A deep learning method and device based on symmetric cross compound training

Abstract

The present application relates to the technical field of deep learning, in particular to a novel deep learning method and device based on symmetric cross compound training, which specifically comprises: obtaining first sample data, dividing the first sample data into a first training set, a first validation set and a first test set, and further dividing the first training set into a first data set and a second data set; taking the first data set and the second data set alternately as a training set and a validation set to perform symmetric cross training, obtaining outlier samples and true samples; performing compound training on the outlier samples and the true samples, simultaneously performing output correction classification, and obtaining a first optimal model. The present application changes the traditional deep learning training method mode, uses the function of deep learning feature extraction to automatically identify and separate the outlier samples in the training set, and further improves the model performance.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a deep learning method and apparatus based on symmetric cross-compound training. Background Technology

[0002] Thanks to its powerful feature extraction capabilities, deep learning has demonstrated outstanding performance in computer vision. Currently, deep learning typically employs traditional training strategies to obtain suitable neural network models, including: 1) first dividing the dataset into training, validation, and test sets; 2) then training the deep neural network on the training set, using gradient optimization to obtain a parameterized model, and selecting the optimal model on the validation set; 3) finally performing specific tasks, such as classification and prediction, on the test set. However, with limited datasets, outliers in the training set can lead to overfitting, meaning outliers cause the neural network to overlearn, resulting in reduced generalization ability. To address this issue, researchers have proposed various methods, such as soft weight sharing, Dropout, and immediately stopping training when the model's performance on the validation set begins to deteriorate. The last method utilizes the validation set to avoid overfitting in deep learning models, thus obtaining a suitable model. While this method prevents performance degradation, it doesn't address the interference caused by outliers leading to overfitting. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning method and apparatus based on symmetric cross-compound training. By changing the traditional deep learning training method, it utilizes the feature extraction function of deep learning to automatically identify and separate outlier samples in the training set, thereby further improving model performance.

[0004] On the one hand, the present invention provides a deep learning method based on symmetric cross-compound training, the method specifically including:

[0005] Obtain the first sample data, divide the first sample data into a first training set, a first validation set and a first test set, and further divide the first training set into a first dataset and a second dataset;

[0006] The first dataset and the second dataset are used alternately as training and validation sets for symmetric cross-training to obtain outlier samples and true samples. The outlier samples include the first outlier sample and the second outlier sample, and the true samples include the first true sample and the second true sample.

[0007] The outlier samples and the true samples are combined for training, and the output is corrected and classified simultaneously to obtain the first optimal model.

[0008] Furthermore, the step of alternately using the first dataset and the second dataset as training and validation sets for symmetric cross-training to obtain outlier samples and true samples specifically includes:

[0009] The first dataset and the second dataset are used as the second training set and the second validation set, respectively, to train and separate the first outlier sample and the first true sample.

[0010] The second dataset and the first dataset are used as the third training set and the third validation set, respectively, to train and separate the second outlier sample and the second true sample.

[0011] Furthermore, the step of using the first dataset and the second dataset as the second training set and the second validation set, respectively, to train and separate the first outlier sample and the first true sample specifically includes:

[0012] The first dataset is designated as the second training set, and the second dataset is designated as the second validation set.

[0013] The second training set is trained based on the first neural network model, and the second optimal model is obtained based on the second validation set.

[0014] The second optimal model is applied to the second training set to determine the first outlier sample and the first true sample.

[0015] Furthermore, the step of using the second dataset and the first dataset as the third training set and the third validation set, respectively, to train and separate the second outlier samples and the second true samples specifically includes:

[0016] The second dataset is designated as the third training set, and the first dataset is designated as the third validation set.

[0017] The third training set is trained based on the second neural network model, and the third optimal model is obtained based on the third validation set;

[0018] The third optimal model is applied to the third training set to determine the second outlier sample and the second true sample.

[0019] Furthermore, the step of performing combined training on the outlier samples and the true samples, while simultaneously correcting and classifying the output to obtain the first optimal model, specifically includes:

[0020] The outlier samples and the true samples are merged, and then labeled and divided according to true positive samples, true negative samples, false positive samples and false negative samples to obtain the fourth training set;

[0021] The fourth training set is trained based on the third neural network model, and the first optimal model is obtained based on the first validation set.

[0022] The first optimal model is evaluated based on the first test set, and the output is corrected and categorized using a classification formula.

