Error detection model training method, medical image report detection method, system and device

Abstract

The application provides a method, system and device for training an error detection model and detecting a medical image report. The training method comprises: determining a first identification model trained based on a plurality of first sample texts of a medical image report; obtaining a plurality of second sample texts of the medical image report, wherein at least one sample sentence with a first label in the second sample texts is an error sample sentence containing an error word; training the first identification model based on the at least one sample sentence to obtain a second identification model; during the training of the first identification model, if it is monitored that the first identification model cannot identify a first error sample sentence in the at least one sample sentence, the first error sample sentence is sent to a user, the user determines a first error word in the first error sample sentence and adds the first error word to an error detection rule set; and determining an error detection model according to the second identification model and the error detection rule set. The present scheme can train an error detection model with better performance at low training cost.

Description

[0001] Cross-references

[0002] This application incorporates Chinese Patent Application No. 202310036606.X, filed on January 10, 2023, entitled “Error Detection Model Training Method, Medical Image Report Detection Method, System and Device”, which is incorporated herein by reference in its entirety. Technical Field

[0003] This application relates to the field of intelligent medical technology, and in particular to a method for training an error detection model, a method for detecting medical image reports, a system, and equipment. Background Technology

[0004] With the development of modern medicine, medical imaging has become an indispensable part of the diagnostic process for doctors. Doctors often write an electronic medical imaging report for each medical image. The medical imaging report mainly includes a quantitative description of whether there are abnormalities in the medical image, and a diagnostic opinion based on the analysis of medical image manifestations and clinical data.

[0005] Since medical imaging reports are written manually, some errors are inevitable, such as spelling errors and vocabulary errors. Currently, the main method used for error detection in medical imaging reports is rule-based. However, rule-based methods require manual creation of a dictionary, are costly, and cannot identify error types outside the dictionary. Summary of the Invention

[0006] To address the problems existing in the error detection of medical image reports, this application provides an error detection model training method, a medical image report detection method, system, and device.

[0007] In one embodiment of this application, a method for training an error detection model is provided, wherein the detection model is used to detect errors in medical image reports. The method includes:

[0008] A first recognition model is determined, which is trained based on multiple first sample texts from medical image reports;

[0009] Obtain multiple second sample texts of a medical imaging report; wherein the second sample texts include at least one sample statement, and the sample statement with a first tag is an erroneous sample statement containing erroneous words, and the sample statement without the first tag is a correct sample statement without erroneous words;

[0010] Based on at least one sample sentence included in the plurality of first sample texts, the first recognition model is trained to obtain a second recognition model;

[0011] During the training of the first recognition model, if it is detected that the first recognition model cannot recognize the first erroneous sample statement in the at least one sample statement, the first erroneous sample statement is sent to the user so that the user can determine the first erroneous word in the first erroneous sample statement and add it to the error detection rule set.

[0012] The error detection model is determined based on the second identification model and the error detection rule set.

[0013] In another embodiment of this application, a medical report detection method is also provided, the method comprising:

[0014] Identify the target medical imaging report to be tested;

[0015] An error detection model is obtained, wherein the error detection model is trained using the error detection model training method provided in one embodiment of this application described above;

[0016] The error detection model is used to detect errors in the target medical image report.

[0017] In another embodiment of this application, a medical image report detection system is also provided, the system comprising:

[0018] The server is used to determine the target medical image report to be detected; obtain an error detection model, wherein the error detection model is trained by the error detection model training method provided in an embodiment of this application; and use the error detection model to detect errors in the target medical image report.

[0019] The client is used to provide an interactive interface on which the error detection results of the target medical image report are displayed.

[0020] In another embodiment of this application, an electronic device is also provided. The electronic device includes: a memory and a processor, wherein the memory is used to store a computer program; the processor, coupled to the memory, is used to execute the computer program stored in the memory to implement steps in the error detection model training method provided in one embodiment of this application, or to implement steps in the medical image report detection method provided in another embodiment of this application.

[0021] The technical solutions provided in the embodiments of this application, when training an error detection model for detecting errors in medical image reports, first determine a first recognition model, which is trained based on multiple first sample texts of the medical image report. Then, acquire multiple second sample texts of the medical image report, each including at least one sample sentence. Among these, sample sentences with a first label are erroneous sample sentences containing incorrect words, and sample sentences without a first label are correct sample sentences without incorrect words. The first recognition model is then trained using the at least one sample sentence from the multiple first sample texts to obtain a second recognition model. Through the training performed in these two stages using at least one sample sentence from the first and second sample texts respectively, the trained second recognition model can have a high ability to understand text semantics, effectively improving the accuracy of error detection. Furthermore, during the training process of the first recognition model, if it is detected that the first recognition model cannot identify the first erroneous sample statement in at least one sample statement, the first erroneous sample statement is sent to the user for identification of the first erroneous word in the first erroneous sample statement and added to the error detection rule set. Finally, the error detection model is determined based on the second recognition model and the error detection rule set. Here, the error detection rule set is used to supplement the difficult errors that the second recognition model trained in the aforementioned two training stages cannot identify, which can further improve the error identification accuracy of the error detection model and avoid the problem of high burden and cost of manually building a dictionary (such as the error detection rule set mentioned above) when directly building an error detection model based on rules. In summary, this solution can train a high-performance error detection model with low training cost. Subsequently, using the error detection model trained above, error detection can be performed on the target medical report to be detected, thus ensuring high error detection accuracy. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a training method for an error detection model provided in an embodiment of this application;

[0024] Figure 2a Example of a medical image report provided in an embodiment of this application;

[0025] Figure 2bA schematic diagram illustrating the principle of training a pre-trained language model according to an embodiment of this application;

[0026] Figure 3 A simplified schematic diagram illustrating the training principle of an error detection model provided in an embodiment of this application;

[0027] Figure 4 A schematic flowchart of a medical image report detection method provided in an embodiment of this application;

[0028] Figure 5 This is a schematic diagram of the structure of a medical image report detection system provided in one embodiment of this application;

[0029] Figure 6 This is a schematic diagram of the structure of an error detection model training device provided in an embodiment of this application;

[0030] Figure 7 This is a schematic diagram of the structure of a medical image report detection device provided in one embodiment of this application;

[0031] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application;

[0032] Figure 9 This is a schematic diagram of the structure of a computer program product provided in an embodiment of this application. Detailed Implementation

[0033] With the widespread adoption of medical imaging equipment, medical imaging reports have increasingly become crucial tools for clinicians in diagnosing diseases. Currently, most medical imaging reports are electronic reports written by radiologists, inevitably containing errors such as spelling mistakes, vocabulary errors, and grammatical errors. Since medical imaging reports serve as a communication tool between radiologists and clinicians and carry certain medical legal responsibilities, errors in these reports can cause confusion for clinicians when making diagnoses. Furthermore, these reports are often provided to patients, and errors can lead to questions about the physician's skill and competence, potentially even causing doctor-patient conflicts. Therefore, error detection in medical imaging reports to control report quality is a critical issue.

