Medical image report splitting method for large model training
By performing sentence-processing and classification model input methods on medical imaging reports, we can accurately split the inspection observations and inspection impressions in medical imaging reports, solving the problem of inaccurate splitting in the existing technology, and improving the training effect of large models and data acquisition efficiency.
Patent Information
- Application Number
- CN202411561213.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2025-03-04
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art is difficult to accurately split the examination observations and inspection impressions in medical imaging reports by categories, resulting in insufficient accuracy in large models when generating targeted reports, high data labeling costs, and it is difficult to obtain training data that meets the requirements.
By reading medical image reports, processing and extracting word vectors and mask vectors, inputting pre-trained classification models for classification, and combining clauses with the same classification results, the accurate split of what is seen and what is inspected is achieved.
It improves the accuracy of medical imaging models when generating targeted reports, reduces the cost of data labeling, and can obtain large batches of training data that meet the requirements in a short period of time.
Smart Images
Figure CN119560089A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to medical image data processing, and in particular to a medical image report splitting method for large model training. Background Art
[0002] A medical imaging report is a detailed report written by a doctor after observation and analysis based on the images generated by medical imaging equipment. It plays an important role in medical diagnosis. Medical imaging reports can provide a basis for diagnosis and detailed information about the patient's condition, which helps doctors make accurate diagnoses and treatment plans. At the same time, medical imaging reports are an important tool for effective communication between doctors and patients, which can help patients better understand their condition, improve treatment compliance, and further promote communication between doctors and patients.
[0003] With the rapid development of computer and big data technology, deep learning has achieved remarkable results in various fields. Researchers have continuously increased the number of model parameters to improve the performance of the model, thus giving birth to the concept of big models. Big models use a large amount of data and computing resources to train neural network models with a large number of parameters, which can achieve good performance in various tasks. Therefore, training big models with medical knowledge and learning to simulate human doctors is a major development trend at present, and medical imaging report splitting belongs to the upstream task of training medical imaging big models.
[0004] At present, in the process of training large medical imaging models, medical imaging reports are not split accordingly, but directly input into the model for training. Although this method can train large medical imaging models, when large models need to generate targeted reports (such as nasopharyngeal reports, etc.), the accuracy is slightly insufficient. At the same time, the cost of data annotation is high, and it is not easy to obtain large quantities of training data that meet the requirements. Summary of the invention
[0005] 1. Technical issues to be resolved
[0006] In view of the above-mentioned shortcomings of the prior art, the present invention provides a medical imaging report splitting method for large model training, which can effectively overcome the defect of the prior art that it is difficult to accurately split the examination findings and examination impressions in the medical imaging report by category.
[0007] (II) Technical solution
[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:
[0009] A medical imaging report splitting method for large model training includes the following steps:
[0010] S1. Read the medical imaging report to be split, and perform sentence processing on the medical imaging report to obtain a sentence combination of examination findings and a sentence combination of examination impressions;
[0011] S2, extracting the word vector and mask vector corresponding to each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination;
[0012] S3, inputting the word vector and mask vector corresponding to each inspection observation clause in the inspection observation clause combination into the pre-trained inspection observation classification model to obtain the inspection observation classification result corresponding to each inspection observation clause;
[0013] S4, inputting the word vector and mask vector corresponding to each inspection impression sentence in the inspection impression sentence combination into the pre-trained inspection impression classification model to obtain the inspection impression classification result corresponding to each inspection impression sentence;
[0014] S5, merging and outputting the inspection finding sentences with the same inspection finding classification results, and merging and outputting the inspection impression sentences with the same inspection impression classification results, thereby completing the medical imaging report splitting;
[0015] Among them, the inspection observation classification model and the inspection impression classification model are both built based on the RoBERTa language model.