[0023] Furthermore, the fourth training set includes

[0024]

[0025]

[0026] in, This represents the total number of true positive samples. Indicates the total number of true negative samples. This represents the total number of false positive samples. Indicates the total number of false negative samples. This represents the true positive samples in the first set of true samples. This represents the true negative samples in the first set of true samples. This represents a false positive sample in the first outlier sample. This represents a false negative sample in the first outlier sample. This represents the true positive samples in the second set of true samples. This represents the true negative samples in the second set of true samples. This indicates a false positive sample in the second outlier sample. This indicates a false negative sample in the second outlier sample.

[0027] Furthermore, the classification formula satisfies

[0028]

[0029] in, This indicates a positive test result. This indicates a negative test result. This indicates a false positive result in the test. A test result indicates an error, while a test result indicates a positive result.

[0030] On the other hand, the present invention also provides a deep learning device based on symmetric cross-compound training, the device specifically comprising:

[0031] The acquisition module is used to acquire the first sample data, divide the first sample data into a first training set, a first validation set and a first test set, and further divide the first training set into a first dataset and a second dataset.

[0032] The symmetric cross-training module is used to perform symmetric cross-training by alternately using the first dataset and the second dataset as the training set and the validation set to obtain outlier samples and true samples. The outlier samples include the first outlier sample and the second outlier sample, and the true samples include the first true sample and the second true sample.

[0033] The composite training module is used to perform composite training on the outlier samples and the true samples, and at the same time perform output correction and classification to obtain the first optimal model.

[0034] Compared with the prior art, the present invention has at least one of the following technical effects:

[0035] 1. This method changes the training mode of deep learning methods, and automatically distinguishes swarm samples by applying the learning characteristics of deep learning networks, thereby avoiding overfitting of deep learning networks.

[0036] 2. Using outliers as new classification inputs for deep learning networks helps the network learn the features of outliers, enhances the model's generalization ability, and further improves model performance.

[0037] 3. Isolate outlier samples and enable deep learning networks to identify them. Attached Figure Description

[0038] In order to more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating a deep learning method based on symmetric cross-compound training provided in an embodiment of the present invention.

[0040] Figure 2 This is another flowchart illustrating a deep learning method based on symmetric cross-compound training provided in an embodiment of the present invention;

[0041] Figure 3 This is a comparison chart of ROC curves between the traditional method based on the CC-CCII dataset and the method provided in this embodiment of the invention.

[0042] Figure 4 This is a comparison chart of the PR curves of the traditional method based on the CC-CCII dataset and the method provided in this embodiment of the invention;

[0043] Figure 5This is a comparison chart of the confusion matrix between the traditional method based on the CC-CCII dataset and the method provided in the embodiments of this invention.

[0044] Figure 6 This is a comparison chart of ROC curves between the traditional method based on the HUST-19 dataset and the method provided in this embodiment of the invention.

[0045] Figure 7 This is a comparison chart of the PR curves of the traditional method based on the HUST-19 dataset and the method provided in the embodiments of this invention.

[0046] Figure 8 This is a comparison of the confusion matrix between the traditional method tested based on the HUST-19 dataset and the method provided in this embodiment of the invention.

[0047] Figure 9 This is a schematic diagram of the structure of a deep learning device based on symmetric cross-compound training provided in an embodiment of the present invention. Detailed Implementation

[0048] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0049] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0050] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0051] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0052] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0053] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0054] Thanks to its powerful feature extraction capabilities, deep learning has demonstrated outstanding performance in computer vision. Currently, deep learning typically employs traditional training strategies to obtain suitable neural network models, including: 1) first dividing the dataset into training, validation, and test sets; 2) then training the deep neural network on the training set, using gradient optimization to obtain a parameterized model, and selecting the optimal model on the validation set; 3) finally performing specific tasks, such as classification and prediction, on the test set. However, with limited datasets, outliers in the training set can lead to overfitting, meaning outliers cause the neural network to overlearn, resulting in reduced generalization ability. To address this issue, researchers have proposed various methods, such as soft weight sharing, Dropout, and immediately stopping training when the model's performance on the validation set begins to deteriorate. The last method utilizes the validation set to avoid overfitting in deep learning models, thus obtaining a suitable model. While this method prevents performance degradation, it doesn't address the interference caused by outliers leading to overfitting.

[0055] Based on this, refer to Figure 1 This invention provides a deep learning method based on symmetric cross-compound training, the method specifically including:

[0056] S101: Obtain the first sample data, divide the first sample data into a first training set, a first validation set and a first test set, and further divide the first training set into a first dataset and a second dataset.