[0034] Currently, error detection in medical imaging reports primarily utilizes artificial intelligence techniques such as Natural Language Processing (NPL), specifically rule-based modeling. However, rule-based modeling requires manually creating a lexicon, which necessitates extensive collection of various error types. This approach is time-consuming, costly, and struggles to encompass all possible error types, while also failing to identify error types outside the lexicon. Furthermore, since most medical imaging reports are unstructured text, different categories exhibit significant differences in corpus scope and semantic logic. While separate modeling for different categories offers high accuracy, it has a narrow applicability and high cost. Conversely, hybrid modeling across broad categories of medical imaging reports, while better balancing applicability across different categories, suffers from limited and lower accuracy.

[0035] To address the aforementioned problems, embodiments of this application provide a technical solution for detecting errors in medical image reports. To enable those skilled in the art to better understand the solutions presented in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0036] In some processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. These operations may be executed out of order or in parallel. Operation numbers such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the terms "first," "second," etc., used herein are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types. Moreover, the following embodiments are merely some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0037] The proposed solution utilizes a well-trained error detection model to perform error detection on medical imaging reports. Figure 1This illustration shows a flowchart of the error detection model training method provided in this application embodiment. The execution subject of this method can be an electronic device with logic processing capabilities, which can be a client or a server. The client can be hardware with embedded programs integrated on a terminal, application software installed on the terminal, or tool software embedded in the terminal's operating system, etc., which are not limited in this embodiment. The terminal can be any device with a certain computing power, such as a smartphone, laptop, smart wearable device, PDA (Personal Digital Assistant), desktop computer, etc. The server can be a single server, a server cluster composed of multiple servers, a cloud or virtual server, etc., which are not specifically limited in this embodiment. In the above, the error detection model is used to detect errors in medical image reports. See also... Figure 1 As shown, the error detection model training method provided in this embodiment includes the following steps:

[0038] 101. Determine the first recognition model, which is trained based on multiple first sample texts of medical image reports;

[0039] 102. Obtain multiple second sample texts of a medical imaging report; wherein the second sample text includes at least one sample statement, and the sample statement with a first tag is an erroneous sample statement containing erroneous words, and the sample statement without the first tag is a correct sample statement without erroneous words;

[0040] 103. Based on at least one sample statement included in the plurality of second samples, the first recognition model is trained to obtain a second recognition model;

[0041] 104. During the training of the first recognition model, if it is detected that the first recognition model cannot recognize the first erroneous sample statement in the at least one sample statement, the first erroneous sample statement is sent to the user so that the user can determine the first erroneous word in the first erroneous sample statement and add it to the error detection rule set.

[0042] 105. Determine the error detection model based on the second identification model and the error detection rule set.

[0043] In the above 101, the first sample text consists of multiple samples, which are generated based on multiple historical medical image report information. A description of the historical medical image report information can be found in the relevant content below. To reduce training costs, this embodiment utilizes the first sample text of the medical image report and trains the first recognition model using a model fine-tuning method. The model fine-tuning method refers to using existing pre-trained model resources for solving problems similar to the new task, training on the sample data of the new task, thereby fine-tuning the parameters of the pre-trained model to adapt it to the new task, thus quickly obtaining a model adapted to the new task; wherein, the pre-trained model is a model trained on a large standard dataset. Using a model fine-tuning method to train a model adapted to the new task avoids training the model required for the new task from scratch, allowing for better model training results with less training sample data, and greatly reducing training costs. Based on the above, in one feasible technical solution, "determining the first recognition model" in the above 101 may specifically include:

[0044] 1011. Obtain information from multiple historical medical imaging reports;

[0045] 1012. Generate the multiple first sample texts based on the multiple historical medical image report information;

[0046] 1013. Using the multiple first sample texts, train the pre-trained language model to be trained to obtain the first recognition model.

[0047] In the above 1011, the information on multiple historical medical imaging reports includes multiple historical medical imaging reports and a second tag corresponding to each historical medical imaging report. The second tag is used to identify whether the corresponding historical medical imaging report is a correct medical imaging report. A correct medical imaging report means that the text content of the medical imaging report is completely correct and there are no errors (such as incorrect words, misspellings, etc.). For example, if the first tag is "1", it means that the historical medical imaging report corresponding to the first tag is a correct medical imaging report; if the first tag is "0", it means that the historical medical imaging report corresponding to the first tag is an incorrect medical imaging report.

[0048] To enable the trained model to be applicable to various types of medical image reports, the aforementioned historical medical image reports can be obtained by sampling from multiple sets of medical image reports. The sampled historical medical image reports can be sent to a client used by a user (e.g., a radiology physician) and displayed on the client's interactive interface, allowing the user to annotate the historical medical image reports and obtain a second tag for each report. That is, a specific feasible solution for the above-mentioned 1011 "obtaining information from multiple historical medical reports" can include the following steps:

[0049] 10111. Obtain multiple historical medical image report sets; where multiple historical medical image reports in one historical medical image report set belong to the same category;

[0050] 10112. Sample from multiple historical image report sets to extract a set number of historical image reports from each historical image report set;

[0051] 10113. A set number of historical image reports will be extracted from each historical image report set and sent to the user, so that the user can label the received historical image reports with the corresponding second tags;

[0052] 10114. Based on the set number of historical image reports and their corresponding second tags extracted from each historical image report set, determine multiple historical medical report information.

