WO2021151338A1 - 医学影像图片分析方法、装置、电子设备及可读存储介质 - Google Patents

医学影像图片分析方法、装置、电子设备及可读存储介质 Download PDF

Info

Publication number
WO2021151338A1
WO2021151338A1 PCT/CN2020/125474 CN2020125474W WO2021151338A1 WO 2021151338 A1 WO2021151338 A1 WO 2021151338A1 CN 2020125474 W CN2020125474 W CN 2020125474W WO 2021151338 A1 WO2021151338 A1 WO 2021151338A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
diagnosis
target
training
disease
Prior art date
Application number
PCT/CN2020/125474
Other languages
English (en)
French (fr)
Inventor
魏文琦
王健宗
贾雪丽
程宁
Original Assignee
平安科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2021151338A1 publication Critical patent/WO2021151338A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • G06T7/0012Biomedical image inspection
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20081Training; Learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20084Artificial neural networks [ANN]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30061Lung
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30068Mammography; Breast

Definitions

  • This application relates to the field of artificial intelligence, and in particular to a medical image picture analysis method, device, electronic equipment, and readable storage medium.
  • training deep learning models usually requires a lot of effort.
  • the hardware threshold cannot be migrated to the mobile terminal or places where computing resources are scarce. If you directly train a lightweight model, not only the feature extraction ability is weak, but also the accuracy is low. Therefore, there is a need to maintain a high accuracy without too much Medical imaging picture analysis method for computing resources.
  • a medical imaging picture analysis method provided by this application includes:
  • the target diagnostic model is used to analyze the medical image picture to be analyzed to obtain an analysis result.
  • the present application also provides a medical image picture analysis device, the device includes:
  • the model generation module is used to obtain a preset disease history picture set, and use the preset disease history picture set to train a pre-built deep learning network model to obtain a disease recognition model;
  • the model distillation module is used to construct a distillation loss function according to the disease recognition model and the pre-built initial diagnosis model; perform distillation training on the initial diagnosis model according to the distillation loss function to obtain the first diagnosis model;
  • the diagnosis target trains and adjusts the output of the first diagnosis model to obtain the target diagnosis model;
  • the picture analysis module is used to analyze the medical image picture to be analyzed by using the target diagnosis model to obtain the analysis result when the medical image picture to be analyzed is received.
  • This application also provides an electronic device, which includes:
  • Memory storing at least one instruction
  • the processor executes the instructions stored in the memory to implement the medical imaging picture analysis method as described below:
  • the target diagnostic model is used to analyze the medical image picture to be analyzed to obtain the analysis result.
  • the present application also provides a computer-readable storage medium in which at least one instruction is stored, and the at least one instruction is executed by a processor in an electronic device to implement the medical imaging picture analysis method described below :
  • the target diagnostic model is used to analyze the medical image picture to be analyzed to obtain an analysis result.
  • FIG. 1 is a schematic flowchart of a method for analyzing medical images according to an embodiment of the application
  • FIG. 2 is a detailed flowchart of obtaining a distillation loss function in a medical image picture analysis method provided by an embodiment of the application;
  • FIG. 3 is a detailed flowchart of obtaining a target diagnosis model in a medical image picture analysis method provided by an embodiment of the application;
  • FIG. 4 is a schematic diagram of modules of a medical image analysis device provided by an embodiment of the application.
  • FIG. 5 is a schematic diagram of the internal structure of an electronic device that implements a method for analyzing medical images according to an embodiment of the application;
  • This application provides a method for analyzing medical images.
  • FIG. 1 it is a schematic flowchart of a medical image picture analysis method provided by an embodiment of this application.
  • the method can be executed by a device, and the device can be implemented by software and/or hardware.
  • the medical imaging picture analysis method includes:
  • the preset disease history picture collection is a medical imaging picture of a patient at a preset position, such as a collection of CXR (chest X-ray, ChestXRay) pictures of the patient's chest, and the preset disease history picture
  • the set can be obtained from the CXR database of an institution, for example: the CXR database of the NIH (National Institutes of Health) Clinical Center.
  • the convolutional neural network model in the embodiment of the present application may be a fully convolutional neural network model.
  • the preset disease history picture set is determined as a training set, and the disease type label is performed on the preset disease history picture set to obtain a label set.
  • the disease types include atelectasis, solidification, infiltration, pneumothorax, edema, emphysema, fibrosis, effusion, pneumonia, pleural thickening, cardiac hypertrophy, nodules, masses, hernias, etc.
  • the LabelMe image labeling tool can be used to manually label the disease type.
  • using the training set and the label set to train the deep learning network model includes:
  • Step A Perform a convolution pooling operation on the training set according to the preset number of times of convolution pooling to obtain a dimensionality reduction data set;
  • Step B Perform a deconvolution operation on the dimensionality reduction data set according to the preset number of deconvolutions to obtain an increased dimensionality data set;
  • Step C Use a preset activation function to calculate a predicted value for the dimension-up data set, and calculate a loss value by using a pre-built loss function according to the predicted value and the label value contained in the label set.
  • Step D Compare the loss value with the preset loss threshold. If the loss value is greater than or equal to the loss threshold, return to A; if the loss value is less than the loss threshold, stop training to obtain the Disease recognition model.
  • the convolution pooling operation includes: convolution operation and pooling operation.
  • G is a convolution data set
  • is the training set
  • k is the size of the preset convolution kernel
  • f is the step size of the preset convolution operation
  • p is the preset data zero-filling matrix
  • the pooling operation in the embodiment of the present application is to perform a maximum pooling operation on the convolutional data set to obtain the dimensionality reduction data set.
  • the preset activation function includes:
  • ⁇ t represents the predicted value
  • s represents the data in the ascending data set.
  • the loss function includes:
  • T represents the loss value
  • z is the number of data in the training set
  • t is a positive integer
  • b t is the label value
  • the disease recognition model is obtained through the above-mentioned model training process, so that the disease recognition model has the ability to accurately extract features of preset parts.
  • the training data of the disease recognition model can be stored in the blockchain.
  • the initial diagnosis model and the disease recognition model have the same dimensional feature layer and output.
  • the initial diagnosis model is a deep learning network model with a simpler structure and fewer parameters.
  • the feature layer is a fully connected layer connected before the softmax classifier in the initial diagnosis model and the disease recognition model.
  • the initial diagnosis model is a ResNet-8 model.
  • the model information in the disease recognition model is subsequently transferred to the initial diagnosis model, for
  • the difference in feature extraction capabilities between the disease recognition model and the pre-built initial diagnosis model is measured, a distillation loss function is constructed according to the disease recognition model and the pre-built initial diagnosis model, and the distillation loss function is used to measure the difference between the disease recognition model and the pre-built initial diagnosis model. Differences in feature extraction capabilities of pre-built initial diagnosis models.
  • the construction of the distillation loss function based on the disease recognition model and the pre-built initial diagnosis model includes:
  • S21 Perform normalization index processing on the feature information of the first feature layer included in the disease recognition model to obtain a probability distribution function of the disease recognition model;
  • the normalized index processing on the feature information of the first feature layer included in the disease recognition model can be performed by the following formula:
  • X represents the probability distribution of the disease recognition model
  • n represents the dimension of the first feature layer
  • i represents the sequence number of the feature node of the first feature layer
  • represents a mapping function
  • S22 Perform normalization index processing on the feature information of the second feature layer included in the initial diagnosis model to obtain a probability distribution function of the initial diagnosis model;
  • the normalized index processing on the feature information of the second feature layer included in the initial diagnosis model can be performed by the following formula:
  • Y represents the probability distribution of the initial diagnosis model
  • m represents the dimension of the second feature layer
  • j represents the sequence number of the feature node of the second feature layer.
  • distillation loss function can be expressed as follows:
  • E represents the distillation loss function
  • H represents the mapping space of the mapping function
  • the disease recognition model has a large amount of parameters and strong feature extraction capabilities, but the disease recognition model also has high requirements for hardware.
  • the initial diagnosis model has There are fewer parameters, and the feature extraction ability of individual training is weaker. Therefore, in order to keep the initial diagnosis model lightweight and have the same feature extraction ability as the disease recognition model, the embodiment of the present application is based on the distillation loss function Using the disease recognition model to perform distillation training on the initial diagnosis model.
  • performing distillation training on the initial diagnosis model according to the distillation loss function includes: using the first feature layer as a training target, continuously changing the initial diagnosis model parameters, and when calculating the When the distillation loss value obtained by the distillation loss function is less than the preset threshold, the training is stopped to obtain the first diagnostic model.
  • this embodiment trains and adjusts the output of the first diagnostic model according to a preset diagnostic target.
  • the training and output adjustment of the first diagnosis model to obtain the target diagnosis model includes:
  • Step I Determine whether the diagnosis target is a newly-added diagnosis target
  • the diagnosis target is a diagnosed disease type
  • the diagnosed disease type is a disease type that occurs at the preset site, and when the diagnosed disease type is a disease that can be identified by the disease recognition model
