Case type prediction method and device, equipment and storage medium

Abstract

The invention relates to a case type prediction method and device, equipment and a storage medium. The method comprises the following steps: acquiring medical examination information of a target part; inputting the medical examination information into a pre-trained type prediction model to obtain a case type prediction result of the target part; according to the type prediction model, a plurality of historical case types extracted from case collection information of a target part and case type detailed information corresponding to each historical case type are adopted to carry out secondary pre-training on a pre-trained generation model; and performing fine adjustment on the generation model after secondary pre-training by adopting each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information. By adopting the method, not only can common case types be predicted, but also many rare case types can be predicted, and the scene applicability of case type prediction is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a case type prediction method, device, equipment and storage medium. Background Art

[0002] Usually, when doctors diagnose a disease, they need to analyze the case type of the patient's examination area, and then comprehensively determine the patient's disease diagnosis results based on the case type.

[0003] In the related art, when analyzing the case type of the examined part, a classification model is usually used for prediction, and the classification model is trained based on the categories of different case types in various parts of the body. However, since the case type samples of many parts are relatively scarce, the labeling work is very difficult in actual operation, resulting in the classification model being able to predict only common case types.

[0004] Therefore, the case type prediction method in the related art has the problem of poor scenario applicability. Summary of the invention

[0005] Based on this, it is necessary to provide a case type prediction method, device, equipment and storage medium to address the above technical problems, which can not only predict common case types, but also many rare case types, thereby improving the scenario applicability of case type prediction.

[0006] In a first aspect, an embodiment of the present application provides a case type prediction method, comprising:

[0007] Obtain medical examination information of the target area;

[0008] The medical examination information is input into a pre-trained type prediction model to obtain a case type prediction result for the target site; the type prediction model is obtained by secondary pre-training a pre-trained generation model using multiple historical case types extracted from the case collection information of the target site and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information.

[0009] In one embodiment, the method further comprises:

[0010] According to the target part, case information related to the target part is retrieved from the case database;

[0011] The case information related to the target site is determined as the case collection information of the target site.

[0012] In one embodiment, a pre-trained generative model is pre-trained twice using a plurality of historical case types extracted from case collection information of a target site and case type detailed information corresponding to each historical case type, including:

[0013] Input each historical case type into the pre-trained generative model to obtain the prediction detailed information output by the pre-trained generative model;

[0014] According to the detailed information of each case type, the prediction detailed information and the preset loss function, the model parameters of the pre-trained generative model are adjusted until the training is completed to obtain the generative model after secondary pre-training.

[0015] In one embodiment, each historical case type and the historical examination information corresponding to each historical case type extracted from the case collection information are used to fine-tune the generative model after the secondary pre-training, including:

[0016] Input each historical inspection information into the generative model after secondary pre-training to obtain the prediction result output by the generative model after secondary pre-training;

[0017] According to the prediction results and the types of historical cases, the type prediction loss of the generative model after secondary pre-training is determined;

[0018] Based on the type prediction loss, the model parameters of the generative model after secondary pre-training are adjusted until the type prediction loss is less than the preset threshold to obtain the type prediction model.

[0019] In one embodiment, the method further comprises:

[0020] According to the pre-trained reward model and sample inspection information, the output result of the type prediction model is scored to obtain the quality score result of the sample inspection information;

[0021] Based on the quality score results, the type prediction model parameters are optimized to obtain an optimized type prediction model.

[0022] In one embodiment, the quality score of the output result of the type prediction model is performed according to the pre-trained reward model and the sample inspection information to obtain the quality score result of the sample inspection information, including:

[0023] Inputting sample inspection information into the type prediction model to obtain sample prediction results output by the type prediction model;

[0024] The sample prediction results are input into the reward model to obtain the quality score results.

[0025] In one embodiment, the training process of the reward model includes:

[0026] Obtain a training data set; the training data set is a plurality of type prediction results generated by the type prediction model for each inspection information;

[0027] Input each type of prediction result into the quality assessment model to obtain the prediction quality ranking result output by the quality assessment model;

[0028] According to the predicted quality ranking results and the standard quality ranking results, the model parameters of the quality assessment model are adjusted until the training is completed to obtain the reward model.

[0029] In a second aspect, the embodiment of the present application further provides a case type prediction device, comprising:

[0030] An information acquisition module, used to acquire medical examination information of the target part;

[0031] The result determination module is used to input the medical examination information into a pre-trained type prediction model to obtain the case type prediction result of the target part; the type prediction model is obtained by secondary pre-training the pre-trained generation model using multiple historical case types extracted from the case collection information of the target part and the case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and the historical examination information corresponding to each historical case type extracted from the case collection information.

[0032] In a third aspect, the present application also provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in any one of the embodiments of the first aspect when executing the computer program.

[0033] In a fourth aspect, the embodiments of the present application further provide a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps in any one of the embodiments of the first aspect are implemented.

[0034] In a fifth aspect, the embodiments of the present application further provide a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps in any one of the embodiments in the first aspect are implemented.

[0035] The above-mentioned case type prediction method, device, equipment and storage medium obtain medical examination information of the target part, and then input the medical examination information into a pre-trained type prediction model to obtain a case type prediction result of the target part. The type prediction model is obtained by secondary pre-training the pre-trained generation model using multiple historical case types extracted from the case collection information of the target part and the case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and the historical examination information corresponding to each historical case type extracted from the case collection information. In this method, a pre-trained generative model is pre-trained twice by using a variety of historical case types and corresponding case type detailed information. In this way, the generative model can learn a large amount of medical knowledge, that is, it has knowledge in the medical field, so that the generative model can better handle various tasks in the medical field; further, a variety of historical case types and corresponding historical examination information are used to fine-tune the generative model after the second pre-training, so that the generative model adapts to new data sets and tasks, and the performance of the generative model on specific tasks is improved by training the new data sets, that is, the adaptability and prediction performance of the generative model in specific tasks such as case type prediction are improved. In this way, the type prediction model has learned the examination information and disease information corresponding to various cases. The relationship between case types, and then the type prediction model can infer the case type corresponding to the medical examination information based on any medical examination information, which is equivalent to inputting any medical examination information into the type prediction model. The type prediction model can infer the case type corresponding to the input medical examination information based on the learned relationship between the medical examination information and the case type. Because the relationship between the medical examination information and the case type has been learned, it has a strong logical inference ability. Therefore, there is no need to mark different case types in advance, and various case types can be predicted. Not only common case types can be predicted, but also many rare case types can be predicted, which improves the scenario applicability of case type prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] FIG1 is a diagram showing the internal structure of a computer device in one embodiment;

