Medical data set optimization method and equipment

Abstract

This application relates to a medical dataset optimization method and device, specifically to the field of computer vision technology. The method includes: obtaining a medical dataset; for each medical sample data, processing the medical sample data using a neural network model to obtain the entropies corresponding to the medical sample data; constructing the entropies corresponding to each medical sample data into an entropy matrix; dividing the medical dataset to obtain medical data subsets; for each medical data subset, calculating the diversity, coverage, and balance of samples in the medical data subset using the entropy matrix, and updating the medical data subset to obtain a target data subset; and constructing an optimized medical dataset based on each target data subset. The above solution improves the training effect of the neural network model training using the optimized medical dataset.

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a medical data set optimization method and device. Background Art

[0002] Deep learning (DL) is a new research direction in the field of machine learning (ML). Deep learning studies the inherent patterns and representational hierarchies of sample data. Its ultimate goal is to enable machines to have human-like analytical learning capabilities and recognize data such as text, images, and sounds.

[0003] Deep learning is a general term for a class of pattern analysis methods. Specifically, it encompasses three main approaches: convolutional neural network systems, known as convolutional neural networks (CNNs); autoencoder neural networks based on multi-layer neurons, including autoencoders and the recently popular sparse coding method (sparse coding); and deep belief networks (DBNs), which use multi-layer autoencoder neural networks for pre-training and then further optimize the neural network weights by incorporating discriminant information. Compared to traditional learning methods, deep learning methods pre-set more model parameters, making model training more difficult. According to the general principles of statistical learning, the more model parameters a model has, the greater the amount of data required for training.

[0004] The use of deep learning technology often relies on a large number of samples to train the model. Given the large size of the required dataset, it is very easy for redundant samples and noise samples to exist, that is, samples that do not contribute to or have a negative impact on the optimization of the deep model, resulting in low training efficiency of the machine learning model. Summary of the Invention

[0005] This application provides a medical dataset optimization method and device to improve the training efficiency of machine learning models. The technical solution is as follows.

[0006] In one aspect, a method for optimizing a medical data set is provided, the method comprising:

[0007] Acquire a medical data set; the medical data set includes various medical sample data;

[0008] For each medical sample data, the medical sample data is processed by each neural network model to obtain each entropy corresponding to the medical sample data; the neural network model is used to perform inference calculation on the medical sample data;

[0009] Constructing the entropies corresponding to the medical sample data into an entropy matrix;

[0010] Dividing the medical data set to obtain various medical data subsets;

[0011] For each medical data subset, the diversity, coverage, and balance of samples in the medical data subset are calculated using the entropy matrix, and the medical data subset is updated based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain a target data subset;

[0012] Based on each of the target data subsets, an optimized medical data set is constructed.

[0013] In another aspect, a medical data set optimization device is provided, the device comprising:

[0014] A medical data set acquisition module is used to acquire a medical data set; the medical data set includes various medical sample data;

[0015] a sample data processing module, configured to process each medical sample data using respective neural network models to obtain respective entropies corresponding to the medical sample data; the neural network models are configured to perform inference calculations on the medical sample data;

[0016] An entropy matrix construction module, configured to construct an entropy matrix from the entropies corresponding to the medical sample data;

[0017] A data subset division module, configured to divide the medical data set into subsets to obtain medical data;

[0018] a data subset updating module, configured to calculate, for each medical data subset, the diversity, coverage, and balance of samples in the medical data subset using the entropy matrix, and update the medical data subset based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain a target data subset;

[0019] The data set optimization module is used to construct an optimized medical data set based on each of the target data subsets.

[0020] In one possible implementation, the entropy matrix building module is used to

[0021] Taking the entropies corresponding to the medical sample data as elements to generate the entropy matrix;

[0022] The number of rows of the entropy matrix is ​​the number of the medical sample data; the number of columns of the entropy matrix is ​​the number of the neural network models.

[0023] In a possible implementation, the data subset updating module is further configured to:

[0024] Reading the element value corresponding to each medical sample data in the medical data subset in the entropy matrix;

[0025] The ratio of the sum of the element values ​​corresponding to each medical sample data in the medical data subset to the number of neural network models is determined as the diversity of the medical sample data in the medical data subset.

[0026] In a possible implementation, the entropy matrix construction module is further used to:

[0027] Converting the elements of the entropy matrix that are greater than the target sliding factor to 1, and converting the elements of the entropy matrix that are less than the target sliding factor to 0, to obtain a change matrix;

[0028] For each neural network model, obtaining, in the change matrix, elements obtained by the neural network model processing each medical sample data in the medical data subset, and summing the elements to obtain a response count;

[0029] Calculating the average of each element in the change matrix to obtain a response average;

[0030] Calculating a standard deviation between the response mean and the response count to obtain a response deviation corresponding to the neural network model;

[0031] summing the ratio of the response count to the response deviation according to the category of the neural network model to obtain the balance of the medical data subset;

[0032] The ratio of the response count to the number of elements of the change matrix is ​​summed according to the category of the neural network model to obtain the coverage of the medical data subset.

[0033] In a possible implementation, the data subset updating module is further configured to:

[0034] Performing weighted summation based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain the fitness corresponding to the medical data subset;

[0035] The medical data subset is iteratively updated according to the fitness of the medical data subset, and the iteratively updated medical data subset is determined as the target data subset.

[0036] In a possible implementation, the data subset updating module is further configured to:

[0037] In the Nth round of iterative updating, at least one medical sample data in the medical data subset is replaced with medical sample data other than the medical data subset to obtain an intermediate medical data subset, where the initial value of N is 1;

[0038] Calculating the fitness of the intermediate medical data subset according to the entropy matrix, and determining the intermediate medical data subset and the medical data subset with the higher fitness as the medical data subset after the Nth round of iterative update;

[0039] If N is less than the preset number of rounds, the operation of replacing at least one medical sample data in the medical data subset with medical sample data outside the medical data subset is performed again;

[0040] If N reaches the preset number of rounds, the medical data subset after the Nth round of iterative update is determined as the target data subset.