  • type it is determined that the diagnostic target is not a newly-added diagnostic target; when the diagnosed disease type is a disease type that cannot be recognized by the disease recognition model, the diagnostic target is determined to be a newly-added diagnostic target.
  • Step II When the diagnosis target is not a newly-added diagnosis target, perform output adjustment on the first diagnosis model to obtain a target diagnosis model;
  • the preset program framework is used to adjust the output of the first diagnostic model.
  • there are 4 outputs in the first diagnostic model including: infiltration, pneumothorax, edema, and emphysema;
  • the diagnosis target is the diagnosis of emphysema, and the output of the first diagnosis model is adjusted to two outputs, including: emphysema and non-emphysema.
  • the program framework is a pytorch framework.
  • Step III When the diagnosis target is a newly-added diagnosis target, the first diagnosis model is adjusted and trained to obtain the target diagnosis model.
  • the diagnosed disease type corresponding to the diagnostic target is the disease identification Types of diseases that cannot be identified by the model, for example: the types of diseases that can be identified by the disease identification model are infiltration, pneumothorax, edema, and emphysema, and the diagnosed disease type corresponding to the diagnosis target is new coronary pneumonia; therefore, this application is implemented For example, it is also necessary to adjust and train the first diagnostic model so that the first diagnostic model is familiar with the picture features of new coronary pneumonia.
  • the adjustment and training of the first diagnosis model in the embodiment of the present application to obtain the target diagnosis model includes:
  • S431 Obtain a diagnostic picture set, and mark the diagnostic picture set to obtain an initial diagnosis picture set;
  • the diagnosis picture set is a set of diagnosis pictures including the diagnosed disease type corresponding to the diagnosis target; the picture format of the diagnosis picture set is consistent with the picture format of the disease history picture set.
  • S432 Perform data enhancement processing on the initial diagnosis picture set to obtain a standard diagnosis picture set
  • data enhancement processing is performed on the initial diagnosis picture set.
  • performing data enhancement processing on the initial diagnosis picture set includes: performing random horizontal flipping and random edge cropping on the pictures included in the diagnosis picture set to obtain the standard diagnosis picture set.
  • the diagnosed disease type corresponding to the diagnosis target is a disease type that cannot be recognized by the first diagnosis model. Therefore, the diagnosed disease type corresponding to the diagnosis target needs to be added to the first diagnosis model. The type of output included.
  • S434 Perform iterative training on the second diagnosis model by using the standard diagnosis picture set until the second diagnosis model converges to obtain a third diagnosis model
  • the preset program framework is used to adjust the output of the third diagnostic model.
  • there are 4 outputs in the third diagnostic model including: infiltration, pneumothorax, edema, and emphysema;
  • the diagnosis target is the diagnosis of emphysema, and the output of the third diagnosis model is adjusted to two outputs, including: emphysema and non-emphysema.
  • the program framework is a pytorch framework.
  • the target diagnosis model is used to analyze the medical image picture to be analyzed to obtain an analysis result.
  • the medical image picture to be analyzed is the medical image picture to be analyzed of the predetermined part.
  • the format of the medical image picture to be analyzed is the same as the picture format included in the disease history picture set.
  • the medical image picture analysis model is used to analyze the medical image picture to be analyzed, and the probability of disease is output; the confidence threshold of the disease is confirmed by the principle of Yoden index; The probability is compared with the confidence threshold; when the disease probability is greater than or equal to the confidence threshold, it is determined to be diseased, and when the disease probability is less than the confidence threshold, it is determined to be unaffected.
  • FIG. 4 it is a functional block diagram of the medical image analysis device of the present application.
  • the medical imaging picture analysis device 100 described in this application can be installed in an electronic device.
  • the medical imaging picture analysis device may include a model generation module 101, a model distillation module 102, and an image analysis module 103.
  • the module described in the present invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can complete fixed functions, and are stored in the memory of the electronic device.
  • each module/unit is as follows:
  • the model generation module 101 is configured to obtain a preset disease history picture set, and use the preset disease history picture set to train a pre-built deep learning network model to obtain a disease recognition model.
  • the preset disease history picture collection is a medical imaging picture of a patient at a preset location, such as a collection of CXR pictures of a patient’s chest.
  • the preset disease history picture collection may be from a CXR of a certain institution. Obtained from the database, such as the CXR database of the NIH Clinical Center.
  • the convolutional neural network model in the embodiment of the present application may be a fully convolutional neural network model.
  • the model generation module 101 in the embodiment of the present application determines the preset disease history picture set as a training set, and performs disease type labeling on the preset disease history picture set to obtain a label set.
  • the disease types include atelectasis, solidification, infiltration, pneumothorax, edema, emphysema, fibrosis, effusion, pneumonia, pleural thickening, cardiac hypertrophy, nodules, masses, hernias, etc.
  • the LabelMe image labeling tool can be used to manually label the disease type.
  • model generation module 101 of the embodiment of the present application uses the training set and the label set to train the deep learning network model, including:
  • Step A Perform a convolution pooling operation on the training set according to the preset number of times of convolution pooling to obtain a dimensionality reduction data set;
  • Step B Perform a deconvolution operation on the dimensionality reduction data set according to the preset number of deconvolutions to obtain an increased dimensionality data set;
  • Step C Use a preset activation function to calculate a predicted value for the dimension-up data set, and calculate a loss value by using a pre-built loss function according to the predicted value and the label value contained in the label set.
  • Step D Compare the magnitude of the loss value with the preset loss threshold, if the loss value is greater than or equal to the loss threshold, return to the step A; if the loss value is less than the loss threshold, stop training, Obtain the disease recognition model.
  • the convolution pooling operation includes: convolution operation and pooling operation.
  • G is a convolution data set
  • is the training set
  • k is the size of the preset convolution kernel
  • f is the step size of the preset convolution operation
  • p is the preset data zero-filling matrix
  • the pooling operation in the embodiment of the present application is that the model generation module 101 performs a maximum pooling operation on the convolutional data set to obtain the dimensionality reduction data set.
  • the preset activation function includes:
  • ⁇ t represents the predicted value
  • s represents the data in the ascending data set.
  • the loss function includes:
  • T represents the loss value
  • z is the number of data in the training set
  • t is a positive integer
  • b t is the label value
  • the disease recognition model is obtained through the above-mentioned model training process, so that the disease recognition model has the ability to accurately extract features of preset parts.
  • the training data of the disease recognition model can be stored in the blockchain.
  • the model distillation module 102 is used to construct a distillation loss function according to the disease recognition model and a pre-built initial diagnosis model; perform distillation training on the initial diagnosis model according to the distillation loss function to obtain a first diagnosis model;
  • the set diagnosis target trains and adjusts the output of the first diagnosis model to obtain the target diagnosis model.
  • the initial diagnosis model is the same dimensional feature layer and output as the disease recognition model, and at the same time, the initial diagnosis model is a deep learning network model with a simpler structure and fewer parameters. Further, the feature layer is a fully connected layer that is connected before the softmax classifier in the disease recognition model as the initial diagnosis model.
  • the initial diagnosis model is a ResNet-8 model.
  • the model information in the disease recognition model is subsequently transferred to the initial diagnosis model, for To measure the difference in feature extraction capabilities between the disease recognition model and the pre-built initial diagnosis model, the model distillation module 102 constructs a distillation loss function according to the disease recognition model and the pre-built initial diagnosis model, and uses the distillation loss function to measure The feature extraction capabilities of the disease recognition model and the pre-built initial diagnosis model are different.
  • the model distillation module 102 uses the following methods to construct a distillation loss function according to the disease recognition model and the pre-built initial diagnosis model, including:
  • Normalized index processing is performed on the feature information of the first feature layer included in the disease recognition model to obtain the probability distribution function of the disease recognition model;
  • the normalized index processing on the feature information of the first feature layer included in the disease recognition model can be performed by the following formula:
  • X represents the probability distribution of the disease recognition model
  • n represents the dimension of the first feature layer
  • i represents the sequence number of the feature node of the first feature layer
  • represents a mapping function
  • Normalized index processing is performed on the feature information of the second feature layer included in the initial diagnosis model to obtain a probability distribution function of the initial diagnosis model;
  • the normalized index processing on the feature information of the second feature layer included in the initial diagnosis model can be performed by the following formula:
  • Y represents the probability distribution of the initial diagnosis model
  • m represents the dimension of the second feature layer
  • j represents the sequence number of the feature node of the second feature layer.
  • the maximum mean difference calculation is performed according to the probability distribution function of the disease recognition model and the probability distribution function of the initial diagnosis model to obtain the distillation loss function.
  • distillation loss function can be expressed as follows:
  • E represents the distillation loss function
  • H represents the mapping space of the mapping function
  • the disease recognition model has a large amount of parameters and strong feature extraction capabilities, but the disease recognition model also has high requirements for hardware.
