A training method, training device, equipment and medium for an image recognition model

Abstract

The present application provides a training method, a training device, a device and a medium for an image recognition model. The method includes: for the first medical image of each training sample, inputting the first medical image into a deep learning recognition model to obtain a plurality of first medical image features, inputting each first medical image feature into a correction model to be trained to obtain a first correction parameter for each first medical image feature, and using the first correction parameter to perform correction processing on each first medical image feature to obtain a corrected first medical image feature; for the first medical image of each training sample, re-inputting the plurality of corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain a first recognition result, and training the correction model to be trained based on the difference between the first recognition result and the second recognition result; and obtaining an image recognition model based on the deep learning recognition model and the correction model.

Description

Technical Field

[0001] This application relates to the field of image recognition technology. Specifically, it relates to a training method, training device, equipment, and medium for an image recognition model. Background Art

[0002] Medical images are widely used in clinical medicine, providing a great deal of scientific and intuitive basis for disease diagnosis. They can better cooperate with clinical symptoms, laboratory tests, etc., and play an irreplaceable role in the final accurate diagnosis of the condition. With the development of technology, in order to diagnose lesions based on medical images faster and relieve the pressure on doctors, in most cases, deep learning models are used to identify medical images and diagnose lesions.

[0003] However, since medical images belong to the privacy of patients, different hospitals will not hand over the medical images of their own patients to other hospitals for data analysis. Therefore, each hospital can only use the medical images of its own hospital to train the deep learning model, which also results in the deep learning models trained by each hospital being unable to be applied to other hospitals. Each hospital needs to train its own medical images. For example, in the intelligent quantitative analysis of acute pulmonary infection diseases such as COVID-19, this situation increases the complexity of data processing. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a training method, training device, equipment, and medium for an image recognition model, which is used to solve the problem of low robustness of the deep recognition model for identifying medical images in the prior art.

[0005] In a first aspect, an embodiment of this application provides a training method for an image recognition model. The training method includes:

[0006] Obtain a training sample set;

[0007] For the first medical image of each training sample, input the first medical image into a trained deep learning recognition model to obtain multiple first medical image features. Input each first medical image feature into a correction model to be trained to obtain a first correction parameter for each first medical image feature, and use the first correction parameter to perform correction processing on each first medical image feature to obtain a corrected first medical image feature;

[0008] For the first medical image of each training sample, re-input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain a first recognition result, and train the to-be-trained correction model based on the difference between the first recognition result and the second recognition result; the second recognition result is obtained by the trained deep learning recognition model from the second medical image corresponding to the first medical image, and the first medical image is obtained by performing data augmentation processing on the second medical image;

[0009] Based on the trained deep learning recognition model and the trained correction model, obtain a trained image recognition model.

[0010] Optionally, the trained deep learning recognition model is trained based on single-center medical image data.

[0011] Optionally, the training sample set is obtained by performing data augmentation processing on the single-center medical image data.

[0012] Optionally, the method further includes:

[0013] For the first medical image of each training sample, input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain multiple first medical image correction features, and input the second medical image corresponding to the first medical image into the trained deep learning recognition model to obtain multiple second medical image features;

[0014] For the first medical image of each training sample, compare the similarity of each first medical image correction feature of the medical image and each second medical image feature of the second medical image according to the corresponding feature level, and train the to-be-trained correction model according to the comparison result.

[0015] Optionally, the training method further includes:

[0016] Input the obtained medical image to be recognized into the deep learning recognition model in the trained image recognition model to obtain multiple third medical image features;

[0017] Input the multiple third medical image features into the correction model in the trained image recognition model respectively to determine the second correction parameter of each third medical image feature;

[0018] For each third medical image feature, correct the third medical image feature by using the second correction parameter to obtain a corrected third medical image feature;

[0019] Input the multiple corrected third medical image features into the deep learning recognition model in the trained image recognition model to determine the recognition result of the medical image to be recognized.

