Artificial intelligence medical image quality control method applied to clinical images

Abstract

The invention discloses an artificial intelligence medical image quality control method applied to clinical images, and relates to the technical field of medical image quality control. The artificial intelligence medical image quality control method applied to the clinical image comprises the following five steps: improving on the basis of a conventional word sequence training model by using a BiLSTM (Bidirectional Circulating Long Short Term Memory) neural network sequence prediction model, and improving input data of sequence prediction from a mode of predicting a next word by word input into a mode of predicting a next word by word input. According to the method, a sentence is input according to sentences, learning, recognition and prediction of a combination sequence between characters are carried out in a mode of predicting a word sequence, a model is trained in a mode of simulating word sequence prediction by taking the sentences as a whole, and the problem that difference between different prediction report groups cannot be unifiedly applicable to error screening index thresholds in practical application is solved; the occurrence of misdiagnosis is reduced, and the efficiency and accuracy of quality control marking are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image quality control, and specifically to an artificial intelligence medical image quality control method applied to clinical images. Background Art

[0002] DR chest radiographs, or Radio Radiography, have become the best clinical diagnostic method for preliminary screening of chest and lung diseases in recent years due to their extremely low radiation dose, low instrument cost, high-density resolution, and fast imaging. With the increase in patient and clinical needs and the trend of digital and intelligent transformation, intelligent quality control of DR chest radiographs is particularly important.

[0003] Focusing on the discussion and research on the future of quality control in imaging departments across the country, most of them focus on the uniformity of quality control standards, consistency of specifications, and objectivity of manual scoring, and develop a series of solutions. The existing DR quality control standards are formulated by the Quality Control Center of the Chinese Society of Radiology and local radiology quality control centers. Therefore, the degree of quality control standardization is low, the quality control standards are inconsistent, and the quality of images or reports is basically completely dependent on manual judgment, which lacks objectivity, is not only time-consuming and labor-intensive, but also has subjective judgment differences.

[0004] Today, when artificial intelligence (AI) empowers all industries, the public's vision is still limited to using various high-tech technologies to conduct intelligent quality assessments of existing medical images in the imaging department and quality management at the image application level, ignoring the quality control of the entire medical imaging process. In fact, it is very necessary to use accurate and effective quality control methods for the entire process from image data collection to diagnosis.

[0005] The invention with publication number CN113555089A discloses an artificial intelligence medical image quality control method applied to clinical images. A technician collects patient images and transmits the images to an artificial intelligence medical image quality control management system. The artificial intelligence medical image quality control management system is used to perform semantic segmentation and classification processing on the images and perform quality control scoring. The score is displayed on the operation interface where the technician previews the image. The technician determines whether to make notes or re-collect the image based on the quality control score and the patient's condition to reduce the generation of low-scoring image quality.

[0006] As shown in the above invention, the existing medical image quality control method has solved the limitations of low efficiency, large workload, small number of quality control samples, and strong subjectivity of traditional quality control. However, in terms of quality control of image reports, there is little application of AI technology or the investment in AI in the detailed field of medical image report quality control is not deep, comprehensive, and specific enough. There is a certain rate of misdiagnosis and missed diagnosis, and the diagnosis reports are not standardized, the quality is uneven, and the quality cannot be guaranteed. Summary of the invention

[0007] In view of the deficiencies of the prior art, the present invention provides an artificial intelligence medical image quality control method applied to clinical images, which solves the existing problems.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: an artificial intelligence medical image quality control method applied to clinical images, comprising the following steps:

[0009] Step 1: Obtain various original medical imaging reports, perform feature analysis and integration, data processing on the original medical imaging reports, and create a report corpus;

[0010] Step 2: Create a quality control knowledge graph based on the processed report corpus, wherein the quality control knowledge graph includes several graph parameters;

[0011] Step 3: Build multiple parallel error detection models, input the quality control knowledge graph into each error detection model to train the model, learn the image report rules, and obtain the trained error detection model;

[0012] Step 4: Collect patient medical information including imaging information, the information collection terminal 1 includes digital imaging equipment such as MRI, CT, DR, etc.;

[0013] Step 5: Upload the patient's imaging information to the AI ​​server for classification diagnosis, and input the diagnosis results into the error detection model for further error detection and diagnosis to obtain the AI ​​diagnosis report and quality control report.