[0038] FIG2 is a schematic diagram of a flow chart of a case type prediction method in one embodiment;

[0039] FIG3 is a schematic diagram of a process for obtaining case collection information in one embodiment;

[0040] FIG4 is a schematic diagram of a process of performing secondary pre-training on a generative model in one embodiment;

[0041] FIG5 is a schematic diagram of a process of fine-tuning a generative model after secondary pre-training in one embodiment;

[0042] FIG6 is a schematic diagram of a process for optimizing a type prediction model in one embodiment;

[0043] FIG7 is a schematic diagram of a process for determining a quality score result in one embodiment;

[0044] FIG8 is a schematic diagram of a process of training a reward model in one embodiment;

[0045] FIG9 is a schematic flow chart of a case type prediction method in another embodiment;

[0046] FIG10 is a schematic diagram of a process for case type prediction in one embodiment;

[0047] FIG. 11 is a schematic diagram of the structure of a case type prediction device in one embodiment. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0049] The case type prediction method provided in the embodiment of the present application can be applied to a computer device. The computer device can be a server, and its internal structure diagram can be shown in Figure 1. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a case type prediction method is implemented. Those skilled in the art will understand that the structure shown in FIG. 1 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0050] Usually, when doctors diagnose a disease, they need to analyze the case type of the patient's examination area, and then comprehensively determine the patient's disease diagnosis results based on the case type.

[0051] Taking the cardiothoracic and pulmonary diseases as an example, there are many types of cardiothoracic and pulmonary disease cases, such as cancer, tumors, inflammation, nodules, fungal and bacterial infections, etc. When doctors diagnose a specific condition, they can first analyze the patient's cardiothoracic and pulmonary disease case type, and then comprehensively determine the final disease diagnosis result based on the case type and the patient's various examination information.

[0052] In the related art, when analyzing the case type of the examined part, a classification model is usually used for prediction, and the classification model is trained based on the categories of different case types in various parts of the body. However, since the case type samples of many parts are relatively scarce, the labeling work is very difficult in actual operation, resulting in the classification model being able to predict only common case types.

[0053] Therefore, the case type prediction method in the related art has the problem of poor scenario applicability.

[0054] Based on this, the present application proposes a case type prediction method, which performs secondary pre-training on a pre-trained generative model by using a variety of historical case types and corresponding case type detailed information, so that the generative model can learn a large amount of medical knowledge, that is, knowledge in the medical field, so that the generative model can better handle various tasks in the medical field; further, a variety of historical case types and corresponding historical examination information are used to fine-tune the generative model after the secondary pre-training, so that the generative model adapts to new data sets and tasks, and improves the performance of the generative model on specific tasks by training the new data sets, that is, improves the adaptability and predictive performance of the generative model in specific tasks such as case type prediction, so that the type prediction model has learned various case pairs. The relationship between the corresponding examination information and the case type can be learned, and then the type prediction model can infer the case type corresponding to any medical examination information based on the medical examination information, which is equivalent to inputting any medical examination information into the type prediction model. The type prediction model can infer the case type corresponding to the input medical examination information based on the learned relationship between the medical examination information and the case type. Because the relationship between the medical examination information and the case type is learned, it has strong logical inference ability. Therefore, there is no need to mark different case types in advance, and various case types can be predicted. Not only common case types can be predicted, but also many rare case types can be predicted, which improves the scenario applicability of case type prediction.

[0055] It should be noted that the beneficial effects brought about by the embodiments of the present application or the technical problems solved are not limited to this one, but may also include other implicit or related problems. For details, please refer to the description of the following embodiments.

[0056] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0057] In an exemplary embodiment, as shown in FIG2 , a case type prediction method is provided, which is described by taking the method applied to the computer device in FIG1 as an example, and includes the following steps 201 to 202. Among them:

[0058] S201, obtaining medical examination information of a target part.

[0059] The target part refers to the patient's examination part, such as the heart, lungs, liver, etc. Medical examination information may include the patient's basic information, medical images, image data, other examination results, etc. Other examination results refer to the examination results before the final diagnosis result is obtained, for example, the examination results given by the radiologist when the medical image is taken.

[0060] In one embodiment, obtaining medical examination information of the target part includes: obtaining basic information of the object to which the target part belongs; and obtaining the medical examination information of the target part from an examination storage repository according to the basic information.

[0061] The patient's medical examination information is stored in the hospital's examination storage. When the medical examination information is needed, it can be obtained from the examination storage repository.

[0062] It is understandable that the examination repository stores medical examination information of multiple patients. Therefore, when obtaining the medical examination information of the target part, it is necessary to first determine the object to which the target part belongs, and then find the medical examination information corresponding to the basic information from the examination repository based on the basic information of the object. The basic information of the object may include the patient's name, ID number, medical card number, etc.

[0063] In another embodiment, medical examination information can also be obtained from the hospital business management system. Various examinations and personal information of patients in the hospital are stored in the hospital business management system. For example, the patient's image data is stored in the medical image archiving and communication system (PACS), and the patient's basic information is stored in the hospital information system (HIS).

[0064] For example, when medical examination information needs to be obtained, corresponding data can be obtained from various hospital business management systems, for example, requesting the image data of the target part from the server where the PACS system is located, and requesting the basic information of the target patient from the server where the HIS system is located.

[0065] S202, inputting the medical examination information into a pre-trained type prediction model to obtain a case type prediction result of the target site.

[0066] In an embodiment of the present application, the type prediction model is obtained by secondary pre-training a pre-trained generation model using multiple historical case types extracted from the case collection information of the target site and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information.