[0041] In one possible implementation, each neural network model includes m small replacement models and n checkpoint models; the m small replacement models are models obtained by reducing the number of convolutional layers and / or pooling layers of the target initial model, or the m small replacement models are network models with a training accuracy less than a preset threshold; the n checkpoint models have the same model structure as the target initial model;

[0042] The target initial model is a neural network model that performs inference calculations on medical sample data.

[0043] In a possible implementation, the apparatus further includes:

[0044] A model reduction module, configured to reduce the number of convolutional layers and / or pooling layers of the target initial model to obtain m small initial models;

[0045] An initial training module, configured to train m small initial models according to the medical dataset to obtain m trained small replacement models;

[0046] A data set selection module is used to perform P rounds of pre-training on the target initial model based on the medical data set, and select a training data set from the medical data set based on the pre-trained target neural network model;

[0047] A model parameter acquisition module is used to perform T rounds of iterative training on the target initial model according to the training data set, and obtain the model parameters of the target initial model in the last n rounds of iterations;

[0048] The checkpoint model acquisition module is used to construct the n checkpoint models based on the model parameters of the target initial model in the last n rounds of iterations.

[0049] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the above-mentioned medical data set optimization method.

[0050] On the other hand, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned medical dataset optimization method.

[0051] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the above-described medical dataset optimization method.

[0052] The technical solution provided by this application may have the following beneficial effects:

[0053] Before training a neural network model for processing medical data, a computer device first obtains each medical sample data in the medical data set, processes each medical sample data using each neural network model, and constructs an entropy matrix based on the processing results. The computer device also divides the medical data set into medical data subsets and calculates the diversity, balance, and coverage of the medical data subsets using the entropy matrix to update each medical data subset, thereby obtaining each target data subset. The target data subsets are then combined to obtain an optimized medical data set. In the above scheme, the medical sample data are processed by multiple neural network models, and the entropy matrices corresponding to each medical sample data and the neural network model are calculated. The diversity, balance, and coverage are calculated based on the information content reflected in the entropy matrices for each medical data subset, thereby better assessing the quality of the medical sample data in each medical data subset. The medical data subsets are then updated to obtain each target data subset, which are then combined to obtain the optimized medical data set. This improves the matching degree between each medical sample data in the optimized medical data set and the neural network model, thereby improving the training effect of the neural network model trained using the optimized medical data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the specific implementation methods or the description of the prior art. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0055] Figure 1 It is a structural diagram of a model training system according to an exemplary embodiment.

[0056] Figure 2 FIG. 4 is a flowchart of a method for optimizing a medical data set according to an exemplary embodiment.

[0057] Figure 3 FIG. 4 is a flowchart of a method for optimizing a medical data set according to an exemplary embodiment.

[0058] Figure 4 A schematic diagram of pre-training of a machine learning model involved in an embodiment of the present application is shown.

[0059] Figure 5 A schematic diagram showing the relationship between training time and accuracy in deep learning is shown.

[0060] Figure 6 A schematic diagram showing the relationship between the number of model network layers and accuracy is shown.

[0061] Figure 7 A logical framework diagram of a data set optimization involved in an embodiment of the present application is shown.

[0062] Figure 8 It is a structural block diagram of a data set optimization device according to an exemplary embodiment.

[0063] Figure 9 It is a schematic diagram of a computer device provided according to an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solution of this application in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0065] It should be understood that the "indication" mentioned in the embodiments of this application can be a direct indication, an indirect indication, or an indication of an association. For example, "A indicates B" can mean that A directly indicates B, for example, B can be obtained through A; it can also mean that A indirectly indicates B, for example, A indicates C, and B can be obtained through C; it can also mean that there is an association between A and B.

[0066] In the description of the embodiments of the present application, the term "corresponding" may indicate a direct or indirect correspondence between the two, or an association relationship between the two, or a relationship between indication and being indicated, configuration and being configured, etc.

[0067] In an embodiment of the present application, "predefinition" can be achieved by pre-saving corresponding codes, tables or other methods that can be used to indicate relevant information in a device (for example, including a terminal device and a network device). This application does not limit its specific implementation method.

[0068] Before describing the various embodiments shown in this application, several concepts involved in this application are first introduced.

[0069] 1) AI (Artificial Intelligence)

[0070] Artificial Intelligence (AI) is a new technical discipline that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. A branch of computer science, AI seeks to understand the essence of intelligence and create new intelligent machines that can respond in a manner similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. Since its inception, AI's theories and technologies have matured, and its application areas have continued to expand. It is conceivable that future AI-powered technological products will serve as "containers" of human wisdom. AI can simulate human consciousness and the information processes of thought. While not human intelligence, AI can think like humans and may even surpass human intelligence.

[0071] Computers are the primary material foundation for AI research and the machines that enable AI technology platforms. In addition to computer science, AI encompasses a wide range of disciplines, including information theory, cybernetics, automation, bionics, biology, psychology, mathematical logic, linguistics, medicine, and philosophy. Key areas of AI research include knowledge representation, automated reasoning and search methods, machine learning and knowledge acquisition, knowledge processing systems, natural language understanding, computer vision, intelligent robotics, and automated programming.

[0072] 2) Machine Learning (ML)

[0073] Machine learning is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0074] The medical data set optimization method provided in the embodiment of the present application can be applied to computer equipment with strong data processing capabilities. The medical data set optimization method can optimize the medical data set before formally training the machine learning model, so as to train the machine learning model through the optimized medical data set. Optionally, the machine learning model can be a medical image classification model that can process the input medical image to obtain a category corresponding to the input medical image. In one possible implementation, the medical image classification model provided in the embodiment of the present application can be applied to a personal computer, a workstation or a server, that is, the medical image classification model is trained or applied by at least one of a personal computer, a workstation and a server.