[0057] In this embodiment, the first sample data is divided into a first training set, a first validation set, and a first test set according to a specific ratio. Then, the first training set is divided into two data subsets in an equal manner, namely a new validation set and a new training set. The case of dividing the first training set into two data subsets in a ratio of 5:5 is preferred, but other ratios are also within the scope of this embodiment.

[0058] S102: The first dataset and the second dataset are used alternately as training and validation sets for symmetric cross-training to obtain outlier samples and true samples. The outlier samples include the first outlier sample and the second outlier sample, and the true samples include the first true sample and the second true sample.

[0059] In some embodiments, the step of alternately using the first dataset and the second dataset as training and validation sets for symmetric cross-training to obtain outlier samples and true samples specifically includes:

[0060] The first dataset and the second dataset are used as the second training set and the second validation set, respectively, to train and separate the first outlier sample and the first true sample.

[0061] The second dataset and the first dataset are used as the third training set and the third validation set, respectively, to train and separate the second outlier sample and the second true sample.

[0062] Specifically, the step of using the first dataset and the second dataset as the second training set and the second validation set for training, respectively, to separate the first outlier sample and the first true sample, specifically includes:

[0063] The first dataset is designated as the second training set, and the second dataset is designated as the second validation set.

[0064] The second training set is trained based on the first neural network model, and the second optimal model is obtained based on the second validation set.

[0065] The second optimal model is applied to the second training set to determine the first outlier sample and the first true sample.

[0066] Specifically, the step of using the second dataset and the first dataset as the third training set and the third validation set, respectively, to train and separate the second outlier samples and the second true samples includes:

[0067] The second dataset is designated as the third training set, and the first dataset is designated as the third validation set.

[0068] The third training set is trained based on the second neural network model, and the third optimal model is obtained based on the third validation set;

[0069] The third optimal model is applied to the third training set to determine the second outlier sample and the second true sample.

[0070] In this embodiment, symmetric cross-training involves selecting the optimal model through iterative training and validation testing based on a new validation set and a new training set. Then, a new validation set and a new training set are exchanged, and the optimal model is selected again through iterative training and validation testing. That is, a deep neural network model is obtained by iterative training on the new training set, and the model performance is verified using the validation set to screen for the optimal model. Outlier samples, i.e. samples that the network model identifies incorrectly, are separated from the new training set using the optimal model. Then, a new training set and a new validation set are exchanged, and the same method is used to screen for the optimal model and separate outlier samples again.

[0071] S103: Perform composite training on the outlier samples and the true samples, and simultaneously perform output correction and classification to obtain the first optimal model.

[0072] In some embodiments, the step of performing combined training on the outlier samples and the true samples, while simultaneously performing output correction and classification to obtain a first optimal model, specifically includes:

[0073] The outlier samples and the true samples are merged, and then labeled and divided according to true positive samples, true negative samples, false positive samples and false negative samples to obtain the fourth training set;

[0074] The fourth training set is trained based on the third neural network model, and the first optimal model is obtained based on the first validation set.

[0075] The first optimal model is evaluated based on the first test set, and the output is corrected and categorized using a classification formula.

[0076] Specifically, the fourth training set includes

[0077]

[0078]

[0079] in, This represents the total number of true positive samples. Indicates the total number of true negative samples. This represents the total number of false positive samples. Indicates the total number of false negative samples. This represents the true positive samples in the first set of true samples. This represents the true negative samples in the first set of true samples. This represents a false positive sample in the first outlier sample. This represents a false negative sample in the first outlier sample. This represents the true positive samples in the second set of true samples. This represents the true negative samples in the second set of true samples. This indicates a false positive sample in the second outlier sample. This indicates a false negative sample in the second outlier sample.

[0080] Specifically, the classification formula satisfies

[0081]

[0082] in, This indicates a positive test result. This indicates a negative test result. This indicates a false positive result in the test. This indicates a negative test result.

[0083] In this embodiment, composite training involves combining outlier samples with correct samples to form a new training set, which is then used by the deep neural network model to learn and acquire the ability to identify outlier samples. Finally, in the output correction and classification stage, the outlier samples identified by the deep neural network need to be correctly classified. That is, these outlier samples and correct samples are combined into a new input to train the deep neural network to obtain a deep learning model with the ability to identify outlier samples. Finally, in the output correction and classification stage, the outlier results (misidentified samples) of the test need to be correctly classified to ensure accurate output.