[0020] In a second aspect, an embodiment of the present application provides a training device for an image recognition model. The training device includes:

[0021] An acquisition module, configured to acquire a training sample set;

[0022] A first correction module, for each first medical image of each training sample, input the first medical image into the trained deep learning recognition model to obtain multiple first medical image features, input each first medical image feature into the correction model to be trained to obtain the first correction parameter of each first medical image feature, and use the first correction parameter to perform correction processing on each first medical image feature to obtain the corrected first medical image feature;

[0023] A first training module, for each first medical image of each training sample, re-input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain a first recognition result, and train the correction model to be trained based on the difference between the first recognition result and the second recognition result; the second recognition result is recognized by the trained deep learning recognition model from the second medical image corresponding to the first medical image, and the first medical image is obtained by performing data augmentation processing on the second medical image;

[0024] A combination module, configured to obtain a trained image recognition model based on the trained deep learning recognition model and the trained correction model.

[0025] Optionally, the trained deep learning recognition model is trained based on single-center medical image data.

[0026] Optionally, the training sample set is obtained by performing data augmentation processing on the single-center medical image data.

[0027] In a third aspect, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above training method are implemented.

[0028] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the steps of the above training method are executed.

[0029] The embodiment of the present application provides a training method for an image recognition model. First, a training sample set is obtained. Secondly, for the first medical image of each training sample, the first medical image is input into a trained deep learning recognition model to obtain multiple first medical image features. Each first medical image feature is input into a correction model to be trained to obtain a first correction parameter for each first medical image feature. The first medical image feature is corrected using the first correction parameter to obtain a corrected first medical image feature. Then, for the first medical image of each training sample, the multiple corrected first medical image features corresponding to the first medical image are re-input into the trained deep learning recognition model to obtain a first recognition result. Based on the difference between the first recognition result and the second recognition result, the correction model to be trained is trained. The second recognition result is obtained by the trained deep learning recognition model from the second medical image corresponding to the first medical image, and the first medical image is obtained by performing data augmentation processing on the second medical image. Finally, a trained image recognition model is obtained based on the trained deep learning recognition model and the trained correction model.

[0030] In some embodiments, after obtaining a deep learning recognition model that has been trained using single-center medical image data, if multi-center medical image data needs to be recognized, it is not necessary to retrain the already trained deep learning recognition model, nor is it necessary to obtain multi-center medical image data. Only the single-center medical image data is subjected to data augmentation processing, and the first medical image after data augmentation processing is used to train the correction model to be trained to obtain a trained correction model. Finally, a trained deep learning recognition model and the trained correction model can be used to obtain an image recognition model for recognizing multi-center medical images, improving the robustness of the image recognition model.

[0031] To make the above objects, features, and advantages of the present application more obvious and understandable, the following specific preferred embodiments are given in conjunction with the accompanying drawings and are described in detail as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] To more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 It is a schematic flowchart of a training method for an image recognition model provided by an embodiment of the present application;

[0034] Figure 2A flowchart showing a method for recognizing an image recognition model provided by an embodiment of the present application;

[0035] Figure 3 A schematic structural diagram of a training device for an image recognition model provided by an embodiment of the present application;

[0036] Figure 4 A schematic structural diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and illustrated herein usually can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application to be protected, but merely represents the selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts fall within the scope of protection of the present application.

[0038] In the prior art, the medical devices used by different hospitals are different, and different medical devices are relatively complex, making it difficult to form unified rules. A deep learning model trained with the data of a certain hospital cannot be used to identify lesions in the medical images of other hospitals. When a patient uses the medical images of a certain hospital to see a doctor in another hospital, the deep learning model cannot be used to identify accurate lesion results.

[0039] Based on the above defects, the embodiments of the present application provide a training method for an image recognition model, as Figure 1 shown, including the following steps:

[0040] S101, obtain a training sample set;

[0041] S102. For the first medical image of each training sample, input the first medical image into a trained deep learning recognition model to obtain a plurality of first medical image features. Input each first medical image feature into a correction model to be trained to obtain a first correction parameter for each first medical image feature, and use the first correction parameter to perform correction processing on each first medical image feature to obtain a corrected first medical image feature;

[0042] S103. For the first medical image of each training sample, re-enter the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain a first recognition result, and train the to-be-trained correction model based on the difference between the first recognition result and the second recognition result; the second recognition result is obtained by the trained deep learning recognition model from the second medical image corresponding to the first medical image, and the first medical image is obtained by performing data augmentation processing on the second medical image;

[0043] S104. Based on the trained deep learning recognition model and the trained correction model, obtain a trained image recognition model.