[0016] Preferably, in S1, the medical imaging report to be split is read, and the medical imaging report is processed into sentences to obtain a combination of examination findings sentences and a combination of examination impression sentences, including:
[0017] S11, reading the medical imaging report to be split;
[0018] S12. Segment the medical imaging report according to a preset text end symbol, and distinguish the sentences corresponding to the examination findings and the examination impressions, to obtain an examination findings sentence combination and an examination impression sentence combination, respectively.
[0019] Preferably, extracting the word vector and mask vector corresponding to each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination in S2 includes:
[0020] S21, counting the maximum number of characters N contained in all the sentences in the inspection-seen sentence combinations and the inspection-impression sentence combinations;
[0021] S22, initialize the BERT word segmenter and set the maximum length of the sentence to the maximum number of characters N;
[0022] S23. Input each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination into the BERT word segmenter respectively to obtain the corresponding word vector and mask vector.
[0023] Preferably, in S3, the word vector and mask vector corresponding to each inspection observation clause in the inspection observation clause combination are input into a pre-trained inspection observation classification model to obtain the inspection observation classification result corresponding to each inspection observation clause, including:
[0024] S31, inputting the word vector and mask vector corresponding to the inspection-what-you-see clause in the inspection-what-you-see clause combination into the pre-trained inspection-what-you-see classification model, and the inspection-what-you-see classification model outputs the probability prediction value of the inspection-what-you-see clause belonging to each category label;
[0025] S32, selecting the category label with the maximum probability prediction value as the candidate classification result, and determining the relationship between the maximum probability prediction value and a preset threshold;
[0026] S33, if the maximum probability prediction value is greater than a preset threshold, the candidate classification result is used as the inspection finding classification result corresponding to the inspection finding sentence, otherwise the corresponding inspection finding classification result is determined according to the inspection finding sentence content;
[0027] S34, repeat S31 to S33 until all inspection observation clauses in the inspection observation clause combination obtain corresponding inspection observation classification results.
[0028] Preferably, in S4, the word vector and mask vector corresponding to each inspection impression sentence in the inspection impression sentence combination are input into a pre-trained inspection impression classification model to obtain the inspection impression classification result corresponding to each inspection impression sentence, including:
[0029] S41, inputting the word vector and mask vector corresponding to the inspection impression sentence in the inspection impression sentence combination into the pre-trained inspection impression classification model, and the inspection impression classification model outputs the probability prediction value of the inspection impression sentence belonging to each category label;
[0030] S42, selecting the category label with the maximum probability prediction value as the candidate classification result, and determining the relationship between the maximum probability prediction value and a preset threshold;
[0031] S43, if the maximum probability prediction value is greater than a preset threshold, the candidate classification result is used as the inspection impression classification result corresponding to the inspection impression sentence, otherwise the corresponding inspection impression classification result is determined according to the inspection impression sentence content;
[0032] S44. Repeat S41 to S43 until all inspection impression sentences in the inspection impression sentence combination obtain corresponding inspection impression classification results.
[0033] Preferably, before reading the medical imaging report to be split in S1 and performing sentence processing on the medical imaging report to obtain the inspection findings sentence combination and the inspection impression sentence combination, the following steps are included:
[0034] S01. Obtain relevant medical imaging reports and the categories of examination items included therein, and set category labels according to the categories of examination items;
[0035] S02. Classify the medical imaging reports according to the categories of the examination items to obtain a medical imaging report data set corresponding to each category label;
[0036] S03, performing sentence processing on the medical imaging report data set corresponding to each category label to obtain an examination finding sentence data set and an examination impression sentence data set corresponding to each category label;
[0037] S04, merging the inspection observation sentence data sets corresponding to each category label into an inspection observation sentence set, and merging the inspection impression sentence data sets corresponding to each category label into an inspection impression sentence set;
[0038] S05, extracting the word vector and mask vector corresponding to each inspection-seen sentence in the inspection-seen sentence set and each inspection-impression sentence in the inspection-impression sentence set, and completing the construction of the inspection-seen sentence historical data set and the inspection-impression sentence historical data set;
[0039] S06, using the historical data set of inspection observations, the inspection observations classification model is trained;
[0040] S07. Use the inspection impression sentence historical data set to train the inspection impression classification model.