[0084] Reference Figure 2 The present invention also provides another flowchart of a deep learning method based on symmetric cross-compound training, the specific implementation process of which is as follows:

[0085] S1, partition the entire dataset according to traditional dataset partitioning methods. Divided into training set Validation set and test set ;

[0086] S2, training set Divide into two data subsets in an equal manner and Validation set and test set Reserve it for use in the next stage;

[0087] S3, the dataset As a training set, the dataset As a validation set, suitable deep learning networks are selected for training, and the optimal model is chosen. This model must satisfy the requirement of working effectively on the dataset. The accuracy rate reached its maximum.

[0088] S4, dataset As input to this optimal model, and then based on the dataset The actual labels are used to mark the output, resulting in true positive samples. True negative samples False positive samples False negative samples Among them, false positive samples and false negative samples Collectively referred to as datasets Outlier samples;

[0089] S5, dataset As a training set, the dataset As a validation set, suitable deep learning networks are selected for training, and the optimal model is chosen. This model must satisfy the requirement of working effectively on the dataset. The accuracy rate reached its maximum.

[0090] S6, dataset As input to this optimal model, and then based on the dataset The actual labels are used to mark the output, resulting in true positive samples. True negative samples False positive samples False negative samples Among them, false positive samples and false negative samples Collectively referred to as datasets Outlier samples;

[0091] S7, integrate all samples output by the model according to type, and divide the training set. Divided into positive samples True negative samples False positive samples False negative samples Among them, false positive samples and false negative samples Collectively referred to as datasets Outlier samples:

[0092]

[0093]

[0094] S8. Using the above four types of samples as input to the deep neural network, retrain the deep neural network and select the optimal model. This model must satisfy the condition of operating on the validation set. The accuracy rate reached its maximum.

[0095] S9, Application Test Set To evaluate the performance of this deep network model, it is necessary to perform classification and correction according to the following formula:

[0096]

[0097] in, This represents the positive outlier results output by the network. This represents the negative outlier sample results output by the network.

[0098] The above-described process changes the training paradigm of deep learning methods, automatically distinguishing outlier samples by applying the learning characteristics of deep learning networks, thus avoiding overfitting. Simultaneously, using outlier samples as new classification inputs helps the network learn the features of outliers, enhancing the model's generalization ability and further improving model performance.

[0099] Reference Figures 3 to 5 , Figures 3 to 5 The figures show a comparison of receiver operating characteristic (ROC) curves, precision-recall (PR) curves, and confusion matrices between the traditional method and the method provided in this embodiment of the invention, based on the CC-CCII dataset (Zhang K, Liu X, Shen J, et al. Clinically applicable AI system for accurate diagnosis, quantitative measurements, and prognosis of COVID-19 pneumonia using computed tomography[J]. Cell, 2020, 181(6): 1423-1433. e11.).

[0100] Figure 3 The AUC (Area Under Curve) is an evaluation metric for multi-class classification models, providing a reference for model performance assessment. Specifically, the AUCs of traditional training methods based on four different deep learning models are 0.9948, 0.9915, 0.9969, and 0.9959, respectively, while the AUCs of the symmetric cross-training method based on the four different deep learning models are 0.9969, 0.9951, 0.9993, and 0.9976, respectively. This demonstrates that the overall performance of the model trained by the method provided in this embodiment is higher than that of the model trained by traditional methods.

[0101] Figure 4 In this context, PRE stands for Average Precision. The PREs for the four tests of the traditional method are 0.9903, 0.9840, 0.9841, and 0.9820, respectively. The PREs for the symmetric cross-training method based on four different deep learning models are 0.9922, 0.9907, 0.9987, and 0.9954, respectively. This demonstrates that the model trained by the method provided in this embodiment of the invention has a higher average precision than the model trained by the traditional method.

[0102] Figure 5 In the figures, CP represents patients with ordinary pneumonia, and Normal represents healthy individuals. In Figure a, the correct prediction counts for CP and Normal using the traditional method and the method provided in this embodiment are 5756 (6076) and 5750 (5718), respectively; in Figure b, the correct prediction counts for CP and Normal using the traditional method and the method provided in this embodiment are 5658 (5868) and 5736 (5643), respectively; in Figure c, the correct prediction counts for CP and Normal using the traditional method and the method provided in this embodiment are 5577 (5800) and 5724 (5579), respectively; and in Figure d, the correct prediction counts for CP and Normal using the traditional method and the method provided in this embodiment are 5640 (5837) and 5750 (5634), respectively. Therefore, it can be seen that the method provided in this embodiment has a higher correct prediction count for CP than the traditional method.