[0053] In practice, the aforementioned historical medical image report set can be constructed by users (such as radiologists) collecting a certain number (e.g., 2,000, 5,000, 10,000, etc.) of historical medical image reports. During collection, these reports can be obtained, but are not limited to, from relevant medical image report databases. The types of historical medical image reports can be categorized by body part, including reports on the head, neck, chest, abdomen, pelvis, and limbs; by imaging method, including reports on X-ray, CT (Computed Tomography), and MRI (Magnetic Resonance Imaging); and by modality, including reports on plain scans or enhanced scans. When sampling from multiple historical image report sets, sampling can be performed at the same proportion to ensure sampling balance among different types of historical image reports. After the extracted historical image reports are sent to the user, the user can use a text annotation tool (such as Brat) to annotate the received historical image reports.

[0054] It should be noted that if each historical medical image report in the aforementioned historical medical image report set has been pre-labeled with the corresponding second label, then step 10113 can be skipped, and multiple historical medical image report sets can be sampled directly to obtain information from multiple historical medical images reports.

[0055] In the above 1012, see, for example Figure 2a The illustrated medical imaging report example shows that a typical medical imaging report includes, but is not limited to: patient information (such as the patient's name, age, gender, etc.), medical images (such as image diagrams), descriptive text related to the medical images (such as image descriptions (also known as diagnostic findings, which are the contents obtained by the doctor based on the images acquired by the machine, such as atrium in positive position #####), diagnostic opinions), examination date, physician information (such as physician's name), etc. Since the descriptive text related to the medical images in the aforementioned medical imaging report is closely related to the patient's condition, it is an important textual content affecting the quality of the medical imaging report. Therefore, in this embodiment, when training the model, the descriptive text related to the medical images in the medical imaging report is mainly used as training samples. Based on this, the above-mentioned 1012 "generating the multiple first sample texts based on the multiple historical medical imaging reports" can be implemented using the following specific steps:

[0056] 10121. Extract descriptive text related to medical images from the multiple historical medical image reports;

[0057] 10122. Process the description text to obtain a processed description text that meets the preset text length;

[0058] 10123. Based on the second tag corresponding to each historical medical image report, determine the third tag of the processed description text corresponding to the multiple historical medical image reports;

[0059] 1024. Based on the third tag, determine the positive sample text and negative sample text from the processed description text corresponding to the plurality of historical medical image reports.

[0060] In practice, relevant text extraction algorithms can be used to extract medical-related descriptive text from historical medical image reports. Since the pre-trained language model often requires a fixed input text length for subsequent training, the extracted descriptive text needs to be aligned to ensure its length matches the pre-trained language model's input length requirement. Specifically, if the descriptive text length exceeds the preset length, it can be truncated, for example, by using the beginning of the text as the starting position. If the descriptive text length is less than the preset length, it can be padded, for example, by adding characters (such as punctuation marks, numbers, letters, spaces, etc.) to the end. If the descriptive text length equals the preset length, no processing is performed.

[0061] Subsequently, the second tag of the historical medical image report can be directly used as the third tag for the processed descriptive text corresponding to the historical medical image report. Alternatively, a new third tag can be automatically generated for the processed descriptive text corresponding to the historical medical image report based on the second tag. The third tag is used to identify whether the corresponding processed descriptive text is correct. For example, if the third tag indicates that the corresponding processed descriptive text is correct text without errors, then the processed descriptive text is a positive sample text; if the third tag indicates that the corresponding processed descriptive text is incorrect text containing errors, then the processed descriptive text is a negative sample text.

[0062] It should be noted that step 10121 above is not mandatory. For example, suppose a user collects historical medical image reports to construct the aforementioned historical medical image report set. If the historical medical image reports are exported from the corresponding medical image report database and the exported report items are only descriptive text related to medical images, meaning the text content of the aforementioned historical medical image reports consists only of descriptive text related to medical images, then step 10121 is not required. Furthermore, since the aforementioned multiple historical medical image reports are sampled from the historical medical image report set, and the historical medical image reports in the set may contain duplicates or incompleteness, preprocessing can be performed before generating the corresponding sample text using the aforementioned multiple historical medical image reports to ensure the quality of the final generated sample text. That is, before step 10121 above, the method provided in this embodiment may also include the following steps:

[0063] 10120. Preprocessing multiple historical medical image reports; wherein the preprocessing includes at least one of the following: cleaning reports (e.g., removing duplicate reports), deleting invalid reports (e.g., deleting incomplete reports, deleting non-medical image reports, etc.).

[0064] In the above 1013, the pre-trained language model can be, but is not limited to, language-related machine learning network models (such as neural network models, deep neural network models, etc.). More specifically, it can be a natural language model such as BERT (Bidirectional Encoder Representation from Transformers), GPT (Generative Pre-Training), ERNIE (Enhanced Language Representation with Informative Entities), etc. Training of the pre-trained language model will stop only after a preset training stopping condition is reached, and the trained pre-trained language model will be determined as the first recognition model. The model's recognition error rate or the number of training iterations can be used as the training stopping condition, but is not limited to. Based on this, in a specific implementable technical solution, the above 1013 "using the multiple first sample texts to train the pre-trained language model to obtain the first recognition model" can be implemented using the following specific steps:

[0065] 10131. Input the plurality of first sample texts into the pre-trained language model, execute the pre-trained language model, and obtain the first predicted label of the plurality of first sample texts;

[0066] 10132. Determine the first recognition error rate based on the first predicted label and the third label corresponding to the plurality of first sample texts;

[0067] 10133. If the first recognition error rate is greater than or equal to the first threshold, then the parameters of the pre-trained language model are fine-tuned, and the process returns to execute step 10131 above.

[0068] 10134. If the first recognition error rate is less than the first threshold, then stop training the pre-trained language model and determine the trained pre-trained language model as the first recognition model.

[0069] In practice, the first sample text mentioned above is the processed descriptive text mentioned earlier, and correspondingly, the third label corresponding to the first sample text refers to the third label of the processed descriptive text mentioned above. When determining the first recognition error rate, different metrics used to represent model performance can be used based on the first predicted labels and third labels corresponding to multiple first samples, such as common metrics like loss and accuracy. The calculation of these metrics can be performed on the training dataset (i.e., the multiple first sample texts mentioned above) to understand the model's "learning" status.