[0041] Preferably, in S01, obtaining the relevant medical imaging report and the categories of the examination items contained therein, and setting the category label according to the category of the examination item includes:
[0042] Obtain relevant medical imaging reports and the categories of examination items contained therein, and number each examination item category starting from 0 as the corresponding category label.
[0043] Preferably, in S03, the medical imaging report data set corresponding to each category label is processed by sentence segmentation to obtain the examination finding sentence data set and the examination impression sentence data set corresponding to each category label, including:
[0044] The medical imaging report dataset corresponding to each category label is processed by sentence processing according to the preset text end symbol, the corresponding category label is set for each sentence, and the sentences corresponding to the examination findings and examination impressions are distinguished, so as to obtain the examination findings sentence dataset and the examination impression sentence dataset corresponding to each category label respectively.
[0045] Preferably, in S06, the inspection observations sentence history data set is used to perform model training on the inspection observations classification model, including:
[0046] S061, dividing the inspection sentence history data set into a training set, a verification set and a test set according to a preset ratio;
[0047] S062. Setting the loss function and optimizer of the classification model observed by the inspection;
[0048] S063, inputting the training set into the classification model observed by the inspection to perform model training;
[0049] S064. Calculate the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;
[0050] S065. If the loss value is less than the preset threshold, the model training ends, and the current inspection classification model is the trained inspection classification model. Otherwise, the process returns to S063 and continues to use the training set for model training.
[0051] S066. Input the validation set into the trained inspection observation classification model, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model;
[0052] S067, input the test set into the optimized inspection classification model to evaluate the model performance;
[0053] Among them, considering the multi-label classification task, the cross entropy loss function BCEWithLogitsLoss containing the Sigmoid function is used as the loss function of the classification model seen in the inspection:
[0054]
[0055] Among them, y i is the category label value of the sentence seen in the i-th inspection, y i ∈{0,1},p i is the probability prediction value of the sentence seen in the i-th inspection, N is the number of sentences seen in the inspection, σ(x) is the Sigmoid function,
[0056] Preferably, in S07, the inspection impression classification model is trained using the inspection impression sentence history data set, including:
[0057] S071. Divide the inspection impression sentence history data set into a training set, a validation set, and a test set according to a preset ratio;
[0058] S072. Setting a loss function and an optimizer for the inspection impression classification model;
[0059] S073, inputting the training set into the inspection impression classification model for model training;
[0060] S074. Calculate the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;
[0061] S075. If the loss value is less than the preset threshold, the model training ends, and the current inspection impression classification model is the trained inspection impression classification model. Otherwise, the process returns to S073 and continues to use the training set for model training.
[0062] S076. Input the validation set into the trained inspection impression classification model, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model;
[0063] S077, input the test set into the optimized inspection impression classification model to evaluate the model performance;
[0064] Among them, considering the multi-label classification task, the loss function of the inspection impression classification model also adopts the cross entropy loss function BCEWithLogitsLoss containing the Sigmoid function.
[0065] (III) Beneficial effects
[0066] Compared with the prior art, the medical imaging report splitting method for large model training provided by the present invention can accurately split the examination findings and examination impressions in the medical imaging report according to the examination item category, so that the medical imaging large model to be trained can make full use of the text report information in the medical imaging data, and conduct targeted learning on the examination findings and examination impressions in the medical imaging report, thereby ensuring the training effect of the medical imaging large model. At the same time, it can effectively reduce the data annotation cost and obtain a large amount of training data that meets the requirements in a short time. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0068] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0069] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.
[0070] A medical imaging report segmentation method for large model training, such as Figure 1 As shown, S1, read the medical imaging report to be split, and perform sentence processing on the medical imaging report to obtain a sentence combination of examination findings and a sentence combination of examination impressions, which specifically include:
[0071] S11, reading the medical imaging report to be split;
[0072] S12. Segment the medical imaging report according to a preset text end symbol, and distinguish the sentences corresponding to the examination findings and the examination impressions, to obtain an examination findings sentence combination and an examination impression sentence combination, respectively.
