Data classification method, device, electronic device and storage device
The data classification method optimizes single-atlas data classification using a student network trained with multi-atlas knowledge distillation, addressing the challenge of applying multi-atlas-trained networks in clinical settings by improving efficiency and accuracy.
Patent Information
- Application Number
- CN202111372403.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-18
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-11-18
AI Technical Summary
Existing classification networks trained on multi-atlas data cannot be directly applied in clinical practice due to the unavailability of multi-atlas data in real-time clinical settings, necessitating a method for classifying single-atlas data to improve efficiency and accuracy in medical image diagnosis.
A data classification method utilizing a student network trained with knowledge distillation from a teacher network, where the student network is optimized using multi-atlas data to classify single-atlas data, employing a multi-layer perceptron neural network with residual connections to retain quantitative information and reduce input dimension.
The method enables efficient classification of single-atlas data, improving generalization performance and reducing time requirements while maintaining accurate diagnostic results, thus enhancing clinical applicability.
Smart Images

Figure CN114240840B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of artificial intelligence data classification technology, and specifically to a data classification method, device, electronic device and storage device. Background Art
[0002] In the classification diagnosis research of Alzheimer's Disease (AD), research institutions from all over the world have proposed their own brain maps. Different research institutions have different brain map partitioning methods. Some brain maps are partitioned according to the spatial structure of the brain, some brain maps are partitioned according to the functional connection of the brain, and so on; the characteristic data of these maps may include characteristic data such as the thickness and volume of each brain partition. At present, clinical brain maps need to be classified in neurological medical diagnosis research. The classification targets are three classification categories: Alzheimer's disease, mild cognitive impairment (Mild Cognitive Impairment) and normal. Researchers can obtain multi-map data of case samples publicly available from multiple medical centers. However, in clinical applications in ordinary hospitals, only single-map data can be obtained.
[0003] In applications such as clinical disease classification, due to real-time requirements and time cost constraints, it is impossible to obtain multi-atlas data like in scientific research scenarios. Therefore, the classification network trained based on multi-atlas data cannot be directly applied in clinical practice. This requires providing a method for classifying single-atlas data of medical images obtained in clinical work in order to improve the efficiency and accuracy of data classification. Summary of the invention
[0004] The present application provides a data classification method, apparatus, electronic device and storage device to solve the problem that the prior art cannot directly classify single atlas data of medical images obtained in clinical work.
[0005] The present application provides a data classification method, comprising:
[0006] Obtain single atlas data of medical images to be classified;
[0007] Obtaining a medical diagnosis classification result for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data;
[0008] The student network is obtained in the following manner:
[0009] Obtaining a medical image multi-atlas data sample set and medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set;
[0010] Training a teacher network according to the medical image multi-atlas data sample set and the medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set to obtain a trained teacher network;
[0011] Obtaining a medical diagnosis category probability vector for the medical image multi-atlas data samples in the medical image multi-atlas data sample set according to the medical image multi-atlas data sample set and the trained teacher network;
[0012] Obtaining a medical image single atlas data sample set from the medical image multi-atlas data sample set;
[0013] The student network is obtained according to the medical image single atlas data sample set and the medical diagnosis category probability vectors for the medical image multi-atlas data samples in the medical image multi-atlas data sample set.
[0014] Optionally, obtaining a medical diagnosis classification result for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data includes:
[0015] Obtaining a medical diagnosis category probability vector for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data;
[0016] The medical diagnosis category corresponding to the medical image single atlas data to be classified is determined according to the medical diagnosis category probability vector for the medical image single atlas data to be classified.
[0017] Optionally, each medical image multi-atlas data sample in the medical image multi-atlas data sample set corresponds to a medical diagnosis classification label, and the number of the medical diagnosis classification labels is the same as the dimension of the medical diagnosis category probability vector for the medical image multi-atlas data sample in the medical image multi-atlas data sample set.
[0018] Optionally, the dimension of the medical diagnosis category probability vector for the medical image multi-atlas data samples in the medical image multi-atlas data sample set is the same as the dimension of the medical diagnosis category probability vector for the medical image single-atlas data to be classified.