[0075] Figure 1 1 is a schematic diagram illustrating the structure of a model training system according to an exemplary embodiment. Optionally, the model training system includes a server 110 and a terminal 120. The terminal 120 and the server 110 communicate data via a communication network, which can be a wired network or a wireless network.

[0076] Optionally, when the model training system is used to identify medical images in the medical field, an application with image processing function is installed in the terminal 120. The application can be a professional image processing application, a medical application, or an AI application with image processing function. The embodiments of the present application are not limited to this.

[0077] Optionally, the terminal 120 can be a terminal device having an image acquisition component, which is used to acquire images and store them in a data storage module in the terminal 120; the terminal 120 can also be a terminal device having a data transmission interface, which is used to receive image data captured by an image acquisition device having an image acquisition component.

[0078] Optionally, the terminal 120 can be a mobile terminal such as a smart phone, a tablet computer, a laptop computer, or a similar mobile terminal, or a terminal such as a desktop computer, a projection computer, or a smart terminal with a data processing component, and the embodiment of the present application does not impose any restrictions on this.

[0079] The server 110 may be implemented as a single server or a server cluster consisting of a group of servers. The server 110 may be a physical server or a cloud server. In one possible implementation, the server 110 is a background server of an application in the terminal 120 .

[0080] In one possible implementation of the present embodiment, server 110 trains a medical image classification model using a pre-set medical dataset (including various medical sample images). The medical dataset may include features extracted from medical sample images of different categories, each of which has its own category annotation information (i.e., label value). After server 110 completes training of the medical image classification model, it transmits the trained medical image classification model to terminal 120 via a wired or wireless network.

[0081] In an embodiment of the present application, before training the medical image classification model, the server also needs to optimize the medical data set composed of each medical sample image, so that the training effect of the machine learning model in the optimized medical data set is better.

[0082] The terminal 120 receives the trained medical image classification model and sends the data information (such as weight information) corresponding to the medical image classification model to an application with recognition and classification functions, so that when the user uses the application, the input features can be recognized and classified.

[0083] Optionally, the above-mentioned server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), as well as big data and artificial intelligence platforms and other technical computing services.

[0084] Optionally, the system may further include a management device for managing the system (such as managing the connection status between each module and the server, etc.), and the management device and the server are connected via a communication network. Optionally, the communication network is a wired network or a wireless network.

[0085] Optionally, the above-mentioned wireless network or wired network uses standard communication technology and / or protocol. The network is typically the Internet, but may also be any other network, including but not limited to any combination of a local area network, a metropolitan area network, a wide area network, a mobile, a limited or wireless network, a private network, or a virtual private network. In some embodiments, the data exchanged over the network is represented using technologies and / or formats including hypertext markup language, extensible markup language, etc. In addition, conventional encryption technologies such as secure socket layer, transport layer security, virtual private network, internet protocol security, etc. may be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies may also be used to replace or supplement the above-mentioned data communication technologies.

[0086] Figure 2 is a flowchart of a method for optimizing a medical data set according to an exemplary embodiment. The method is executed by a computer device, which may be a computer device such as Figure 1 The server or terminal in the model training system shown in FIG. Figure 2 As shown, the medical data set optimization method may include the following steps:

[0087] Step 201: Acquire a medical data set.

[0088] In the embodiment of the present application, the medical data set includes various medical sample data.

[0089] Optionally, the medical sample data may be at least one of medical image data and scan video data.

[0090] When the medical sample data is medical image data, according to the scheme shown in the embodiment of the present application, the medical data set composed of the medical image data can be optimized so that the neural network model can be trained based on the optimized medical data set, so that the trained neural network model can process the medical image data (for example, to perform classification, labeling and other tasks).

[0091] Similarly, when the medical sample data is scanned video data, according to the scheme shown in the embodiment of the present application, the medical data set composed of the scanned video data can be optimized so that the neural network model can be trained based on the optimized medical data set, so that the trained neural network model can process the scanned video data (for example, to achieve classification, labeling and other tasks).

[0092] Step 202 : For each medical sample data, process the medical sample data through each neural network model to obtain each entropy corresponding to the medical sample data.

[0093] The neural network model is used to perform inference calculations on the medical sample data. For example, the neural network model can be a ResNet (residual network) model, such as a Resnet50 model.

[0094] In one possible implementation, since ResNet (residual network) can have different numbers of model layers, and residual networks with different numbers of layers have different feature extraction depths, at least two neural network models with different numbers of layers can be used to process the medical sample data separately, thereby obtaining the processing results of the medical sample data in neural network models with different structures, and calculating the entropy obtained after the medical sample data is processed by neural network models with different structures.

[0095] In one possible implementation, the neural network model may be a neural network model that needs to be trained. The neural network model may be a neural network model with initial parameters set, or the neural network model may be a pre-trained neural network model.

[0096] At this time, after the medical sample data is processed by the neural network model, the entropy of each medical sample data obtained represents the amount of information obtained by the neural network model when extracting the medical sample data.

[0097] Optionally, the entropy of the medical sample data is calculated as follows:

[0098]

[0099] Where x is the data sample, is the probability distribution, Represents the machine learning model participating in training.

[0100] Step 203: construct an entropy matrix by using the entropies corresponding to the medical sample data.

[0101] Optionally, when obtaining the entropy values ​​of each medical sample data after processing it through each neural network model, for one of the medical sample data, the ten entropy values ​​obtained by passing the medical sample data through each neural network model (taking the number of models as 10 as an example) can be used as the ten elements of one row of the entropy matrix. Therefore, ten entropy values ​​can be calculated for each medical sample data. At this time, the entropy values ​​are sorted according to the number of the medical sample data and the number of the neural network model (for example, the medical sample data number is used as the row number and the neural network model number is used as the column number), and the entropy matrix can be obtained.

[0102] Step 204: divide the medical data set to obtain various medical data subsets.

[0103] The computer device contains a medical data set, which contains various medical sample data. At this time, the computer device arbitrarily obtains N medical sample data from the medical data set to construct a medical sample sequence, that is, a medical data subset.