[0014] Preferably, the method for creating the quality control knowledge graph comprises the following steps:

[0015] S1: Obtain the initial quality control knowledge graph through the medical imaging parameters in the report corpus;

[0016] S2: Obtain information parameters of the initial quality control knowledge graph, including: several entity parameters of the initial quality control knowledge graph, attributes of each entity parameter of the initial quality control knowledge graph, and relationships between each entity parameter of the initial quality control knowledge graph;

[0017] S3: performing fusion processing on the information parameters of the initial quality control knowledge graph to obtain the ontology of the initial quality control knowledge graph;

[0018] S4: Perform quality assessment on the ontology of the initial quality control knowledge graph, perform implicit information reasoning on the initial quality control knowledge graph, and finally obtain the quality control knowledge graph.

[0019] Preferably, the medical imaging parameters include: image type, acquisition device, acquisition parameters, image size and resolution, number of images and acquisition time.

[0020] Preferably, the several entity parameters of the initial quality control knowledge graph include: anatomic, examination, feature, location, modality, quality, and root.

[0021] Preferably, the fusion processing method includes: unification of entity parameters, reference resolution, and disambiguation of entity parameters.

[0022] Preferably, the methods of creating the quality control knowledge graph include: manual construction, automatic construction, semi-automatic construction and fusion construction.

[0023] Preferably, the error detection model is a BiLSTM neural network sequence prediction model.

[0024] Preferably, the neural network sequence prediction model adopts a 2-layer bidirectional recurrent long short-term memory neural network BiLSTM, adopts a word-based unsupervised sequence prediction task to detect and correct errors, uses unlabeled data input, predicts the following words through the previous words, and generates the word probability distribution of the language model.

[0025] Preferably, the neural network sequence prediction model includes a first data processing unit, a model training unit and a first prediction error detection unit;

[0026] The first data processing unit is used to process various medical imaging reports or medical imaging reports to be inspected;

[0027] The model training unit is used to input the processed medical imaging report data corpus into the bidirectional recurrent long short-term memory neural network model sequence prediction model for training;

[0028] The first prediction and error detection unit is used to input the processed medical image report to be inspected into the score language model for query.

[0029] Beneficial Effects

[0030] The present invention provides an artificial intelligence medical image quality control method applied to clinical images. Compared with the prior art, it has the following beneficial effects:

[0031] 1. This artificial intelligence medical image quality control method applied to clinical images uses the BiLSTM (bidirectional recurrent long short-term memory neural network model) neural network sequence prediction model to improve the conventional word sequence training model. The input data of the sequence prediction is improved from the method of predicting the next word by word input to the method of predicting the reorganized sentence after one word is misplaced by sentence input. The sequence order of the combination of words is learned, recognized, and predicted. The model is trained by simulating word sequence prediction with the sentence as a whole. In actual application, the problem of the inability to uniformly apply the threshold of the screening error index due to differences between different prediction report groups is solved, the occurrence of misdiagnosis is reduced, and the efficiency and accuracy of quality control labeling are improved.

[0032] 2. This artificial intelligence medical imaging quality control method applied to clinical imaging evaluates manual imaging diagnosis reports through intelligent prompts and supervision integrated with medical big data, solves the problems of incompleteness and subjectivity in traditional spot-check quality control, and provides a quantitative basis for the optimization of imaging examination processes and team management, effectively improving the homogeneity and standardization of regional imaging examination technology and diagnostic quality, while promoting the further integration and development of medical communities and health communities, and contributing to the construction of universal health. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the structure of the method flow of the present invention;

[0034] Figure 2 Schematic diagram of the creation process of the quality control knowledge graph of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] See Figure 1-2 , the present invention provides the following three technical solutions:

[0037] The first implementation method: an artificial intelligence medical image quality control method applied to clinical images, comprising the following steps:

[0038] Step 1: Obtain various original medical imaging reports, perform feature analysis and integration, data processing on the original medical imaging reports, and create a report corpus;

[0039] Step 2: Create a quality control knowledge graph based on the processed report corpus, wherein the quality control knowledge graph includes several graph parameters;

[0040] Step 3: Build multiple parallel error detection models, input the quality control knowledge graph into each error detection model to train the model, learn the image report rules, and obtain the trained error detection model;

[0041] Step 4: Collect patient medical information including imaging information, the information collection terminal 1 includes digital imaging equipment such as MRI, CT, DR, etc.;

[0042] Step 5: Upload the patient's imaging information to the AI ​​server for classification diagnosis, and input the diagnosis results into the error detection model for further error detection and diagnosis to obtain the AI ​​diagnosis report and quality control report.

