of the structured information processing method of the imaging report text of the present application, the lung disease monitoring method and system based on the imaging report text, and the electronic device has the following beneficial effects: the structured information processing method of the imaging report text according to the embodiment of the present application distinguishes between negative and positive descriptions of entity labels when performing named entity recognition, and can efficiently filter out redundant word descriptions and negative feature expressions in the report text, while retaining positive features; the structured information processing method of the imaging report text according to the embodiment of the present application further calculates the imaging report text sentence vector based on the weighted type of named entity label to obtain a more expressive sentence vector. The lung disease monitoring method according to the embodiment of the present application uses the above-mentioned structured information processing method of the imaging report text to process a large number of lung imaging report texts, and can obtain more accurate similar case clusters for spatiotemporal distribution feature analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present application will be further described below with reference to the accompanying drawings and embodiments, in which:

[0039] Figure 1This is a flowchart of a method for processing structured information of an imaging report text according to an embodiment of the present application;

[0040] Figure 2 This is a schematic diagram of the BERT-BiLSTM-CRF model used in this application;

[0041] Figure 3 This is a flow chart of a lung disease monitoring method based on imaging report text according to an embodiment of the present application;

[0042] Figure 4 This is a logical structure diagram of a lung disease monitoring system based on imaging report text according to an embodiment of the present application;

[0043] Figure 5 This is a schematic diagram of the logical structure of an electronic device according to an embodiment of the present application;

[0044] Figure 6 It is a schematic diagram of the logical structure of an electronic device according to another embodiment of the present application. DETAILED DESCRIPTION

[0045] To make the objectives, technical solutions, and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely intended to explain this application and are not intended to limit this application. Furthermore, the embodiments and features of the embodiments in this application may be combined with each other unless there is a conflict.

[0046] This application relates to a structured information processing method for imaging report text. This method uses named entity recognition within the text to extract entity information containing the imaging characteristics of the disease, thereby constructing a more representative sentence vector for the imaging report text. This application also relates to a lung disease monitoring method based on imaging report text. This method monitors the similarity of text in a large number of imaging reports involving lung X-rays, CT scans, and MRI scans, and uses this to analyze the spatiotemporal distribution characteristics of similar cases.

[0047] Figure 1 FIG1 is a flow chart showing a method 10 for processing structured information of an imaging report text according to an embodiment of the present application. Figure 1 As shown, the structured information processing method 10 of the imaging report text includes the following steps:

[0048] Step S11: further classify the parts and morphological features in the medical imaging professional entity into two categories: negative and positive, and obtain eight entity labels, namely, blank filler, sentence starter, sentence terminator, part-negative, part-positive, morphology-negative, morphology-positive, and disease name. Based on the eight entity labels, named entity recognition is performed on the imaging report text to extract and obtain named entity output in BIO format;

[0049] Step S12: Based on the named entities extracted in step S11, redundant information marked as O and entity labels as part-negative and morphology-negative are filtered out, and the sentence vector of the imaging report text is calculated and stored in the database.

[0050] According to a specific embodiment of the present application, the structured information processing method 10 of the above-mentioned imaging report text performs named entity recognition of the imaging report text in step S11. The named entities that need to be recognized are labels for the data. These labels include sentence symbols PAD, CLS and SEP, where PAD represents a placeholder, CLS is a sentence starter, and SEP is a sentence terminator, and the imaging medical professional entities of interest include part, morphology, and disease name. For medical entities, because there are not only descriptions of symptoms in the imaging report, but also descriptions of normal tissues and organs, and these descriptions of normal tissues and organs will affect the characteristics of the symptoms, the above-mentioned embodiment of the present application adopts the method of expert annotation to further classify the text labels, and further divides the part and morphological features in the imaging report into two categories: negative and positive. Therefore, there are a total of eight named entity labels, namely PAD, CLS, SEP, part-negative (par-neg), part-positive (par-neg), morphology-negative (mor-pos), morphology-positive (mor-pos) and disease name (dis). In addition, this application uses the BIO format for entity annotation, with one mark for each word, where B indicates that the word is a start word or a single word body, I indicates that the word is an intermediate word or an end word, and O indicates a non-target word.

[0051] According to a specific embodiment of the present application, the above-mentioned structured information processing method 10 of the radiology report text uses the BERT-BiLSTM-CRF model to perform named entity recognition on the radiology report text based on the eight entity tags in step S11. The structural diagram of the BERT-BiLSTM-CRF model can be found in Figure 2As shown. BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained model. For the input radiology report text, it first searches the model's vocab file to obtain the ID number corresponding to each word, number, or symbol, and realizes the conversion from input text to input digital sequence for subsequent calculations. The BERT model converts each word ID in the input sequence of text with a length of n into a 768-dimensional vector, and finally inputs the ID sequence of [n, 1] and outputs a word embedding representation of [n, 768]. The word vector sequence output by the BERT model is further input into the BiLSTM (Bidirectional Long Short-Term Memory) model for further processing and extraction of contextual information.