[0067] Among them, the pre-trained generative model can be a pre-trained large language model, etc. The case collection information of the target part refers to the case information of multiple historical cases related to the target part. In the embodiment of the present application, the case collection information of the target part includes information on common cases and rare cases of the target part collected from various channels. Among them, the various channels may include books, hospital networks, etc., among which the hospital network may include but is not limited to websites such as Baidu Library, Renren Library, and Yimaitong.

[0068] Optionally, before training the type prediction model, first obtain the case collection information of the target part, collect various common cases and rare cases of the target part from multiple channels, and obtain the case information corresponding to various common cases and various rare cases, then perform data cleaning on the case information, remove low-quality and repeated case information, and obtain the case collection information. Among them, the case information includes the medical examination information, case type and case type detailed information corresponding to each case. In this way, the case collection information of the target part includes multiple historical case types of the target part, case type detailed information corresponding to each historical case type, and corresponding historical examination information.

[0069] After obtaining the case collection information of the target part, the training data for secondary pre-training (multiple historical case types and detailed information on case types corresponding to each historical case type) and fine-tuning data for fine-tuning (each historical case type and historical examination information corresponding to each historical case type) are extracted from the case collection information. Then, the pre-trained generative model is secondary pre-trained using the multiple historical case types and detailed information on case types corresponding to each historical case type. Furthermore, the generative model after the secondary pre-training is fine-tuned using each historical case type and historical examination information corresponding to each historical case type to obtain a type prediction model.

[0070] Among them, the case type detailed information refers to the professional detailed introduction information of the case type. For example, if the historical case type is nodule, the corresponding case type detailed information is that the nodule is a palpable localized substantial lesion; the formation of nodules can be inflammatory or non-inflammatory, and can involve the epidermis, dermis or subcutaneous tissue. For example, if the historical case type is pulmonary nodule, the corresponding case type detailed information is that the pulmonary nodule is a focal, rounded, solid or subsolid lung shadow with a diameter of ≤3cm, which can be isolated or multiple, without atelectasis, hilar lymphadenopathy and pleural effusion. Among them, the case type detailed information of the historical case type can be obtained from books, hospital networks, etc. In addition, in the embodiments of the present application, the scope of the case type is not limited. It can be a large range of types, such as cancer, tumors, inflammation, nodules, fungal and bacterial infections, etc., or a small range of types, such as sclerosing pneumocystis carcinoma, malignant pleural mesothelioma, pulmonary nodules, myocarditis, etc. The historical examination information corresponding to each historical case type refers to the relevant examination information of the historical patients corresponding to each historical case type, which may include basic information, medical images, imaging data, other examination results, etc. of the historical patients.

[0071] In this embodiment, there is a certain relationship between the case type detailed information and the historical examination information. The case type detailed information includes the imaging manifestations of the medical images in the historical examination information, that is, the case type detailed information in the embodiment of the present application includes not only the professional medical description information of the case type in the above example, but also the imaging manifestations. For example, when the medical image included in the historical examination information is an X-ray image, the case type detailed information can be expressed as the specific manifestation of a certain type of disease in the X-ray image. For example, when the case type detailed information is a certain type of disease, the grayscale value of the X-ray image increases (decreases), etc. In the secondary pre-training process, the pre-trained generative model can obtain the imaging manifestations of the case type by learning the case type detailed information, paving the way for the subsequent fine-tuning process, thereby improving the accuracy of case type prediction.

[0072] Usually, the pre-trained large language model is obtained by training the large language model through various types of texts such as web page data, books, news, scientific articles, etc. In the embodiment of the present application, in order to apply the pre-trained large language model to medicine, it is necessary to train the pre-trained generative model to learn medical knowledge, and then train the generative model after learning medical knowledge to predict case types. Based on this, in the embodiment of the present application, multiple historical case types can be used as inputs of the pre-trained generative model, and the case type detailed information corresponding to each historical case type can be used as the prediction standard of the pre-trained generative model, and the pre-trained generative model is pre-trained twice to enable the pre-trained generative model to learn medical knowledge.

[0073] Furthermore, the historical examination information corresponding to each historical case type is used as the input of the generative model after the second pre-training, and each historical case type is used as the prediction standard of the generative model after the second pre-training. The generative model after the second pre-training is fine-tuned so that the generative model after the second pre-training has the ability to logically infer between the historical medical examination information and the historical case types. Once the training is completed, the type prediction model can be obtained.

[0074] It can be seen that in the embodiment of the present application, when training the type prediction model, different case types are not labeled in advance, but the inference ability of the generative model is used to learn the relationship between medical examination information and case types. In this way, the type prediction model can infer the case type corresponding to any medical examination information based on the learned relationship between medical examination information and case types.

[0075] Exemplarily, after obtaining a trained type prediction model, the medical examination information of the target site is input into the type prediction model, and the type prediction model can directly output the case type prediction result of the target site.

[0076] In the case type prediction method provided in the embodiment of the present application, medical examination information of the target part is obtained, and then the medical examination information is input into a pre-trained type prediction model to obtain a case type prediction result of the target part. The type prediction model is obtained by secondary pre-training the pre-trained generation model using multiple historical case types extracted from the case collection information of the target part and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information. In this method, a pre-trained generative model is pre-trained twice by using a variety of historical case types and corresponding case type detailed information. In this way, the generative model can learn a large amount of medical knowledge, that is, it has knowledge in the medical field, so that the generative model can better handle various tasks in the medical field; further, a variety of historical case types and corresponding historical examination information are used to fine-tune the generative model after the second pre-training, so that the generative model adapts to new data sets and tasks, and the performance of the generative model on specific tasks is improved by training the new data sets, that is, the adaptability and prediction performance of the generative model in specific tasks such as case type prediction are improved. In this way, the type prediction model has learned the examination information and disease information corresponding to various cases. The relationship between case types, and then the type prediction model can infer the case type corresponding to the medical examination information based on any medical examination information, which is equivalent to inputting any medical examination information into the type prediction model. The type prediction model can infer the case type corresponding to the input medical examination information based on the learned relationship between the medical examination information and the case type. Because the relationship between the medical examination information and the case type has been learned, it has a strong logical inference ability. Therefore, there is no need to mark different case types in advance, and various case types can be predicted. Not only common case types can be predicted, but also many rare case types can be predicted, which improves the scenario applicability of case type prediction.