  • the initial diagnosis model has There are fewer parameters, and the feature extraction capability of individual training is weaker. Therefore, in order to keep the initial diagnosis model lightweight and have the same feature extraction capabilities as the disease recognition model, the model distillation module 102 described in the embodiment of the present application Perform distillation training on the initial diagnosis model by using the disease recognition model according to the distillation loss function.
  • the model distillation module 102 in the embodiment of the present application uses the following methods to perform distillation training on the initial diagnosis model according to the distillation loss function, including: using the first feature layer as a training target, and continuously changing the Initial diagnosis model parameters, when the distillation loss value obtained by calculating the distillation loss function is less than a preset threshold, stop training to obtain the first diagnosis model.
  • this embodiment trains and adjusts the output of the first diagnosis model according to a preset diagnosis target.
  • the model distillation module 102 uses the following methods to train and output the first diagnostic model to obtain the target diagnostic model, including:
  • the diagnosis target is a diagnosed disease type
  • the diagnosed disease type is a disease type that occurs at the preset site, and when the diagnosed disease type is a disease that can be identified by the disease recognition model
  • type it is determined that the diagnostic target is not a newly-added diagnostic target; when the diagnosed disease type is a disease type that cannot be recognized by the disease recognition model, the diagnostic target is determined to be a newly-added diagnostic target.
  • diagnosis target is not a newly-added diagnosis target
  • output adjustment is performed on the first diagnosis model to obtain a target diagnosis model
  • a preset program framework is used to adjust the output of the first diagnostic model.
  • the first diagnostic model including: infiltration, pneumothorax, edema, and emphysema
  • the diagnosis target is the diagnosis of emphysema
  • the output of the first diagnosis model is adjusted to two outputs including: emphysema and non-emphysema.
  • the program framework is a pytorch framework.
  • the first diagnosis model is adjusted and trained to obtain the target diagnosis model.
  • the diagnosed disease type corresponding to the diagnostic target is the disease identification Types of diseases that cannot be identified by the model, for example: the types of diseases that can be identified by the disease identification model are infiltration, pneumothorax, edema, and emphysema, and the diagnosed disease type corresponding to the diagnosis target is new coronary pneumonia; therefore, this application is implemented For example, it is also necessary to adjust and train the first diagnostic model so that the first diagnostic model is familiar with the picture features of new coronary pneumonia.
  • model distillation module 102 uses the following methods to adjust and train the first diagnosis model to obtain the target diagnosis model, including:
  • the diagnosis picture set is a set of diagnosis pictures including the diagnosed disease type corresponding to the diagnosis target; the picture format of the diagnosis picture set is consistent with the picture format of the disease history picture set.
  • data enhancement processing is performed on the initial diagnosis picture set.
  • performing data enhancement processing on the initial diagnosis picture set includes: performing random horizontal flipping and random edge cropping on the pictures included in the diagnosis picture set to obtain the standard diagnosis picture set.
  • the diagnosed disease type corresponding to the diagnosis target is a disease type that cannot be recognized by the first diagnosis model. Therefore, the diagnosed disease type corresponding to the diagnosis target needs to be added to the first diagnosis model. The type of output included.
  • the preset program framework is used to adjust the output of the third diagnostic model.
  • there are 4 outputs in the third diagnostic model including: infiltration, pneumothorax, edema, and emphysema;
  • the diagnosis target is the diagnosis of emphysema, and the output of the third diagnosis model is adjusted to two outputs, including: emphysema and non-emphysema.
  • the program framework is a pytorch framework.
  • the picture analysis module 103 is configured to use the target diagnosis model to analyze the medical image picture to be analyzed when receiving the medical image picture to be analyzed to obtain the analysis result.
  • the medical image picture to be analyzed is the medical image picture to be analyzed of the predetermined part.
  • the format of the medical image picture to be analyzed is the same as the picture format included in the disease history picture set.
  • the picture analysis module 103 of the embodiment of the present application analyzes the medical image picture to be analyzed using the medical image picture analysis model, and outputs the disease probability; uses the Youden index principle to confirm the confidence threshold of the disease Comparing the disease probability with the confidence threshold; when the disease probability is greater than or equal to the confidence threshold, it is determined to be diseased; when the disease probability is less than the confidence threshold, it is determined to be Not sick.
  • FIG. 5 it is a schematic diagram of the structure of an electronic device that implements the medical image picture analysis method according to the present application.
  • the electronic device 1 may include a processor 10, a memory 11, and a bus, and may also include a computer program stored in the memory 11 and running on the processor 10, such as a medical image analysis program.
  • the memory 11 includes at least one type of readable storage medium, and the readable storage medium may be volatile or nonvolatile.
  • the readable storage medium includes flash memory, mobile hard disk, multimedia card, card-type memory (for example: SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, etc.
  • the memory 11 may be an internal storage unit of the electronic device 1 in some embodiments, for example, a mobile hard disk of the electronic device 1.
  • the memory 11 may also be an external storage device of the electronic device 1, such as a plug-in mobile hard disk, a smart memory card (SmartMediaCard, SMC), and a secure digital (SecureDigital, SD) equipped on the electronic device 1. Card, flash card (FlashCard), etc.
  • the memory 11 may also include both an internal storage unit of the electronic device 1 and an external storage device.
  • the memory 11 can be used not only to store application software and various data installed in the electronic device 1, such as the code of a medical image picture analysis program, etc., but also to temporarily store data that has been output or will be output.
  • the processor 10 may be composed of integrated circuits in some embodiments, for example, may be composed of a single packaged integrated circuit, or may be composed of multiple integrated circuits with the same function or different functions, including one or more Central Processing Unit (CPU), microprocessor, digital processing chip, graphics processor and a combination of various control chips, etc.
  • the processor 10 is the control core (ControlUnit) of the electronic device, which uses various interfaces and lines to connect the various components of the entire electronic device, and runs or executes programs or modules (such as medical imaging) stored in the memory 11 Picture analysis programs, etc.), and call data stored in the memory 11 to execute various functions of the electronic device 1 and process data.
  • ControlUnit the control core of the electronic device, which uses various interfaces and lines to connect the various components of the entire electronic device, and runs or executes programs or modules (such as medical imaging) stored in the memory 11 Picture analysis programs, etc.), and call data stored in the memory 11 to execute various functions of the electronic device 1 and process data.
  • the bus may be a peripheral component interconnection standard (PCI) bus or an extended industry standard architecture (EISA) bus or the like.
  • PCI peripheral component interconnection standard
  • EISA extended industry standard architecture
  • the bus can be divided into address bus, data bus, control bus and so on.
  • the bus is configured to implement connection and communication between the memory 11 and at least one processor 10 and the like.
  • FIG. 3 only shows an electronic device with components. Those skilled in the art can understand that the structure shown in FIG. 3 does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown in the figure. Components, or combinations of certain components, or different component arrangements.
  • the electronic device 1 may also include a power source (such as a battery) for supplying power to various components.
  • the power source may be logically connected to the at least one processor 10 through a power management device, thereby controlling power
  • the device implements functions such as charge management, discharge management, and power consumption management.
  • the power supply may also include any components such as one or more DC or AC power supplies, recharging devices, power failure detection circuits, power converters or inverters, and power status indicators.
  • the electronic device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be repeated here.
  • the electronic device 1 may also include a network interface.
  • the network interface may include a wired interface and/or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is usually used in the electronic device 1 Establish a communication connection with other electronic devices.
  • the electronic device 1 may also include a user interface.
  • the user interface may be a display (Display) and an input unit (such as a keyboard (Keyboard)).
  • the user interface may also be a standard wired interface or a wireless interface.
  • the display may be an LED display, a liquid crystal display, a touch liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light emitting diode) touch device, etc.
  • the display can also be appropriately called a display screen or a display unit, which is used to display the information processed in the electronic device 1 and to display a visualized user interface.
  • the medical image picture analysis program 12 stored in the memory 11 in the electronic device 1 is a combination of multiple instructions. When running in the processor 10, it can realize:
  • the target diagnostic model is used to analyze the medical image picture to be analyzed to obtain an analysis result.
  • the integrated module/unit of the electronic device 1 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory).
  • the computer usable storage medium may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required by at least one function, etc.; the storage data area may store a block chain node Use the created data, etc.
  • modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the goals of the solutions of the embodiments.
  • the functional modules in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit may be implemented in the form of hardware, or may be implemented in the form of hardware plus software functional modules.
  • the blockchain referred to in this application is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
  • Blockchain essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information for verification. The validity of the information (anti-counterfeiting) and the generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Biology (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Image Analysis (AREA)

