WO2021097675A1 - 一种基于医学图像的智能辅助诊断方法及终端 - Google Patents

一种基于医学图像的智能辅助诊断方法及终端 Download PDF

Info

Publication number
WO2021097675A1
WO2021097675A1 PCT/CN2019/119491 CN2019119491W WO2021097675A1 WO 2021097675 A1 WO2021097675 A1 WO 2021097675A1 CN 2019119491 W CN2019119491 W CN 2019119491W WO 2021097675 A1 WO2021097675 A1 WO 2021097675A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
preset
image
classification
feature map
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2019/119491
Other languages
English (en)
French (fr)
Inventor
王书强
余雯
申妍燕
陈卓
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Institute of Advanced Technology of CAS
Original Assignee
Shenzhen Institute of Advanced Technology of CAS
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 Shenzhen Institute of Advanced Technology of CAS filed Critical Shenzhen Institute of Advanced Technology of CAS
Priority to EP19953002.3A priority Critical patent/EP4064124B1/en
Priority to PCT/CN2019/119491 priority patent/WO2021097675A1/zh
Priority to US17/763,513 priority patent/US12254684B2/en
Publication of WO2021097675A1 publication Critical patent/WO2021097675A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0475—Generative networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0495—Quantised networks; Sparse networks; Compressed networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images

Definitions

  • This application belongs to the field of computer technology, and in particular relates to an intelligent auxiliary diagnosis method and terminal based on medical images.
  • One of the purposes of the embodiments of this application is to provide an intelligent auxiliary diagnosis method and terminal based on medical images, so as to solve the problem of losing a large amount of internal structure information and internal correlations of images when the traditional deep network model processes medical images. Information, leading to inaccurate classification results.
  • an intelligent auxiliary diagnosis method based on medical images including:
  • the preprocessed image is input into a trained classification model for classification processing to obtain the classification category corresponding to the preprocessed image; wherein, the classification model includes a network layer after quantization and a second-level pooling module;
  • the classification model is based on a preset generator model, a preset discriminator model, and a preset classifier model.
  • the sample image and the classification category corresponding to the sample image are trained on a ternary generative confrontation network.
  • the preprocessed image is input into the trained classification model for classification processing, and the result is
  • the classification category corresponding to the preprocessed image includes: normalizing the preprocessed image by using the classifier model to obtain a target image;
  • the classifier model is used to obtain the classification category corresponding to the global high-order feature map.
  • using the classifier model to extract key features in the target image, and obtaining a global high-order feature map includes: The quantized network layer in the model extracts features in the target image to obtain a first feature map;
  • the first feature map is weighted based on the weight vector to obtain the global high-order feature map.
  • this application further includes:
  • the sample image and the classification category corresponding to the sample image are trained to obtain a ternary generative confrontation network
  • the sample image and the classification category corresponding to the sample image are trained to obtain
  • the ternary generative confrontation network includes: generating synthetic image annotation pairs based on preset classification annotations, one-dimensional Gaussian random vectors and the preset generator model;
  • the sample image annotation pair, the preset real image annotation pair, and the synthetic image annotation are input to the preset discriminator model for discrimination processing, and the first discrimination result corresponding to the sample image annotation pair, the Preset the second discrimination result corresponding to the real image annotation pair and the third discrimination result corresponding to the synthetic image annotation pair;
  • the first loss function corresponding to the preset generator model and the second loss function corresponding to the preset discriminator model are calculated And a third loss function corresponding to the preset classifier model;
  • the preset generator model Based on the first loss function, the second loss function, and the third loss function, the preset generator model, the preset discriminator model, and the preset The network parameters corresponding to each classifier model;
  • a synthetic image with classification annotations can be generated through a preset generator model, based on the preset classification annotations, one-dimensional Gaussian random vectors, and the preset generator model.
  • Setting a generator model to generate a synthetic image annotation pair includes: cascading the preset classification annotations to the quantized network layer, and generating a target feature map based on the one-dimensional Gaussian random vector;
  • the composite image with the classification label is generated.