[0070] For example, after each iteration of training, the pre-trained language model can calculate the corresponding loss value as the first recognition error rate based on the first predicted label and third label corresponding to multiple first sample texts using a loss function. The loss function can be, but is not limited to, the cross-entropy loss function (also called the logarithmic loss function), and its corresponding expression is as follows:

[0071]

[0072] Where N is the total number of samples in the first sample text; y i Let represent the third label of the i-th first sample text. The third label indicates whether the corresponding first sample text is a positive or negative sample text. For example, if the third label is 1, it means that the corresponding first sample text is a positive sample text; if the third label is 0, it means that the corresponding first sample text is a negative sample text. q represents the probability distribution of the sample labels (i.e., the third labels mentioned above) of the N first sample texts; P represents the probability distribution of the predicted labels (i.e., the first predicted labels mentioned above) of the N first sample texts. Specifically, p(y i ) represents the i-th first sample text y i H represents the predicted probability (i.e., predicted label) of the positive sample text. p (q) represents the total loss value. Further, assuming the pre-trained language model has a total of L network layers, then the loss value of the j-th network layer is Loss_L. j for:

[0073] Loss_L j =-W j [Y j *logX j +(1-Y j )*log(1-X j )]

[0074] Among them, X j Y is the input to the j-th network layer. j For the predicted label output by the j-th network layer, W jThese are the parameters (weights) of the j-th network layer.

[0075] If the calculated H above p When the loss value is greater than or equal to the preset loss threshold (i.e., the first threshold), it indicates that the performance of the pre-trained language model has not met the requirements. In this case, the parameters of the pre-trained language model will be fine-tuned. Specifically, depending on the actual training situation, the parameters (weights) of all network layers of the pre-trained language model can be adjusted, or the parameters of some network layers can be adjusted; this is not limited here. During the fine-tuning of the pre-trained language model parameters, gradient descent can be used, but is not limited to. Afterwards, based on the fine-tuned pre-trained language model, the above step 10131 is returned to be executed. If H... p If the loss is less than the preset threshold, it indicates that the performance of the pre-trained language model has met the requirements. At this point, training of the pre-trained language model will be stopped, and the pre-trained language model will be designated as the first recognition model.

[0076] For example, the predicted labels of each first sample text can be compared with the corresponding third labels to determine the number n of correctly identified sample texts among the multiple first sample texts. Then, the ratio of the number n of correctly identified sample texts to the total number N of the multiple first sample texts can be calculated to obtain the recognition accuracy (i.e., n / N). The aforementioned first recognition error rate is the difference between a set value (such as an integer 1) and the recognition accuracy, and thus, based on the first recognition error rate, it can be determined whether to stop training the pre-trained language model.

[0077] Figure 2b An example is shown of training a pre-trained language model using multiple first sample texts.

[0078] In the above 102, multiple second sample texts can be generated from a certain number (e.g., 500) of historical medical image reports containing various errors collected by a user (e.g., a radiologist). When annotating these historical medical image reports, the user can only annotate the erroneous statements, without needing to annotate the specific error content. This is because the training method used to train the first recognition model using the second sample texts is weak supervision. Weak supervision is a training method that provides the model with a small amount of annotation or incomplete annotation, allowing the model to learn autonomously based on this limited annotation information, thereby inferring the overall sample characteristics. Therefore, due to the annotation method described above, the sample statements with labels (i.e., the first labels) in each of the generated second sample texts are erroneous sample statements containing incorrect words, while the unlabeled sample statements are correct sample statements without incorrect words.

[0079] For details on the specific implementation of generating the second sample text, please refer to the relevant content on the specific implementation of generating the first sample text described above.

[0080] After obtaining multiple second sample texts, the sample sentences contained in the multiple second sample texts can be used to train the first recognition model. The training process can be found in the pre-trained language model training process described above. Based on this, in one possible technical solution, the above-mentioned 103 "training the first recognition model based on at least one sample sentence included in the multiple second sample texts to obtain a second recognition model" can specifically include:

[0081] 1031. Input the at least one sample statement into the first recognition model, and execute the first recognition model to obtain the second predicted label of the at least one sample statement;

[0082] 1032. Determine the second recognition error rate based on the second predicted label and the first label information of the at least one sample statement;

[0083] 1033. If the second recognition error rate is greater than or equal to the second threshold, then the parameters of the first recognition model are fine-tuned, and the process returns to the step of inputting the at least one sample statement into the first recognition model and executing the first recognition model to obtain the predicted label of the at least one sample statement.

[0084] 1034. If the second recognition error rate is less than the second threshold, then stop training the first recognition model and determine the trained first recognition model as the second recognition model.

[0085] In practice, the second recognition error rate can be determined using different metrics to represent model performance, such as common metrics like loss and accuracy, based on the second predicted label and the first label information corresponding to at least one sample statement. For the calculation of the second recognition error rate, please refer to the example of calculating the first recognition error rate given above. Furthermore, the description of performing steps 1033-1034 based on the second recognition error rate can also be found in the example given above regarding the first recognition error rate.

[0086] Furthermore, during the training process, at least one sample sentence that was not correctly identified can be sent to the user so that the user can manually extract the erroneous words from the sample sentences that the model failed to identify, and establish a corresponding vocabulary (i.e., the error detection rule set described below). This facilitates the subsequent supplementation of errors that the aforementioned model still cannot identify after training, using a rule-based method. That is, step 103 above may also include the following specific steps:

[0087] 1035. Compare the first predicted label and the first label corresponding to the first erroneous sample statement in the at least one sample statement;

[0088] 1036. If the second predicted label corresponding to the first erroneous sample statement is consistent with the first label, then the first erroneous sample statement is identified;

[0089] 1037. If the second predicted label corresponding to the first erroneous sample statement is inconsistent with the first label, then the first erroneous sample statement is not identified, triggering the operation of "sending the first erroneous sample statement to the user" in 104 above.

[0090] In the above 104-105, the error detection rule set can be added to the second recognition model to obtain the error detection model, thereby improving the performance of the error detection model.