[0043] Second implementation method: The method for creating the quality control knowledge graph includes the following steps:

[0044] S1: Obtaining the initial quality control knowledge graph through medical imaging parameters in the report corpus

[0045] S2: Obtain information parameters of the initial quality control knowledge graph, including: several entity parameters of the initial quality control knowledge graph, attributes of each entity parameter of the initial quality control knowledge graph, and relationships between each entity parameter of the initial quality control knowledge graph;

[0046] S3: performing fusion processing on the information parameters of the initial quality control knowledge graph to obtain the ontology of the initial quality control knowledge graph;

[0047] S4: Perform quality assessment on the ontology of the initial quality control knowledge graph, perform implicit information reasoning on the initial quality control knowledge graph, and finally obtain the quality control knowledge graph.

[0048] The medical imaging parameters include: image type, acquisition equipment, acquisition parameters, image size and resolution, number of images and acquisition time.

[0049] The several entity parameters of the initial quality control knowledge graph include: anatomic, examination, feature, location, modality, quality, and root.

[0050] The fusion processing method includes: unification of entity parameters, reference resolution, and disambiguation of entity parameters.

[0051] The methods for creating the quality control knowledge graph include: manual construction, automatic construction, semi-automatic construction and fusion construction.

[0052] Manual construction is to establish a quality control knowledge graph by manually collecting, organizing, classifying and describing knowledge points in the quality control field. It is suitable for situations where the data scale is small.

[0053] Automatic construction utilizes technologies such as natural language processing, machine learning, and quality control knowledge graph extraction to automatically extract and construct knowledge graphs from relevant literature, networks, databases, etc. It is suitable for situations where the data volume is large and the update speed is fast.

[0054] Semi-automatic construction combines manual construction and automatic construction, using a combination of manual and automated methods to build a quality control knowledge graph more accurately and efficiently.

[0055] Fusion construction combines multiple quality control knowledge graph sources to build a more comprehensive and accurate quality control knowledge graph, and uses data quality control, expert knowledge addition and multi-party confirmation to improve the accuracy and practicality of the quality control knowledge graph.

[0056] The quality control knowledge graph is continuously optimized and improved. Methods for optimizing the quality control knowledge graph include: improving the data quality of the initial examination report and regularly updating medical images.

[0057] Improving the data quality of the initial inspection report includes: ensuring the accuracy, completeness and consistency of the data as much as possible when collecting data, and avoiding errors in the quality control knowledge graph due to data problems.

[0058] Regularly updating medical images includes: timely adding new medical imaging data and knowledge points, and paying attention to the latest developments and changes in the field of quality control.

[0059] Methods for optimizing and improving the quality control knowledge graph and improving the accuracy of the quality control knowledge graph also include: adding expert knowledge, multi-party confirmation, and combining practical applications.

[0060] The purpose of adding expert knowledge is to add expert knowledge and professional knowledge in the field of quality control to enhance the accuracy and practicality of the quality control knowledge graph.

[0061] The role of multi-party confirmation is to evaluate and revise through expert argumentation, experimental verification, etc.