[0077] Based on the above embodiment, an embodiment is provided to illustrate the process of obtaining the case collection information of the above target part.

[0078] In an exemplary embodiment, as shown in FIG3 , the method further includes:

[0079] S301, according to the target part, case information related to the target part is retrieved from a case database.

[0080] In the embodiment of the present application, the case database is a database storing case information of various parts, wherein the case database includes case information of common cases and rare cases of various parts.

[0081] Exemplarily, according to the target part, retrieving case information related to the target part from the case database may be to obtain part identification information of the target part, and then to retrieve from the case database according to the part identification information to obtain case information related to the target part.

[0082] For example, case information of different parts in the case database is stored with identification information of each part, such as each part is identified starting with the number 0, and the identification information of each part is associated with the case information of each part and stored in the case database. When obtaining case collection information of the target part, the digital identification information of the target part is first obtained, and then the case information associated with the digital identification information is matched from the case database according to the digital identification information of the target part, that is, the case information related to the target part.

[0083] S302: Determine the case information related to the target part as the case collection information of the target part.

[0084] After the case information related to the target part is retrieved, the case information related to the target part is used as case collection information of the target part.

[0085] In the case type prediction method provided in the embodiment of the present application, case information related to the target part is retrieved from the case database according to the target part, and then the case information related to the target part is determined as the case collection information of the target part. In this method, an optional way to obtain the case collection information of the target part is provided; by introducing a case database, which includes case information of multiple parts, case information related to the target part can be quickly obtained based on the target part. The case information related to the target part is the case collection information of the target part, which provides data support for the training of the type prediction model.

[0086] In order to make the trained type prediction model more suitable for medical prediction tasks, when training the type prediction model, the generated model can be first trained for medical knowledge learning, and then case type inference training can be performed. Based on this, the following embodiment describes the process of secondary pre-training.

[0087] In an exemplary embodiment, as shown in FIG4 , a pre-trained generative model is pre-trained twice using a plurality of historical case types extracted from case collection information of a target site and case type detailed information corresponding to each historical case type, including:

[0088] S401, input each historical case type into the pre-trained generation model to obtain the prediction detailed information output by the pre-trained generation model.

[0089] The prediction details refer to the detailed information corresponding to the historical case type predicted by the pre-trained generative model.

[0090] Each historical case type is input into the pre-trained generative model, and the pre-trained generative model can output detailed information corresponding to the predicted historical case type, that is, the predicted detailed information.

[0091] S402, determining the detailed information prediction loss of the pre-trained generation model according to the detailed information of each case type, the prediction detailed information and a preset loss function.

[0092] S403, based on the detailed information prediction loss, adjusting the model parameters of the pre-trained generation model until the detailed information prediction loss is less than a preset threshold, thereby obtaining a generation model after secondary pre-training.

[0093] The loss value can be calculated based on the prediction details output by the pre-trained generative model and the case type details, and then the model parameters of the pre-trained generative model are continuously adjusted based on the loss value to obtain the generative model after secondary pre-training.

[0094] Exemplarily, according to a preset loss function, the gap between the predicted detailed information and the case type detailed information, i.e., the detailed information prediction loss, is determined, and the model parameters of the pre-trained generative model are continuously adjusted based on the detailed information prediction loss until the detailed information prediction loss is less than a preset threshold, thereby obtaining a secondary pre-trained generative model.

[0095] For example, the historical case type is "Sclerosing pneumocystoma", and its corresponding case type detailed information is "Sclerosing pneumocystoma is a rare benign lung tumor, which currently tends to originate from the respiratory epithelium. This tumor is located in...". By inputting "Sclerosing pneumocystoma" into the pre-trained generative model, the prediction detailed information output by the pre-trained generative model can be obtained, and then the loss value between the prediction detailed information output by the model and the case type detailed information can be calculated. The model parameters are adjusted based on the loss value to obtain a secondary pre-trained generative model. Among them, the gradient descent method can be used to make the prediction detailed information output by the model closer and closer to the case type detailed information.

[0096] The last layer of the pre-trained generative model is a probability distribution function, which calculates the token with the highest probability as the generated token, and then generates it step by step. The cross entropy loss function is used to calculate the loss between the generated token and the corresponding position token of the case type detailed information, and then the loss calculated between each corresponding position is averaged and weighted. Therefore, the overall optimization target L during the training process is shown in the following formula (1):

[0097] (1)

[0098] Among them, N represents the maximum number of tokens generated by the model (if the sentence is less than N, blank characters will be filled to N), i represents the positional encoding, M represents the number of token categories, and c represents the token category encoding. Represents the symbolic function of the token of category c at the ith position. If the generated token is the same as the label, then is equal to 1, otherwise it is equal to 0. Represents the probability of generating a token of category c at the i-th position.

[0099] In the case type prediction method provided in the embodiment of the present application, each historical case type is input into the pre-trained generation model to obtain the prediction detailed information output by the pre-trained generation model, and the detailed information prediction loss of the pre-trained generation model is determined according to the detailed information of each case type, the prediction detailed information and the preset loss function; based on the detailed information prediction loss, the model parameters of the pre-trained generation model are adjusted until the detailed information prediction loss is less than the preset threshold value, and the generation model after secondary pre-training is obtained. In this method, the loss value is calculated by the prediction detailed information output by the pre-trained generation model and the case type detailed information to adjust the model parameters of the pre-trained generation model, so that the prediction detailed information output by the model is closer and closer to the case type detailed information. In this process, the pre-trained generation model can learn more medical knowledge so that a better case type prediction effect can be achieved.

[0100] Based on the above embodiment, an embodiment is provided to illustrate the process of fine-tuning the generative model after secondary pre-training using each historical case type and the historical examination information corresponding to each historical case type extracted from the case collection information.

[0101] In an exemplary embodiment, as shown in FIG5 , each historical case type and the historical examination information corresponding to each historical case type extracted from the case collection information are used to fine-tune the generative model after secondary pre-training, including:

[0102] S501, input each historical inspection information into the generation model after secondary pre-training, and obtain the prediction result output by the generation model after secondary pre-training.