[0019] Optionally, a medical image multi-atlas data sample in the medical image multi-atlas data sample set is medical image atlas data with a first quantitative dimension characteristic obtained after a medical imaging examination of a patient, and the medical image single atlas data to be classified is medical image atlas data with a second quantitative dimension characteristic obtained after a medical imaging examination of a patient to be diagnosed, and the first quantity is higher than the second quantity; the medical diagnosis classification label corresponding to the medical image multi-atlas data sample is obtained by being marked by doctors and experts after reaching a consensus on the diagnosis of the medical image multi-atlas data sample.
[0020] Optionally, the medical image multi-atlas data sample set is a neuromedical image multi-atlas data sample set, and the medical image single-atlas data sample set is a neuromedical image single-atlas data sample set;
[0021] Optionally, obtaining a medical image single atlas data sample set from the medical image multi-atlas data sample set includes:
[0022] Select any one of the single atlas data sample sets from the medical image multi-atlas data sample set as the neuromedical image single atlas data sample set;
[0023] The feature content of the single atlas data sample set is the same as the feature content of the single atlas data of the medical image to be classified; the dimension of the single atlas data sample set is the same as the dimension of the single atlas data of the medical image to be classified, and the components of the dimension of the single atlas data sample set and the dimension of the single atlas data of the medical image to be classified refer to the same features.
[0024] Optionally, the medical image multi-atlas data sample set is formed by splicing multiple neuromedical image single-atlas data sample sets; the number of dimensions of the medical image multi-atlas data is the sum of the number of dimensions of the multiple medical image single-atlas data.
[0025] Optionally, the medical image single atlas data to be classified is neuromedical image single atlas data to be classified, and the medical diagnosis classification result for the medical image single atlas data to be classified includes neuromedical diagnosis categories corresponding to the neuromedical image single atlas data to be classified, and the neuromedical diagnosis categories include Alzheimer's disease category, mild cognitive impairment category and no obvious abnormality category.
[0026] Optionally, the method is applied to a data server;
[0027] The obtaining of the medical image single atlas data to be classified includes: obtaining the medical image single atlas data to be classified sent by the first data terminal;
[0028] The method further comprises: returning the medical diagnosis classification result for the medical image single atlas data to be classified to the second data terminal;
[0029] The first data terminal is a terminal used by medical personnel or a medical imaging examination device, and the second data terminal is a terminal used by medical personnel.
[0030] Optionally, the medical image multi-atlas data sample set is a medical image multi-atlas data sample set amplified by adding clinically collected multi-atlas data.
[0031] Preferably, the student network and the teacher network adopt a multi-layer perceptron neural network with residual connections.
[0032] Preferably, the multilayer perceptron neural network with residual connections includes a first network layer with residual connections and a second network layer with residual connections, and the residual connection of the multilayer perceptron with residual connections is the addition of the outputs of the first network layer with residual connections and the second network layer with residual connections, and the sum is used as the output of the second network layer with residual connections; wherein the number of neurons in the first network layer with residual connections is equal to the number of neurons in the second network layer with residual connections; the output of the first network layer with residual connections does not pass the drop out function, and the output of the second network layer with residual connections passes the drop out function.
[0033] Preferably, when the network depth is greater than a set threshold, the network layer does not contain a drop out function, and when the network depth is less than or equal to the set threshold, the network layer contains a drop out function.
[0034] Preferably, the set threshold is determined by calculating the following formula:
[0035]
[0036] Wherein, d is a set threshold, L is the total number of network layers of the multi-layer perceptron neural network with residual connections, and z is an integer function.
[0037] The present application also provides a data classification device, the device comprising:
[0038] The acquisition module is used to obtain the single atlas data of medical images to be classified.