[0104] In a possible implementation, the medical sample data in each medical data subset are different from each other, that is, each medical sample sequence does not overlap with each other.

[0105] Step 205: For each medical data subset, the diversity, coverage and balance of the samples in the medical data subset are calculated using the entropy matrix, and the medical data subset is updated based on the diversity, coverage and balance of the medical sample data in the medical data subset to obtain the target data subset.

[0106] Since in the above steps, the entropy obtained after each medical sample data in the medical data set is processed by each neural network model is calculated, the entropy matrix formed is a matrix in which each element can actually represent the amount of information obtained after the neural network model processes the medical sample data.

[0107] At this time, since each medical data subset contains different medical sample data, based on the medical sample data, the corresponding value of each medical sample data in each medical data subset in the entropy matrix can be queried, thereby determining the diversity, coverage and balance of the medical sample data in the medical data subset, and comprehensively considering the quality of the samples in the medical data subset to determine whether the samples in the medical data subset need to be updated.

[0108] For example, when the diversity, coverage, and balance corresponding to a certain medical data subset are obtained, and the sum of the values ​​corresponding to the diversity, coverage, and balance indicates that the samples in the medical data subset are not in good condition, then some samples in the medical data subset can be replaced with other samples (such as random replacement), thereby updating the samples in the medical data subset.

[0109] Step 206: construct an optimized medical data set based on each of the target data subsets.

[0110] After obtaining each target data subset, the target data subset is now a set optimized based on the neural network model. Therefore, the medical data set merged from each target data subset is used to train the machine learning model, which has a better data training effect than the original medical data set, thereby improving the training efficiency of the machine learning model.

[0111] In summary, before training a neural network model for processing medical data, a computer device first obtains each medical sample data in the medical data set, processes each medical sample data separately through each neural network model, and constructs an entropy matrix based on the processing results. The computer device also divides the medical data set into medical data subsets and calculates the diversity, balance, and coverage of the medical data subsets using the entropy matrix to update each medical data subset, thereby obtaining each target data subset. At this point, the target data subsets are combined to obtain an optimized medical data set. In the above scheme, the medical sample data are processed by multiple neural network models, and the entropy matrices corresponding to each medical sample data and the neural network model are calculated. The diversity, balance, and coverage are calculated based on the information content reflected in the entropy matrices for each medical data subset, thereby better evaluating the quality of the medical sample data in each medical data subset. The medical data subsets are then updated to obtain each target data subset, which are then combined to obtain the optimized medical data set. This improves the matching degree between each medical sample data in the optimized medical data set and the neural network model, thereby improving the training effect of the neural network model training using the optimized medical data set.

[0112] Figure 3 is a flowchart of a method for optimizing a medical data set according to an exemplary embodiment. The method is executed by a computer device, which may be a computer device such as Figure 1 The server or terminal in the model training system shown in FIG. Figure 3 As shown, the medical data set optimization method may include the following steps:

[0113] Step 301: Acquire a medical data set.

[0114] In an embodiment of the present application, the medical data set is an initial set of medical sample data that needs to be used to train the neural network model. Before training the neural network model using the individual medical sample data in the medical data set, the initial medical data set can be split into individual medical sample subsets, so that each medical sample subset can be processed through subsequent steps, thereby optimizing the initial medical data set and improving the training efficiency of the neural network model.

[0115] In step 302 , each medical sample data is processed by each neural network model to obtain entropies corresponding to the medical sample data.

[0116] In a possible implementation of an embodiment of the present application, each neural network model includes m small replacement models and n checkpoint models; the m small replacement models are models obtained by reducing the number of convolutional layers and / or pooling layers of the target initial model, or the m small replacement models are network models with a training accuracy less than a preset threshold; the n checkpoint models have the same model structure as the target initial model; the target initial model is a neural network model that performs inference calculations on medical sample data.

[0117] Since the increase in the number of layers of the ResNet model has little effect on the accuracy, it takes a lot of time to train the model. Figure 4 As shown, it shows a schematic diagram of the relationship between the number of model network layers and accuracy.

[0118] like Figure 5 As shown, it shows a schematic diagram of the relationship between training time and accuracy in deep learning. In the actual training process of the neural network model, the model can reach a top1 error rate of 9% in the 14th minute, while the next 12 minutes of training only reduces the top1 error rate of the model to 7.6%. Therefore, the longer the training time, the more iterations there are. Although the training accuracy is improved, the training efficiency is significantly reduced. Therefore, a training method for the target initial model is also proposed in the embodiment of the present application, which can ensure the training effect as much as possible while reducing the number of training rounds and improve the training efficiency.

[0119] Therefore, in a possible implementation of the embodiment of the present application, the neural network model is designed to include two types of neural network models, module A and module B, and module A is integrated into a small replacement model of m ResNets through the above two methods. This significantly reduces the training time of the integrated model on the entire dataset and also allows us to obtain the information responses of m small surrogate models to each sample in the dataset. Module B is the checkpoint integration of a single large ResNet model.

[0120] When the neural network models are of different types, that is, multiple neural network models, when calculating the entropy of each medical sample data, it is necessary to calculate the entropy values ​​obtained when the medical sample data is processed by multiple neural network models. That is, the information amount α of each sample corresponding to different models is obtained. j,e Construct an entropy matrix, namely Ent=[α j,e ] N,(m+n) , where each row is a sample in the dataset D, with a total of N rows, and each column is a model of the integrated model in step 1, with a total of m+n columns.

[0121] According to the rank-order correlation between the original model and its small replacement model, in the embodiment of the present application, a trade-off between time and model inference accuracy can be achieved by constructing a small replacement model, that is, m small replacement models are constructed, thereby reducing the accuracy of a small part but saving a lot of training time.

[0122] In one possible implementation, the number of convolutional layers and / or pooling layers of the target initial model is reduced to obtain m small initial models;

[0123] According to the medical data set, m small initial models are trained respectively to obtain m trained small replacement models;

[0124] Based on the medical dataset, the target initial model is pre-trained for P rounds, and a training dataset is selected from the medical dataset based on the pre-trained target neural network model;

[0125] Based on the training data set, the target initial model is trained for T rounds of iterative training, and the model parameters of the target initial model in the last n rounds of iterations are obtained;

[0126] The n checkpoint models are constructed based on the model parameters of the target initial model in the last n iterations.