[0103] The prediction result refers to the case type corresponding to each historical examination information predicted by the generative model after the second pre-training. Optionally, one historical examination information corresponds to a prediction result of a case type.

[0104] In an embodiment of the present application, after each piece of historical inspection information is input into the generation model after secondary pre-training, the generation model after secondary pre-training analyzes and processes each piece of historical inspection information to predict the case type corresponding to each piece of historical inspection information, and outputs the predicted case type for each piece of historical inspection information. The predicted case type for each piece of historical inspection information is the prediction result output by the generation model after secondary pre-training.

[0105] S502, determining the type prediction loss of the generated model after secondary pre-training according to the prediction results and the types of each historical case.

[0106] Exemplarily, the method of determining the type prediction loss of the generative model after secondary pre-training according to the prediction results and each historical case type can be to obtain a preset prediction loss function, and then determine the type prediction loss of the generative model after secondary pre-training according to the preset prediction loss function, the prediction results and each historical case type. For example, after obtaining the prediction loss function, the prediction results and each historical case type are substituted into the prediction loss function to obtain the type prediction loss of the generative model after secondary pre-training.

[0107] S503, based on the type prediction loss, adjusting the model parameters of the generative model after the secondary pre-training until the type prediction loss is less than a preset threshold, thereby obtaining a type prediction model.

[0108] In an embodiment of the present application, based on the type prediction loss determined above, the model parameters of the generated model after secondary pre-training are adjusted, and then the adjusted model is used for training, and the current type prediction loss is continued to be calculated. If the current type prediction loss is less than a preset threshold, the current model is determined as a type prediction model. If the current type prediction loss is greater than the preset threshold, the model parameters of the current model are continued to be adjusted until the calculated type prediction loss is less than the preset threshold, and the training is stopped to obtain a type prediction model.

[0109] It should be noted that, in the embodiment of the present application, each historical inspection information and each historical case type can also be divided into a training set and a test set, and the generative model after the secondary pre-training can be fine-tuned. For example, each historical inspection information is divided into training inspection information and test inspection information, and each historical case type is divided into a training case type and a test case type. The training set is used to train the generative model after the secondary pre-training to obtain an initial prediction model, and then the test set is used to test the initial prediction model to obtain a type prediction model.

[0110] For the division rules of the training set (training examination information and training case types) and the test set (fine-tuning examination information and fine-tuning case types), since there are many rare cases, it is necessary to ensure that the diseases in the test set have at least appeared in the training set. According to this rule, the training set and the test set are divided, and the ratio is controlled at about 10:1. Then the training examination information is used as input, and the training case type is used as the standard case type to fine-tune the generative model after the second pre-training. For example, the training examination information of a case is: "Female, 53 years old. Chest pain and chest tightness for more than half a year. Multiple huge mass shadows in both lungs...", and the training case type is: "Malignant pleural mesothelioma".

[0111] Among them, fine-tuning can use global fine-tuning, that is, all parameters of the model are updated. The optimization function of global fine-tuning is the same as the optimization function of the previous secondary pre-training, except that the learning rate will be reduced. Fine-tuning can also use the prompt tuning method, by fixing the parameters of the pre-trained model, and then adding learnable prompts to the input of the model to adjust some parameters. Alternatively, Lora (Low-Rank Adaptation of Large Language Models) can be used for fine-tuning that assists in dimensionality reduction, which is not specifically limited here.

[0112] In the case type prediction method provided in the embodiment of the present application, by inputting each historical inspection information into the generation model after secondary pre-training, the prediction result output by the generation model after secondary pre-training is obtained; according to the prediction result and each historical case type, the type prediction loss of the generation model after secondary pre-training is determined; based on the type prediction loss, the model parameters of the generation model after secondary pre-training are adjusted until the type prediction loss is less than a preset threshold, thereby obtaining a type prediction model. In this method, by inputting each historical inspection information into the generation model after secondary pre-training, the prediction result is output, and the loss value can be calculated based on the prediction result combined with each historical case type, and then the model parameters of the generation model after secondary pre-training are adjusted according to the loss value, and the optimal model, namely the type prediction model, is obtained when the loss value reaches a certain threshold; by adjusting the model based on the calculated prediction loss, the performance of the model can be optimized, so that the obtained type prediction model can better adapt to different data and task requirements, thereby improving the prediction performance of the type prediction model.

[0113] Based on the above embodiment, an embodiment is provided to illustrate the process of optimizing the type prediction model.

[0114] In an exemplary embodiment, as shown in FIG6 , the method further includes:

[0115] S601, performing a quality score on the output result of the type prediction model according to the pre-trained reward model and sample inspection information, and obtaining a quality score result of the sample inspection information.

[0116] In an embodiment of the present application, a pre-trained reward model can perform a quality score on the prediction results of the type prediction model. The sample inspection information is input into the type prediction model, and the type prediction model predicts the case type corresponding to the sample inspection information to obtain a sample prediction result, and then the sample prediction result is input into the pre-trained reward model, and the reward model performs a quality score on the sample prediction result to obtain a quality score result of the sample inspection information. Among them, the sample inspection information refers to the inspection information used to evaluate the prediction ability of the type prediction model.

[0117] S602, based on the quality scoring result, optimizing the parameters of the type prediction model to obtain an optimized type prediction model.

[0118] The type prediction model has a certain ability to predict case types, but it has certain limitations in natural language measurement. It can only be infinitely close to the so-called "gold standard" but cannot reflect the alignment with human intentions. Therefore, the type prediction model can be self-feedback and error correction to obtain a type prediction model with higher prediction performance.

[0119] Among them, the self-feedback and error correction process needs to first evaluate the output results of the type prediction model, and then optimize the parameters of the type prediction model based on the evaluation results to obtain the optimized type prediction model.

[0120] In an embodiment of the present application, both the quality score results and the sample inspection information can be input into the type prediction model, and the proximal policy optimization (PPO) algorithm is used to optimize the parameters of the type prediction model to obtain an optimized type prediction model. Among them, the PPO algorithm adjusts the policy parameters of the model through score feedback, aiming to enhance the model's ability to generate higher quality case type prediction results in future tasks. This reward-based feedback mechanism enables the model to continuously improve itself during iterative training and gradually improve its performance in medical diagnosis tasks.