[0052] The calculation function of a common LSTM model is as follows:

[0053] i t =σ(W ix x t +W ir y t-1 +W ic c t-1 +b i )

[0054] f t =σ(W fx x t +W fr y t-1 +W fc c t-1 +b f )

[0055] g t =σ(W cx x t +W cr y t-1 +b c )

[0056] c t =f t ⊙c t-1 +g t ⊙i t

[0057] o t =σ(W ox x t +W or y t-1 +W oc c t+b o )

[0058] m t =o t ⊙h(c t )

[0059] y t =W ym m t

[0060] Among them, {W ix , W ir , W ic , W fx , W fr , W fc , W cx , W cr , W ox , W or , W oc , W ym}、{b i , b f , b c , b o} is the weight matrix and bias vector parameter of the neural network, which is calculated and iteratively updated by the gradient descent method during the training process; x t Represents the input information at the current position t, corresponding to the word vector information of length 768 at n=t in the input sequence [n, 768]; t-1 is the output at position t-1, and the length of the output feature vector is the same as the input, both are 768; c t-1 is the mainline memory feature vector of LSTM, with a vector length of 768; i t 、f t 、g t 、c t 、o t 、m t are all intermediate calculation vectors; y t is the calculated output of the LSTM module; σ is the sigmoid activation function, and ⊙ is the element-by-element matrix multiplication. This LSTM module is called a forward LSTM, meaning that the feature vector at the current position t is related to the information at the previous position t-1.

[0061] The BiLSTM model used in this application is a bidirectional LSTM model, which includes a forward and a backward LSTM model. For the backward LSTM, the feature vector at position t is related to the information of the next adjacent position t-1. On the other hand, the output information length of each submodule in the BiLSTM model is 768 / 2=384, and the outputs of the two LSTM models are finally combined to form a feature vector of length 768. This sequence dependency is shown in Figure 2 In the BiLSTM layer shown in the figure, a linear layer is added after the input of the BiLSTM layer to reduce the dimension of each 768-length feature vector to 8 dimensions, which corresponds to the named entities of the eight labels.

[0062] BiLSTM outputs a sequence of [n, 8], where the length of the feature vector at each position is 8, corresponding to the scores of 8 labels. CRF further processes this sequence to generate labels in BIO format. CRF (Conditional Random Fields) is a probabilistic graphical model that takes a given random variable as input and solves the distribution of the conditional probability of the output random variable. The scores of the 8 labels at each position of BiLSTM output do not take into account the connection between words, while CRF establishes the connection between adjacent words by learning the transfer matrix between labels. This transfer matrix is marked as T, with a dimension of 8x8, and T ij Represents the transition probability from label i to label j. Define P as the output of the BiLSTM layer, its element P ij Represents the probability of the jth label of the i-th word in a sentence. For a sentence of length m, S=(S1,S2,…,S m ), assuming that the label sequence to be predicted for this sentence is y=(y1,y2,…,y m ), the following formula calculates the score of the label sequence:

[0063]

[0064] The score of the entire sequence is equal to the sum of the scores of each word in the sentence, where P is the score from the BiLSTM output and T is the transfer matrix of the CRF layer. The probability is calculated using the softmax function, and let y' represent all possible labels. The calculation form is:

[0065]

[0066] During the model training process, the maximum likelihood function method is used, which is expressed as follows:

[0067] logP(y|S)=score(S,y)-log∑ y′ exp(score(S,y′))

[0068] Through the above steps, the output of named entities from a text of length n to the corresponding BIO format can be completed.

[0069] According to a specific embodiment of the present application, the structured information processing method 10 of the above-mentioned imaging report text filters out redundant information in the imaging report based on the extracted named entities in step S12, calculates the sentence vector of the imaging report text and stores it in the database. Among them, redundant information refers to the words marked as O in the named entities and the entity labels "part-negative (par-neg)" and "morphology-negative (mor-neg)" classified as negative, and finally only retains the disease name information and the entity labels "part-positive (par-pos)" and "morphology-positive (mor_pos)" marked as positive. According to a specific embodiment of the present application, the word vectors corresponding to these words and the output of the BERT model in the aforementioned step S11 are used to calculate the sentence vector, and respective weights are set for the word vectors corresponding to different entity labels. Use W i,l Represents the word vector of the i-th position marked with label l, and the calculation formula of the sentence vector is:

[0070] Sentence=∑ i λ l W i,l

[0071] Among them, Sentence is a sentence vector, which is used to represent all the information about the disease description in the entire sentence; l is the sentence vector weight corresponding to label l, and its weight value is set by experts based on statistical experience. Finally, the entities and sentence vectors segmented by named entity recognition are stored in the database.