[0091] The technical solution provided in this embodiment, when training an error detection model for detecting errors in medical image reports, first determines a first recognition model, which is trained based on multiple first sample texts of the medical image report. Then, it acquires multiple second sample texts of the medical image report, each including at least one sample statement. Among these at least one sample statement, those with a first label are erroneous sample statements containing incorrect words, while those without a first label are correct sample statements without incorrect words. The first recognition model is then trained using the at least one sample statement from the multiple first sample texts to obtain a second recognition model. Through the training performed in these two stages using at least one sample statement from the first and second sample texts respectively, the trained second recognition model possesses a high level of text semantic understanding, effectively improving the accuracy of error detection. Furthermore, during the training process of the first recognition model, if it is detected that the first recognition model cannot identify the first erroneous sample statement in at least one sample statement, the first erroneous sample statement is sent to the user for identification of the first erroneous word in the first erroneous sample statement and added to the error detection rule set. Finally, the error detection model is determined based on the second recognition model and the error detection rule set. Here, the error detection rule set is used to supplement the difficult errors that the second recognition model trained in the aforementioned two training stages cannot identify, which can further improve the error identification accuracy of the error detection model and avoid the problem of high burden and cost of manually constructing a dictionary (such as the error detection rule set mentioned above) when directly constructing an error detection model based on rules. In summary, this solution can train a high-performance error detection model with low training cost.

[0092] Furthermore, to verify the performance of the error detection model, it can be tested using a test sample text set of corresponding medical image reports. If the performance of the error detection model does not meet the requirements based on the test results, the model can be optimized. That is, the method provided in this embodiment may also include the following steps:

[0093] 106. Obtain a test sample text set of medical imaging reports;

[0094] 107. Based on the test sample text set, test the detection model to optimize the error detection model.

[0095] In specific implementation, the aforementioned test sample text set includes multiple third sample texts and fourth tag information for each third sample text. The multiple test sample texts are generated based on multiple test sample medical image reports. For specific implementation details regarding test sample generation, please refer to the relevant descriptions of the generation of the first or second sample texts. The fourth tag information of a third sample text may include at least one fourth tag, which is used to identify any of the following: the third sample text is a positive or negative sample text, an erroneous statement, an erroneous word, etc. When testing the error detection model using the test sample text set, the model performance can be determined based on the error detection rate of the error detection model to determine whether optimization is needed. Based on this, in one implementable technical solution, the aforementioned 107 "testing the detection model based on the test sample text set to optimize the error detection model" may specifically include:

[0096] 1071. Input the plurality of third sample texts into the error detection model, execute the test results of the label information into the error detection model, and obtain test results containing the third predicted label information corresponding to each third sample text;

[0097] 1072. Using the test results and the fourth label information, determine the error rate of the error detection model;

[0098] 1073. If the error rate of the error detection is greater than or equal to the third threshold, then based on the test results and the fourth label information, optimize the error detection model, and return to the step of testing the error detection model based on the test sample text set to optimize the error detection model;

[0099] 1074. If the error rate of the error detection is less than the third threshold, then the testing of the error detection model shall be stopped.

[0100] In specific implementation, the values ​​of indicators such as loss and Accuray can be calculated based on the third predicted label information and the fourth label information corresponding to each third sample text contained in the test results, which serve as the error rate of the error detection model. For specific implementation details regarding the calculation of the values ​​of loss, Accuray, etc., please refer to the example given above for the first recognition error rate. When the error rate is greater than or equal to the third threshold value, the detection model can be optimized by adding erroneous words that the error detection model cannot recognize from the third sample text to the error detection rule set. That is, in a specific implementable scheme, the "optimizing the error detection model based on the test results and the third label information" in section 1073 above can be achieved using the following specific steps:

[0101] 10731. Based on the test results and the fourth tag information, determine the second erroneous words in the plurality of third sample texts that the error detection model cannot recognize and add the error detection rule set;

[0102] 10732. Based on the added error detection rule set, trigger 105 above.

[0103] In practice, the determination of the second erroneous word and its addition to the error detection rule set can both be completed by the user. Therefore, the above-mentioned 10731 "based on the test results and the third tag information, determine the second erroneous word in the plurality of third sample texts that the error detection model cannot recognize and add it to the error detection rule set" can be implemented using the following specific steps:

[0104] 107311. Based on the test results and the fourth label information, determine the target third sample text in the plurality of third sample texts where the third predicted label information differs from the corresponding fourth label information;

[0105] 107312. The target third sample text, the fourth predicted label information corresponding to the target third sample text, and the difference label information between the fourth label information are sent to the user so that the user can determine the second error word that the error detection model cannot recognize from the target third sample text based on the difference label information, and add the second error word to the error detection rule set.

[0106] For example, if there is a difference between the third predicted label information corresponding to the third sample text and its corresponding fourth label information, then the difference label information between the third predicted label information corresponding to the third sample text and its corresponding fourth label information can be determined. For example, if the third predicted label 'a' in the third predicted label information... 11 With the fourth tag a in the fourth tag information 21If both are labels of sample statement 'a' in the target third sample text A, and they are inconsistent (i.e., there is a difference between them), then the difference label information can contain the third predicted label 'a'. 11 and the fourth label a 21 Next, the difference label information can be included in the corresponding target third sample text A and sent to the user's client. This information is then displayed on the client's interactive interface, allowing the user to quickly locate the position in the target third sample text that the error detection model cannot recognize. This identifies the second erroneous word that the error detection model cannot identify and adds it to the error detection rule set. In response to the user-triggered completion of the addition operation, the system returns to step 105 based on the added error detection rule set to obtain the optimized error detection model. Furthermore, based on the optimized error detection model, the system returns to step 1071 to continue testing and optimizing the error detection model until the testing stop condition is met.

[0107] It should be noted that the thresholds mentioned above, such as the first threshold, the second threshold, and the third threshold, are flexibly set according to the actual situation. For example, the first threshold can be 50%, meaning that training the pre-trained language model is stopped only when the recognition error rate is less than 50%, thus obtaining the first recognition model; the second threshold can be 30%, meaning that training the first recognition model is stopped only when the recognition error rate is less than 30%, thus obtaining the second recognition model; the third threshold can be 10%, meaning that testing the error detection model is stopped only when the recognition error rate is less than 10%, and optimization is carried out in sequence to obtain the final error detection model.

[0108] In summary, the error detection model training process provided in the above-described embodiments can be simply summarized as follows: Figure 3 The three stages shown:

[0109] Phase 1: Fine-tuning training based on the pre-trained language model

[0110] Because human language is extremely complex, training a text recognition model from scratch is very difficult. Currently, the mainstream technique involves using publicly available pre-trained preliminary models (i.e., pre-trained language models) and further training these models in sub-domains to obtain models capable of handling new tasks. Therefore, since medical image reports are part of Chinese big data, containing both general Chinese features and medical-specific vocabulary and expressions, this application, to reduce training costs, utilizes a pre-trained language model in the first training stage. A fine-tuning training method is employed, using multiple first-sample texts generated from a large number of medical image reports to train the pre-trained language model. This allows the trained first recognition model to gain a preliminary understanding of the combination patterns of commonly used characters and words in medical imaging.