[0062] The third implementation mode: the error detection model is a BiLSTM neural network sequence prediction model, which adopts a 2-layer bidirectional cyclic long short-term memory neural network BiLSTM, adopts an unsupervised sequence prediction task based on words to detect and correct errors, uses unlabeled data input, predicts the following words through the previous words, and generates the word probability distribution of the language model. In this embodiment, the output probability of the last hidden layer is taken, the mean absolute deviation (MAD, Mean Absolute Deviatio average difference) of the words in the sentence is calculated, and a threshold is set to find abnormal words. BiLSTM bidirectional cyclic long short-term memory neural network principle: The forget gate, input gate, and output gate structure of the LSTM network (Long Short-Term Memory cyclic neural network LSTM, LongShort-TermMemory) can control the selective transmission of sentence sequence information, has long and short-term memory ability, can selectively transmit useful information of historical moments to the final hidden layer, extract probability for use, and use it as a language model that describes the distribution law of words or words in the input sequence sentence for sequence prediction, detection of abnormal discrete points in sentences and other tasks. The concept of language model is: calculate the probability distribution of language sequence, that is, the probability of a combination of words appearing as a normal sentence. For example, the probability of the sentence "the patient suffers from diabetes" is greater than the probability of "the patient suffers from glycosuria", and the threshold, that is, the boundary, is defined based on this to find abnormal words. Forget gate: used to control the amount of information at the previous moment, and choose to remember or forget. Input gate: used to input the information at this moment and the memory of the previous moment. Input node: equivalent to a switch, such as using the sigmoid binary classification activation function to control whether to input. Selective memory is reflected here. Output gate: outputs the information at this moment and the historical memory of the previous moment. There is also an activation function to control the switch to achieve selective output. Each unit has the same structure, thereby achieving selective information transmission. BiLSTM, bidirectional LSTM, includes prediction from front to back and prediction from back to front, making the model more informative and more accurate. The difference between bidirectional LSTM and LSTM is that a copy of the information is retained during the forward transmission process for back propagation, realizing the transmission process of sentence information from front to back and then from back to front. In this embodiment, the word-to-word prediction of the bidirectional recurrent long short-term memory neural network word sequence prediction model is improved to a method in which a sentence is fed into the model as a whole, and the new sentence reorganized after one word is misplaced is predicted for training as a whole, and word sequence prediction is simulated using the sentence as a whole. This can overcome the memory consumption problem of simple word sequence prediction and achieve the purpose of more easily expanding the training corpus and input dimension.For example: several x (x0, x1, ..., x3) combinations are used as a sentence input model. For example, "Density shadows are seen in the lower lobes of both lungs" is input into the neural network model, and the next word combination sentence "Density shadows are seen in the lower lobes of the lungs e" is used, where: e is the end of the sentence, as the result label of this sentence, to predict the result sentence, and use this as a standard to judge the accuracy of the calculation model prediction. The specific implementation is: when training the corpus, the neural network model is used to extract features and calculate, and the model memorizes and learns the sentence rules to form a probability output y (y0, y1, ..., y3) combination of all words or characters in the corpus. With this predicted probabilistic language model, the combination order of the new sentence can be predicted. After entering the new sentence, the previous word can be used to predict the next word. The training method of taking the sentence as a whole and taking the new sentence as a whole after a word is misplaced as the prediction result overcomes the memory consumption problem of simple word sequence prediction and achieves the purpose of expanding the training corpus and input dimension more easily.

[0063] The neural network sequence prediction model includes a first data processing unit, a model training unit and a first prediction error detection unit;

[0064] The first data processing unit is used to process various types of medical imaging reports or medical imaging reports to be inspected, extract data from medical imaging report data by vertical classification, horizontal classification and different time periods, and randomly extract original corpus data in layers; process high-frequency words, low-frequency words, stop words, and long-tail data (extreme reports) in the extracted original corpus data; split the inspection reports of multiple categories according to the inspection categories; mark the start mark of the split inspection sentences with punctuation marks; perform word segmentation and word mapping ID numbering on the marked data; convert Chinese characters into mathematical vector representations;

[0065] The model training unit is used to input various types of processed medical imaging report data corpus into the bidirectional cyclic long short-term memory neural network model sequence prediction model for training, learn the rules, and obtain a scoring language model of normal sentence combination, wherein improvements are made to the problem that word-to-word sequence prediction consumes memory and is not suitable for expanding the dimension of training data, that is, the prediction of a single word to a word is changed to a method of predicting a new sentence obtained by misplacing a word in the sentence as a whole, and the sequence order of the combination between words is learned, recognized, and predicted. This training method of simulating word sequence prediction in units of sentences greatly improves the efficiency of model training;

[0066] The first prediction and error detection unit is used to input the processed medical imaging report to be inspected into the scoring language model for query, generate a predicted sentence score, and obtain the probability score threshold of each word in the combined sentence based on the mean absolute deviation, standard deviation dedimensionalization and probability mean. The abnormal score words are screened out according to the threshold and output as the wrong word candidate set.