[0044] In the above step S101, the training sample set includes multiple training samples, and each training sample is a first medical image. Each first medical image in the training sample set is obtained by performing data augmentation processing on the second medical image in the single-center medical image data. The single-center medical image data can be medical images with a unified specification. For example, medical images produced by the same medical device, or medical images produced by multiple medical devices controlled by the same hospital according to the same specification. The unified specification can include any one or more of the following specifications: image size, brightness, contrast, noise, etc. The data augmentation processing specifically refers to performing data augmentation processing on the pixel values in the medical image, and the specific data augmentation processing methods include any one or more of the following methods: contrast change, brightness change, random noise, equalization, and non-linear function change, etc.

[0045] In the above step S102, the deep learning recognition model is trained based on the single-center medical image data mentioned above.

[0046] The deep learning recognition model is trained through the following steps:

[0047] Obtain single-center medical image data, and the single-center medical image data includes multiple second medical images;

[0048] Perform labeling processing on each second medical image in the single-center medical image data, and label the target object of each second medical image; wherein, the target object can be artificially specified, and the target object can include any one or more of the following: lesions, organs;

[0049] For each second medical image, input the second medical image labeled with the target object into the to-be-trained deep learning recognition model to train the to-be-trained deep learning recognition model.

[0050] The first medical image feature is the feature extracted according to a preset rule during the process of feature extraction of the first medical image by a trained deep learning recognition model after the first medical image is input into the trained deep learning recognition model. The deep learning recognition model includes multiple layers of structures (possibly dozens of layers). Each layer will recognize corresponding features from the first medical image, and the difference in the features recognized between every two adjacent layers is not too large. To save computing resources, a small number of features can be extracted from the features recognized by the multiple layers of structures according to the preset rule. For example, the multiple layers of structures are divided into multiple groups according to the preset feature levels, and one feature is extracted from each group, that is, the first medical image feature.

[0051] For example, the trained deep learning recognition model has a total of m layers of structures. After the first medical image is input into the trained deep recognition model, according to the preset rule, n first medical image features can be extracted from the features recognized by the m layers of structures of the trained deep learning recognition model. The extracted first medical image features can be represented by F1 to Fn, where m is greater than n.

[0052] The correction model to be trained is composed of learnable vectors and a fully connected layer with the same number as the number of the first medical image features. Among them, the learnable vectors can be represented by D1 to Dn.

[0053] In this application, for the convenience of calculation, both the first medical image feature and the learnable vector are set to form a unified feature scale, that is, the length, width, and height of the vector are kept consistent.

[0054] The first correction parameter is used to correct the first medical image feature so that the corrected first medical image feature can more truly restore the features of the second medical image corresponding to the first medical image before data augmentation processing.

[0055] In specific implementation, for each first medical image in the training sample set, the first medical image is input into the trained deep learning recognition model. Among the multiple layers of features extracted from the first medical image by the trained deep learning recognition model, multiple first medical image features are extracted according to the preset rule, and corresponding serial numbers are set for each first medical image feature according to the hierarchical relationship.

[0056] For each first medical image feature, input the first medical image feature into the correction model to be trained. Set serial numbers for the multiple learnable vectors in the correction model to be trained (the setting rule of the serial numbers of the learnable vectors is the same as that of the serial numbers of the first medical image features. For example, if the serial numbers of the first medical image features are 1 - n, then the serial numbers of the learnable vectors are also 1 - n). Perform matrix multiplication on the first medical image feature and the learnable vector with the same serial number to obtain a first feature vector. Input the first feature vector into the fully connected layer in the correction model to be trained, and two correction parameters can be calculated. The correction parameters include the scaling value for feature correction and the offset value for feature correction. Use the two correction parameters to perform correction processing on the first medical image feature, and the corrected first medical image feature can be obtained.

[0057] For example, there is a first medical image feature Fa with serial number a. Input the first medical image feature Fa into the correction model to be trained. Find the learnable vector Da with serial number a among the multiple learnable vectors in the correction model to be trained. Perform matrix multiplication on the first medical image feature Fa and the learnable vector Da to obtain a first feature vector Ga. Input the first feature vector Ga into the fully connected layer in the correction model to be trained, and two correction parameters for the first medical image feature Fa can be calculated, which are respectively the scaling value gama for the correction of the first medical image feature Fa a and the offset value beta for the correction of the first medical image feature Fa a . Use the two correction parameters to perform correction processing on the first medical image feature, and the corrected first medical image feature Ca can be obtained.