  • the sample image annotation pair, the preset real image annotation pair, and the synthetic image annotation pair are input to the preset discriminator
  • the model performs discrimination processing to obtain the first discrimination result corresponding to the sample image annotation pair, the second discrimination result corresponding to the preset real image annotation pair, and the third discrimination result corresponding to the synthetic image annotation pair.
  • the sample feature map, the real feature map, and the synthetic feature map are respectively subjected to discrimination processing to obtain the first discrimination result, the second discrimination result, and the third discrimination result. Determine the result.
  • an intelligent auxiliary diagnosis terminal based on medical images includes:
  • a preprocessing unit configured to preprocess the medical image to be classified to obtain a preprocessed image
  • the classification unit is configured to input the preprocessed image into a trained classification model for classification processing to obtain the classification category corresponding to the preprocessed image; wherein the classification model includes a network layer after quantization and a second-order pool
  • the classification model is based on a preset generator model, a preset discriminator model, and a preset classifier model, and a ternary generative confrontation network obtained by training the sample image and the classification category corresponding to the sample image.
  • the trained classification model includes a trained classifier model.
  • classification unit includes:
  • a processing unit configured to use the classifier model to normalize the preprocessed image to obtain a target image
  • An extraction unit configured to use the classifier model to extract key features in the target image to obtain a global high-order feature map
  • the classification category obtaining unit is configured to use the classifier model to obtain the classification category corresponding to the global high-order feature map.
  • extraction unit is specifically configured to:
  • the first feature map is weighted based on the weight vector to obtain the global high-order feature map.
  • the terminal also includes:
  • the training unit is used to train the sample image and the classification category corresponding to the sample image based on the preset generator model, the preset discriminator model, and the preset classifier model to obtain a ternary generative confrontation network;
  • the model acquisition unit is configured to acquire the trained classifier model from the ternary generative confrontation network.
  • the training unit includes:
  • a generating unit configured to generate a synthetic image annotation pair based on a preset classification label, a one-dimensional Gaussian random vector, and the preset generator model;
  • a determining unit configured to predict a sample image annotation pair corresponding to the sample image based on the sample image and the preset classifier model
  • the discrimination unit is configured to perform discrimination processing on the sample image tagging pair, the preset real image tagging pair, and the synthetic image tagging pair input to the preset discriminator model to obtain the first corresponding to the sample image tagging pair A judgment result, a second judgment result corresponding to the preset real image annotation pair, and a third judgment result corresponding to the synthetic image annotation pair;
  • the calculation unit is configured to calculate the first loss function corresponding to the preset generator model and the corresponding to the preset discriminator model based on the first discrimination result, the second discrimination result, and the third discrimination result The second loss function of and the third loss function corresponding to the preset classifier model;
  • An update unit configured to update the preset generator model and the preset discriminator model by using a backpropagation algorithm gradient descent based on the first loss function, the second loss function, and the third loss function And network parameters corresponding to each of the preset classifier models;
  • the network generation unit is configured to stop training when the first loss function, the second loss function, and the third loss function all converge to obtain the ternary generation confrontation network.
  • the preset generator model includes the quantized network layer.
  • the generating unit is specifically configured to:
  • the preset discriminator model includes a dense convolutional neural network after quantization.
  • discrimination unit is specifically used for:
  • the sample feature map, the real feature map, and the synthetic feature map are respectively subjected to discrimination processing to obtain the first discrimination result, the second discrimination result, and the third discrimination result. Determine the result.
  • another terminal including a processor, an input device, an output device, and a memory.
  • the processor, input device, output device, and memory are connected to each other, wherein the memory is used to store and support the terminal to execute the above
  • a computer program of the method the computer program including program instructions, and the processor is configured to call the program instructions to perform the following steps:
  • the preprocessed image is input into a trained classification model for classification processing to obtain the classification category corresponding to the preprocessed image; wherein, the classification model includes a network layer after quantization and a second-level pooling module;
  • the classification model is based on a preset generator model, a preset discriminator model, and a preset classifier model.