[0111] The first stage mentioned above can be understood as training the model from the medical image report as a whole, so that the first recognition model trained can simply identify whether there are errors in the medical image report as a whole, and is not required to be able to specifically identify the specific errors in the medical image report.

[0112] Phase Two: Weakly Supervised Learning Training Based on Incomplete Error Labeling

[0113] Weakly supervised learning is a training method that provides a model with limited or incomplete annotations, enabling it to learn autonomously based on this limited information and infer the characteristics of the overall sample. In this stage, using weakly supervised learning to train the first recognition model to obtain the second recognition model allows physicians to collect samples by simply annotating the entire sentence or paragraph containing the error in the medical image report, without needing to standardize specific erroneous keywords or characters. During training, the model can learn the semantic features of the erroneous sentences or paragraphs on its own.

[0114] The second stage described above can be understood as training the model using sentences from medical imaging reports, enabling the trained second recognition model to specifically identify the errors in the medical imaging reports. For example, during training, the input to the first recognition model is a sentence such as "left lung nodule, follow-up treatment for discomfort," rather than partial words such as "follow-up treatment for discomfort."

[0115] The third stage: a rule-based recognition method to supplement the second recognition model with rules.

[0116] Rule-based error detection methods (i.e., the rule-based modeling methods mentioned above) refer to defining certain specific keywords or words, and judging an error as occurring when such words or words appear in a sentence. The limitation of this rule-based method is that it can only identify fixed combinations of incorrect words, and the number and types of incorrect words are impossible to exhaustively enumerate or estimate. In this application, for errors that cannot be identified after training the aforementioned two models, a rule-based method is used to supplement the second detection model, ultimately obtaining a valid error detection model.

[0117] In summary, this application's solution is based on natural language processing technology and trains the required error detection model through three stages. Specifically, in the first two stages, the model is trained using fine-tuning and weakly supervised learning methods to improve its ability to understand text semantics. This allows the trained model (the second recognition model) to identify not only the features of erroneous words but also the overall semantic and word order features of sentences or paragraphs containing erroneous words. Then, based on the second recognition model trained in the aforementioned two stages, a rule-based recognition method is used to supplement the model with difficult errors that it could not identify, thus constructing the error detection model. This avoids the problem of large-scale rule annotation by physicians required to directly use rule-based recognition methods to build the required error detection model, reducing the workload of manually constructing a lexicon.

[0118] Figure 4 A flowchart illustrating a medical image report detection method according to another embodiment of this application is shown. The execution entity of this method can be an electronic device with logic processing capabilities, which can be a client or a server. For detailed descriptions of the client or server, please refer to the relevant content in other embodiments above. Figure 4 As shown, the method provided in this embodiment includes the following steps:

[0119] 201. Identify the target medical imaging report to be tested;

[0120] 202. Obtain the error detection model, wherein the monitoring model is obtained through, for example... Figure 1 The error detection model shown in the embodiment of this application is trained using the training method provided.

[0121] 203. Using the error detection model, perform error detection on the target medical image report.

[0122] In specific implementation, the aforementioned target medical image report can be an electronic image report recently completed by a user (such as a radiologist). After the user completes the corresponding target medical image report, they can click "Confirm." The executing entity responds to the user-triggered operation, obtains an error detection model, and uses the model to detect errors in the target medical image report. Furthermore, the error detection results can be sent to the user for display on their client application. If the displayed error detection results indicate an error in the target medical image report, the user can quickly modify the erroneous content based on the displayed error detection results. Therefore, the method provided in this embodiment may further include:

[0123] 204. Display the error detection results to the user;

[0124] 205. In response to the user's modification operation on the target medical report based on the error detection results, update the target medical image report.

[0125] Figure 5 This application illustrates a medical image report detection system according to another embodiment. The system includes a server 32 and a client 31.

[0126] Server 32 determines the target medical image report to be detected; obtains an error detection model, wherein the error detection model is trained by an error detection model training method provided in an embodiment of this application; and uses the error detection model to detect errors in the target medical image report.

[0127] Client 31 is used to provide an interactive interface, on which the error detection results of the target medical image report are displayed.

[0128] For a detailed description of the specific forms of the server and client described above, please refer to the relevant content in other embodiments above, which will not be repeated here. Furthermore, for any undescribed functions performed by the server and client, please refer to the relevant content in other embodiments above. In addition to the steps described above, the medical image report detection system provided in this application may also include other parts or all of the steps in the above embodiments; please refer to the corresponding content in the above embodiments for details, which will not be repeated here.

[0129] Figure 6 A schematic diagram of the structure of an error detection model training device according to an embodiment of this application is shown, wherein the error detection model is used to detect errors in medical reports. Figure 6 As shown, the error detection model training device includes: a determination module 41, an acquisition module 42, a training module 43, and a monitoring and transmission module 44; wherein,

[0130] The determination module 41 is used to determine the first recognition model, which is trained based on multiple first sample texts of medical image reports;

[0131] The acquisition module 42 is used to acquire multiple second sample texts of a medical image report; wherein, the second sample text includes at least one sample statement, and the sample statement with a first tag is an erroneous sample statement containing erroneous words, and the sample statement without the first tag is a correct sample statement without erroneous words;

[0132] Training module 43 is used to train the first recognition model based on at least one sample sentence included in the plurality of first sample texts to obtain a second recognition model;

[0133] The monitoring and sending module 44 is used to send the first erroneous sample statement to the user if it is detected that the first recognition model cannot recognize the first erroneous sample statement in the at least one sample statement during the training process of the first recognition model, so that the user can determine the first erroneous word in the first erroneous sample statement and add it to the error detection rule set.

[0134] The determining module 41 is further configured to determine the error detection model based on the second identification model and the error detection rule set.

[0135] Furthermore, the aforementioned determining module 41, when used to determine the first recognition model, is specifically used for: acquiring multiple historical medical image report information; generating multiple first sample texts based on the multiple historical medical image report information; and training the pre-trained language model to be trained using the multiple first sample texts to obtain the first recognition model.