[0073] S2. Extract the word vector and mask vector corresponding to each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination, specifically including:
[0074] S21, counting the maximum number of characters N contained in all the sentences in the inspection-seen sentence combinations and the inspection-impression sentence combinations;
[0075] S22, initialize the BERT word segmenter and set the maximum length of the sentence to the maximum number of characters N;
[0076] S23. Input each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination into the BERT word segmenter respectively to obtain the corresponding word vector and mask vector.
[0077] S3. Input the word vector and mask vector corresponding to each inspection observation clause in the inspection observation clause combination into the pre-trained inspection observation classification model to obtain the inspection observation classification result corresponding to each inspection observation clause, specifically including:
[0078] S31, inputting the word vector and mask vector corresponding to the inspection-what-you-see clause in the inspection-what-you-see clause combination into the pre-trained inspection-what-you-see classification model, and the inspection-what-you-see classification model outputs the probability prediction value of the inspection-what-you-see clause belonging to each category label;
[0079] S32, selecting the category label with the maximum probability prediction value as the candidate classification result, and determining the relationship between the maximum probability prediction value and a preset threshold value (0.5);
[0080] S33, if the maximum probability prediction value is greater than a preset threshold, the candidate classification result is used as the inspection finding classification result corresponding to the inspection finding sentence, otherwise the corresponding inspection finding classification result is determined according to the inspection finding sentence content;
[0081] S34, repeat S31 to S33 until all inspection observation clauses in the inspection observation clause combination obtain corresponding inspection observation classification results.
[0082] S4. Input the word vector and mask vector corresponding to each inspection impression sentence in the inspection impression sentence combination into the pre-trained inspection impression classification model to obtain the inspection impression classification result corresponding to each inspection impression sentence, specifically including:
[0083] S41, inputting the word vector and mask vector corresponding to the inspection impression sentence in the inspection impression sentence combination into the pre-trained inspection impression classification model, and the inspection impression classification model outputs the probability prediction value of the inspection impression sentence belonging to each category label;
[0084] S42, selecting the category label with the maximum probability prediction value as the candidate classification result, and determining the relationship between the maximum probability prediction value and a preset threshold value (0.5);
[0085] S43, if the maximum probability prediction value is greater than a preset threshold, the candidate classification result is used as the inspection impression classification result corresponding to the inspection impression sentence, otherwise the corresponding inspection impression classification result is determined according to the inspection impression sentence content;
[0086] S44. Repeat S41 to S43 until all inspection impression sentences in the inspection impression sentence combination obtain corresponding inspection impression classification results.
[0087] S5. The inspection finding sentences with the same inspection finding classification results are merged and output, and the inspection impression sentences with the same inspection impression classification results are merged and output, so as to complete the splitting of the medical imaging report.
[0088] In the technical solution of the present application, the inspection observation classification model and the inspection impression classification model are both built based on the RoBERTa language model.
[0089] In the technical solution of the present application, S1 reads the medical imaging report to be split, and performs sentence processing on the medical imaging report to obtain the combination of examination findings and examination impressions, including:
[0090] S01. Obtain relevant medical imaging reports and the categories of examination items included therein, and set category labels according to the categories of examination items;
[0091] S02. Classify the medical imaging reports according to the categories of the examination items to obtain a medical imaging report data set corresponding to each category label;
[0092] S03, performing sentence processing on the medical imaging report data set corresponding to each category label to obtain an examination finding sentence data set and an examination impression sentence data set corresponding to each category label;
[0093] S04, merging the inspection observation sentence data sets corresponding to each category label into an inspection observation sentence set, and merging the inspection impression sentence data sets corresponding to each category label into an inspection impression sentence set;
[0094] S05, extracting the word vector and mask vector corresponding to each inspection-seen sentence in the inspection-seen sentence set and each inspection-impression sentence in the inspection-impression sentence set, and completing the construction of the inspection-seen sentence historical data set and the inspection-impression sentence historical data set;
[0095] S06, using the historical data set of inspection observations, the inspection observations classification model is trained;
[0096] S07. Use the inspection impression sentence historical data set to train the inspection impression classification model.