[0127] That is to say, in the process of generating the Checkpoint model, since the number of model convolution layers and / or pooling layers in the Checkpoint model is large, it takes a lot of time to train n Checkpoint models in sequence. Therefore, in an embodiment of the present application, a training data set can be selected for the Checkpoint model, and the neural network model can be iteratively trained for T rounds based on the training data set, and the model parameters of the target initial model in the last n rounds of iterative training can be saved. For the target initial model in the last n rounds of iterative training, the model parameters will be updated in each round of iteration. Therefore, based on the model parameters after each update, a checkpoint model with the same structure as the target initial model but different parameters can be constructed.

[0128] In a possible implementation, a first number of randomly selected medical data in the medical dataset is moved out to a training dataset;

[0129] Training the target initial model using the first amount of medical data;

[0130] Obtain the information volume of each medical data in the medical data set;

[0131] According to the acquired information volume, medical data that meets a preset condition in the medical data set is removed from the medical data set to the training data set, and the target initial model is retrained using the medical data that meets the preset condition, where the preset condition includes that the information volume ranks in the top m or the information volume is greater than a predetermined threshold;

[0132] If the sample size of the training data set is less than the preset value, the step of obtaining the information volume of each medical data in the medical data set is performed again;

[0133] If the sample size of the training data set reaches the preset value, the current training data set is determined as the final selected training data set, and the preset value is less than the total amount of medical data in the medical data set.

[0134] The above iterative training process is described below by taking the first number Ns / 8 as an example.

[0135] Please refer to Figure 6 , which shows a schematic diagram of a neural network model pre-training involved in the embodiment of the present application. Figure 6 As shown, in order to reduce the training time of the model on the entire data set, the embodiment of the present application adopts an incremental data reading method. where each x j ∈X is a data sample, each y j ∈Y is its corresponding label. acq It is composed of a single ResNet model n Checkpoints (a set of n different model parameters ), that is, the query model obtained by integrating N neural network models with the same structure but different parameters (that is, m small replacement models and n checkpoint models), The training set selected for the above-mentioned incremental data reading method, where N s For training M acq First, N samples are uniformly randomly sampled from the medical dataset D. s / 8 samples as the initial training sample S, and use the above samples to train the model Entropy is used as the acquisition function, that is, Calculate the information amount α of each sample in the data set D, and select the N samples with the largest α in the data set D. s / 8 samples are moved into the training set S. If the sample size of the dataset S is less than N s , then continue to the next iteration until the sample size of the data set S is greater than or equal to N s Stop. By using the above data set reading method, it is only necessary to iterate twice on the basis of the initial training sample S to obtain the final data set S. The final model is trained based on this data set S. For example, in one possible embodiment, the first number is Ns / 8, and the numbers selected in each subsequent iteration are Ns / 8, Ns / 4, and Ns / 2, respectively. After a cumulative total of four iterations, the final dataset S is obtained. The above is merely an example of the iterative approach. In actual implementation, the number of selected datasets may be greater than S at the end of the iteration, and this embodiment does not impose any limitation thereto.

[0136] Optionally, the first number in the above scheme can also be Ns / 4, Ns / 2 or other numbers. Taking the first number as Ns / 4 as an example, Ns / 4 samples are selected as initial training samples through the above scheme, and the model is trained using the above samples. Entropy is used as the acquisition function, that is, Calculate the information amount α of each sample in the data set D, and select the N samples with the largest α in the data set D. s / 4 samples are moved into the training set S. At this time, only two more iterations are required based on the initial training sample S to obtain the final data set S.

[0137] Similarly, when the first number in the above solution is other numbers, it is only necessary to iterate the above steps until the sample size of the training data set meets the requirements (such as the number is greater than the product of Ns and the specified ratio).

[0138] Step 303: construct an entropy matrix by using the entropies corresponding to the medical sample data.

[0139] In a possible implementation, the entropies corresponding to the medical sample data are used as elements to generate the entropy matrix; the number of rows of the entropy matrix is ​​the number of the medical sample data; and the number of columns of the entropy matrix is ​​the number of the neural network models.

[0140] Alternatively, in another possible implementation, the number of rows of the entropy matrix is ​​the number of the neural network models, in which case the row numbers of the entropy matrix are the numbers of the neural network models, and the number of columns of the entropy matrix is ​​the number of the medical sample data, in which case the column numbers of the entropy matrix are the numbers of the medical sample data.

[0141] Step 304: divide the medical data set to obtain various medical data subsets.

[0142] In one possible implementation, the various medical data subsets are divided according to a specified number of subsets, that is, before the medical data set is divided, the number of divided medical data subsets can be set in advance, and the various medical sample data in the medical data set can be randomly assigned to the various medical data subsets, thereby realizing the division of the medical data set.

[0143] In another possible implementation, the number of divisions of each medical data subset is determined by the number of medical sample data in the medical data set. When the computer device needs to divide the medical data set, it can randomly select medical sample data, and when a specified number of medical sample data are taken out, the taken out medical sample data are constructed into a medical data subset; and the computer restarts a new round of medical sample data extraction operation until the number in the medical data set is less than the specified value.

[0144] Step 305 : For each medical data subset, the diversity, coverage, and balance of samples in the medical data subset are calculated using the entropy matrix.

[0145] In one possible implementation, the element values ​​corresponding to the individual medical sample data in the medical data subset are read from the entropy matrix; and the ratio of the sum of the element values ​​corresponding to the individual medical sample data in the medical data subset to the number of neural network models is determined as the diversity of the medical sample data in the medical data subset.

[0146] That is, for the calculation of sample diversity representation, the computer equipment can use the formula Get medical sample data x j representation of diversity.