[0039] A processing module, which obtains a medical diagnosis classification result for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data;
[0040] The student network is obtained in the following manner:
[0041] Obtaining a medical image multi-atlas data sample set and medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set;
[0042] Training a teacher network according to the medical image multi-atlas data sample set and the medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set to obtain a trained teacher network;
[0043] Obtaining a medical diagnosis category probability vector for the medical image multi-atlas data samples in the medical image multi-atlas data sample set according to the medical image multi-atlas data sample set and the trained teacher network;
[0044] Obtaining a medical image single atlas data sample set from the medical image multi-atlas data sample set;
[0045] The student network is obtained according to the medical image single atlas data sample set and the medical diagnosis category probability vectors for the medical image multi-atlas data samples in the medical image multi-atlas data sample set.
[0046] The present application also provides an electronic device, comprising:
[0047] a memory on which a computer program is stored;
[0048] A processor is used to execute the above data classification method after running the computer program in the memory.
[0049] The present application also provides a storage device, including:
[0050] The storage device stores a computer program, and after the computer program is executed by the processor, the processor executes the above data classification method.
[0051] Compared with the prior art, the data classification method provided by the present application has the following beneficial effects: first, by inputting single atlas data of clinical medical images into the student network, the medical diagnosis classification results can be obtained, which greatly saves the time of acquiring atlas data and improves the actual operability; at the same time, the distillation knowledge of multi-atlas data is utilized to improve the generalization performance of the classification network. Secondly, the existing knowledge distillation usually only reduces the network model and does not change the input dimension. The knowledge distillation method provided by the present application that uses multi-atlas big data to optimize the training of the classification network can reduce the input dimension and meet the application occasions where the input dimension is reduced. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1This is a flowchart of a medical image data classification method provided in this application.
[0053] Figure 2 This is a working principle diagram of a multi-layer perceptron neural network with residual connections provided in this application.
[0054] Figure 3 It is a schematic diagram of a medical image data classification device provided in this application.
[0055] Figure 4 This is a schematic diagram of an electronic device provided by this application DETAILED DESCRIPTION
[0056] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application. Therefore, the present application is not limited to the specific implementation disclosed below.
[0057] In recent years, neural network models have achieved superior performance in solving classification problems and have been widely used. Through the method of knowledge distillation, the small network model learns the knowledge distilled from the large network model, hoping that the small network model can achieve the effect of the large network model. Among them, the small network model is called the student network, and the large network model is called the teacher network.
[0058] The existing knowledge distillation technology is mainly used for the purpose of lightweight models. Since the student network model has a smaller structure and fewer parameters than the teacher network model, the student network reasoning speed is faster than the teacher network reasoning speed, the hardware computing power requirements are small, and it is easy to deploy. However, in multi-graph data classification applications, the existing student network still needs to input multi-graph data. Therefore, a knowledge distillation method is needed to reduce the dimension of the student network input parameters, that is, the dimension of the student network input parameters should be equal to the dimension of the single-graph data, so that the single-graph input requirements can be met.
[0059] Therefore, there is a need for a method that can utilize multi-atlas big data to optimize the training of classification networks and use clinically acquired single atlases and trained classification networks to obtain medical diagnosis classification.
[0060] The present application provides a data classification method, which is described in detail below in conjunction with the drawings of the embodiments provided by the present application, and each step of the method is described.
[0061] The first embodiment of the present application provides a data classification method, please refer to the attached Figure 1 , which shows a flow chart of a data classification method provided by the present application, including:
[0062] Step S101: Obtain single atlas data of medical images to be classified.
[0063] For example, the medical image single atlas data to be classified can be obtained by using medical image data of a patient acquired in real time by a medical imaging examination device, or medical image data of a patient stored in an electronic device, or medical image data of a patient stored in the cloud.
[0064] After obtaining the single atlas data of the medical image to be classified, it is necessary to perform standardization preprocessing on each feature data f in the single atlas data to obtain the mean and standard deviation of the single atlas data. The formula for standardization preprocessing is:
[0065]
[0066] Among them, f represents the characteristic data, μ represents the mean of the characteristic data, σ represents the standard deviation of the characteristic data; f′ represents the standardized characteristic data;
[0067] Step S1035: The single atlas data after the standardized preprocessing in step S1011 is input into a student network used to obtain a medical diagnosis classification result for the medical image single atlas data based on the medical image single atlas data. The student network is a trained neural network. In order to obtain the trained student network, a neural network can be pre-constructed or a neural network can be selected. In this application, the neural network is trained by using a sample set of medical image single atlas data. The trained neural network can be used as a student network for medical diagnosis classification.