[0097] 1) In S01, relevant medical imaging reports and the categories of examination items included are obtained, and category labels are set according to the categories of examination items, including:
[0098] Obtain relevant medical imaging reports and the categories of examination items contained therein, and number each examination item category starting from 0 as the corresponding category label.
[0099] 2) In S03, the medical imaging report dataset corresponding to each category label is processed by sentence segmentation to obtain the inspection finding sentence dataset and the inspection impression sentence dataset corresponding to each category label, including:
[0100] The medical imaging report dataset corresponding to each category label is processed by sentence processing according to the preset text end symbol, the corresponding category label is set for each sentence, and the sentences corresponding to the examination findings and examination impressions are distinguished, so as to obtain the examination findings sentence dataset and the examination impression sentence dataset corresponding to each category label respectively.
[0101] 3) In S06, the inspection observations sentence history data set is used to train the inspection observations classification model, including:
[0102] S061, dividing the inspection sentence history data set into a training set, a validation set and a test set according to a preset ratio (7:2:1);
[0103] S062. Setting the loss function and optimizer of the classification model observed by the inspection;
[0104] S063, inputting the training set into the classification model observed by the inspection to perform model training;
[0105] S064. Calculate the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;
[0106] S065. If the loss value is less than the preset threshold, the model training ends, and the current inspection classification model is the trained inspection classification model. Otherwise, the process returns to S063 and continues to use the training set for model training.
[0107] S066. Input the validation set into the trained inspection observation classification model, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model;
[0108] S067, input the test set into the optimized inspection classification model to evaluate the model performance;
[0109] Among them, considering the multi-label classification task, the cross entropy loss function BCEWithLogitsLoss containing the Sigmoid function is used as the loss function of the classification model seen in the inspection:
[0110]
[0111] Among them, y i is the category label value of the sentence seen in the i-th inspection, y i ∈{0,1},p i is the probability prediction value of the sentence seen in the i-th inspection, N is the number of sentences seen in the inspection, σ(x) is the Sigmoid function,
[0112] 4) In S07, the inspection impression classification model is trained using the inspection impression sentence history data set, including:
[0113] S071, dividing the inspection impression sentence history data set into a training set, a validation set, and a test set according to a preset ratio (7:2:1);
[0114] S072. Setting a loss function and an optimizer for the inspection impression classification model;
[0115] S073, inputting the training set into the inspection impression classification model for model training;
[0116] S074. Calculate the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information;
[0117] S075. If the loss value is less than the preset threshold, the model training ends, and the current inspection impression classification model is the trained inspection impression classification model. Otherwise, the process returns to S073 and continues to use the training set for model training.
[0118] S076. Input the validation set into the trained inspection impression classification model, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model;
[0119] S077, input the test set into the optimized inspection impression classification model to evaluate the model performance;
[0120] Among them, considering the multi-label classification task, the loss function of the inspection impression classification model also adopts the cross entropy loss function BCEWithLogitsLoss containing the Sigmoid function.
[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A medical image report splitting method for large model training, characterized by: The following steps are involved: S1. Read the medical imaging report to be split, and perform sentence processing on the medical imaging report to obtain a sentence combination of examination findings and a sentence combination of examination impressions; S2, extracting the word vector and mask vector corresponding to each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination; S3, inputting the word vector and mask vector corresponding to each inspection observation clause in the inspection observation clause combination into the pre-trained inspection observation classification model to obtain the inspection observation classification result corresponding to each inspection observation clause; S4, inputting the word vector and mask vector corresponding to each inspection impression sentence in the inspection impression sentence combination into the pre-trained inspection impression classification model to obtain the inspection impression classification result corresponding to each inspection impression sentence; S5, merging and outputting the inspection finding sentences with the same inspection finding classification results, and merging and outputting the inspection impression sentences with the same inspection impression classification results, thereby completing the medical imaging report splitting; Among them, the inspection observation classification model and the inspection impression classification model are both built based on the RoBERTa language model.