Abstract

一种医学影像图片分析方法,包括:获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型(S1);根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数(S2);根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型(S3);根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型(S4);当接收到待分析医学影像图片时,利用目标诊断模型对待分析医学影像图片进行分析,得到分析结果(S5)。训练模型的数据可以存储在区块链中。还包括一种医学影像图片分析装置、电子设备以及一种计算机可读存储介质。可以降低医学影像图片分析的模型计算资源消耗。

Description

医学影像图片分析方法、装置、电子设备及可读存储介质
本申请要求于2020年09月22日提交中国专利局、申请号为202011003181.5,发明名称为“医学影像图片分析方法、装置、电子设备及可读存储介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能领域,尤其涉及一种医学影像图片分析方法、装置、电子设备及可读存储介质。
背景技术
发明人意识到,随着人工智能的发展,利用基于深度学习模型训练的医学影像图片分析的模型对医学影像图片进行分析从而辅助疾病诊断已经越来越普遍,但是训练深度学习模型通常需要很高的硬件门槛,不能迁移到移动端或者计算资源匮乏的地方,若直接训练轻量化的模型,不仅特征提取能力弱,且准确度低,因此,需要一种保持高准确度的且不需要太多计算资源的医学影像图片分析方法。
发明内容
本申请提供的一种医学影像图片分析方法,包括:
获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
本申请还提供一种医学影像图片分析装置,所述装置包括:
模型生成模块,用于获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
模型蒸馏模块,用于根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
图片分析模块,用于当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
本申请还提供一种电子设备,所述电子设备包括:
存储器,存储至少一个指令;及
处理器,执行所述存储器中存储的指令以实现如下所述的医学影像图片分析方法:
获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片 进行分析,得到分析结果。
本申请还提供一种计算机可读存储介质,所述计算机可读存储介质中存储有至少一个指令,所述至少一个指令被电子设备中的处理器执行以实现如下所述的医学影像图片分析方法:
获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
附图说明
图1为本申请一实施例提供的医学影像图片分析方法的流程示意图;
图2为本申请一实施例提供的医学影像图片分析方法中得到蒸馏损失函数的详细流程示意图;
图3为本申请一实施例提供的医学影像图片分析方法中得到目标诊断模型的详细流程示意图;
图4为本申请一实施例提供的医学影像图片分析装置的模块示意图;
图5为本申请一实施例提供的实现医学影像图片分析方法的电子设备的内部结构示意图;
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
具体实施方式
应当理解,此处所描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
本申请提供一种医学影像图片分析方法。参照图1所示,为本申请一实施例提供的医学影像图片分析方法的流程示意图。该方法可以由一个装置执行,该装置可以由软件和/或硬件实现。
在本实施例中,医学影像图片分析方法包括:
S1、获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
本申请实施例中,所述预设部位疾病历史图片集为预设部位的病人的医学影像图片,如病人胸部的CXR(胸部X射线,ChestXRay)图片的集合,所述预设部位疾病历史图片集可以从某机构的CXR数据库中进行获取,例如:NIH(美国国立卫生研究院,NationalInstitutesofHealth)临床中心的CXR数据库。
较佳地,本申请实施例中所述卷积神经网络模型可以为全卷积神经网络模型。
详细地,本申请实施例中将所述预设部位疾病历史图片集确定为训练集,对所述预设部位疾病历史图片集进行疾病类型标记得到标签集。其中,所述疾病类型为肺不张、变实、浸润、气胸、水肿、肺气肿、纤维变性、积液、肺炎、胸膜增厚、心脏肥大、结节、肿块和疝气等。较佳地,本申请实施例可使用LabelMe图片标注工具人工进行疾病类型标记。
进一步地,本申请实施例利用所述训练集及所述标签集训练所述深度学习网络模型,包括:
步骤A:根据预设的卷积池化次数,对所述训练集进行卷积池化操作得到得到降维数据集;
步骤B:根据预设的反卷积次数,对所述降维数据集进行反卷积操作得到升维数据集;
步骤C:利用预设的激活函数对所述升维数据集进行计算得到预测值,根据所述预测值和所述标签集包含的标签值,利用预构建的损失函数计算得到损失值。
步骤D:对比所述损失值与预设的损失阈值的大小,若所述损失值大于或等于所述损失阈值,返回A;若所述损失值小于所述损失阈值,停止训练,得到所述疾病识别模型。
详细地,所述卷积池化操作包括:卷积操作和池化操作。
进一步地,所述卷积操作为:
Figure PCTCN2020125474-appb-000001
其中,G为卷积数据集,ω为所述训练集,k为预设卷积核的大小,f为预设卷积操作的步幅,p为预设数据补零矩阵。
较佳地,本申请实施例中所述池化操作为对所述卷积数据集进行最大池化操作得到所述降维数据集。
进一步地,所述预设的激活函数包括:
Figure PCTCN2020125474-appb-000002
其中,μ t表示所述预测值,s表示所述升维数据集中的数据。
详细地,所述损失函数包括:
Figure PCTCN2020125474-appb-000003
其中,T表示所述损失值,z为所述训练集的数据数目,t为正整数,b t为所述标签值。
通过上述模型训练的过程得到所述疾病识别模型,使所述疾病识别模型具备对预设部位准确的特征提取能力。
本申请的另一实施例中,为了保证病人的数据隐私,所述疾病识别模型的训练数据可以存储在区块链中。
S2、根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
本申请实施例中,所述初始诊断模型与所述疾病识别模型具有相同维度特征层及输出,同时,所述初始诊断模型为结构更简单、参数更少的深度学习网络模型。进一步地,所述特征层为所述初始诊断模型与所述疾病识别模型中softmax分类器前连接的全连接层。
较佳地,所述初始诊断模型为ResNet-8模型。
进一步地,本申请实施例中,为了让所述初始诊断模型拥有与所述疾病识别模型相同的特征提取能力,后续将所述疾病识别模型中的模型信息转移到所述初始诊断模型中,为了衡量所述疾病识别模型与预构建的初始诊断模型的特征提取能力差异,根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数,利用所述蒸馏损失函数衡量所述疾病识别模型与预构建的初始诊断模型的特征提取能力差异。
详细地,本申请实施例中,参照图2所示,所述根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数,包括:
S21、对所述疾病识别模型中包含的第一特征层的特征信息进行归一化指数处理,得到疾病识别模型概率分布函数;
详细地,本申请实施例中,所述对所述疾病识别模型中包含的第一特征层的特征信息进行归一化指数处理可用如下公式进行:
Figure PCTCN2020125474-appb-000004
其中,X表示所述疾病识别模型概率分布,n表示所述第一特征层的维度,i表示所述第一特征层的特征节点序号,φ表示映射函数。
S22、对所述初始诊断模型中包含的第二特征层的特征信息进行归一化指数处理,得到初始诊断模型概率分布函数;
详细地,本申请实施例中,所述对所述初始诊断模型中包含的第二特征层的特征信息进行归一化指数处理可用如下公式进行:
Figure PCTCN2020125474-appb-000005
其中,Y表示所述初始诊断模型概率分布,m表示所述第二特征层的维度,j表示所述第二特征层的特征节点序号。
S23、根据所述疾病识别模型概率分布函数与所述初始诊断模型概率分布函数进行最大均值差异计算,得到所述蒸馏损失函数。
详细地,所述蒸馏损失函数可用如下公示表示:
Figure PCTCN2020125474-appb-000006
其中,E表示所述蒸馏损失函数,H表示所述映射函数的映射空间。
S3、根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
本申请实施例中,所述疾病识别模型的参数量大,特征提取能力强,但所述疾病识别模型对硬件的要求也很高,和所述疾病识别模型相比,所述初始诊断模型的参数量更少,单独训练的特征提取能力更弱,因此,为了让所述初始诊断模型保持轻量化的同时具有所述疾病识别模型一样的特征提取能力,本申请实施例根据所述蒸馏损失函数利用所述疾病识别模型对所述初始诊断模型进行蒸馏训练。
详细地,本申请实施例中根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,包括:利用所述第一特征层作为训练目标,不断更迭所述初始诊断模型参数,当计算所述蒸馏损失函数得到的蒸馏损失值小于预设阈值时,停止训练,得到所述第一诊断模型。
S4、根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
为了模型进一步地轻量化,并提高所述诊断模型的可迁移性,本实施例根据预设的诊断目标对所述第一诊断模型进行训练及输出调整。