[0147] In a possible implementation, among the elements of the entropy matrix, the elements that are greater than the target sliding factor are converted to 1, and among the elements of the entropy matrix, the elements that are less than the target sliding factor are converted to 0, so as to obtain the change matrix.

[0148] For each neural network model, obtaining, in the change matrix, elements obtained by the neural network model processing each medical sample data in the medical data subset, and summing the elements to obtain a response count;

[0149] Calculate the mean of each element in the change matrix to obtain the response mean;

[0150] Calculate the standard deviation between the response mean and the response count to obtain the response deviation corresponding to the neural network model;

[0151] The ratio of the response count to the response deviation is summed according to the category of the neural network model to obtain the balance of the medical data subset;

[0152] The ratio of the response count to the number of elements of the change matrix is ​​summed according to the category of the neural network model to obtain the coverage of the medical data subset.

[0153] That is, the computer device converts the Ent matrix into a 0-1 matrix, namely Mat by setting a sliding factor th.01 =[zo j,e ] N,(m+n) When the elements in the Ent matrix are greater than or equal to the sliding factor, then Mat 01 The corresponding element in the is 1, otherwise it is 0.

[0154] And the computer device can also set C model (seq,e)=∑ seq Mat 01 (seq,e) is the response count of the e-th model for the medical data subset seq, and is its response mean, then the response deviation is It can be obtained that the balance of the selected medical data subset seq for the e-th model can be obtained by For the e-th model, the coverage of the medical data subset seq can be obtained by

[0155] Step 306 : updating the medical data subset based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain a target data subset.

[0156] In a possible implementation, a weighted sum is performed based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain the fitness corresponding to the medical data subset;

[0157] According to the fitness of the medical data subset, the medical data subset is iteratively updated, and the iteratively updated medical data subset is determined as the target data subset.

[0158] In other words, the computer device can calculate the fitness of each medical sample data through the fitness function. The three items of the fitness function Fitness(seq) are: ∑ j＝seq Diversity(x j ),∑ e Balance(seq,e) and ∑ e Coverage(seq,e). The fitness function is expressed as: (the amount of information of each sample in the medical data subset, the relevance of the medical data subset to each model, and the coverage of the medical data subset)

[0159]

[0160] Schematically, That is, the three attributes have equal importance.

[0161] In a possible implementation, during the Nth round of iterative updating, at least one medical sample data in the medical data subset is replaced with medical sample data other than the medical data subset to obtain an intermediate medical data subset, where the initial value of N is 1;

[0162] Calculating the fitness of the intermediate medical data subset according to the entropy matrix, and determining the intermediate medical data subset and the one with higher fitness among the medical data subsets as the medical data subset after the Nth round of iterative update;

[0163] If N is less than the preset number of rounds, the operation of replacing at least one medical sample data in the medical data subset with medical sample data outside the medical data subset is performed again;

[0164] If N reaches the preset number of rounds, the medical data subset after the Nth round of iterative update is determined as the target data subset.

[0165] That is, in the process of iteratively updating the medical data subset, the medical sample data in the medical data subset can be continuously replaced with other medical sample data outside the medical data subset, thereby obtaining an intermediate medical data subset corresponding to the medical data subset;

[0166] At this time, the fitness calculation method in the above steps is used to calculate the fitness of the medical data subset and the fitness of the intermediate medical data subset respectively. After obtaining the fitness of the medical data subset and the fitness of the intermediate medical data subset, the fitness of the medical data subset in this iteration round can be compared with the fitness of the intermediate medical data subset.

[0167] When the fitness of the medical data subset is higher than that of the intermediate medical data subset, it means that replacing the medical data subset actually reduces the matching degree between the medical data subset and the neural network model. In this case, the medical data subset remains unchanged.

[0168] Correspondingly, when the fitness of the intermediate medical data subset is higher than that of the medical data subset, it means that after the replacement operation of the medical data subset, the matching degree between the medical data subset and the machine learning model is improved. At this time, the intermediate medical data subset can be directly updated to the medical data subset.

[0169] In a possible implementation, the computer device divides the medical data set D into mutually disjoint medical data subsets sub according to the labels. k For each medical data subset sub k ,k∈[1,2,...,K],from sub k Randomly select m (m greater than or equal to 1) samples from seqk As a subset of the data to be tested, and That is, first copy the corresponding candidate data subset according to the data subset to be tested, and then k -seq k Randomly select a sample to replace A sample is selected, thereby completing the replacement of the medical sample data in the copied candidate data subset.

[0170] At this time, the fitness of the data subset to be tested and the candidate data subset are compared, and the one with higher fitness is updated as the data subset to be tested. When the iteration step is less than the specified number, the iteration of the data subset to be tested continues. When the iteration step reaches the specified number, the data subset to be tested is used as the target data subset.

[0171] Step 307: combine the target data subsets into an optimized medical data set.

[0172] Please refer to Figure 7 , which shows a logical framework diagram of a data set optimization involved in an embodiment of the present application. Figure 7 When optimizing the data set, each alternative model (i.e., small alternative model) and each checkpoint model can be constructed first, thereby constructing a query integration model. The computer device processes each medical sample data through the query integration model, thereby constructing an entropy matrix Ent = [α j,e ] N,(m+n) ; At this time, the computer device converts the Ent matrix into a 0-1 matrix, namely Mat, through the sliding factor th 01 =[zo j,e ] N,(m+n) .

[0173] After completing the above calculations, for different medical data subsets, their corresponding sample diversity representations, response counts, and coverage can be calculated respectively, and through the optimization algorithm shown in step 305, the target data subsets composed of the medical sample data are obtained, and the various target data subsets are merged as the final medical data set.