[0068] Step S102: obtaining a medical diagnosis classification result for the medical image single atlas data to be classified based on the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data based on the medical image single atlas data.
[0069] Step S103: The student network is obtained in the following manner:
[0070] Step S1031: obtaining a medical image multi-atlas data sample set and medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set;
[0071] Step S1032: training a teacher network according to the medical image multi-atlas data sample set and the medical diagnosis classification labels of the medical image multi-atlas data samples in the medical image multi-atlas data sample set to obtain a trained teacher network;
[0072] It should be noted that the untrained teacher network can be a pre-built neural network or a selected neural network, which needs to be trained with a large amount of sample data. In this application, a medical imaging multi-atlas data sample set is used to train the neural network, and the trained teacher network is called a trained teacher network.
[0073] Step S1033: obtaining a medical diagnosis category probability vector for the medical image multi-atlas data samples in the medical image multi-atlas data sample set according to the medical image multi-atlas data sample set and the trained teacher network;
[0074] Step S1034: obtaining a single atlas data sample set of medical images from the multi-atlas data sample set of medical images; the single atlas data sample set is obtained by intercepting from the multi-atlas data sample set of medical images, that is, the multi-atlas data sample set can intercept multiple single atlas data sample sets.
[0075] A student network is trained according to the medical image single atlas data sample set and the medical diagnosis category probability vectors for the medical image multi-atlas data samples in the medical image multi-atlas data sample set to obtain the trained student network.
[0076] Preferably, the student network and the teacher network can adopt a multi-layer perceptron neural network, and the calculation operation of the linear weighted accumulation of neurons of the multi-layer perceptron neural network can obtain and retain a large amount of quantitative information data. Among them, the multi-layer perceptron can extract quantitative information data, and is more suitable for applications where quantitative information data and classification results are related. In the application of Alzheimer's disease brain atlas classification, quantitative information data has important value, such as the volume change of each brain partition, the change of total volume and other quantitative information data are important basis for diagnostic classification. However, when the number of multi-layer perceptron layers increases, the quantitative information data will be lost after multiple nonlinear activation calculations and dropout (dropout refers to the temporary discarding of a part of the neural network units from the network according to a certain probability during the training process of the deep learning network, which is equivalent to finding a thinner network from the original network) calculations. Therefore, a modified multi-layer perceptron network is needed to retain more quantitative information data.
[0077] In view of the above problems, in the solution of the present application, the student network and the teacher network can adopt a multi-layer perceptron neural network with residual connections; the residual connections can be selected so that all network layers contain residual connections or some network layers contain residual connections; the residual connections can alleviate the problem of quantitative information data loss caused by the increase in the number of network layers.
[0078] The following describes the principle and working process of the multilayer perceptron neural network with residual connections:
[0079] Preferably, the multilayer perceptron neural network with residual connection includes a first network layer with residual connection and a second network layer with residual connection, and the residual connection of the multilayer perceptron with residual connection is the addition of the outputs of the first network layer with residual connection and the second network layer with residual connection, and the sum is used as the output of the second network layer with residual connection; wherein the number of neurons of the first network layer with residual connection is equal to the number of neurons of the second network layer with residual connection; the output of the first network layer with residual connection does not pass the drop out function, and the output of the second network layer with residual connection passes the drop out function. That is, the residual connection of the multilayer perceptron with residual connection is the addition of the outputs of the front and rear different network layers; specifically, the output of the front layer is used as the input of the rear layer, the output of the front layer is copied and added to the output of the rear layer, and the sum is used as the output of the rear layer; wherein the number of neurons of the two different network layers is equal; the output of the front network layer of the two network layers with residual connection does not pass the drop out function, and the output of the rear network layer of the two network layers with residual connection passes the drop out function.