详细地,本申请实施例中,所述对所述第一诊断模型进行训练及输出调整,得到目标诊断模型,包括:
步骤I、判断所述诊断目标是否为新增诊断目标;
本申请实施例中,所述诊断目标为诊断的疾病类型,所述诊断的疾病类型为所述预设部位发生的疾病类型,当所述诊断的疾病类型为所述疾病识别模型可以识别的疾病类型时,确定所述诊断目标不为新增诊断目标;当所述诊断的疾病类型为所述疾病识别模型无法识别的疾病类型时,确定所述诊断目标为新增诊断目标。
步骤II、当所述诊断目标不为新增诊断目标时,对所述第一诊断模型进行输出调整,得到目标诊断模型;
本申请实施例中,利用预设的程序框架对所述第一诊断模型进行输出调整,例如:所述第一诊断模型中共有4路输出,包括:浸润、气胸、水肿、肺气肿;所述诊断目标为肺 气肿诊断,将所述第一诊断模型的输出调整为两路输出,包括:肺气肿、非肺气肿。较佳地,所述程序框架为pytorch框架。
步骤III、当所述诊断目标为新增诊断目标时,调整并训练所述第一诊断模型,得到目标诊断模型。
本申请实施例中,当所述诊断目标为新增诊断目标时,所述第一诊断模型虽然具有预设部位的特征提取能力,但所述诊断目标对应的诊断的疾病类型为所述疾病识别模型不可以识别的疾病类型,例如:所述疾病识别模型可以识别的疾病类型为浸润、气胸、水肿、肺气肿,所述诊断目标对应的诊断的疾病类型为新冠肺炎;因此,本申请实施例还需要调整并训练所述第一诊断模型,使得所述第一诊断模型熟悉新冠肺炎的图片特征。
详细地,参照图3所示,本申请实施例中所述调整并训练所述第一诊断模型,得到目标诊断模型,包括:
S431、获取诊断图片集,对所述诊断图片集进行标注,得到初始诊断图片集;
本申请实施例中,所述诊断图片集为包含所述诊断目标对应的诊断的疾病类型的诊断图片的集合;所述诊断图片集与所述疾病历史图片集的图片格式一致。
S432、对所述初始诊断图片集进行数据增强处理,得到标准诊断图片集;
本申请实施例中,为了增强后续模型的泛化能力,对所述初始诊断图片集进行数据增强处理。
详细地,本申请实施例中,对所述初始诊断图片集进行数据增强处理,包括:对所述诊断图片集中包含的图片进行随机水平翻转及随机边缘裁剪,得到所述标准诊断图片集。
S433、根据所述诊断目标对所述第一诊断模型进行输出添加,得到第二诊断模型;
本申请实施例中,所述诊断目标对应的诊断的疾病类型是所述第一诊断模型不能识别的疾病类型,因此需要将所述诊断目标对应的诊断的疾病类型添加进所述第一诊断模型包含的输出类型。
S434、利用所述标准诊断图片集对所述第二诊断模型进行迭代训练,直至所述第二诊断模型收敛,得到第三诊断模型;
S435、对所述第三诊断模型进行输出调整,得到目标诊断模型;
本申请实施例中,利用预设的程序框架对所述第三诊断模型进行输出调整,例如:所述第三诊断模型中共有4路输出,包括:浸润、气胸、水肿、肺气肿;所述诊断目标为肺气肿诊断,将所述第三诊断模型的输出调整为两路输出,包括:肺气肿、非肺气肿。较佳地,所述程序框架为pytorch框架。
S5、当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
本申请实施例中,所述待分析医学影像图片为所述预设部位的待分析医学影像图片。所述待分析医学影像图片的格式与所述疾病历史图片集包含的图片格式相同。
进一步地,本申请实施例利用所述医学影像图片分析模型对所述待分析医学影像图片进行分析,输出患病概率;利用约登指数原则确认所述患病的置信阈值;将所述患病概率与所述置信阈值进行比较;当所述患病概率大于或等于所述置信阈值时,判断为患病,当所述患病概率小于所述置信阈值时,判断为未患病。
如图4所示,是本申请医学影像图片分析装置的功能模块图。
本申请所述医学影像图片分析装置100可以安装于电子设备中。根据实现的功能,所述医学影像图片分析装置可以包括模型生成模块101、模型蒸馏模块102、图片分析模块103。本发所述模块也可以称之为单元,是指一种能够被电子设备处理器所执行,并且能够完成固定功能的一系列计算机程序段,其存储在电子设备的存储器中。
在本实施例中,关于各模块/单元的功能如下:
所述模型生成模块101用于获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型。
本申请实施例中,所述预设部位疾病历史图片集为预设部位的病人的医学影像图片,如病人胸部的CXR图片的集合,所述预设部位疾病历史图片集可以从某机构的CXR数据库中进行获取,例如:NIH临床中心的CXR数据库。
较佳地,本申请实施例中所述卷积神经网络模型可以为全卷积神经网络模型。
详细地,本申请实施例中所述模型生成模块101将所述预设部位疾病历史图片集确定为训练集,对所述预设部位疾病历史图片集进行疾病类型标记得到标签集。其中,所述疾病类型为肺不张、变实、浸润、气胸、水肿、肺气肿、纤维变性、积液、肺炎、胸膜增厚、心脏肥大、结节、肿块和疝气等。较佳地,本申请实施例可使用LabelMe图片标注工具人工进行疾病类型标记。
进一步地,本申请实施例所述模型生成模块101利用所述训练集及所述标签集训练所述深度学习网络模型,包括:
步骤A:根据预设的卷积池化次数,对所述训练集进行卷积池化操作得到得到降维数据集;
步骤B:根据预设的反卷积次数,对所述降维数据集进行反卷积操作得到升维数据集;
步骤C:利用预设的激活函数对所述升维数据集进行计算得到预测值,根据所述预测值和所述标签集包含的标签值,利用预构建的损失函数计算得到损失值。
步骤D:对比所述损失值与预设的损失阈值的大小,若所述损失值大于或等于所述损失阈值,返回所述步骤A;若所述损失值小于所述损失阈值,停止训练,得到所述疾病识别模型。
详细地,所述卷积池化操作包括:卷积操作和池化操作。
进一步地,所述卷积操作为:
Figure PCTCN2020125474-appb-000007
其中,G为卷积数据集,ω为所述训练集,k为预设卷积核的大小,f为预设卷积操作的步幅,p为预设数据补零矩阵。
较佳地,本申请实施例中所述池化操作为所述模型生成模块101对所述卷积数据集进行最大池化操作得到所述降维数据集。
进一步地,所述预设的激活函数包括:
Figure PCTCN2020125474-appb-000008
其中,μ t表示所述预测值,s表示所述升维数据集中的数据。
详细地,所述损失函数包括:
Figure PCTCN2020125474-appb-000009
其中,T表示所述损失值,z为所述训练集的数据数目,t为正整数,b t为所述标签值。
通过上述的模型训练的过程得到所述疾病识别模型,使所述疾病识别模型具备对预设部位准确的特征提取能力。
本申请的另一实施例中,为了保证病人的数据隐私,所述疾病识别模型的训练数据可以存储在区块链中。
所述模型蒸馏模块102用于根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏 损失函数;根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型。
本申请实施例中,所述初始诊断模型为与所述疾病识别模型具有相同维度特征层及输出,同时,所述初始诊断模型为结构更简单、参数更少的深度学习网络模型。进一步地,所述特征层为所述初始诊断模型为与所述疾病识别模型中softmax分类器前连接的的全连接层。
较佳地,所述初始诊断模型为ResNet-8模型。
进一步地,本申请实施例中,为了让所述初始诊断模型拥有与所述疾病识别模型相同的特征提取能力,后续将所述疾病识别模型中的模型信息转移到所述初始诊断模型中,为了衡量所述疾病识别模型与预构建的初始诊断模型的特征提取能力差异,所述模型蒸馏模块102根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数,利用所述蒸馏损失函数衡量所述疾病识别模型与预构建的初始诊断模型的特征提取能力差异。
详细地,本申请实施例中,所述模型蒸馏模块102根据所述疾病识别模型与预构建的初始诊断模型利用如下手段构建蒸馏损失函数,包括:
对所述疾病识别模型中包含的第一特征层的特征信息进行归一化指数处理,得到疾病识别模型概率分布函数;
详细地,本申请实施例中,所述对所述疾病识别模型中包含的第一特征层的特征信息进行归一化指数处理可用如下公式进行:
Figure PCTCN2020125474-appb-000010
其中,X表示所述疾病识别模型概率分布,n表示所述第一特征层的维度,i表示所述第一特征层的特征节点序号,φ表示映射函数。
对所述初始诊断模型中包含的第二特征层的特征信息进行归一化指数处理,得到初始诊断模型概率分布函数;
详细地,本申请实施例中,所述对所述初始诊断模型中包含的第二特征层的特征信息进行归一化指数处理可用如下公式进行:
Figure PCTCN2020125474-appb-000011
其中,Y表示所述初始诊断模型概率分布,m表示所述第二特征层的维度,j表示所述第二特征层的特征节点序号。
根据所述疾病识别模型概率分布函数与所述初始诊断模型概率分布函数进行最大均值差异计算,得到所述蒸馏损失函数。
详细地,所述蒸馏损失函数可用如下公示表示:
Figure PCTCN2020125474-appb-000012
其中,E表示所述蒸馏损失函数,H表示所述映射函数的映射空间。
本申请实施例中,所述疾病识别模型的参数量大,特征提取能力强,但所述疾病识别模型对硬件的要求也很高,和所述疾病识别模型相比,所述初始诊断模型的参数量更少,单独训练的特征提取能力更弱,因此,为了让所述初始诊断模型保持轻量化的同时具有所述疾病识别模型一样的特征提取能力,本申请实施例所述模型蒸馏模块102根据所述蒸馏损失函数利用所述疾病识别模型对所述初始诊断模型进行蒸馏训练。
详细地,本申请实施例中所述模型蒸馏模块102根据所述蒸馏损失函数利用如下手段对所述初始诊断模型进行蒸馏训练,包括:利用所述第一特征层作为训练目标,不断更迭所述初始诊断模型参数,当计算所述蒸馏损失函数得到的蒸馏损失值小于预设阈值时,停止训练,得到所述第一诊断模型。
本申请实施例中,为了模型进一步地轻量化,并提高所述诊断模型的可迁移性,本实施例根据预设的诊断目标对所述第一诊断模型进行训练及输出调整。
详细地,本申请实施例中,所述模型蒸馏模块102利用如下手段对所述第一诊断模型进行训练及输出调整,得到目标诊断模型,包括:
判断所述诊断目标是否为新增诊断目标;