[0072] According to the structured information processing method 10 for imaging report texts in the above-mentioned embodiment of the present application, when performing named entity recognition, the method distinguishes between negative and positive descriptions of entity tags, thereby efficiently filtering out redundant word descriptions and negative feature expressions in the report text, while retaining positive features. According to the structured information processing method 10 for imaging report texts in the above-mentioned embodiment of the present application, the method further calculates the imaging report text sentence vector based on the weighted type of the named entity tag to obtain a more expressive sentence vector.

[0073] Based on the structured information processing method of the imaging report text introduced above, this application also proposes a lung disease monitoring method based on the imaging report text. Figure 3 FIG2 shows a flow chart of a lung disease monitoring method 20 based on an imaging report text according to an embodiment of the present application. Figure 3 As shown, the lung disease monitoring method 20 based on imaging report text includes the following steps:

[0074] Step S21, processing the lung imaging report text using the structured information processing method 10 of the imaging report text as described above to construct a database;

[0075] Step S22: searching the database, calculating the similarity between the sentence vectors of each radiology report text, and classifying all radiology report texts with a similarity greater than a threshold as similar cases;

[0076] Step S23: analyzing the spatiotemporal distribution characteristics of all imaging report texts classified as similar cases.

[0077] According to a specific embodiment of the present application, the above-mentioned lung disease monitoring method 20 based on the imaging report text adopts the above-mentioned combination in step S21. Figure 1 and Figure 2 The structured information processing method 10 of the described imaging report text processes a large number of lung imaging report texts (such as lung X-ray, CT, MRI test reports, etc.), obtains sentence vectors and constructs a structured information database. The specific implementation process will not be repeated here.

[0078] According to a specific embodiment of the present application, the above-mentioned lung disease monitoring method 20 based on imaging report text compares text similarity in step S22 and classifies similar cases. Specifically, the calculation of text similarity can be based on the sentence vectors of each imaging report text to calculate the cosine angle between them as a representation of text similarity. The calculation formula of cosine similarity is:

[0079]

[0080] Among them, cos(θ ij ) is the cosine similarity between corresponding texts i and j, and the || symbol represents the modulo operation on the sentence vector. All radiology report texts with similarity greater than a certain threshold are classified as similar cases for subsequent spatiotemporal distribution analysis.

[0081] According to a specific embodiment of the present application, the above-mentioned lung disease monitoring method 20 based on imaging report text further includes in step S22:

[0082] Step S221, obtain n radiology report texts within a period of time. Assume that there are n texts in total, marked as [0, 1, 2, ..., n-1].

[0083] Step S222, calculate the cosine similarity between the n radiology report texts to obtain an nxn cosine similarity matrix. The diagonal elements of the matrix are 0, indicating that each text does not perform similarity calculation with itself. The cosine similarity matrix is expressed as follows:

[0084]

[0085] Step S223: Take each radiology report text as a graph node and construct a weighted undirected graph representation between the radiology report texts according to the above cosine similarity matrix and the cosine similarity threshold. The specific operation is: take the cosine similarity matrix that is greater than the threshold cos(θ thresh ) is set to cos(θ ij ), indicating that there is a connection between the two texts, and the weight of the edge is cos(θ ij ); otherwise, it is set to 0, indicating that the two texts are unrelated. Thus, the weighted adjacency matrix A of the radiology report text can be obtained, A ij The calculation formula is:

[0086]

[0087] Then in step S224, the Karplast matrix L of the weighted undirected graph between the radiology report texts is calculated, L = D – A, where D is the degree matrix, and D is a diagonal matrix with a dimension of nxn. The calculation formula is:

[0088]

[0089] Then in step S225, the normalized Laplace matrix L' is calculated, where L'=D -1 / 2 LD -1 / 2 Spectral clustering is then used to identify radiology reports with similar symptoms, resulting in clusters of similar text cases. Specifically, eigenvalue decomposition is first used to extract the eigenvalues and eigenvectors of the normalized Laplace matrix L'. The eigenvectors are equivalent to low-dimensional embedding representations of the text graph. K-means is then used to cluster the eigenvectors, resulting in clusters of similar text cases.