2. The medical image report splitting method for large model training according to claim 1 is characterized in that: S1 reads the medical imaging report to be split, and performs sentence processing on the medical imaging report to obtain a sentence combination of examination findings and a sentence combination of examination impressions, including: S11, reading the medical imaging report to be split; S12. Segment the medical imaging report according to a preset text end symbol, and distinguish the sentences corresponding to the examination findings and the examination impressions, to obtain an examination findings sentence combination and an examination impression sentence combination, respectively.
3. The medical image report splitting method for large model training according to claim 2 is characterized in that: S2 extracts the word vectors and mask vectors corresponding to each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination, including: S21, counting the maximum number of characters N contained in all the sentences in the inspection-seen sentence combinations and the inspection-impression sentence combinations; S22, initialize the BERT word segmenter and set the maximum length of the sentence to the maximum number of characters N; S23. Input each inspection observation sentence in the inspection observation sentence combination and each inspection impression sentence in the inspection impression sentence combination into the BERT word segmenter respectively to obtain the corresponding word vector and mask vector.
4. The medical image report splitting method for large model training according to claim 3 is characterized in that: In S3, the word vector and mask vector corresponding to each inspection observation clause in the inspection observation clause combination are input into the pre-trained inspection observation classification model to obtain the inspection observation classification results corresponding to each inspection observation clause, including: S31, inputting the word vector and mask vector corresponding to the inspection-what-you-see clause in the inspection-what-you-see clause combination into the pre-trained inspection-what-you-see classification model, and the inspection-what-you-see classification model outputs the probability prediction value of the inspection-what-you-see clause belonging to each category label; S32, selecting the category label with the maximum probability prediction value as the candidate classification result, and determining the relationship between the maximum probability prediction value and a preset threshold; S33, if the maximum probability prediction value is greater than a preset threshold, the candidate classification result is used as the inspection finding classification result corresponding to the inspection finding sentence, otherwise the corresponding inspection finding classification result is determined according to the inspection finding sentence content; S34, repeat S31 to S33 until all inspection observation clauses in the inspection observation clause combination obtain corresponding inspection observation classification results.
5. The medical image report splitting method for large model training according to claim 3 is characterized in that: In S4, the word vector and mask vector corresponding to each inspection impression sentence in the inspection impression sentence combination are input into the pre-trained inspection impression classification model to obtain the inspection impression classification results corresponding to each inspection impression sentence, including: S41, inputting the word vector and mask vector corresponding to the inspection impression sentence in the inspection impression sentence combination into the pre-trained inspection impression classification model, and the inspection impression classification model outputs the probability prediction value of the inspection impression sentence belonging to each category label; S42, selecting the category label with the maximum probability prediction value as the candidate classification result, and determining the relationship between the maximum probability prediction value and a preset threshold; S43, if the maximum probability prediction value is greater than a preset threshold, the candidate classification result is used as the inspection impression classification result corresponding to the inspection impression sentence, otherwise the corresponding inspection impression classification result is determined according to the inspection impression sentence content; S44. Repeat S41 to S43 until all inspection impression sentences in the inspection impression sentence combination obtain corresponding inspection impression classification results.