[0058] Specifically, use the following formula to calculate the corrected first medical image feature Ca:

[0059] Ca = Fa * gama a + beta a ;

[0060] where Ca is the corrected first medical image feature, Fa is the first medical image feature, gama a is the scaling value for the correction of the first medical image feature Fa, and beta a is the offset value for the correction of the first medical image feature Fa.

[0061] In the above step S103, since the first medical image is obtained by performing data augmentation on the corresponding second medical image, there may be differences between the first medical image and the second medical image in terms of brightness, contrast, etc. However, the human physiological features represented in the two medical images are the same. This application mainly adjusts the recognition result of the first medical image after data augmentation to a situation where it has a small difference from the recognition result of the first medical image. Therefore, the difference between the recognition result obtained after correcting the first medical image and the recognition result of the second medical image should be relatively small. If the difference between the recognition result of the first medical image and the recognition result of the second medical image is relatively large, it indicates that the recognition result obtained after correcting the first medical image is inaccurate. Therefore, the correction model to be trained is trained.

[0062] Specifically, for each first medical image, input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain the first recognition result corresponding to the first medical image. Obtain the second recognition result obtained by recognizing the second medical image corresponding to the first medical image through the trained deep learning model. Compare the first recognition result and the second recognition result (wherein, a loss function can be used for comparison). Based on the comparison result, train the correction model to be trained. The training process includes adjusting the parameters of the learnable vector and adjusting the parameters in the fully connected layer.

[0063] In the above step S104, the trained deep learning recognition model and the trained correction model can form an image recognition model for recognizing the target object in the medical image.

[0064] In the training method of the image recognition model provided by this application, after obtaining the trained deep learning recognition model using single-center medical image data, if it is necessary to recognize multi-center medical image data, it is not necessary to retrain the already trained deep learning recognition model, nor is it necessary to obtain multi-center medical image data. Only perform data augmentation on the single-center medical image data, and use the first medical image after data augmentation to train the correction model to be trained to obtain the trained correction model. Finally, the trained deep learning recognition model and the trained correction model can be used to obtain an image recognition model for recognizing multi-center medical images, improving the robustness of the image recognition model.

[0065] In the process of training the correction model in this application, in addition to correcting it with the first recognition result and the second recognition result, the correction model to be trained can also be trained using the feature space similarity metric loss. That is, the training method of this application further includes:

[0066] Step 105: For the first medical image of each training sample, input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain multiple first medical image correction features, and input the second medical image corresponding to the first medical image into the trained deep learning recognition model to obtain multiple second medical image features;

[0067] Step 106: For the first medical image of each training sample, compare the similarity of each first medical image correction feature of the medical image and each second medical image feature of the second medical image according to the corresponding feature level, and train the correction model to be trained according to the comparison result.

[0068] In the above Step 105, for the first medical image of each training sample, input the corrected first medical image features into the trained deep learning recognition model, and still extract multiple first medical image correction features using a preset rule, and input the second medical image corresponding to the first medical image into the trained deep learning recognition model, and extract multiple second medical image features using a preset rule. Among them, each first medical image correction feature and each second medical image feature have corresponding feature levels. Of course, corresponding serial numbers can also be set for each first medical image correction feature and each second medical image feature according to the size of the feature level.

[0069] In the above Step 106, for the first medical image of each training sample, find the corresponding first medical image correction feature and each second medical image feature using the feature level or serial number, calculate the feature space similarity metric loss between each pair of first medical image correction features and each second medical image feature using a loss function, and then train the correction model to be trained using the above feature space similarity metric loss, that is, adjust the parameters in the learnable vector and fully connected layer of the correction model to be trained.

[0070] In this application, the trained deep learning recognition model and the trained correction model can form an image recognition model for recognizing target objects in medical images. For example, Figure 2 As shown, the specific recognition process of the trained image recognition model includes:

[0071] S107: Input the obtained medical image to be recognized into the deep learning recognition model in the trained image recognition model to obtain multiple third medical image features;

[0072] S108: Input the multiple third medical image features into the correction model in the trained image recognition model respectively to determine the second correction parameters of each third medical image feature;

[0073] S109. For each third medical image feature, correct the third medical image feature by using the second correction parameter to obtain a corrected third medical image feature.