  • the sample image and the classification category corresponding to the sample image are trained on a ternary generative confrontation network.
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:
  • the preprocessed image is input into a trained classification model for classification processing to obtain the classification category corresponding to the preprocessed image; wherein, the classification model includes a network layer after quantization and a second-level pooling module;
  • the classification model is based on a preset generator model, a preset discriminator model, and a preset classifier model.
  • the sample image and the classification category corresponding to the sample image are trained on a ternary generative confrontation network.
  • the medical image to be classified is acquired through the terminal; the medical image to be classified is preprocessed to obtain the preprocessed image; the preprocessed image is classified based on the trained classification model to obtain the corresponding classification result.
  • the trained classification model contains the network layer after tensor decomposition and the second-order pooling module, when medical image processing is based on the classification model, the internal structure information and internal correlation of the medical image are retained, and the second-order pooling
  • the transformation module makes the important feature channels have a large weight and the unimportant channels have a small weight under the action of the self-attention mechanism, so as to extract the information related to the disease.
  • the image is classified based on the ternary generative confrontation network, and the classifier network model is added to the generator and discriminator of the traditional binary generative confrontation network, and the loss function of the compatibility is designed to alleviate the traditional two-dimensional confrontation network.
  • the problem of instability of meta-generative confrontation network training is solved, and the problem of inconsistent target convergence points for the dual-generative confrontation network discriminator to complete classification and distinguish true and false at the same time, so that the generator model and the classifier model can be replaced by collaborative training
  • the traditional binary generative confrontation network is a training method of generative confrontation, so that the generator model and the classifier model can reach the optimal at the same time, speed up the convergence, and make it easier for the generative confrontation network to reach the Nash equilibrium.
  • each network layer in the ternary generation confrontation network is compressed by the Zhang quantization method instead of the traditional vectorization method, which reduces the parameters and has a regularization effect on the network model, which solves the problem of high
  • the problem of excessive parameter and over-fitting in the classification and recognition of resolution images and the internal spatial structure information and the internal correlation between different voxels can be maintained through the Zhang quantization method, which solves the internal structure of the vectorized network layer Information loss problem
  • second-order pooling is used to replace traditional first-order pooling (maximum pooling or average pooling), making full use of the second-order information of the overall image to automatically extract more discriminative under the action of the self-attention mechanism
  • the characteristics of nature improve the accuracy of the classification of the classifier model
  • the ternary generative confrontation network in this application adopts a semi-supervised learning method through a preset generator model, a preset discriminator model, and a preset classifier model to cooperate with the training, weakening
  • the network model ’s requirements for
  • FIG. 1 is an implementation flowchart of a method for intelligently assisted diagnosis based on medical images according to an embodiment of the present application
  • Figure 2 is a schematic diagram of the structure of the trained classifier model provided by the present application.
  • FIG. 3 is a schematic diagram of the structure of the second-level pooling module provided by the present application.
  • FIG. 4 is an implementation flowchart of a method for intelligently assisted diagnosis based on medical images according to another embodiment of the present application
  • FIG. 5 is a schematic diagram of the structure of the ternary generative confrontation network provided by this application.
  • FIG. 6 is a schematic diagram of an intelligent auxiliary diagnosis terminal based on medical images provided by an embodiment of the present application.
  • FIG. 7 is a schematic diagram of a terminal provided by another embodiment of the present application.
  • the embodiment of the present application provides an intelligent auxiliary diagnosis method based on medical images, which can be applied to classify medical images.
  • the invention can efficiently process high-dimensional input data while ensuring optimal classification performance, and has strong practicability and popularization. It is suitable for all disease classification tasks that can be diagnosed with medical images.
  • this patent uses Alzheimer's disease as an example.