[0080] For example, the multilayer perceptron neural network with residual connection includes two different network layers l1 and l2, and the residual connection of the multilayer perceptron with residual connection is to add the outputs of the two different network layers l1 and l2, and the sum is used as the output of the rear layer l2. The output dimensions of the two layers l1 and l2 of the residual connection are the same, wherein the number of neurons of the two different network layers l1 and l2 is equal, and the corresponding single neuron is calculated according to the following formula:
[0081]
[0082] y=D(v L2 )+v L1 ;
[0083] Among them, x is the input value of a neuron; v n is the output value of a neuron; is the neuron activation function; m is the input dimension; w is the weighting coefficient; b is the bias; D is the dropout function; v L1 is the output vector of all neurons in layer L1, v L2 is the output vector of all neurons in the L2 layer. y is the output vector of the residual connection and serves as the input of the next layer of neurons.
[0084] Preferably, when the network depth is greater than a set threshold, the network layer does not contain a drop out function, and when the network depth is less than or equal to the set threshold, the network layer contains a drop out function.
[0085] Preferably, the set threshold is determined by calculating the following formula:
[0086]
[0087] Wherein, d is a set threshold, L is the total number of network layers of the multi-layer perceptron neural network with residual connections, and z is an integer function.
[0088] The softmax (softmax is a probability mapping function) output layer of the teacher network uses the knowledge distillation method to perform knowledge distillation on the multi-atlas data;
[0089] The softmax output layer trains the teacher network according to the following formula to generate a trained teacher network:
[0090]
[0091] Among them, q i is the component of the probability vector output by the teacher network, which is equal to the classification probability of the i-th classification category; z i is the output of the previous network layer of the softmax layer; j is the number of classification categories; T is the distillation temperature.
[0092] like Figure 2 As shown, the embodiment of the present application exemplarily shows that the teacher network and the student network in the data classification method use a 6-layer perceptron neural network with residual connections. The total number of multi-atlas data features in the medical imaging multi-atlas data sample set is fn, and the output dimension of each layer of the 6-layer perceptron neural network is l1, l2, l3, l4, l5, l6. The input dimension of each layer is the output dimension of the previous layer; the output dimension of the previous layer of the softmax layer is the number of classification categories. The residual connection is the addition of the outputs of two different layers. The output dimensions of the two layers of the residual connection must be the same. For example, Figure 2 Here l2 is equal to l3.
[0093] Specifically, the input of the first layer of the neural network is the feature vector (x1, x2, ..., x m ), the first layer of the neural network each neuron input is a vector (x1, x2, ..., x m ), the output of the nth neuron in the neural network layer is v n ; After weighted accumulation, activation function, and drop-out function operations are performed on each neuron in the first layer, the first layer output vector D (v L1 ); the first layer output vector D(v L1) as the input of the second layer of neural network, and so on. The output of the network layer with residual connection input is the sum of the outputs of the two residual connection network layers, and is used as the input of the next layer of neural network. When the network depth is greater than the threshold d, the network layer does not contain a dropout function. For example, Figure 2 Where d=4.
[0094] Preferably, in step S102, the medical image single atlas data to be classified and the student network for obtaining the medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data, obtaining the medical diagnosis classification result for the medical image single atlas data to be classified, includes:
[0095] Obtaining a medical diagnosis category probability vector for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data;
[0096] According to the medical diagnosis category probability vector for the medical image single atlas data to be classified, the probability vector is input into the argmax (argmax is a function for solving the maximum value position) function to determine the medical diagnosis category corresponding to the medical image single atlas data to be classified.
[0097] Preferably, each medical image multi-atlas data sample in the medical image multi-atlas data sample set corresponds to a medical diagnosis classification label, and the number of the medical diagnosis classification labels is the same as the dimension of the medical diagnosis category probability vector for the medical image multi-atlas data sample in the medical image multi-atlas data sample set.
[0098] Furthermore, a medical image multi-atlas data sample in the medical image multi-atlas data sample set is medical image atlas data with a first quantitative dimension characteristic obtained after a medical imaging examination of a patient, and the medical image single atlas data to be classified is medical image atlas data with a second quantitative dimension characteristic obtained after a medical imaging examination of a patient to be diagnosed, and the first quantity is higher than the second quantity; the medical diagnosis classification label corresponding to the medical image multi-atlas data sample is obtained by being marked by doctors and experts after reaching a consensus on the diagnosis of the medical image multi-atlas data sample.