[0067] Error screening method of neural network sequence prediction model: In the above error detection unit, MAD (mean absolute deviation, the absolute value of the difference between the probability of each word and the mean probability of all words in the sentence, accumulated and averaged. An indicator to measure the degree of difference.) is used in combination with the standard deviation (the square of the difference between each value and the mean in a set of data is accumulated and averaged, an indicator to measure the degree of discreteness of the set of data sets.) and the mean comprehensive evaluation index to determine the threshold. By finding outliers (abnormal words) in the prediction results (the output result of the neural network model is the word probability score of a normal sentence obtained after the model automatically learns the word combination sequence in the sentence), the wrong words are determined and the correct candidate words are recommended. This embodiment adds the averaging processing of the standard deviation based on the test feedback of the prediction results, and divides it by the standard deviation to achieve the effect of dedimensionalization. In actual application, it solves the problem that the difference between different prediction report groups cannot be uniformly applied to the threshold of the error screening indicator. Probability mean: In the specific implementation, combined with the actual situation, for the sake of efficiency, the part above the mean is filtered out (because the combination rules of abnormal words rarely appear in the word database, it is judged that abnormal words can only be distributed in the small probability part), and only the MAD (mean absolute deviation) threshold and standard deviation of the words below the mean are retained. The error detection unit obtains the predicted sentence score based on the score language model, obtains the probability score based on the mean absolute deviation, standard deviation and mean, and screens out abnormal words based on the probability score and threshold, and detects abnormal score words.

[0068] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art, and the model parameters of each electrical appliance are not specifically limited, and conventional equipment can be used.

[0069] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0070] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence medical image quality control method applied to clinical images, characterized in that: The following steps are involved: Step 1: Obtain various original medical imaging reports, perform feature analysis and integration, data processing on the original medical imaging reports, and create a report corpus; Step 2: Create a quality control knowledge graph based on the processed report corpus, wherein the quality control knowledge graph includes several graph parameters; Step 3: Build multiple parallel error detection models, input the quality control knowledge graph into each error detection model to train the model, learn the image report rules, and obtain the trained error detection model; Step 4: Collect patient medical information including imaging information, the information collection terminal 1 includes digital imaging equipment such as MRI, CT, DR, etc.; Step 5: Upload the patient's imaging information to the AI ​​server for classification diagnosis, and input the diagnosis results into the error detection model for further error detection and diagnosis to obtain the AI ​​diagnosis report and quality control report.

2. The artificial intelligence medical image quality control method applied to clinical images according to claim 1, characterized in that: The method for creating the quality control knowledge graph comprises the following steps: S1: Obtain the initial quality control knowledge graph through the medical imaging parameters in the report corpus; S2: Obtain information parameters of the initial quality control knowledge graph, including: several entity parameters of the initial quality control knowledge graph, attributes of each entity parameter of the initial quality control knowledge graph, and relationships between each entity parameter of the initial quality control knowledge graph; S3: performing fusion processing on the information parameters of the initial quality control knowledge graph to obtain the ontology of the initial quality control knowledge graph; S4: Perform quality assessment on the ontology of the initial quality control knowledge graph, perform implicit information reasoning on the initial quality control knowledge graph, and finally obtain the quality control knowledge graph.

3. The artificial intelligence medical image quality control method applied to clinical images according to claim 2, characterized in that: The medical imaging parameters include: image type, acquisition equipment, acquisition parameters, image size and resolution, number of images and acquisition time.

4. The artificial intelligence medical image quality control method applied to clinical images according to claim 2, characterized in that: The several entity parameters of the initial quality control knowledge graph include: anatomic, examination, feature, location, modality, quality, and root.

5. The artificial intelligence medical image quality control method applied to clinical images according to claim 2, characterized in that: The fusion processing method includes: unification of entity parameters, reference resolution, and disambiguation of entity parameters.

6. The artificial intelligence medical image quality control method applied to clinical images according to claim 2, characterized in that: The methods for creating the quality control knowledge graph include: manual construction, automatic construction, semi-automatic construction and fusion construction.

7. The artificial intelligence medical image quality control method applied to clinical images according to claim 1, characterized in that: The error detection model is a BiLSTM neural network sequence prediction model.

8. The artificial intelligence medical image quality control method applied to clinical images according to claim 7, characterized in that: The neural network sequence prediction model adopts a 2-layer bidirectional recurrent long short-term memory neural network BiLSTM, adopts a word-based unsupervised sequence prediction task to detect and correct errors, uses unlabeled data input, predicts the next word through the previous word, and generates the word probability distribution of the language model.

9. The artificial intelligence medical image quality control method applied to clinical images according to claim 1, characterized in that: The neural network sequence prediction model includes a first data processing unit, a model training unit and a first prediction error detection unit; The first data processing unit is used to process various medical imaging reports or medical imaging reports to be inspected; The model training unit is used to input the processed medical imaging report data corpus into the bidirectional recurrent long short-term memory neural network model sequence prediction model for training; The first prediction and error detection unit is used to input the processed medical image report to be inspected into the score language model for query.

Images