  • This method can be used to classify magnetic resonance imaging (MRI) of the brain. According to the classification results, such as normal elderly, mild cognitive impairment, and Alzheimer's disease, the diagnosis of Alzheimer's disease (AD ) Perform intelligent auxiliary diagnosis.
  • MRI magnetic resonance imaging
  • the sample image and the classification category corresponding to the sample image are trained to obtain the ternary generative confrontation network
  • the trained ternary generative confrontation network Contains the trained generator model, the trained discriminator model, and the trained classifier model.
  • the MRI image is classified based on the trained classifier model to obtain the corresponding classification result.
  • the trained ternary generative confrontation network contains the network layer after tensor decomposition and the second-order pooling module.
  • the classification of images based on the ternary generative confrontation network compared with the traditional binary generative confrontation network, alleviates the problem of instability in the training of the traditional generative confrontation network, and solves the problem of the discriminator in the traditional binary generative confrontation network at the same time.
  • the quantization method is used to replace the traditional vectorization method to compress each network layer in the ternary generation confrontation network, which reduces At the same time, it has a regularization effect on the network model, which solves the problem of excessive parameter and over-fitting in the classification and recognition of high-resolution images; and the internal spatial structure information and the difference between different voxels can be maintained through the Zhang quantization method.
  • the internal correlation between the two solves the problem of the internal structure loss of the vectorized network layer.
  • second-order pooling is used to replace traditional first-order pooling (maximum pooling or average pooling), making full use of the second-order information of the overall image to automatically extract more discriminative features under the action of the self-attention mechanism.
  • the ternary generative confrontation network in this application uses a semi-supervised learning method to cooperate with the training of the preset generator model, the preset discriminator model, and the preset classifier model, which reduces the network model’s need for image annotation information and makes full use of Annotate data to achieve a high-precision and high-robust intelligent network model; reduce the number of parameters, improve calculation efficiency, help reduce the performance requirements of the terminal, accelerate the speed of medical image classification, and improve disease diagnosis effectiveness.
  • FIG. 1 is a schematic flowchart of a method for intelligently assisted diagnosis based on medical images according to an embodiment of the present application.
  • the execution body of the intelligent assisted diagnosis method is a terminal, which includes but is not limited to mobile terminals such as smart phones, tablet computers, and personal digital assistants (Personal Digital Assistant, PDA), and may also include terminals such as desktop computers.
  • the intelligent auxiliary diagnosis method shown in Fig. 1 may include:
  • S101 Acquire medical images to be classified.
  • the terminal When the terminal detects the medical image classification instruction, it acquires the medical image to be classified.
  • the medical image classification instruction is an instruction used to instruct the terminal to classify medical images.
  • the image classification instruction can be triggered by the user, for example, the doctor clicks on the image classification option in the terminal.
  • Obtaining the medical image to be classified can be the medical image to be classified uploaded by the user to the terminal, or the terminal can obtain the text file corresponding to the file ID according to the file identification contained in the image classification instruction, and extract the medical image to be classified in the text file .
  • S102 Perform preprocessing on the medical image to be classified to obtain a preprocessed image.
  • the terminal preprocesses the medical image to be classified to obtain the preprocessed image. Specifically, the terminal processes the medical image to be classified into a single-color channel image, and stitches the single-color channel image to obtain a preprocessed image.
  • a single color channel image is a color channel image composed of information of one color element. The channel that stores the color information of the image is called the color channel, and each color channel stores the information of the color elements in the image. For example, in the RGB color mode (RGB color mode, RGB), R represents a red channel, G represents a green channel, and B represents a blue channel.
  • the terminal can convert the channel mode of the medical image to be classified into multiple single-color channel images by calling a preset function; stitch multiple single-color channel images through the called preset function to obtain a preprocessed image.
  • S103 Input the preprocessed image into a trained classification model for classification processing to obtain a classification category corresponding to the preprocessed image; wherein the classification model includes a network layer after quantization and a second-level pooling module;
  • the classification model is a ternary generative confrontation network obtained by training the sample image and the classification category corresponding to the sample image based on a preset generator model, a preset discriminator model, and a preset classifier model.