[0136] Furthermore, the aforementioned information on multiple historical medical imaging reports includes: multiple historical medical imaging reports, and a second tag corresponding to each historical medical imaging report; the second tag is used to identify whether the corresponding historical medical imaging report is a correct medical imaging report; and

[0137] The aforementioned determining module 41, when used to generate multiple first sample texts based on the multiple historical medical image report information, is specifically used for: extracting descriptive text related to medical images from the multiple medical image reports; processing the descriptive text to obtain processed descriptive text that meets a preset text length; determining a third tag for the processed descriptive text corresponding to the multiple historical medical image reports based on the second tag corresponding to each historical medical image report; and determining positive sample text and negative sample text from the processed descriptive text corresponding to the multiple historical medical image reports based on the third tag.

[0138] Further, the aforementioned determining module 41, when used to train the pre-trained language model to be trained using the plurality of first sample texts to obtain the first recognition model, is specifically configured to: input the plurality of first sample texts into the pre-trained language model, execute the pre-trained language model, and obtain the first predicted label of the plurality of first sample texts; determine the first recognition error rate based on the first predicted label and the third label corresponding to the plurality of first sample texts; if the first recognition error rate is greater than or equal to the first threshold, fine-tune the parameters of the pre-trained language model and return to execute the step of inputting the plurality of first sample texts into the pre-trained language model and executing the pre-trained language model to obtain the predicted label of the plurality of first sample texts; if the first recognition error rate is less than the first threshold, stop training the pre-trained language model and determine the trained pre-trained language model as the first recognition model.

[0139] Further, the training module 43 described above, when used to train the first recognition model to obtain a second recognition model based on at least one sample statement included in the plurality of first texts, specifically includes: inputting the at least one sample statement into the first recognition model, executing the first recognition model to obtain a second predicted label for the at least one sample statement; determining a second recognition error rate based on the second predicted label and the first label information of the at least one sample statement; if the second recognition error rate is greater than or equal to a second threshold, fine-tuning the parameters of the first recognition model and returning to the step of inputting the at least one sample statement into the first recognition model and executing the first recognition model to obtain a predicted label for the at least one sample statement; if the second recognition error rate is less than the second threshold, stopping the training of the first recognition model and determining the trained first recognition model as the second recognition model.

[0140] Furthermore, the error detection model training device provided in this embodiment also includes:

[0141] The comparison module is used to compare the second predicted label and the first label corresponding to the first erroneous sample statement in the at least one sample statement to obtain a comparison result;

[0142] The aforementioned determining module 41 is further configured to determine, based on the comparison result, whether the second predicted label corresponding to the first erroneous sample statement is consistent with the first label; if the second predicted label corresponding to the first erroneous sample statement is consistent with the first label, then the first erroneous sample statement is identified; if the second predicted label corresponding to the first erroneous sample statement is inconsistent with the first label, then the first erroneous sample statement is not identified, triggering the operation of sending the first erroneous sample statement to the user.

[0143] Furthermore, the acquisition module 42 is also used to acquire a test sample text set of medical image reports; and the error detection training device provided in this embodiment further includes: a testing module, used to test the error detection model based on the test sample text set to optimize the error detection model.

[0144] Further, the aforementioned test sample set includes: multiple third sample texts and fourth label information for each third sample text; and, when the aforementioned testing module is used to test the error detection model based on the test sample text set to optimize the error detection model, it is specifically used to: input the multiple third sample texts into the error detection model, execute the error detection model, and obtain test results containing third prediction label information corresponding to each third sample text; determine the error detection error rate of the error detection model based on the test results and the fourth label information; if the error detection error rate is greater than or equal to a third threshold, optimize the error detection model based on the test results and the third label information, and return to execute the step of testing the error detection model based on the test sample text set to optimize the error detection model; if the error detection error rate is less than the third threshold, stop testing the error detection model.

[0145] Furthermore, the aforementioned testing module, when used to optimize the error detection model based on the test results and the fourth tag information, is specifically used to: determine, based on the test results and the fourth tag information, identify second erroneous words in the plurality of third sample texts that the error detection model cannot recognize and add them to the error detection rule set; and, based on the added error detection rule set, trigger the operation of determining the error detection model according to the second recognition model and the error detection rule set.

[0146] It should be noted that the error detection model training device provided in this embodiment can achieve the above. Figure 1 The technical solutions described in the embodiments of the detection model training method shown above, and the specific implementation principles of each module or unit can be found in the above. Figure 1 The relevant content in the illustrated detection model training method embodiment will not be repeated here.

[0147] Figure 7A schematic diagram of the structure of a medical image report detection device according to another embodiment of this application is shown. Figure 7 As shown, the medical image report detection device includes: a determination module 51, an acquisition module 52, and a detection module 53; wherein,

[0148] Module 51 is used to determine the target medical image report to be detected;

[0149] Acquisition module 52 is used to acquire an error detection model, wherein the error detection model is acquired through methods such as... Figure 1 The error detection model training method provided in one embodiment of this application is shown to be used for training.

[0150] The detection module 53 is used to perform error detection on the target medical image report using the error detection model.

[0151] It should be noted that the medical image report detection device provided in this embodiment can achieve the above-mentioned... Figure 4 The technical solutions described in the illustrated medical image report detection method embodiments, and the specific implementation principles of each module or unit can be found in the above description. Figure 4 The relevant content in the illustrated medical image reporting method embodiments will not be repeated here.

[0152] Figure 8 A schematic diagram of the structure of an electronic device provided according to an embodiment of this application is shown. Figure 8 As shown, the electronic device includes a memory 61 and a processor 62. The memory 61 stores one or more computer instructions; the processor 62, coupled to the memory 61, is used to execute one or more computer instructions (such as computer instructions for implementing data storage logic) to implement the error detection model training method provided in the embodiments of this application, or to implement the medical image report detection method provided in the embodiments of this application.

[0153] The memory 61 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0154] Furthermore, such as Figure 8 As shown, the electronic device may also include other components such as a communication component 63, a power supply component 64, an audio component 65, and a display 66. Figure 8 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 8 The components shown.

[0155] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, can implement the steps or functions in the error detection model training method or medical image report detection method provided in the above embodiments.