[0174] In summary, before training a neural network model for processing medical data, a computer device first obtains each medical sample data in the medical data set, processes each medical sample data separately through each neural network model, and constructs an entropy matrix based on the processing results. The computer device also divides the medical data set into medical data subsets and calculates the diversity, balance, and coverage of the medical data subsets using the entropy matrix to update each medical data subset, thereby obtaining each target data subset. At this point, the target data subsets are combined to obtain an optimized medical data set. In the above scheme, the medical sample data are processed by multiple neural network models, and the entropy matrices corresponding to each medical sample data and the neural network model are calculated. The diversity, balance, and coverage are calculated based on the information content reflected in the entropy matrices for each medical data subset, thereby better evaluating the quality of the medical sample data in each medical data subset. The medical data subsets are then updated to obtain each target data subset, which are then combined to obtain the optimized medical data set. This improves the matching degree between each medical sample data in the optimized medical data set and the neural network model, thereby improving the training effect of the neural network model training using the optimized medical data set.

[0175] Figure 8 1 is a block diagram of a data set optimization device according to an exemplary embodiment. The device includes:

[0176] The medical data set acquisition module 801 is used to acquire a medical data set; the medical data set includes various medical sample data;

[0177] The sample data processing module 802 is configured to process each medical sample data using a neural network model to obtain entropies corresponding to the medical sample data; the neural network model is configured to perform inference calculations on the medical sample data;

[0178] An entropy matrix construction module 803 is used to construct an entropy matrix from the entropies corresponding to the medical sample data;

[0179] A data subset division module 804 is used to divide the medical data set into various medical data subsets;

[0180] a data subset updating module 805 for calculating, for each medical data subset, the diversity, coverage, and balance of samples in the medical data subset using the entropy matrix, and updating the medical data subset based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain a target data subset;

[0181] The data set optimization module 806 is configured to construct an optimized medical data set based on each of the target data subsets.

[0182] In one possible implementation, the entropy matrix construction module is used to:

[0183] Taking the entropies corresponding to the medical sample data as elements to generate the entropy matrix;

[0184] The number of rows of the entropy matrix is ​​the number of the medical sample data; the number of columns of the entropy matrix is ​​the number of the neural network models.

[0185] In a possible implementation, the data subset updating module is further configured to:

[0186] Reading the element value corresponding to each medical sample data in the medical data subset in the entropy matrix;

[0187] The ratio of the sum of the element values ​​corresponding to each medical sample data in the medical data subset to the number of neural network models is determined as the diversity of the medical sample data in the medical data subset.

[0188] In a possible implementation, the entropy matrix construction module is further used to:

[0189] Converting the elements of the entropy matrix that are greater than the target sliding factor to 1, and converting the elements of the entropy matrix that are less than the target sliding factor to 0, to obtain a change matrix;

[0190] For each neural network model, obtaining, in the change matrix, elements obtained by the neural network model processing each medical sample data in the medical data subset, and summing the elements to obtain a response count;

[0191] Calculate the mean of each element in the change matrix to obtain the response mean;

[0192] Calculating a standard deviation between the response mean and the response count to obtain a response deviation corresponding to the neural network model;

[0193] summing the ratio of the response count to the response deviation according to the category of the neural network model to obtain the balance of the medical data subset;

[0194] The ratio of the response count to the number of elements of the change matrix is ​​summed according to the category of the neural network model to obtain the coverage of the medical data subset.

[0195] In a possible implementation, the data subset updating module is further configured to:

[0196] Performing weighted summation based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain the fitness corresponding to the medical data subset;

[0197] The medical data subset is iteratively updated according to the fitness of the medical data subset, and the iteratively updated medical data subset is determined as the target data subset.

[0198] In a possible implementation, the data subset updating module is further configured to:

[0199] During the Nth round of iterative updating, at least one medical sample data in the medical data subset is replaced with medical sample data other than the medical data subset to obtain an intermediate medical data subset;

[0200] Calculating the fitness of the intermediate medical data subset according to the entropy matrix, and determining the intermediate medical data subset and the medical data subset with the higher fitness as the medical data subset after the Nth round of iterative update;

[0201] If N is less than the preset number of rounds, the operation of replacing at least one medical sample data in the medical data subset with medical sample data outside the medical data subset is performed again;

[0202] If N reaches the preset number of rounds, the medical data subset after the Nth round of iterative update is determined as the target data subset.

[0203] In one possible implementation, each neural network model includes m small replacement models and n checkpoint models; the m small replacement models are models obtained by reducing the number of convolutional layers and / or pooling layers of the target initial model, or the m small replacement models are network models with a training accuracy less than a preset threshold; the n checkpoint models have the same model structure as the target initial model;

[0204] The target initial model is a neural network model that performs inference calculations on medical sample data.

[0205] In a possible implementation, the apparatus further includes:

[0206] A model reduction module, configured to reduce the number of convolutional layers and / or pooling layers of the target initial model to obtain m small initial models;

[0207] An initial training module, configured to train m small initial models according to the medical dataset to obtain m trained small replacement models;

[0208] A data set selection module is used to perform P rounds of pre-training on the target initial model based on the medical data set, and select a training data set from the medical data set based on the pre-trained target neural network model;

[0209] A model parameter acquisition module is used to perform T rounds of iterative training on the target initial model according to the training data set, and obtain the model parameters of the target initial model in the last n rounds of iterations;

[0210] The checkpoint model acquisition module is used to construct the n checkpoint models based on the model parameters of the target initial model in the last n rounds of iterations.

[0211] In summary, before training a neural network model for processing medical data, a computer device first obtains each medical sample data in the medical data set, processes each medical sample data separately through each neural network model, and constructs an entropy matrix based on the processing results. The computer device also divides the medical data set into medical data subsets and calculates the diversity, balance, and coverage of the medical data subsets using the entropy matrix to update each medical data subset, thereby obtaining each target data subset. At this point, the target data subsets are combined to obtain an optimized medical data set. In the above scheme, the medical sample data are processed by multiple neural network models, and the entropy matrices corresponding to each medical sample data and the neural network model are calculated. The diversity, balance, and coverage are calculated based on the information content reflected in the entropy matrices for each medical data subset, thereby better evaluating the quality of the medical sample data in each medical data subset. The medical data subsets are then updated to obtain each target data subset, which are then combined to obtain the optimized medical data set. This improves the matching degree between each medical sample data in the optimized medical data set and the neural network model, thereby improving the training effect of the neural network model training using the optimized medical data set.