  • the terminal inputs the preprocessed image into the trained classification model for classification processing, and obtains the classification category corresponding to the preprocessed image.
  • the trained classification model is a ternary generative confrontation network obtained by training the sample image and the classification category corresponding to the sample image based on the preset generator model, the preset discriminator model, and the preset classifier model; the trained The classification model of includes the network layer after tensor decomposition and the second-order pooling module.
  • the training data may include sample images, classification categories corresponding to the sample images (ie, preset real image annotations), and unlabeled sample images; specifically, the terminal is based on preset classification annotations, one-dimensional Gaussian random vectors, and a preset generator Model, generate a composite image with classification annotations, and finally generate a composite image annotation pair; based on the sample image and the preset classifier model, predict the sample image annotation pair corresponding to the sample image; label the sample image pair and preset the real image annotation Perform discrimination processing on the composite image annotation and input the preset discriminator model to obtain the first discrimination result corresponding to the sample image annotation pair, the second discrimination result corresponding to the preset real image annotation pair, and the synthetic image annotation pair corresponding The third discrimination result; based on the first discrimination result, the second discrimination
  • the preprocessed image is input to the trained classifier model, and the trained classifier model normalizes the preprocessed image to obtain the target image; the trained classifier model is used to extract the key features in the target image to obtain the global height Order feature map; the trained classifier model obtains the classification category corresponding to the global high-order feature map and outputs the classification category, that is, the classification category corresponding to the preprocessed image is obtained.
  • S103 may include S1031-S1033, as follows:
  • S1031 Use the classifier model to perform normalization processing on the preprocessed image to obtain a target image.
  • the trained classification model includes a trained generator model, a trained discriminator model, and a trained classifier model.
  • the preprocessed image can be classified through the trained classifier model. Specifically, the preprocessed image is input into the trained classifier model, and the preprocessed image is normalized to obtain the target image. For example, the data corresponding to the preprocessed image is acquired, and the data is linearly changed, so that the voxel value corresponding to the preprocessed image is between [-1, 1].
  • S1032 Use the classifier model to extract key features in the target image to obtain a global high-order feature map.
  • the terminal uses the trained classifier model to extract the key features in the target image to obtain a global high-level feature map; the trained classifier model includes a network layer after tensor decomposition and a second-order pooling module.
  • Figure 2 is a schematic diagram of the structure of the trained classifier model provided by this application. As shown in Figure 2, the trained classifier model includes a 3D convolutional layer, a 3D average pooling layer, and a 3D dense connection Block 1, second-order pooling module 1, transition layer 1, 3D densely connected block 2, second-order pooling module 2, transition layer 2, 3D densely connected block 3, second-order pooling module 3, fully connected layer.
  • each network layer in the preset generator model, the preset discriminator model, and the preset classifier model is quantified.
  • the network layers such as the 3D convolutional layer, the 3D average pooling layer, and the fully connected layer in the preset classifier model are quantified.
  • the terminal extracts the key features in the target image based on the network layer after tensor decomposition and the second-order pooling module in the trained classifier model to obtain a global high-order feature map.
  • each network layer in the classifier model is quantized, and each network layer in the classifier model is compressed by the quantization method instead of the traditional vectorization method, and the network model is reduced while reducing the parameters.
  • the regularization effect solves the problem of excessive parameter and over-fitting in the classification and recognition of high-resolution images; in this embodiment, the traditional first-order pooling is replaced by the second-order pooling to make full use of the input pre-processing.
  • the classifier model uses a 3D convolution layer instead of 2D convolution
  • the layer allows the input image to be input in the form of a tensor without any dimensionality reduction, which preserves the spatial information of the image and reduces the loss of spatial information caused by the 2D network layer.
  • S1032 may include S10321-S10324, as follows:
  • the terminal extracts the features in the target image through the quantized network layer in the trained classifier model to obtain the first feature map.