[0099] For example, fn represents the medical image atlas data of the first quantitative dimension feature, and fn′ represents the medical image atlas data of the second quantitative dimension feature, then fn′ <fn。
[0100] Preferably, the medical image multi-atlas data sample set is a neuromedical image multi-atlas data sample set, and the medical image single-atlas data sample set is a neuromedical image single-atlas data sample set;
[0101] The step of obtaining a single medical image atlas data sample set from the multi-atlas data sample set of medical images includes:
[0102] Select any single atlas data sample set from the medical image multi-atlas data sample set as the medical image single atlas data sample set; the feature content of the single atlas data sample set is the same as the feature content of the medical image single atlas data to be classified; the dimension of the single atlas data sample set is the same as the dimension of the medical image single atlas data to be classified, and the features referred to by the components of the dimensions of the single atlas data sample set are the same as the features referred to by the components of the dimensions of the medical image single atlas data to be classified.
[0103] The medical image multi-atlas data sample set is formed by splicing multiple neuromedical image single-atlas data sample sets; the dimension of the medical image multi-atlas data is obtained by summing up the dimensions of multiple single-atlas data. For example: the dimension of the medical image multi-atlas data sample set is M, and the dimensions of the single-atlas data sample sets are M1, M2, etc.; then the dimension of the medical image multi-atlas data M = M1+M2+, etc.; if the 1 to M1 dimensions of the medical image multi-atlas data sample set are the data of the first medical image single-atlas data sample set, the M1+1 to M2 dimensions are the data of the second medical image single-atlas data sample set, and so on;
[0104] For example, in actual medical image atlas classification research, if the format of the medical image single atlas data to be classified adopts the medical image single atlas data format provided by research institution B, with a dimension of 100, then it is known that a medical image multi-atlas data sample set X composed of multiple medical image single atlas data sample sets includes the medical image single atlas data sample set in the format of research institution A with a dimension of M1=52, the medical image single atlas data sample set in the format of research institution B with a dimension of M2=100, and the data dimension of the medical image multi-atlas data sample set X is M, that is, M=152; when selecting a single atlas data sample set from the medical image multi-atlas data sample set X to train the student network, it is necessary to select the medical image single atlas data sample set in the format of research institution B, that is, select the sample data of dimensions 53-152 in the medical image multi-atlas data sample set.
[0105] Preferably, the medical image single atlas data to be classified is neuromedical image single atlas data to be classified, and the medical diagnosis classification result for the medical image single atlas data to be classified includes the neuromedical diagnosis category corresponding to the neuromedical image single atlas data to be classified, and the neuromedical diagnosis category includes Alzheimer's disease category, mild cognitive impairment category and no obvious abnormality category.
[0106] The data classification method provided in this application is applied to the data service end; specifically includes:
[0107] The obtaining of the medical image single atlas data to be classified includes: obtaining the medical image single atlas data to be classified sent by the first data terminal;
[0108] The method further comprises: returning the medical diagnosis classification result for the medical image single atlas data to be classified to the second data terminal;
[0109] The first data terminal is a terminal used by medical personnel or a medical imaging examination device, and the second data terminal is a terminal used by medical personnel.
[0110] In the data classification method provided by the present application, the student network learns the multi-atlas data knowledge distilled by the teacher network, and the trained student network only needs to input the single atlas data of the medical image to be classified to obtain the classification result, which not only improves the classification generalization performance, but also the multi-layer perceptron neural network with residual connection provided in the present application can properly retain quantitative information, and alleviate the problem of quantitative information data loss caused by the increase in the number of network layers of the multi-layer perceptron. Therefore, in the actual medical clinical application, it is not necessary to obtain multi-atlas data, which saves the time of clinical diagnosis, avoids the problems of low efficiency and low accuracy of manual data classification, and improves feasibility.