[0156] The methods in this application can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as a computer program product. Figure 9 A block diagram of a computer program product provided in this application is schematically shown. The computer program product includes a computer program / instructions 71, which, when the computer program / instructions 71 are executed by, for example... Figure 7 When executed by a processor such as processor 62, the processes or functions in the error detection model training method or medical image report detection method described in this application can be performed in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, network equipment, user equipment, core network equipment, OAM, or other programmable device.

[0157] The computer program or instructions may be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions may be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium may be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; or an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both volatile and non-volatile types of storage media.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for training an error detection model, characterized in that, The error detection model is used to detect errors in medical image reports; the method includes: A first recognition model is determined, which is trained based on multiple first sample texts from medical image reports; Obtain multiple second sample texts of a medical imaging report; wherein the second sample texts include at least one sample statement, and the sample statement with a first tag is an erroneous sample statement containing erroneous words, and the sample statement without the first tag is a correct sample statement without erroneous words; The first recognition model is trained based on at least one sample sentence included in the plurality of second sample texts to obtain a second recognition model; During the training of the first recognition model, if it is detected that the first recognition model cannot recognize the first erroneous sample statement in the at least one sample statement, the first erroneous sample statement is sent to the user so that the user can determine the first erroneous word in the first erroneous sample statement and add it to the error detection rule set. The error detection model is determined based on the second identification model and the error detection rule set.

2. The method according to claim 1, characterized in that, The first recognition model is determined, including: Retrieve information from multiple historical medical imaging reports; Based on the information from the multiple historical medical image reports, the multiple first sample texts are generated; The first recognition model is obtained by training the pre-trained language model to be trained using the multiple first sample texts.

3. The method according to claim 2, characterized in that, The multiple historical medical imaging report information includes: multiple historical medical imaging reports and a second tag corresponding to each historical medical imaging report; the second tag is used to identify whether the corresponding historical medical imaging report is a correct medical imaging report; and Based on the aforementioned historical medical image report information, multiple first sample texts are generated, including: Extract descriptive text related to medical images from the multiple historical medical image reports; The description text is processed to obtain a processed description text that meets the preset text length. Based on the second tag corresponding to each historical medical image report, determine the third tag of the processed descriptive text corresponding to the multiple historical medical image reports; Based on the third tag, positive and negative sample texts are determined from the processed description texts corresponding to the plurality of historical medical image reports.

4. The method according to claim 3, characterized in that, Using the multiple first sample texts, a pre-trained language model to be trained is trained to obtain the first recognition model, including: The plurality of first sample texts are input into the pre-trained language model, and the pre-trained language model is executed to obtain the first predicted labels of the plurality of first sample texts. The first recognition error rate is determined based on the first predicted label and the third label corresponding to the plurality of first sample texts; If the first recognition error rate is greater than or equal to the first threshold, the parameters of the pre-trained language model are fine-tuned, and the process returns to the step of inputting the plurality of first sample texts into the pre-trained language model and executing the pre-trained language model to obtain the predicted labels of the plurality of first sample texts. If the first recognition error rate is less than the first threshold, then training the pre-trained language model is stopped, and the trained pre-trained language model is determined as the first recognition model.

5. The method according to any one of claims 1 to 4, characterized in that, Based on at least one sample statement included in the plurality of second sample texts, the first recognition model is trained to obtain a second recognition model, including: The at least one sample statement is input into the first recognition model, and the first recognition model is executed to obtain the second predicted label of the at least one sample statement; Based on the second predicted label and the first label information of at least one sample statement, a second recognition error rate is determined; If the second recognition error rate is greater than or equal to the second threshold, the parameters of the first recognition model are fine-tuned, and the process returns to the step of inputting the at least one sample statement into the first recognition model and executing the first recognition model to obtain the predicted label of the at least one sample statement. If the second recognition error rate is less than the second threshold, then training the first recognition model is stopped, and the trained first recognition model is determined as the second recognition model.

6. The method according to claim 5, characterized in that, Also includes: Compare the second predicted label and the first label corresponding to the first erroneous sample statement in the at least one sample statement; If the second predicted label corresponding to the first erroneous sample statement is consistent with the first label, then the first erroneous sample statement is identified; If the second predicted label corresponding to the first erroneous sample statement is inconsistent with the first label, then the first erroneous sample statement is not identified, triggering the operation of sending the first erroneous sample statement to the user.

7. The method according to any one of claims 1 to 4, characterized in that, Also includes: Obtain a test sample text set of medical imaging reports; Based on the test sample text set, the error detection model is tested to optimize it.

8. The method according to claim 7, characterized in that, The test sample text set includes: multiple third sample texts and fourth tag information for each third sample text; and Based on the test sample text set, the error detection model is tested to optimize it, including: The multiple third sample texts are input into the error detection model, the error detection model is executed, and a test result containing the third predicted label information corresponding to each third sample text is obtained. Based on the test results and the fourth label information, the error rate of the error detection model is determined. If the error rate is greater than or equal to the third threshold, then based on the test results and the fourth label information, the error detection model is optimized, and the process of testing the error detection model based on the test sample text set to optimize the error detection model is returned. If the error rate of the error detection is less than the third threshold, then the testing of the error detection model is stopped.

9. The method according to claim 8, characterized in that, Based on the test results and the fourth label information, the error detection model is optimized, including: Based on the test results and the fourth tag information, the second erroneous words that the error detection model cannot recognize in the multiple third sample texts are identified and added to the error detection rule set; Based on the added error detection rule set, the operation of determining the error detection model according to the second identification model and the error detection rule set is triggered.

10. A method for detecting medical image reports, characterized in that, include: Identify the target medical imaging report to be tested; An error detection model is obtained, wherein the error detection model is trained by the error detection model training method as described in any one of claims 1 to 9; The error detection model is used to detect errors in the target medical image report.

11. A medical image report detection system, characterized in that, include: The server is used to determine the target medical image report to be detected; obtain an error detection model, wherein the error detection model is trained by the error detection model training method as described in any one of claims 1 to 9; and use the error detection model to detect errors in the target medical image report. A client is used to provide an interactive interface; on the interactive interface, the error detection results of the target medical image report are displayed.

12. An electronic device, characterized in that, include: Memory and processor, among which, The memory is used to store computer programs; The processor, coupled to the memory, is used to execute the computer program stored in the memory to implement the error detection model training method as described in any one of claims 1 to 9, or to implement the medical image report detection method as described in claim 10.

Images