本申请实施例中,所述诊断目标为诊断的疾病类型,所述诊断的疾病类型为所述预设部位发生的疾病类型,当所述诊断的疾病类型为所述疾病识别模型可以识别的疾病类型时,确定所述诊断目标不为新增诊断目标;当所述诊断的疾病类型为所述疾病识别模型无法识别的疾病类型时,确定所述诊断目标为新增诊断目标。
当所述诊断目标不为新增诊断目标时,对所述第一诊断模型进行输出调整,得到目标诊断模型;
本申请实施例中,利用预设的程序框架对对所述第一诊断模型进行输出调整,例如:所述第一诊断模型中共有4路输出包括:浸润、气胸、水肿、肺气肿;所述诊断目标为肺气肿诊断,将所述第一诊断模型的输出调整为两路输出包括:肺气肿、非肺气肿。较佳地,所述程序框架为pytorch框架。
当所述诊断目标为新增诊断目标时,调整并训练所述第一诊断模型,得到目标诊断模型。
本申请实施例中,当所述诊断目标为新增诊断目标时,所述第一诊断模型虽然具有预设部位的特征提取能力,但所述诊断目标对应的诊断的疾病类型为所述疾病识别模型不可以识别的疾病类型,例如:所述疾病识别模型可以识别的疾病类型为浸润、气胸、水肿、肺气肿,所述诊断目标对应的诊断的疾病类型为新冠肺炎;因此,本申请实施例还需要调整并训练所述第一诊断模型,使得所述第一诊断模型熟悉新冠肺炎的图片特征。
详细地,本申请实施例中所述模型蒸馏模块102利用如下手段调整并训练所述第一诊断模型,得到目标诊断模型,包括:
获取诊断图片集,对所述诊断图片集进行标注,得到初始诊断图片集;
本申请实施例中,所述诊断图片集为包含所述诊断目标对应的诊断的疾病类型的诊断图片的集合;所述诊断图片集与所述疾病历史图片集的图片格式一致。
对所述初始诊断图片集进行数据增强处理,得到标准诊断图片集;
本申请实施例中,为了增强后续模型的泛化能力,对所述初始诊断图片集进行数据增强处理。
详细地,本申请实施例中,对所述初始诊断图片集进行数据增强处理,包括:对所述诊断图片集中包含的图片进行随机水平翻转及随机边缘裁剪,得到所述标准诊断图片集。
根据所述诊断目标对所述第一诊断模型进行输出添加,得到第二诊断模型;
本申请实施例中,所述诊断目标对应的诊断的疾病类型是所述第一诊断模型不能识别的疾病类型,因此需要将所述诊断目标对应的诊断的疾病类型添加进所述第一诊断模型包含的输出类型。
利用所述标准诊断图片集对所述第二诊断模型进行迭代训练,直至所述第二诊断模型收敛,得到第三诊断模型;
对所述第三诊断模型进行输出调整,得到目标诊断模型;
本申请实施例中,利用预设的程序框架对所述第三诊断模型进行输出调整,例如:所述第三诊断模型中共有4路输出,包括:浸润、气胸、水肿、肺气肿;所述诊断目标为肺 气肿诊断,将所述第三诊断模型的输出调整为两路输出,包括:肺气肿、非肺气肿。较佳地,所述程序框架为pytorch框架。
所述图片分析模块103用于当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
本申请实施例中,所述待分析医学影像图片为所述预设部位的待分析医学影像图片。所述待分析医学影像图片的格式与所述疾病历史图片集包含的图片格式相同。
进一步地,本申请实施例所述图片分析模块103利用所述医学影像图片分析模型对所述待分析医学影像图片进行分析,输出患病概率;利用约登指数原则确认所述患病的置信阈值;将所述患病概率与所述置信阈值进行比较;当所述患病概率大于或等于所述置信阈值时,判断为患病,当所述患病概率小于所述置信阈值时,判断为未患病。
如图5所示,是本申请实现医学影像图片分析方法的电子设备的结构示意图。
所述电子设备1可以包括处理器10、存储器11和总线,还可以包括存储在所述存储器11中并可在所述处理器10上运行的计算机程序,如医学影像图片分析程序。
其中,所述存储器11至少包括一种类型的可读存储介质,可读存储介质可以是易失性的,也可以是非易失性的。具体的,所述可读存储介质包括闪存、移动硬盘、多媒体卡、卡型存储器(例如:SD或DX存储器等)、磁性存储器、磁盘、光盘等。所述存储器11在一些实施例中可以是电子设备1的内部存储单元,例如该电子设备1的移动硬盘。所述存储器11在另一些实施例中也可以是电子设备1的外部存储设备,例如电子设备1上配备的插接式移动硬盘、智能存储卡(SmartMediaCard,SMC)、安全数字(SecureDigital,SD)卡、闪存卡(FlashCard)等。进一步地,所述存储器11还可以既包括电子设备1的内部存储单元也包括外部存储设备。所述存储器11不仅可以用于存储安装于电子设备1的应用软件及各类数据,例如医学影像图片分析程序的代码等,还可以用于暂时地存储已经输出或者将要输出的数据。
所述处理器10在一些实施例中可以由集成电路组成,例如可以由单个封装的集成电路所组成,也可以是由多个相同功能或不同功能封装的集成电路所组成,包括一个或者多个中央处理器(CentralProcessingunit,CPU)、微处理器、数字处理芯片、图形处理器及各种控制芯片的组合等。所述处理器10是所述电子设备的控制核心(ControlUnit),利用各种接口和线路连接整个电子设备的各个部件,通过运行或执行存储在所述存储器11内的程序或者模块(例如医学影像图片分析程序等),以及调用存储在所述存储器11内的数据,以执行电子设备1的各种功能和处理数据。
所述总线可以是外设部件互连标准(peripheralcomponentinterconnect,简称PCI)总线或扩展工业标准结构(extendedindustrystandardarchitecture,简称EISA)总线等。该总线可以分为地址总线、数据总线、控制总线等。所述总线被设置为实现所述存储器11以及至少一个处理器10等之间的连接通信。
图3仅示出了具有部件的电子设备,本领域技术人员可以理解的是,图3示出的结构并不构成对所述电子设备1的限定,可以包括比图示更少或者更多的部件,或者组合某些部件,或者不同的部件布置。
例如,尽管未示出,所述电子设备1还可以包括给各个部件供电的电源(比如电池),优选地,电源可以通过电源管理装置与所述至少一个处理器10逻辑相连,从而通过电源管理装置实现充电管理、放电管理、以及功耗管理等功能。电源还可以包括一个或一个以上的直流或交流电源、再充电装置、电源故障检测电路、电源转换器或者逆变器、电源状态指示器等任意组件。所述电子设备1还可以包括多种传感器、蓝牙模块、Wi-Fi模块等,在此不再赘述。
进一步地,所述电子设备1还可以包括网络接口,可选地,所述网络接口可以包括有 线接口和/或无线接口(如WI-FI接口、蓝牙接口等),通常用于在该电子设备1与其他电子设备之间建立通信连接。
可选地,该电子设备1还可以包括用户接口,用户接口可以是显示器(Display)、输入单元(比如键盘(Keyboard)),可选地,用户接口还可以是标准的有线接口、无线接口。可选地,在一些实施例中,显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(OrganicLight-EmittingDiode,有机发光二极管)触摸器等。其中,显示器也可以适当的称为显示屏或显示单元,用于显示在电子设备1中处理的信息以及用于显示可视化的用户界面。
应该了解,所述实施例仅为说明之用,在专利申请范围上并不受此结构的限制。
所述电子设备1中的所述存储器11存储的医学影像图片分析程序12是多个指令的组合,在所述处理器10中运行时,可以实现:
获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
具体地,所述处理器10对上述指令的具体实现方法可参考图1对应实施例中相关步骤的描述,在此不赘述。
进一步地,所述电子设备1集成的模块/单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储介质中。所述计算机可读介质可以包括:能够携带所述计算机程序代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM,Read-OnlyMemory)。
进一步地,所述计算机可用存储介质可主要包括存储程序区和存储数据区,其中,存储程序区可存储操作系统、至少一个功能所需的应用程序等;存储数据区可存储根据区块链节点的使用所创建的数据等。
在本申请所提供的几个实施例中,应该理解到,所揭露的设备,装置和方法,可以通过其它的方式实现。例如,以上所描述的装置实施例仅仅是示意性的,例如,所述模块的划分,仅仅为一种逻辑功能划分,实际实现时可以有另外的划分方式。
所述作为分离部件说明的模块可以是或者也可以不是物理上分开的,作为模块显示的部件可以是或者也可以不是物理单元,即可以位于一个地方,或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目标。
另外,在本申请各个实施例中的各功能模块可以集成在一个处理单元中,也可以是各个单元单独物理存在,也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现,也可以采用硬件加软件功能模块的形式实现。
对于本领域技术人员而言,显然本申请不限于上述示范性实施例的细节,而且在不背离本申请的精神或基本特征的情况下,能够以其他的具体形式实现本申请。
因此,无论从哪一点来看,均应将实施例看作是示范性的,而且是非限制性的,本申请的范围由所附权利要求而不是上述说明限定,因此旨在将落在权利要求的等同要件的含义和范围内的所有变化涵括在本申请内。不应将权利要求中的任何附关联图标记视为限制所涉及的权利要求。
本申请所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain),本质上是一个去中心化的数据库,是一串使用密码学方法相关联产生的数据块,每一个数据块中包含了一批次网络交易的信息,用于验 证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
此外,显然“包括”一词不排除其他单元或步骤,单数不排除复数。系统权利要求中陈述的多个单元或装置也可以由一个单元或装置通过软件或者硬件来实现。第二等词语用来表示名称,而并不表示任何特定的顺序。
最后应说明的是,以上实施例仅用以说明本申请的技术方案而非限制,尽管参照较佳实施例对本申请进行了详细说明,本领域的普通技术人员应当理解,可以对本申请的技术方案进行修改或等同替换,而不脱离本申请技术方案的精神和范围。