[0111] Based on the same inventive concept, the second embodiment of the present application further provides a data classification device, which can be part or all of the electronic device through software, hardware or a combination of both. Figure 3 , the data classification device 200 may include:
[0112] The acquisition module 201 is used to acquire the single atlas data of the medical image to be classified.
[0113] The processing module 202 obtains the medical diagnosis classification result for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and the student network used to obtain the medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data;
[0114] The student network is obtained in the following manner:
[0115] Obtaining a medical image multi-atlas data sample set and medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set;
[0116] Training a teacher network according to the medical image multi-atlas data sample set and the medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set to obtain a trained teacher network;
[0117] Obtaining a medical diagnosis category probability vector for the medical image multi-atlas data samples in the medical image multi-atlas data sample set according to the medical image multi-atlas data sample set and the trained teacher network;
[0118] Obtaining a medical image single atlas data sample set from the medical image multi-atlas data sample set;
[0119] The student network is obtained according to the medical image single atlas data sample set and the medical diagnosis category probability vectors for the medical image multi-atlas data samples in the medical image multi-atlas data sample set.
[0120] Based on the same inventive concept, the present application embodiment also provides an electronic device 300, please refer to Figure 4 Since the device embodiment is basically similar to the method embodiment, the description is relatively simple, and the relevant parts refer to the partial description of the method embodiment. The device embodiment described below is only illustrative.
[0121] An electronic device of this embodiment, the electronic device 300 includes: a processor 301; and a memory 302, the memory 302 having a computer program stored thereon; the processor 301 is used to execute the method of the above embodiment after running the computer program in the memory. It may also include one or more of a multimedia component 303, an input / output (I / O) interface 304, and a communication component 305.
[0122] The processor 301 is used to control the overall operation of the electronic device 300 to complete all or part of the steps in the above-mentioned data classification method. The memory 302 is used to store various types of data to support the operation of the electronic device 300, and these data may include, for example, instructions for any application or method used to operate on the electronic device 300, and application-related data, such as data classification results, etc. The memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Read-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, disk or optical disk. The multimedia component 303 may include a screen. The screen may be, for example, a touch screen. For example, researchers can obtain the classification results of the atlas data by selecting the atlas data information displayed on the screen, and the signal instructions received by the screen are input into the processor 301 to perform related operations. The I / O interface 304 provides an interface between the processor 301 and other interface modules, and the above-mentioned other interface modules can be keyboards, mice, buttons, etc. These buttons can be virtual buttons or physical buttons. The communication component 305 is used for wired or wireless communication between the electronic device 300 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, etc., or a combination of one or more of them, is not limited here. Therefore, the corresponding communication component 305 may include: Wi-Fi module, Bluetooth module, NFC module, etc.
[0123] In an exemplary embodiment, the electronic device 300 can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned data classification method.
[0124] In another exemplary embodiment, a computer program product is also provided. The computer program product includes a computer program that can be executed by a programmable device. The computer program has a code portion for executing the above data classification method when executed by the programmable device.
[0125] An embodiment of the present application further provides a storage device, wherein the storage device stores a computer program, and after the computer program is executed by a processor, the processor executes the method as described in the foregoing embodiment.
[0126] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.
Claims
1. A data classification method, characterized in that: include: Obtain single atlas data of medical images to be classified; Obtaining a medical diagnosis classification result for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data; The student network is obtained in the following manner: Obtaining a medical image multi-atlas data sample set and medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set; Training a teacher network according to the medical image multi-atlas data sample set and the medical diagnosis classification labels for the medical image multi-atlas data samples in the medical image multi-atlas data sample set to obtain a trained teacher network; Obtaining a medical diagnosis category probability vector for the medical image multi-atlas data samples in the medical image multi-atlas data sample set according to the medical image multi-atlas data sample set and the trained teacher network; Obtaining a medical image single atlas data sample set from the medical image multi-atlas data sample set; The student network is obtained according to the medical image single atlas data sample set and the medical diagnosis category probability vectors for the medical image multi-atlas data samples in the medical image multi-atlas data sample set.