[0121] In the case type prediction method provided in the embodiment of the present application, the quality score of the output result of the type prediction model is scored according to the pre-trained reward model and sample inspection information to obtain the quality score result of the sample inspection information, and finally, based on the quality score result, the type prediction model is parameter optimized to obtain the optimized type prediction model. In this method, on the basis that the type prediction model has a certain prediction ability, in order to improve the prediction performance of the type prediction model, the quality score result is further used to optimize the parameters of the type prediction model to obtain a type prediction model with better performance, which lays the foundation for accurately determining the case type prediction results of the target site.

[0122] Based on the above embodiment, an embodiment is provided to illustrate the process of performing quality scoring on the output result of the type prediction model according to the pre-trained reward model and sample inspection information to obtain the quality scoring result of the sample inspection information.

[0123] In an exemplary embodiment, as shown in FIG7 , the quality score of the output result of the type prediction model is performed according to the pre-trained reward model and the sample inspection information to obtain the quality score result of the sample inspection information, including:

[0124] S701, inputting sample inspection information into a type prediction model to obtain a sample prediction result output by the type prediction model.

[0125] In the embodiment of the present application, in order to perform a quality score on the output result of the type prediction model, the output result of the type prediction model needs to be obtained first. Based on this, in the present embodiment, the sample inspection information can be input into the type prediction model, and the type prediction model can analyze and process the input sample inspection information and output the sample prediction result.

[0126] S702: Input the sample prediction result into the reward model to obtain the quality score result.

[0127] After obtaining the sample prediction results output by the type prediction model, the sample prediction results are input into the reward model, which analyzes and evaluates the sample prediction results to obtain the quality score results of the sample inspection information.

[0128] In the case type prediction method provided in the embodiment of the present application, the sample inspection information is input into the type prediction model to obtain the sample prediction result output by the type prediction model, and then the sample prediction result is input into the reward model to obtain the quality score result. In this method, an optional way to quickly determine the quality score result is provided; by inputting the sample inspection information into the type prediction model, the output result of the type prediction model, that is, the sample prediction result of the sample inspection information can be obtained, and then the reward model evaluates the sample prediction result output by the type prediction model, and the quality score result of the sample inspection information can be obtained, which provides data support for the subsequent fine-tuning of the type prediction model.

[0129] Based on the above embodiment, an embodiment is provided to illustrate the training process of the above reward model.

[0130] In an exemplary embodiment, as shown in FIG8 , the training process of the reward model includes:

[0131] S801, obtaining a training data set.

[0132] In the embodiment of the present application, the training data set is a plurality of type prediction results generated by the type prediction model for each inspection information.

[0133] Optionally, when training the reward model, some inspection information is selected, and the type prediction model predicts these inspection information to obtain multiple type prediction results for each inspection information.

[0134] S802, input each type of prediction result into the quality assessment model to obtain the prediction quality ranking result output by the quality assessment model.

[0135] Among them, the quality assessment model may include the Tongyi Qianwen model, etc.

[0136] The multiple type prediction results of each inspection information are input into the quality assessment model, and the quality assessment model ranks the multiple type prediction results of each inspection information by quality to determine the better and worse type prediction results, that is, the prediction quality ranking results.

[0137] S803, adjusting the model parameters of the quality assessment model according to the predicted quality ranking result and the standard quality ranking result until the training is completed to obtain the reward model.

[0138] After obtaining the predicted quality ranking results output by the quality assessment model, the loss value can be calculated based on the predicted quality ranking results and the standard quality ranking results, and the model parameters of the quality assessment model can be adjusted based on the loss value until the training is completed to obtain the reward model.

[0139] Among them, the standard quality ranking results can be obtained by doctors and experts ranking the multiple type prediction results output by the type prediction model. For example, the multiple type prediction results output by the type prediction model and the better and sub-optimal type prediction results screened by the quality assessment model are anonymized and randomized to form 50 questions, and then junior doctors and experts in the radiology department can be invited to rank by preference, that is, to obtain the standard quality ranking results.

[0140] In addition, it should be noted that in this application, the pre-trained generative model is pre-trained and fine-tuned twice to obtain a type prediction model. In order to verify whether the type prediction model in the embodiment of this application is effective, a comparison scheme is also designed. One is to directly use the pre-trained generative model; the other is to fine-tune the pre-trained generative model; and the other is the one in this application, in which the pre-trained generative model is fine-tuned after the second pre-training. The pre-trained generative models all use the same dialogue language model, and the fine-tuning method uses the global Finetune. For the second pre-training, the SGD optimizer with a learning rate of 1e-3, a momentum of 0.9, and a weight decay of 1e-5 is used for training. The batch size is set to 8, the maximum input token length is 50 (because the final diagnosis is very short), and the maximum output token length is 2048. For the global Finetune, the SGD optimizer with a learning rate of 1e-4, a momentum of 0.9, and a weight decay of 1e-6 is used for training. The batch size is also set to 8, the maximum input token length is 2048, and the maximum output token length is 50. Both the secondary pre-training and fine-tuning were performed using 4 NVidia A800 GPUs.

[0141] The test indicators used include the natural language processing indicators bleu (Bilingual Evaluation Understudy, evaluating the quality of machine translation results) and rouge (Recall-Oriented Understudy for Gisting Evaluation, evaluating the quality of text automatic summarization), as well as the two indicators of complete prediction accuracy and major category prediction accuracy. Complete prediction accuracy means that the model can completely and accurately predict the specific name corresponding to the case type; major category prediction accuracy means that the major categories such as cancer, tumor, inflammation, nodule, fungal and bacterial infection are predicted correctly, but the specific name may be predicted incorrectly. These indicators are used to judge the effect of the model. The comparison of experimental results indicators can be seen in Table 1.