[0212] See also Figure 9 , which is a schematic diagram of a computer device provided according to an exemplary embodiment of the present application, the computer device includes a memory and a processor, the memory is used to store a computer program, and when the computer program is executed by the processor, the above method is implemented.

[0213] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0214] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods described in the embodiments of the present invention. The processor executes the non-transitory software programs, instructions, and modules stored in the memory to perform various processor functions and data processing, thereby implementing the methods described in the aforementioned method embodiments.

[0215] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include a memory remotely located relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0216] In an exemplary embodiment, a computer-readable storage medium is further provided, storing at least one computer program, which is loaded and executed by a processor to implement all or part of the steps of the above method. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, or an optical data storage device.

[0217] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0218] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A medical data set optimization method, characterized in that: The method comprises: Acquire a medical data set; the medical data set includes various medical sample data; For each medical sample data, the medical sample data is processed by each neural network model to obtain each entropy corresponding to the medical sample data; the neural network model is used to perform inference calculation on the medical sample data; Taking the entropies corresponding to the medical sample data as elements to generate an entropy matrix; Dividing the medical data set to obtain various medical data subsets; For each medical data subset, the diversity, coverage, and balance of samples in the medical data subset are calculated using the entropy matrix, and the medical data subset is updated based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain a target data subset; constructing an optimized medical data set based on each of the target data subsets; For each medical data subset, calculating the diversity, coverage, and balance of samples in the medical data subset using the entropy matrix includes: reading, from the entropy matrix, an element value corresponding to each medical sample data in the medical data subset; and determining a ratio of a sum of the element values ​​corresponding to each medical sample data in the medical data subset to the number of neural network models as the diversity of the medical sample data in the medical data subset; Converting the elements of the entropy matrix that are greater than the target sliding factor to 1, and converting the elements of the entropy matrix that are less than the target sliding factor to 0, to obtain a change matrix; For each neural network model, obtaining, in the change matrix, elements obtained by the neural network model processing each medical sample data in the medical data subset, and summing the elements to obtain a response count; Calculating the average of each element in the change matrix to obtain a response average; Calculating a standard deviation between the response mean and the response count to obtain a response deviation corresponding to the neural network model; summing the ratio of the response count to the response deviation according to the category of the neural network model to obtain the balance of the medical data subset; The ratio of the response count to the number of elements of the change matrix is ​​summed according to the category of the neural network model to obtain the coverage of the medical data subset.

2. The method according to claim 1, characterized in that The number of rows of the entropy matrix is ​​the number of the medical sample data; the number of columns of the entropy matrix is ​​the number of the neural network models.

3. The method according to claim 1 or 2, characterized in that The updating of the medical data subset based on the diversity, coverage and balance of the medical sample data in the medical data subset to obtain the target data subset includes: Performing weighted summation based on the diversity, coverage, and balance of the medical sample data in the medical data subset to obtain the fitness corresponding to the medical data subset; The medical data subset is iteratively updated according to the fitness of the medical data subset, and the iteratively updated medical data subset is determined as the target data subset.

4. The method according to claim 3, characterized in that The iteratively updating the medical data subset according to the fitness of the medical data subset, and determining the iteratively updated medical data subset as the target data subset, includes: In the Nth round of iterative updating, at least one medical sample data in the medical data subset is replaced with medical sample data other than the medical data subset to obtain an intermediate medical data subset, where the initial value of N is 1; Calculating the fitness of the intermediate medical data subset according to the entropy matrix, and determining the intermediate medical data subset and the medical data subset with the higher fitness as the medical data subset after the Nth round of iterative update; If N is less than the preset number of rounds, the operation of replacing at least one medical sample data in the medical data subset with medical sample data outside the medical data subset is performed again; If N reaches the preset number of rounds, the medical data subset after the Nth round of iterative update is determined as the target data subset.

5. The method according to any one of claims 1 or 2, characterized in that: Each neural network model includes m small replacement models and n checkpoint models; the m small replacement models are models obtained by reducing the number of convolutional layers and / or pooling layers of the target initial model, or the m small replacement models are network models with a training accuracy less than a preset threshold; the n checkpoint models have the same model structure as the target initial model; The target initial model is a neural network model that performs inference calculations on medical sample data.

6. The method according to claim 5, characterized in that Before processing each medical sample data by using each neural network model, the method further includes: Reducing the number of convolutional layers and / or pooling layers of the target initial model to obtain m small initial models; Training m small initial models respectively according to the medical data set to obtain m trained small replacement models; Performing P rounds of pre-training on the target initial model according to the medical data set, and selecting a training data set from the medical data set according to the pre-trained target neural network model; Performing T rounds of iterative training on the target initial model according to the training data set, and obtaining model parameters of the target initial model in the last n rounds of iterations; The n checkpoint models are constructed based on the model parameters of the target initial model in the last n rounds of iterations.

7. The method according to claim 6, characterized in that The step of performing P rounds of pre-training on the target initial model based on the medical data set, and selecting a training data set from the medical data set based on the pre-trained neural network model, includes: moving a first number of randomly selected medical data from the medical data set to a training data set; Training the target initial model using the first amount of medical data; Obtaining the information volume of each medical data in the medical data set; According to the acquired information volume, medical data that meets preset conditions in the medical data set is removed from the medical data set to the training data set, and the target initial model is retrained using the medical data that meets the preset conditions, wherein the preset conditions include that the information volume ranks in the top m positions, or the information volume is greater than a predetermined threshold; If the sample size of the training data set is less than the preset value, performing the step of obtaining the information volume of each medical data in the medical data set again; If the sample size of the training data set reaches the preset value, the current training data set is determined as the final selected training data set, and the preset value is less than the total amount of medical data in the medical data set.

8. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the medical dataset optimization method according to any one of claims 1 to 7.

Images