[0074] S110. Input multiple corrected third medical image features into the deep learning recognition model in the trained image recognition model to determine the recognition result of the medical image to be recognized.

[0075] In the above step S107, the medical image to be recognized is the image from which the recognition result needs to be determined. The third medical image feature is the feature recognized by the deep learning recognition model in the trained image recognition model from the medical image to be recognized.

[0076] In the above step S108, during the prediction process, input multiple third medical image features recognized from the medical image to be recognized into the correction model of the trained image recognition model respectively. Each third medical image feature will be multiplied by the corresponding learnable vector in the correction model respectively, and the obtained vector after multiplication will be input into the fully connected layer in the correction model, and finally the second correction parameter corresponding to each third medical image feature is obtained.

[0077] In the above steps S109 and S110, according to each third medical image feature and the corresponding correction parameter respectively, multiple corrected third medical image features are determined, and the multiple corrected medical image features are input into the deep learning recognition model of the trained image recognition model. The result output by this deep learning recognition model is the recognition result of the medical image to be recognized.

[0078] An embodiment of the present application provides a training device for an image recognition model, as Figure 3 shown. The training device includes:

[0079] An acquisition module 301, configured to acquire a training sample set;

[0080] A first correction module 302, configured to, for the first medical image of each training sample, input the first medical image into the trained deep learning recognition model to obtain multiple first medical image features, input each first medical image feature into the correction model to be trained to obtain the first correction parameter of each first medical image feature, and use the first correction parameter to perform a correction process on each first medical image feature to obtain a corrected first medical image feature;

[0081] The first training module 303 is configured to, for the first medical image of each training sample, re-input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain a first recognition result, and train the to-be-trained correction model based on the difference between the first recognition result and the second recognition result; the second recognition result is recognized by the trained deep learning recognition model from the second medical image corresponding to the first medical image, and the first medical image is obtained by performing data augmentation processing on the second medical image;

[0082] The combination module 304 is configured to obtain a trained image recognition model based on the trained deep learning recognition model and the trained correction model.

[0083] Optionally, the trained deep learning recognition model is trained using the single-center medical image data.

[0084] Optionally, the training sample set is obtained by performing data augmentation processing on the single-center medical image data.

[0085] Optionally, the training device further includes:

[0086] The first feature recognition module is configured to, for the first medical image of each training sample, input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain multiple first medical image correction features, and input the second medical image corresponding to the first medical image into the trained deep learning recognition model to obtain multiple second medical image features;

[0087] The second training module is configured to, for the first medical image of each training sample, compare the similarity of each first medical image correction feature of the medical image and each second medical image feature of the second medical image according to the corresponding feature levels, and train the to-be-trained correction model according to the comparison result.

[0088] Optionally, the training device further includes:

[0089] The second feature recognition module is configured to input the obtained medical image to be recognized into the deep learning recognition model in the image recognition model to obtain multiple third medical image features;

[0090] The parameter determination module is configured to input the multiple third medical image features into the correction model in the image recognition model respectively to determine the second correction parameter of each third medical image feature;

[0091] A second correction module, configured to correct each third medical image feature by using the second correction parameter to obtain a corrected third medical image feature;

[0092] An identification module, configured to input multiple corrected third medical image features into a deep learning identification model in the image identification model, and determine an identification result of the medical image to be identified.

[0093] Corresponding to Figure 1 In the training method of the image identification model, an embodiment of the present application further provides a computer device 400, as Figure 4 shown. The device includes a memory 401, a processor 402, and a computer program stored on the memory 401 and executable on the processor 402. When the processor 402 executes the computer program, the training method of the image identification model is implemented.

[0094] Specifically, the memory 401 and the processor 402 can be general memories and processors, which are not specifically limited here. When the processor 402 runs the computer program stored in the memory 401, the training method of the image identification model can be executed, solving the problem of low robustness of the deep identification model for identifying medical images in the prior art.

[0095] Corresponding to Figure 1 In the training method of the image identification model, an embodiment of the present application further provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, the steps of the training method of the image identification model are executed.