[0103] Reference Figures 6 to 8 , Figures 6 to 8 The figures show a comparison of receiver operating characteristic (ROC) curves, precision-recall (PR) curves, and confusion matrices between the traditional method and the method provided in this embodiment of the invention, based on the HUST-19 dataset (Ning W, Lei S, Yang J, et al. Open resource of clinical data from patients with pneumonia for the prediction of COVID-19 outcomes via deep learning[J]. Nature biomedicalengineering, 2020, 4(12): 1197-1207.).

[0104] Figure 6In the above, the AUCs of traditional training methods based on five different deep learning models are 0.9991, 0.9663, 0.9997, 0.9948 and 0.9981, respectively, while the AUCs of symmetric cross-training methods based on five different deep learning models are 1.0000, 0.9800, 0.9999, 0.9999 and 0.9998, respectively. It can be seen that the overall performance of the model trained by the method provided in this embodiment of the invention is higher than that of the model trained by traditional methods.

[0105] Figure 7 In the traditional training methods based on five different deep learning models, the PREs were 0.9982, 0.9370, 0.9994, 0.9899, ​​and 0.9965, respectively. The PREs of the method provided in this embodiment of the invention were 0.9999, 0.9639, 0.9998, 0.9999, and 0.9996, respectively. It can be seen that the model obtained by the symmetric cross-training method based on five different deep learning models has a higher average accuracy than the model trained by the traditional method.

[0106] Figure 8 In the diagram, pCT represents a positive CT image, nCT represents a negative CT image, and NiCT represents a CT image with no information. In Figure a, the correct prediction counts for pCT, nCT, and NiCT provided by the traditional method and the method provided by the embodiments of the present invention are 789 (798), 1901 (1909), and 1091 (1127), respectively; in Figure b, the correct prediction counts for pCT, nCT, and NiCT provided by the traditional method and the method provided by the embodiments of the present invention are 647 (654), 1567 (1720), and 1119 (1120), respectively; in Figure c, the correct prediction counts for pCT, nCT, and NiCT provided by the traditional method and the method provided by the embodiments of the present invention are... The correct prediction counts for pCT, nCT, and NiCT were 788 (798), 1897 (1908), and 1115 (1128), respectively. In Figure d, the correct prediction counts for pCT, nCT, and NiCT provided by the traditional method and the method provided by the present invention were 744 (797), 1843 (1911), and 1132 (1125), respectively. In Figure e, the correct prediction counts for pCT, nCT, and NiCT provided by the traditional method and the method provided by the present invention were 739 (788), 1904 (1905), and 1124 (1122), respectively. It can be seen that the method provided by the present invention has a higher correct prediction count for pCT and nCT than the traditional method, and its correct prediction count for NiCT is also higher than the traditional method in 3 out of 5 tests.

[0107] Reference Figure 9This invention also provides a deep learning device 9 based on symmetric cross-training, the device 9 specifically comprising: an acquisition module 901, used to acquire first sample data, divide the first sample data into a first training set, a first validation set, and a first test set, and further divide the first training set into a first dataset and a second dataset; a symmetric cross-training module 902, used to alternately use the first dataset and the second dataset as training and validation sets for symmetric cross-training to obtain outlier samples and true samples, the outlier samples including a first outlier sample and a second outlier sample, and the true samples including a first true sample and a second true sample; and a composite training module 903, used to perform composite training on the outlier samples and the true samples, and simultaneously perform output correction and classification to obtain a first optimal model.

[0108] It is understandable that, such as Figure 1 and Figure 2 The content of the deep learning method embodiment based on symmetric cross-compound training shown is applicable to the deep learning device embodiment based on symmetric cross-compound training. The specific functions implemented by the deep learning device embodiment based on symmetric cross-compound training are as follows: Figure 1 and Figure 2 The illustrated deep learning method based on symmetric cross-compound training is the same as the one shown, and achieves the same beneficial effects. Figure 1 and Figure 2 The beneficial effects achieved by the deep learning method based on symmetric cross-compound training shown in the example are also the same.

[0109] It should be noted that the information interaction and execution process between the above-mentioned devices are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0111] This invention also provides a computer-readable storage medium, characterized in that it stores a computer program thereon, which, when executed by a processor, implements the deep learning method based on symmetric cross-compound training as described in any of the above methods.