6. The medical image report splitting method for large model training according to claim 3 is characterized in that: S1 reads the medical imaging report to be split, and performs sentence processing on the medical imaging report to obtain the combination of examination findings and examination impressions, including: S01. Obtain relevant medical imaging reports and the categories of examination items included therein, and set category labels according to the categories of examination items; S02. Classify the medical imaging reports according to the categories of the examination items to obtain a medical imaging report data set corresponding to each category label; S03, performing sentence processing on the medical imaging report data set corresponding to each category label to obtain an examination finding sentence data set and an examination impression sentence data set corresponding to each category label; S04, merging the inspection observation sentence data sets corresponding to each category label into an inspection observation sentence set, and merging the inspection impression sentence data sets corresponding to each category label into an inspection impression sentence set; S05, extracting the word vector and mask vector corresponding to each inspection-seen sentence in the inspection-seen sentence set and each inspection-impression sentence in the inspection-impression sentence set, and completing the construction of the inspection-seen sentence historical data set and the inspection-impression sentence historical data set; S06, using the historical data set of inspection observations, the inspection observations classification model is trained; S07. Use the inspection impression sentence historical data set to train the inspection impression classification model.
7. The medical image report splitting method for large model training according to claim 6 is characterized in that: S01 obtains relevant medical imaging reports and the categories of examination items included, and sets category labels according to the categories of examination items, including: Obtain relevant medical imaging reports and the categories of examination items contained therein, and number each examination item category starting from 0 as the corresponding category label.
8. The medical image report splitting method for large model training according to claim 6 is characterized in that: In S03, the medical imaging report dataset corresponding to each category label is processed by sentence segmentation to obtain the inspection finding sentence dataset and the inspection impression sentence dataset corresponding to each category label, including: The medical imaging report dataset corresponding to each category label is processed by sentence processing according to the preset text end symbol, the corresponding category label is set for each sentence, and the sentences corresponding to the examination findings and examination impressions are distinguished, so as to obtain the examination findings sentence dataset and the examination impression sentence dataset corresponding to each category label respectively.
9. The medical image report splitting method for large model training according to claim 6, characterized in that: In S06, the inspection observations sentence history data set is used to train the inspection observations classification model, including: S061, dividing the inspection sentence history data set into a training set, a verification set and a test set according to a preset ratio; S062. Setting the loss function and optimizer of the classification model observed by the inspection; S063, inputting the training set into the classification model observed by the inspection to perform model training; S064. Calculate the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information; S065. If the loss value is less than the preset threshold, the model training ends, and the current inspection classification model is the trained inspection classification model. Otherwise, the process returns to S063 and continues to use the training set for model training. S066. Input the validation set into the trained inspection observation classification model, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model; S067, input the test set into the optimized inspection classification model to evaluate the model performance; Among them, considering the multi-label classification task, the cross entropy loss function BCEWithLogitsLoss containing the Sigmoid function is used as the loss function of the classification model seen in the inspection: Among them, y i is the category label value of the sentence seen in the i-th inspection, y i ∈{0,1},p i is the probability prediction value of the sentence seen in the i-th inspection, N is the number of sentences seen in the inspection, σ(x) is the Sigmoid function, 10. The medical image report splitting method for large model training according to claim 9, characterized in that: In S07, the inspection impression classification model is trained using the inspection impression sentence history dataset, including: S071. Divide the inspection impression sentence history data set into a training set, a validation set, and a test set according to a preset ratio; S072. Setting a loss function and an optimizer for the inspection impression classification model; S073, inputting the training set into the inspection impression classification model for model training; S074. Calculate the loss value based on the loss function, and the optimizer updates the model parameters according to the loss value and the network gradient information; S075. If the loss value is less than the preset threshold, the model training ends, and the current inspection impression classification model is the trained inspection impression classification model. Otherwise, the process returns to S073 and continues to use the training set for model training. S076. Input the validation set into the trained inspection impression classification model, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model; S077, input the test set into the optimized inspection impression classification model to evaluate the model performance; Among them, considering the multi-label classification task, the loss function of the inspection impression classification model also adopts the cross entropy loss function BCEWithLogitsLoss containing the Sigmoid function.
Citation Information
Patent Citations
Medical image report information extraction method and device, equipment and storage medium
CN112712879A
Medical image diagnosis method and device based on film reading knowledge graph
CN115062165A
Medical image report recommendation method and device, equipment and storage medium
CN118506959A
Medical image project management platform
US20230197268A1