[0090] According to a specific embodiment of the present application, the above-mentioned lung disease monitoring method 20 based on imaging report text analyzes the spatiotemporal distribution characteristics of all imaging report texts classified as similar cases in step S23, specifically including: extracting the imaging report texts classified into the same case cluster, analyzing the set of words corresponding to the location-positive and morphology-positive entity labels extracted in the aforementioned step S21, as the imaging feature representation of this type of disease; then, based on the case's consultation time and hospital location, respectively draw the spatial distribution and temporal distribution of the cases corresponding to the imaging report texts classified into the same case cluster as the spatiotemporal distribution characteristics of this type of disease. If the disease is contagious, its spatiotemporal distribution will show a strong correlation.

[0091] According to the lung disease monitoring method 20 based on imaging report text of the above embodiment of the present application, the aforementioned structured information processing method 10 of imaging report text is used to perform structured information processing on a large number of lung imaging report texts, so as to obtain more accurate similar case clusters for spatiotemporal distribution feature analysis.

[0092] Based on the structured information processing method of imaging report text and the lung disease monitoring method based on imaging report text introduced above, this application also proposes a structured information processing system of imaging report text and a lung disease monitoring system based on imaging report text that uses the system to construct a structured information database. Figure 4 FIG1 shows a logical structure diagram of a lung disease monitoring system 40 based on imaging report text according to an embodiment of the present application. Figure 4 As shown, the lung disease monitoring system 40 based on the imaging report text includes a structured information processing system 30 of the imaging report text, a similar case classification module 41 and a spatiotemporal feature analysis module 42. The structured information processing system 30 of the imaging report text is used to perform structured information processing on the lung imaging report text and build a database. The structured information processing system 30 of the imaging report text further includes a named entity recognition module 31 and a sentence vector construction module 32. Among them, the named entity recognition module 31 is used to further divide the parts and morphological features in the imaging medical professional entities into two categories: negative and positive, and obtain eight entity labels, namely, blank fillers, sentence starters, sentence terminators, part-negative, part-positive, morphology-negative, morphology-positive and disease names. Based on the eight entity labels, the imaging report text is subjected to named entity recognition and the named entity output in BIO format is extracted. The sentence vector construction module 32 is used to filter out redundant information marked as O and entity labels as part-negative and morphology-negative based on the named entities extracted by the named entity recognition module 31, calculate the sentence vector of the imaging report text and store the sentence vector in the database. For the specific implementation of the named entity recognition module 31 and the sentence vector construction module 32, please refer to the specific description of steps S11 and S12 of the aforementioned structured information processing method 10 for imaging report texts, which will not be repeated here. The similar case classification module 41 is used to search the database constructed by the structured information processing system 30 of the above-mentioned imaging report text, calculate the similarity between the sentence vectors of each imaging report text, and classify all imaging report texts with a similarity greater than a threshold as similar cases. The spatiotemporal feature analysis module 42 is used to analyze the spatiotemporal distribution characteristics of all imaging report texts classified as similar cases. For the specific implementation of the similar case classification module 41 and the spatiotemporal feature analysis module 42, please refer to the specific description of steps S22 and S23 of the aforementioned lung disease monitoring method 20 based on imaging report texts, which will not be repeated here.

[0093] Based on the structured information processing method 10 of the imaging report text described above, the present application also proposes an electronic device 50. Figure 5 As shown, the electronic device 50 includes a processor 51 and a memory 52, which are in communication with each other. The memory 52 stores a computer program, which, when executed by the processor 51, enables the processor 51 to implement the structured information processing method 10 of the imaging report text of the aforementioned embodiment of the present application.

[0094] Based on the above-mentioned lung disease monitoring method 20 based on imaging report text, this application also proposes an electronic device 60. Figure 6 As shown, the electronic device 60 includes a processor 61 and a memory 62, which are communicatively connected to each other. The memory 62 stores a computer program, which, when executed by the processor 61, enables the processor 61 to implement the lung disease monitoring method 20 based on the imaging report text of the aforementioned embodiment of the present application.