[0112] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0113] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0114] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0115] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0116] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A deep learning method based on symmetric cross-compound training, characterized in that, The method specifically includes: Obtain the first sample data, divide the first sample data into a first training set, a first validation set and a first test set, and further divide the first training set into a first dataset and a second dataset. The first sample data is either the CC-CCII dataset or the HUST-19 dataset. The first training set includes positive CT images, negative CT images and CT images with no information. The first dataset and the second dataset are used alternately as training and validation sets for symmetric cross-training to obtain outlier samples and true samples. The outlier samples include the first outlier sample and the second outlier sample, and the true samples include the first true sample and the second true sample. The model is trained by combining the outlier samples and the true samples, and the output is corrected and classified simultaneously to obtain the first optimal model. Specifically, this includes: The outlier samples and the true samples are merged, and then labeled and divided according to true positive samples, true negative samples, false positive samples and false negative samples to obtain the fourth training set; The fourth training set is trained based on the third neural network model, and the first optimal model is obtained based on the first validation set. The first optimal model is evaluated based on the first test set, and the output is corrected and categorized using a classification formula. Wherein, the classification formula satisfies ； ； in, This indicates a positive test result. This indicates a negative test result. This represents the positive outlier results output by the network. This represents the negative outlier sample results output by the network.

2. The method according to claim 1, characterized in that, The step of alternately using the first dataset and the second dataset as training and validation sets for symmetrical cross-training to obtain outlier samples and true samples specifically includes: The first dataset and the second dataset are used as the second training set and the second validation set, respectively, to train and separate the first outlier sample and the first true sample. The second dataset and the first dataset are used as the third training set and the third validation set, respectively, to train and separate the second outlier sample and the second true sample.

3. The method according to claim 2, characterized in that, The step of using the first dataset and the second dataset as the second training set and the second validation set, respectively, to train and separate the first outlier sample and the first true sample specifically includes: The first dataset is designated as the second training set, and the second dataset is designated as the second validation set. The second training set is trained based on the first neural network model, and the second optimal model is obtained based on the second validation set. The second optimal model is applied to the second training set to determine the first outlier sample and the first true sample.

4. The method according to claim 2, characterized in that, The step of using the second dataset and the first dataset as the third training set and the third validation set, respectively, to train and separate the second outlier sample and the second true sample specifically includes: The second dataset is designated as the third training set, and the first dataset is designated as the third validation set. The third training set is trained based on the second neural network model, and the third optimal model is obtained based on the third validation set; The third optimal model is applied to the third training set to determine the second outlier sample and the second true sample.

5. The method according to claim 1, characterized in that, The fourth training set includes ； ； in, This represents the total number of true positive samples. Indicates the total number of true negative samples. This represents the total number of false positive samples. Indicates the total number of false negative samples. This represents the true positive samples in the first set of true samples. This represents the true negative samples in the first set of true samples. This represents a false positive sample in the first outlier sample. This represents a false negative sample in the first outlier sample. This represents the true positive samples in the second set of true samples. This represents the true negative samples in the second set of true samples. This indicates a false positive sample in the second outlier sample. This indicates a false negative sample in the second outlier sample.

6. A deep learning device based on symmetric cross-compound training, characterized in that, The device specifically includes: The acquisition module is used to acquire first sample data, divide the first sample data into a first training set, a first validation set and a first test set, and further divide the first training set into a first dataset and a second dataset. The first sample data is either the CC-CCII dataset or the HUST-19 dataset. The first training set includes positive CT images, negative CT images and CT images without information. The symmetric cross-training module is used to perform symmetric cross-training by alternately using the first dataset and the second dataset as the training set and the validation set to obtain outlier samples and true samples. The outlier samples include the first outlier sample and the second outlier sample, and the true samples include the first true sample and the second true sample. The composite training module is used to perform composite training on the outlier samples and the true samples, while simultaneously correcting and classifying the output to obtain the first optimal model. Specifically, it includes: The outlier samples and the true samples are merged, and then labeled and divided according to true positive samples, true negative samples, false positive samples and false negative samples to obtain the fourth training set; The fourth training set is trained based on the third neural network model, and the first optimal model is obtained based on the first validation set. The first optimal model is evaluated based on the first test set, and the output is corrected and categorized using a classification formula. Wherein, the classification formula satisfies ； ； in, This indicates a positive test result. This indicates a negative test result. This represents the positive outlier results output by the network. This represents the negative outlier sample results output by the network.

Images