2. The method according to claim 1, characterized in that The medical image single atlas data to be classified and the student network for obtaining the medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data, obtaining the medical diagnosis classification result for the medical image single atlas data to be classified, includes: Obtaining a medical diagnosis category probability vector for the medical image single atlas data to be classified according to the medical image single atlas data to be classified and a student network for obtaining a medical diagnosis classification result for the medical image single atlas data according to the medical image single atlas data; The medical diagnosis category corresponding to the medical image single atlas data to be classified is determined according to the medical diagnosis category probability vector for the medical image single atlas data to be classified.
3. The method according to claim 1, characterized in that: Each medical image multi-atlas data sample in the medical image multi-atlas data sample set corresponds to a medical diagnosis classification label, and the number of the medical diagnosis classification labels is the same as the dimension of the medical diagnosis category probability vector for the medical image multi-atlas data sample in the medical image multi-atlas data sample set.
4. The method according to claim 2, characterized in that: The dimension of the medical diagnosis category probability vector for the medical image multi-atlas data samples in the medical image multi-atlas data sample set is the same as the dimension of the medical diagnosis category probability vector for the medical image single-atlas data to be classified.
5. The method according to claim 1, characterized in that A medical image multi-atlas data sample in the medical image multi-atlas data sample set is medical image atlas data with a first quantity dimension characteristic obtained after a medical imaging examination of a patient, and the medical image single atlas data to be classified is medical image atlas data with a second quantity dimension characteristic obtained after a medical imaging examination of a patient to be diagnosed, and the first quantity is higher than the second quantity.
6. The method according to claim 1, characterized in that Obtaining a medical image single atlas data sample set from the medical image multi-atlas data sample set includes: Select any single atlas data sample set from the medical image multi-atlas data sample set as the medical image single atlas data sample set; the feature content of the single atlas data sample set is the same as the feature content of the medical image single atlas data to be classified; the dimension of the single atlas data sample set is the same as the dimension of the medical image single atlas data to be classified, and the features referred to by the components of the dimensions of the single atlas data sample set are the same as the features referred to by the components of the dimensions of the medical image single atlas data to be classified.
7. The method according to claim 1, characterized in that The medical image multi-atlas data sample set is formed by splicing multiple medical image single-atlas data sample sets; the number of dimensions of the medical image multi-atlas data is the sum of the number of dimensions of the multiple medical image single-atlas data.
8. The method according to claim 1, characterized in that The medical image single atlas data to be classified is neuromedical image single atlas data to be classified, and the medical diagnosis classification result for the medical image single atlas data to be classified includes the neuromedical diagnosis category corresponding to the neuromedical image single atlas data to be classified, and the neuromedical diagnosis category includes Alzheimer's disease category, mild cognitive impairment category and no obvious abnormality category.
9. The method according to claim 1, characterized in that: The student network and the teacher network both use a multilayer perceptron neural network with residual connections; the multilayer perceptron neural network with residual connections includes a first network layer with residual connections and a second network layer with residual connections, the residual connection of the multilayer perceptron with residual connections is the addition of the outputs of the first network layer with residual connections and the second network layer with residual connections, and the sum is used as the output of the second network layer with residual connections; wherein the number of neurons in the first network layer with residual connections is equal to the number of neurons in the second network layer with residual connections; the output of the first network layer with residual connections does not pass the drop out function, and the output of the second network layer with residual connections passes the drop out function.
10. The method according to claim 9, characterized in that When the network depth of the multilayer perceptron neural network with residual connection is greater than the integer of the set threshold, the network layer does not contain a drop out function. When the network depth is less than or equal to the set threshold, the network layer contains a drop out function. The set threshold is calculated and determined by the following formula: ; in, d To set the threshold, L is the total number of network layers of the multilayer perceptron neural network with residual connections, z is the integer function.
Citation Information
Patent Citations
Knowledge distillation-based hyperspectral image unsupervised waveband selection method and system
CN113470036A
Attribute recognition method based on knowledge distillation, terminal equipment and storage medium
CN113515614A
Eye fundus image classification method and device based on multi-task curriculum type learning, equipment and medium
CN113610118A
Structuring free text medical reports with forced taxonomies
US10692602B1