[0095] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A lung disease monitoring method based on imaging report text, characterized in that: The steps include: S21. Process the lung imaging report text using a structured information processing method for imaging report text to construct a database; S22. Search the database, calculate the similarity between the sentence vectors of each imaging report text, and classify all imaging report texts with similarity greater than a threshold as similar cases; S23. Analyze the spatiotemporal distribution characteristics of all imaging report texts classified as similar cases; The method for processing structured information of the imaging report text includes the following steps: S11. Further classify the part and morphological features in the medical imaging professional entity into two categories, negative and positive, to obtain eight entity labels, namely, a blank filler, a sentence starter, a sentence terminator, part-negative, part-positive, morphology-negative, morphology-positive, and a disease name; perform named entity recognition on the imaging report text based on the eight entity labels, and extract and obtain named entity output in BIO format; S12, filtering out redundant information marked as O and entity labels as part-negative and morphology-negative based on the named entities extracted in step S11, calculating a sentence vector for the radiology report text, and storing the sentence vector in a database; The step S22 further includes: Get n imaging report texts within a period of time; Calculating the cosine similarity between the n radiology report texts to obtain an nxn cosine similarity matrix; Taking each radiology report text as a graph node, constructing a weighted undirected graph representation between the radiology report texts according to the cosine similarity matrix and the cosine similarity threshold, and obtaining a weighted adjacency matrix A of the radiology report texts; Calculate the Karplast matrix L of the weighted undirected graph representation between the texts of the radiology report, L = DA, where D is the degree matrix, and D is a diagonal matrix with dimension nxn; Calculate the normalized Laplace matrix L', L' = D -1 / 2 LD -1 / 2 ,The spectral clustering method is used to find the imaging report texts with similar symptoms and obtain several case clusters of similar texts; The step S23 further includes: Extract the word sets corresponding to the location-positive and morphology-positive entity labels in the imaging report texts that are grouped in the same case cluster as the imaging feature representation of this type of disease; According to the consultation time and hospital location of the cases, the spatial distribution and temporal distribution of the cases corresponding to the imaging report texts of the same case cluster are drawn respectively.

2. The lung disease monitoring method based on imaging report text according to claim 1, characterized in that: In step S11, the BERT-BiLSTM-CRF model is used to perform named entity recognition on the radiology report text.

3. The lung disease monitoring method based on imaging report text according to claim 2, characterized in that: Calculating the sentence vector in step S12 further includes: using the word vectors of the words corresponding to the three entity labels of part-positive, morphology-positive, and disease name retained after removing redundant information to calculate the sentence vector, and setting respective weights for the word vectors corresponding to different entity labels when calculating the sentence vector, wherein the word vector is the output of the BERT model in step S11.

4. A lung disease monitoring system based on imaging report text, characterized in that: include: The structured information processing system for imaging report text is used to process the structured information of lung imaging report text and build a database; The similar case classification module is used to search the database, calculate the similarity between the sentence vectors of each imaging report text, and classify all imaging report texts with similarity greater than a threshold as similar cases; The spatiotemporal feature analysis module is used to analyze the spatiotemporal distribution characteristics of all imaging report texts classified as similar cases; The structured information processing system for the imaging report text further includes: The named entity recognition module is used to further classify the parts and morphological features in the medical imaging professional entities into two categories: negative and positive. The module obtains eight entity labels, namely, the blank filler, the sentence starter, the sentence terminator, the part-negative, the part-positive, the morphology-negative, the morphology-positive, and the disease name. The named entity recognition module is used to extract the named entity output in BIO format based on the eight entity labels. a sentence vector construction module, configured to filter out redundant information marked as O and entity labels as part-negative and morphology-negative based on the named entities extracted by the named entity recognition module, calculate a sentence vector for the radiology report text, and store the sentence vector in a database; The similar case classification module is further used to: Get n imaging report texts within a period of time; Calculating the cosine similarity between the n radiology report texts to obtain an nxn cosine similarity matrix; Taking each radiology report text as a graph node, constructing a weighted undirected graph representation between the radiology report texts according to the cosine similarity matrix and the cosine similarity threshold, and obtaining a weighted adjacency matrix A of the radiology report texts; Calculate the Karplast matrix L of the weighted undirected graph representation between the texts of the radiology report, L = DA, where D is the degree matrix, and D is a diagonal matrix with dimension nxn; Calculate the normalized Laplace matrix L', L' = D -1 / 2 LD -1 / 2 ,The spectral clustering method is used to find the imaging report texts with similar symptoms and obtain several case clusters of similar texts; The spatiotemporal feature analysis module is further used to: Extract the word sets corresponding to the location-positive and morphology-positive entity labels in the imaging report texts that are grouped in the same case cluster as the imaging feature representation of this type of disease; According to the consultation time and hospital location of the cases, the spatial distribution and temporal distribution of the cases corresponding to the imaging report texts of the same case cluster are drawn respectively.

5. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, wherein: When the computer program is executed by a processor, the steps of the lung disease monitoring method based on imaging report text as described in any one of claims 1 to 3 are implemented.

Images