Claims (20)

  1. 一种医学影像图片分析方法,其中,所述方法包括:
    获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
    根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
    根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
    根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
    当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
  2. 如权利要求1所述的医学影像图片分析方法,其中,所述利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型,包括:
    将所述预设部位疾病历史图片集确定为训练集;
    对所述预设部位疾病历史图片集进行疾病类型标记得到标签集;
    利用所述训练集及所述标签集训练所述深度学习网络模型,得到所述疾病识别模型。
  3. 如权利要求2所述的医学影像图片分析方法,其中,所述利用所述训练集及所述标签集训练所述深度学习网络模型,得到所述疾病识别模型,包括:
    降维步骤:根据预设的卷积池化次数,对所述训练集进行卷积池化操作,得到降维数据集;
    升维步骤:根据预设的反卷积次数,对所述降维数据集进行反卷积操作,得到升维数据集;
    损失值计算步骤:利用预设的激活函数对所述升维数据集进行计算得到预测值,根据所述预测值和所述标签集包含的标签值,利用预构建的损失函数的输入参数计算得到损失值;
    损失值对比步骤:对比所述损失值与预设的损失阈值的大小,若所述损失值大于或等于所述损失阈值,返回所述降维步骤;若所述损失值小于所述损失阈值,停止训练,得到所述疾病识别模型。
  4. 如权利要求1所述的医学影像图片分析方法,其中,所述根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数,包括:
    对所述疾病识别模型中包含的第一特征层的特征信息进行归一化指数处理,得到疾病识别模型概率分布函数;
    对所述初始诊断模型中包含的第二特征层的特征信息进行归一化指数处理,得到初始诊断模型概率分布函数;
    根据所述疾病识别模型概率分布函数与所述初始诊断模型概率分布函数进行最大均值差异计算,得到所述蒸馏损失函数。
  5. 如权利要求4所述的医学影像图片分析方法,其中,所述根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型,包括:
    利用所述第一特征层作为训练目标,不断更迭所述初始诊断模型的参数,当计算所述蒸馏损失函数得到的蒸馏损失值小于预设阈值时,停止训练,得到所述第一诊断模型。
  6. 如权利要求1所述的医学影像图片分析方法,其中,所述根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型,包括:
    判断所述诊断目标是否为新增诊断目标;
    当所述诊断目标不为新增诊断目标时,对所述第一诊断模型进行输出调整,得到所述目标诊断模型;或者
    当所述诊断目标为新增诊断目标时,调整并训练所述第一诊断模型,得到所述目标诊 断模型。
  7. 如权利要求6所述的医学影像图片分析方法,其中,所述调整并训练所述第一诊断模型,得到目标诊断模型,包括:
    获取诊断图片集,对所述诊断图片集进行标注,得到初始诊断图片集;
    对所述初始诊断图片集进行数据增强处理,得到标准诊断图片集;
    根据所述诊断目标对所述第一诊断模型进行输出添加,得到第二诊断模型;
    利用所述标准诊断图片集对所述第二诊断模型进行迭代训练,直至所述第二诊断模型收敛,得到第三诊断模型;
    对所述第三诊断模型进行输出调整,得到所述目标诊断模型。
  8. 一种医学影像图片分析装置,其中,所述装置包括:
    模型生成模块,用于获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
    知识蒸馏模块,用于根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
    图片分析模块,用于当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
  9. 一种电子设备,其中,所述电子设备包括:
    至少一个处理器;以及,
    与所述至少一个处理器通信连接的存储器;其中,
    所述存储器存储有可被所述至少一个处理器执行的指令,所述指令被所述至少一个处理器执行,以使所述至少一个处理器能够执行如下所述的医学影像图片分析方法:
    获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
    根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
    根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
    根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
    当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
  10. 如权利要求9所述的电子设备,其中,所述利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型,包括:
    将所述预设部位疾病历史图片集确定为训练集;
    对所述预设部位疾病历史图片集进行疾病类型标记得到标签集;
    利用所述训练集及所述标签集训练所述深度学习网络模型,得到所述疾病识别模型。
  11. 如权利要求10所述的电子设备,其中,所述利用所述训练集及所述标签集训练所述深度学习网络模型,得到所述疾病识别模型,包括:
    降维步骤:根据预设的卷积池化次数,对所述训练集进行卷积池化操作,得到降维数据集;
    升维步骤:根据预设的反卷积次数,对所述降维数据集进行反卷积操作,得到升维数据集;
    损失值计算步骤:利用预设的激活函数对所述升维数据集进行计算得到预测值,根据所述预测值和所述标签集包含的标签值,利用预构建的损失函数的输入参数计算得到损失值;
    损失值对比步骤:对比所述损失值与预设的损失阈值的大小,若所述损失值大于或等于所述损失阈值,返回所述降维步骤;若所述损失值小于所述损失阈值,停止训练,得到 所述疾病识别模型。
  12. 如权利要求9所述的电子设备,其中,所述根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数,包括:
    对所述疾病识别模型中包含的第一特征层的特征信息进行归一化指数处理,得到疾病识别模型概率分布函数;
    对所述初始诊断模型中包含的第二特征层的特征信息进行归一化指数处理,得到初始诊断模型概率分布函数;
    根据所述疾病识别模型概率分布函数与所述初始诊断模型概率分布函数进行最大均值差异计算,得到所述蒸馏损失函数。
  13. 如权利要求9所述的电子设备,其中,所述根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型,包括:
    判断所述诊断目标是否为新增诊断目标;
    当所述诊断目标不为新增诊断目标时,对所述第一诊断模型进行输出调整,得到所述目标诊断模型;或者
    当所述诊断目标为新增诊断目标时,调整并训练所述第一诊断模型,得到所述目标诊断模型。
  14. 如权利要求13所述的电子设备,其中,所述调整并训练所述第一诊断模型,得到目标诊断模型,包括:
    获取诊断图片集,对所述诊断图片集进行标注,得到初始诊断图片集;
    对所述初始诊断图片集进行数据增强处理,得到标准诊断图片集;
    根据所述诊断目标对所述第一诊断模型进行输出添加,得到第二诊断模型;
    利用所述标准诊断图片集对所述第二诊断模型进行迭代训练,直至所述第二诊断模型收敛,得到第三诊断模型;
    对所述第三诊断模型进行输出调整,得到所述目标诊断模型。
  15. 一种计算机可读存储介质,存储有计算机程序,其中,所述计算机程序被处理器执行时实现如下所述的医学影像图片分析方法:
    获取预设部位疾病历史图片集,利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型;
    根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数;
    根据所述蒸馏损失函数对所述初始诊断模型进行蒸馏训练,得到第一诊断模型;
    根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型;
    当接收到待分析医学影像图片时,利用所述目标诊断模型对所述待分析医学影像图片进行分析,得到分析结果。
  16. 如权利要求15所述的计算机可读存储介质,其中,所述利用所述预设部位疾病历史图片集训练预构建的深度学习网络模型得到疾病识别模型,包括:
    将所述预设部位疾病历史图片集确定为训练集;
    对所述预设部位疾病历史图片集进行疾病类型标记得到标签集;
    利用所述训练集及所述标签集训练所述深度学习网络模型,得到所述疾病识别模型。
  17. 如权利要求16所述的计算机可读存储介质,其中,所述利用所述训练集及所述标签集训练所述深度学习网络模型,得到所述疾病识别模型,包括:
    降维步骤:根据预设的卷积池化次数,对所述训练集进行卷积池化操作,得到降维数据集;
    升维步骤:根据预设的反卷积次数,对所述降维数据集进行反卷积操作,得到升维数据集;
    损失值计算步骤:利用预设的激活函数对所述升维数据集进行计算得到预测值,根据 所述预测值和所述标签集包含的标签值,利用预构建的损失函数的输入参数计算得到损失值;
    损失值对比步骤:对比所述损失值与预设的损失阈值的大小,若所述损失值大于或等于所述损失阈值,返回所述降维步骤;若所述损失值小于所述损失阈值,停止训练,得到所述疾病识别模型。
  18. 如权利要求15所述的计算机可读存储介质,其中,所述根据所述疾病识别模型与预构建的初始诊断模型构建蒸馏损失函数,包括:
    对所述疾病识别模型中包含的第一特征层的特征信息进行归一化指数处理,得到疾病识别模型概率分布函数;
    对所述初始诊断模型中包含的第二特征层的特征信息进行归一化指数处理,得到初始诊断模型概率分布函数;
    根据所述疾病识别模型概率分布函数与所述初始诊断模型概率分布函数进行最大均值差异计算,得到所述蒸馏损失函数。
  19. 如权利要求15所述的计算机可读存储介质,其中,所述根据预设的诊断目标对所述第一诊断模型进行训练及输出调整,得到目标诊断模型,包括:
    判断所述诊断目标是否为新增诊断目标;
    当所述诊断目标不为新增诊断目标时,对所述第一诊断模型进行输出调整,得到所述目标诊断模型;或者
    当所述诊断目标为新增诊断目标时,调整并训练所述第一诊断模型,得到所述目标诊断模型。
  20. 如权利要求19所述的计算机可读存储介质,其中,所述调整并训练所述第一诊断模型,得到目标诊断模型,包括:
    获取诊断图片集,对所述诊断图片集进行标注,得到初始诊断图片集;
    对所述初始诊断图片集进行数据增强处理,得到标准诊断图片集;
    根据所述诊断目标对所述第一诊断模型进行输出添加,得到第二诊断模型;
    利用所述标准诊断图片集对所述第二诊断模型进行迭代训练,直至所述第二诊断模型收敛,得到第三诊断模型;
    对所述第三诊断模型进行输出调整,得到所述目标诊断模型。
PCT/CN2020/125474 2020-09-22 2020-10-30 医学影像图片分析方法、装置、电子设备及可读存储介质 WO2021151338A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011003181.5 2020-09-22
CN202011003181.5A CN111932534B (zh) 2020-09-22 2020-09-22 医学影像图片分析方法、装置、电子设备及可读存储介质