[0096] Specifically, the storage medium can be a general storage medium, such as a mobile disk, a hard disk, etc. When the computer program on the storage medium is run, the training method of the image identification model can be executed, solving the problem of low robustness of the deep identification model for identifying medical images in the prior art. After obtaining a deep learning identification model that has been trained using single-center medical image data, if multi-center medical image data needs to be identified, the trained deep learning identification model does not need to be retrained, nor does it need to obtain multi-center medical image data. Only the single-center medical image data is subjected to data augmentation processing, and the first medical image after data augmentation processing is used to train the correction model to be trained to obtain a trained correction model. Finally, the trained deep learning identification model and the trained correction model can be used to obtain an image identification model for identifying multi-center medical images, improving the robustness of the image identification model.

[0097] In the embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling, direct coupling, or communication connection between each other can be through some communication interfaces. The indirect coupling or communication connection of the devices or units can be in electrical, mechanical, or other forms.

[0098] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0099] In addition, each functional unit in the embodiments provided in this application can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0100] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0101] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In addition, the terms "first", "second", "third", etc. are only used for descriptive distinction and cannot be understood as indicating or implying relative importance.

[0102] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present application, used to illustrate the technical solutions of the present application, rather than limiting it. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any person skilled in the technical field can still modify the technical solutions described in the foregoing embodiments or easily conceive of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. All should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A training method for an image recognition model, characterized in that, The training method includes: Obtain a training sample set; For the first medical image of each training sample, input the first medical image into the trained deep learning recognition model to obtain multiple first medical image features. Input each first medical image feature into the correction model to be trained to obtain the first correction parameter of each first medical image feature, and use the first correction parameter to perform correction processing on each first medical image feature to obtain the corrected first medical image feature; wherein, the trained deep learning recognition model is trained based on single-center medical image data; the training sample set is obtained by performing data augmentation processing on the single-center medical image data; For the first medical image of each training sample, input the multiple corrected first medical image features corresponding to the first medical image back into the trained deep learning recognition model to obtain a first recognition result, and train the correction model to be trained based on the difference between the first recognition result and the second recognition result; the second recognition result is recognized by the trained deep learning recognition model from the second medical image corresponding to the first medical image, and the first medical image is obtained by performing data augmentation processing on the second medical image; Based on the trained deep learning recognition model and the trained correction model, obtain a trained image recognition model.

2. The training method according to claim 1, wherein The method further includes: For the first medical image of each training sample, input the multiple corrected first medical image features corresponding to the first medical image into the trained deep learning recognition model to obtain multiple first medical image correction features, and input the second medical image corresponding to the first medical image into the trained deep learning recognition model to obtain multiple second medical image features; For the first medical image of each training sample, compare the similarity of each first medical image correction feature of the medical image and each second medical image feature of the second medical image according to the corresponding feature level, and train the correction model to be trained according to the comparison result.

3. The training method according to claim 1, wherein The training method further includes: Input the obtained medical image to be recognized into the deep learning recognition model in the trained image recognition model to obtain multiple third medical image features; Input the multiple third medical image features into the correction model in the trained image recognition model respectively to determine the second correction parameter of each third medical image feature; For each third medical image feature, use the second correction parameter to correct the third medical image feature to obtain the corrected third medical image feature; Input the multiple corrected third medical image features into the deep learning recognition model in the trained image recognition model to determine the recognition result of the medical image to be recognized.

4. A training device for an image recognition model, characterized in that, The training device includes: An acquisition module, configured to acquire a training sample set; The first correction module is configured to input the first medical image of each training sample into the trained deep learning recognition model to obtain a plurality of first medical image features, input each first medical image feature into the correction model to be trained to obtain a first correction parameter for each first medical image feature, and use the first correction parameter to perform correction processing on each first medical image feature to obtain the corrected first medical image feature; wherein, the trained deep learning recognition model is trained based on single-center medical image data; the training sample set is obtained by performing data augmentation processing on the single-center medical image data; The first training module is configured to input the plurality of corrected first medical image features corresponding to the first medical image of each training sample back into the trained deep learning recognition model to obtain a first recognition result, and train the correction model to be trained based on the difference between the first recognition result and the second recognition result; the second recognition result is recognized by the trained deep learning recognition model from the second medical image corresponding to the first medical image, and the first medical image is obtained by performing data augmentation processing on the second medical image; The combination module is configured to obtain a trained image recognition model based on the trained deep learning recognition model and the trained correction model.

5. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1-3 above.

6. A computer-readable storage medium, on which a computer program is stored, characterized in that, When the computer program is run by the processor, it executes the steps of the method described in any one of claims 1-3 above.

Images