[0142] Table 1

[0143]

[0144] From the table, we can see that the prediction effect of directly using the pre-trained generative model is very poor. The reason is that there are many types of cases, and many difficult and complicated diseases are included, so the direct prediction effect is very poor. The prediction effect of using secondary pre-training and then fine-tuning is better than that of direct fine-tuning in all indicators, which shows that the secondary pre-training method designed in this application has played a role and improved the prediction effect of the model.

[0145] In the case type prediction method provided in the embodiment of the present application, a training data set is obtained, and then each type of prediction result is input into the quality assessment model to obtain the prediction quality ranking result output by the quality assessment model, and then the model parameters of the quality assessment model are adjusted according to the prediction quality ranking result and the standard quality ranking result until the training is completed to obtain a reward model, wherein the training data set is a plurality of type prediction results generated by the type prediction model for each inspection information. In this method, by inputting a plurality of type prediction results output by the type prediction model into the quality assessment model, the ranking information of the plurality of type prediction results by the quality assessment model can be obtained, and then based on the ranking result output by the quality assessment model and the standard ranking result, the model parameters of the quality assessment model are continuously optimized and adjusted to obtain a reward model, which provides an optional method for training the reward model.

[0146] In addition, in an exemplary embodiment, the present application also provides an optional example of a case type prediction method, as shown in FIG9 , which may include the following steps:

[0147] S901, obtaining case collection information, and extracting multiple historical case types, case type detailed information corresponding to each historical case type, and historical examination information corresponding to each historical case type from the case collection information.

[0148] S902, input each historical case type into the pre-trained generation model to obtain the prediction detailed information output by the pre-trained generation model.

[0149] S903, adjusting the model parameters of the pre-trained generation model according to the detailed information of each case type, the prediction detailed information and the preset loss function until the training is completed, and obtaining the generation model after the second pre-training.

[0150] S904, input each historical inspection information into the generation model after secondary pre-training, and obtain the prediction result output by the generation model after secondary pre-training.

[0151] S905, determining the type prediction loss of the generated model after secondary pre-training according to the prediction results and the types of each historical case.

[0152] S906, based on the type prediction loss, adjusting the model parameters of the generated model after the secondary pre-training until the type prediction loss is less than a preset threshold, thereby obtaining a type prediction model.

[0153] S907, obtaining medical examination information of the target part.

[0154] S908, inputting the medical examination information into a pre-trained type prediction model to obtain a case type prediction result of the target site.

[0155] The above-mentioned processes of S901-S908 can refer to the description of the above-mentioned method embodiment, and the implementation principles and technical effects thereof are similar and will not be repeated here.

[0156] In addition, in one embodiment, the process of case type prediction is explained in detail.

[0157] As shown in Figure 10, first obtain case information from case books, hospitals, etc., extract the basic information of the patient, imaging manifestations, and other examination results in the case as the first part (historical examination information), and use the final diagnosis result as the second part (historical case type). Then extract detailed descriptions of the case from books, the Internet, etc., and use the obtained detailed descriptions as the third part (detailed information on case types). Then select the pre-trained large language model, and perform secondary pre-training on the large language model. The secondary pre-training is to use the second part as input and the third part as the prediction label for training. After the secondary pre-training, divide the first and second parts into training sets and test sets according to certain rules, use the model after secondary pre-training on the training set to fine-tune the case type prediction model, and test the prediction effect of the model on the test set.

[0158] In the case type prediction method provided in the embodiment of the present application, a pre-trained generative model is pre-trained for a second time by using a plurality of historical case types and corresponding case type detailed information. In this way, the generative model can learn a large amount of medical knowledge, that is, it has knowledge in the medical field, so that the generative model can better handle various tasks in the medical field; further, a plurality of historical case types and corresponding historical examination information are used to fine-tune the generative model after the second pre-training, so that the generative model adapts to new data sets and tasks, and the performance of the generative model on specific tasks is improved by training the new data sets, that is, the adaptability and prediction performance of the generative model in specific tasks such as case type prediction are improved. In this way, the type prediction model has learned the corresponding information of various cases. The relationship between the medical examination information and the case type can be learned, and then the type prediction model can infer the case type corresponding to any medical examination information based on the medical examination information, which is equivalent to inputting any medical examination information into the type prediction model. The type prediction model can infer the case type corresponding to the input medical examination information based on the learned relationship between the medical examination information and the case type. Because the relationship between the medical examination information and the case type has been learned, it has a strong logical inference ability. Therefore, there is no need to mark different case types in advance, and various case types can be predicted. Not only common case types can be predicted, but also many rare case types can be predicted, which improves the scenario applicability of case type prediction.

[0159] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0160] Based on the same inventive concept, the embodiment of the present application also provides a case type prediction device for implementing the case type prediction method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more case type prediction device embodiments provided below can refer to the limitations of the case type prediction method above, and will not be repeated here.

[0161] In an exemplary embodiment, as shown in FIG11 , a case type prediction device 1 is provided, comprising: an information acquisition module 10 and a result determination module 20, wherein:

[0162] An information acquisition module 10 is used to acquire medical examination information of a target part;

[0163] The result determination module 20 is used to input the medical examination information into a pre-trained type prediction model to obtain the case type prediction result of the target part; the type prediction model is obtained by secondary pre-training the pre-trained generation model using multiple historical case types extracted from the case collection information of the target part and the case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and the historical examination information corresponding to each historical case type extracted from the case collection information.

[0164] In one embodiment, the case type prediction device 1 further includes:

[0165] The case collection information acquisition module is used to retrieve case information related to the target part from the case database according to the target part; and determine the case information related to the target part as the case collection information of the target part.

[0166] In one embodiment, the result determination module 20 is further configured to:

[0167] Input each historical case type into the pre-trained generative model to obtain the prediction detailed information output by the pre-trained generative model; determine the detailed information prediction loss of the pre-trained generative model based on the detailed information of each case type, the prediction detailed information and the preset loss function; based on the detailed information prediction loss, adjust the model parameters of the pre-trained generative model until the detailed information prediction loss is less than the preset threshold, and obtain the generative model after secondary pre-training.

[0168] In one embodiment, the result determination module 20 is further configured to:

[0169] Input each historical inspection information into the generative model after secondary pre-training to obtain the prediction result output by the generative model after secondary pre-training; determine the type prediction loss of the generative model after secondary pre-training based on the prediction result and each historical case type; based on the type prediction loss, adjust the model parameters of the generative model after secondary pre-training until the type prediction loss is less than the preset threshold, and obtain the type prediction model.