Publications (1)

Publication Number Publication Date
WO2021151338A1 true WO2021151338A1 (zh) 2021-08-05

Family

ID=73333978

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/125474 WO2021151338A1 (zh) 2020-09-22 2020-10-30 医学影像图片分析方法、装置、电子设备及可读存储介质

Country Status (2)

Country Link
CN (1) CN111932534B (zh)
WO (1) WO2021151338A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115100189A (zh) * 2022-07-28 2022-09-23 上海品课信息科技有限公司 三维医疗影像的切割方法、电子设备及存储介质
CN116740220A (zh) * 2023-08-16 2023-09-12 海马云(天津)信息技术有限公司 模型构建的方法和装置、照片生成方法和装置

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112464760A (zh) * 2020-11-16 2021-03-09 北京明略软件系统有限公司 一种目标识别模型的训练方法和装置
CN112545452B (zh) * 2020-12-07 2021-11-30 南京医科大学眼科医院 高度近视眼底病变图像识别装置
CN112819043B (zh) * 2021-01-19 2023-10-10 东北大学 一种基于arm的深度学习轻量化肺结节分类方法
CN114913371A (zh) * 2022-05-10 2022-08-16 平安科技(深圳)有限公司 多任务学习模型训练方法、装置、电子设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130129159A1 (en) * 2011-11-22 2013-05-23 Ronald Huijgens Face recognition method and apparatus
CN107368670A (zh) * 2017-06-07 2017-11-21 万香波 基于大数据深度学习的胃癌病理诊断支持系统和方法
CN107977668A (zh) * 2017-07-28 2018-05-01 北京物灵智能科技有限公司 一种机器人图像识别方法及系统
CN110837846A (zh) * 2019-10-12 2020-02-25 深圳力维智联技术有限公司 一种图像识别模型的构建方法、图像识别方法及装置
CN111563152A (zh) * 2020-06-19 2020-08-21 平安科技(深圳)有限公司 智能问答语料分析方法、装置、电子设备及可读存储介质

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107133942A (zh) * 2017-04-24 2017-09-05 南京天数信息科技有限公司 一种基于深度学习的医疗图像处理方法
US10152571B1 (en) * 2017-05-25 2018-12-11 Enlitic, Inc. Chest x-ray differential diagnosis system
CN109815976A (zh) * 2018-12-14 2019-05-28 深圳壹账通智能科技有限公司 一种证件信息识别方法、装置及设备
CN109544585A (zh) * 2018-12-19 2019-03-29 中国石油大学(华东) 一种基于轻量化卷积神经网络的胆石病ct医疗图像数据增强方法
CN110619947A (zh) * 2019-09-19 2019-12-27 南京工程学院 基于轻量级深度学习的肺部ct辅助筛查系统及筛查方法
CN110689089A (zh) * 2019-10-12 2020-01-14 电子科技大学 用于深度学习多类别医疗图像分类的主动增量训练方法
CN111027508B (zh) * 2019-12-23 2022-09-06 电子科技大学 一种基于深层神经网络的遥感图像覆被变化检测方法
CN111460797B (zh) * 2020-06-09 2021-01-15 平安国际智慧城市科技股份有限公司 关键字抽取方法、装置、电子设备及可读存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130129159A1 (en) * 2011-11-22 2013-05-23 Ronald Huijgens Face recognition method and apparatus
CN107368670A (zh) * 2017-06-07 2017-11-21 万香波 基于大数据深度学习的胃癌病理诊断支持系统和方法
CN107977668A (zh) * 2017-07-28 2018-05-01 北京物灵智能科技有限公司 一种机器人图像识别方法及系统
CN110837846A (zh) * 2019-10-12 2020-02-25 深圳力维智联技术有限公司 一种图像识别模型的构建方法、图像识别方法及装置
CN111563152A (zh) * 2020-06-19 2020-08-21 平安科技(深圳)有限公司 智能问答语料分析方法、装置、电子设备及可读存储介质

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115100189A (zh) * 2022-07-28 2022-09-23 上海品课信息科技有限公司 三维医疗影像的切割方法、电子设备及存储介质
CN115100189B (zh) * 2022-07-28 2023-07-18 上海品课信息科技有限公司 三维医疗影像的切割方法、电子设备及存储介质
CN116740220A (zh) * 2023-08-16 2023-09-12 海马云(天津)信息技术有限公司 模型构建的方法和装置、照片生成方法和装置
CN116740220B (zh) * 2023-08-16 2023-10-13 海马云(天津)信息技术有限公司 模型构建的方法和装置、照片生成方法和装置

Also Published As

Publication number Publication date
CN111932534A (zh) 2020-11-13
CN111932534B (zh) 2021-07-06

Similar Documents

Publication Publication Date Title
WO2021151338A1 (zh) 医学影像图片分析方法、装置、电子设备及可读存储介质
WO2021217851A1 (zh) 异常细胞自动标注方法、装置、电子设备及存储介质
WO2021189912A1 (zh) 图像中目标物的检测方法、装置、电子设备及存储介质
WO2021189910A1 (zh) 图片识别方法、装置、电子设备及计算机可读存储介质
CN110059697B (zh) 一种基于深度学习的肺结节自动分割方法
WO2022213465A1 (zh) 基于神经网络的图像识别方法、装置、电子设备及介质
WO2021151345A1 (zh) 识别模型的参数获取方法、装置、电子设备及存储介质
WO2021189909A1 (zh) 病灶检测分析方法、装置、电子设备及计算机存储介质
WO2022116424A1 (zh) 交通流预测模型训练方法、装置、电子设备及存储介质
WO2022121156A1 (zh) 图像中目标物检测方法、装置、电子设备及可读存储介质
WO2021189913A1 (zh) 图像中目标物的分割方法、装置、电子设备及存储介质
WO2022116423A1 (zh) 物体位姿估计方法、装置、电子设备及计算机存储介质
WO2022105179A1 (zh) 生物特征图像识别方法、装置、电子设备及可读存储介质
WO2021189901A1 (zh) 图像分割方法、装置、电子设备及计算机可读存储介质
WO2022048209A1 (zh) 车牌识别方法、装置、电子设备及存储介质
CN111862096B (zh) 图像分割方法、装置、电子设备及存储介质
WO2021189855A1 (zh) 基于ct序列的图像识别方法、装置、电子设备及介质
US12009104B2 (en) Medical image processing apparatus, medical image processing method, machine learning system, and program
WO2021135474A1 (zh) 多数据来源的数据融合方法、装置、电子设备及存储介质
WO2021189856A1 (zh) 证件校验方法、装置、电子设备及介质
WO2022194062A1 (zh) 疾病标签检测方法、装置、电子设备及存储介质
WO2021184576A1 (zh) 医学图像的生成方法、装置、电子设备及介质
CN111933274A (zh) 疾病分类诊断方法、装置、电子设备及存储介质
WO2022227192A1 (zh) 图像分类方法、装置、电子设备及介质
CN113157739B (zh) 跨模态检索方法、装置、电子设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20917001

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20917001

Country of ref document: EP

Kind code of ref document: A1