[0170] In one embodiment, the case type prediction device 1 further includes:

[0171] A quality scoring module is used to perform a quality scoring on the output result of the type prediction model according to the pre-trained reward model and sample inspection information, and obtain a quality scoring result of the sample inspection information;

[0172] The parameter optimization module is used to optimize the parameters of the type prediction model based on the quality scoring results to obtain an optimized type prediction model.

[0173] In one embodiment, the quality scoring module is further used to:

[0174] The sample inspection information is input into the type prediction model to obtain the sample prediction result output by the type prediction model; the sample prediction result is input into the reward model to obtain the quality score result.

[0175] In one embodiment, the case type prediction device further includes:

[0176] The data set acquisition module is used to acquire a training data set; the training data set is a plurality of type prediction results generated by the type prediction model for each inspection information;

[0177] The ranking result is determined, which is used to input each type of prediction result into the quality assessment model to obtain the prediction quality ranking result output by the quality assessment model;

[0178] The reward model determination module is used to adjust the model parameters of the quality assessment model according to the predicted quality ranking results and the standard quality ranking results until the training is completed to obtain the reward model.

[0179] Each module in the above-mentioned case type prediction device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0180] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0181] Obtain medical examination information of the target area;

[0182] The medical examination information is input into a pre-trained type prediction model to obtain a case type prediction result for the target site; the type prediction model is obtained by secondary pre-training a pre-trained generation model using multiple historical case types extracted from the case collection information of the target site and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information.

[0183] The implementation principles and technical effects of each step implemented by the processor in the embodiment of the present application are similar to the principles of the above-mentioned case type prediction method, and will not be repeated here.

[0184] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0185] Obtain medical examination information of the target area;

[0186] The medical examination information is input into a pre-trained type prediction model to obtain a case type prediction result for the target site; the type prediction model is obtained by secondary pre-training a pre-trained generation model using multiple historical case types extracted from the case collection information of the target site and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information.

[0187] The implementation principles and technical effects of the various steps implemented when the computer program in the embodiment of the present application is executed by the processor are similar to the principles of the above-mentioned case type prediction method, and will not be repeated here.

[0188] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0189] Obtain medical examination information of the target area;

[0190] The medical examination information is input into a pre-trained type prediction model to obtain a case type prediction result for the target site; the type prediction model is obtained by secondary pre-training a pre-trained generation model using multiple historical case types extracted from the case collection information of the target site and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information.

[0191] The implementation principles and technical effects of the various steps implemented when the computer program in the embodiment of the present application is executed by the processor are similar to the principles of the above-mentioned case type prediction method, and will not be repeated here.

[0192] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0193] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0194] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0195] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A case type prediction method, characterized in that: The method comprises: Obtain medical examination information of the target area; The medical examination information is input into a pre-trained type prediction model to obtain a case type prediction result of the target part; the type prediction model is obtained by secondary pre-training a pre-trained generation model using multiple historical case types extracted from the case collection information of the target part and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information.

2. The method according to claim 1, characterized in that The method further comprises: According to the target part, retrieving case information related to the target part from a case database; The case information related to the target part is determined as the case collection information of the target part.

3. The method according to claim 1 or 2, characterized in that: The method of performing secondary pre-training on the pre-trained generation model using a plurality of historical case types extracted from the case collection information of the target part and case type detailed information corresponding to each of the historical case types comprises: Inputting each of the historical case types into the pre-trained generation model to obtain the prediction detailed information output by the pre-trained generation model; Determining the detailed information prediction loss of the pre-trained generation model according to the case type detailed information, the prediction detailed information and a preset loss function; Based on the detailed information prediction loss, the model parameters of the pre-trained generation model are adjusted until the detailed information prediction loss is less than a preset threshold, thereby obtaining the generation model after the secondary pre-training.

4. The method according to claim 1 or 2, characterized in that: The method of fine-tuning the generative model after the secondary pre-training by using each of the historical case types and the historical examination information corresponding to each of the historical case types extracted from the case collection information comprises: Input each of the historical inspection information into the generation model after the secondary pre-training to obtain a prediction result output by the generation model after the secondary pre-training; Determining the type prediction loss of the generative model after the secondary pre-training according to the prediction result and each of the historical case types; Based on the type prediction loss, the model parameters of the generative model after the secondary pre-training are adjusted until the type prediction loss is less than a preset threshold, thereby obtaining the type prediction model.

5. The method according to claim 1 or 2, characterized in that: The method further comprises: According to the pre-trained reward model and sample inspection information, a quality score is performed on the output result of the type prediction model to obtain a quality score result of the sample inspection information; Based on the quality score result, the type prediction model is optimized in parameters to obtain an optimized type prediction model.

6. The method according to claim 5, characterized in that The method of performing a quality score on the output result of the type prediction model according to the pre-trained reward model and the sample inspection information to obtain a quality score result of the sample inspection information includes: Inputting the sample inspection information into the type prediction model to obtain a sample prediction result output by the type prediction model; The sample prediction result is input into the reward model to obtain the quality score result.

7. The method according to claim 5, characterized in that The training process of the reward model includes: Acquire a training data set; the training data set is a plurality of type prediction results generated by the type prediction model for each inspection information; Inputting each type of prediction result into a quality assessment model to obtain a prediction quality ranking result output by the quality assessment model; According to the predicted quality ranking result and the standard quality ranking result, the model parameters of the quality assessment model are adjusted until the training is completed to obtain the reward model.

8. A case type prediction device, characterized in that: The device comprises: An information acquisition module, used to acquire medical examination information of the target part; A result determination module is used to input the medical examination information into a pre-trained type prediction model to obtain a case type prediction result of the target part; the type prediction model is obtained by secondary pre-training a pre-trained generation model using multiple historical case types extracted from the case collection information of the target part and case type detailed information corresponding to each historical case type, and fine-tuning the secondary pre-trained generation model using each historical case type and historical examination information corresponding to each historical case type extracted from the case collection information.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images