  • the features in the target image are extracted through the 3D convolutional layer and the 3D average pooling layer in the trained classifier model to obtain the first feature map.
  • S10322 Perform channel dimensionality reduction on the first feature map by using the second-order pooling module in the classifier model to obtain a dimensionality-reduced second feature map.
  • FIG. 3 is a schematic structural diagram of the second-level pooling module provided by the present application.
  • the second-level pooling module includes a pre-shrinking module and a calibration module. Specifically, the input 4-dimensional feature map (that is, the first feature map) is reduced in channel dimension through 1 ⁇ 1 ⁇ 1 convolution to obtain the second feature map after dimensionality reduction.
  • the terminal calculates the weight vector corresponding to the second feature map through the trained classifier model.
  • Figure 3 specifically, calculate the covariance information of the two channels between different channels in the second feature map after dimensionality reduction to obtain the covariance matrix; according to the covariance matrix, group convolution and 1 ⁇ 1 ⁇ 1 volume
  • the product yields the same weight vector as the number of channels in the 4-dimensional feature map.
  • the terminal weights the first feature map based on the calculated weight vector, so that the important channel in the first feature map has a large weight, and the unimportant channel has a small weight, and a more representative global high-order feature map is obtained.
  • the trained classifier model uses the backpropagation algorithm to make the important channels in the first feature map have a large weight, and the unimportant channels have a small weight, so as to extract more representative feature information to obtain a global high-order feature map .
  • S1033 Use the classifier model to obtain the classification category corresponding to the global high-order feature map.
  • the trained classifier model obtains the classification category corresponding to the global high-order feature map and outputs the classification category to obtain the classification category corresponding to the preprocessed image. Further, the classification category can be used to assist disease diagnosis.
  • the medical image to be classified is obtained through the terminal; the medical image to be classified is preprocessed to obtain the preprocessed image; the preprocessed image is classified based on the trained classification model , Get the corresponding classification result.
  • the trained classification model contains the network layer after tensor decomposition and the second-order pooling module
  • the transformation module makes the important feature channels have a large weight and the unimportant channels have a small weight under the action of the self-attention mechanism, so as to extract the information related to the disease. Relevant and more discriminative features improve the accuracy of intelligently assisted diagnosis of diseases.
  • the image is classified based on the ternary generative confrontation network, and the classifier network model is added to the generator and discriminator of the traditional binary generative confrontation network, and the loss function of the compatibility is designed to alleviate the traditional two-dimensional confrontation network.
  • the problem of instability of meta-generative confrontation network training is solved, and the problem of inconsistent target convergence points for the dual-generative confrontation network discriminator to complete classification and distinguish true and false at the same time, so that the generator model and the classifier model can be replaced by collaborative training
  • the traditional binary generative confrontation network is a training method of generative confrontation, so that the generator model and the classifier model can reach the optimal at the same time, speed up the convergence, and make it easier for the generative confrontation network to reach the Nash equilibrium.
  • each network layer in the ternary generation confrontation network is compressed by the Zhang quantization method instead of the traditional vectorization method, which reduces the parameters and has a regularization effect on the network model, which solves the problem of high
  • the problem of excessive parameter and over-fitting in the classification and recognition of resolution images and the internal spatial structure information and the internal correlation between different voxels can be maintained through the Zhang quantization method, which solves the internal structure of the vectorized network layer Loss problem:
  • second-order pooling is used to replace traditional first-order pooling (maximum pooling or average pooling), making full use of the second-order information of the overall image to automatically extract more discriminative under the action of the self-attention mechanism
  • the characteristics of the classifier can improve the accuracy of the classification of the classifier model
  • the ternary generative confrontation network in this application adopts a semi-supervised learning method to cooperate with the training through the preset generator model, the preset discriminator model and the preset classifier model, which reduces The network model's demand for image
  • FIG. 4 is a schematic flowchart of a method for intelligently assisted diagnosis based on medical images according to another embodiment of the present application.
  • the execution subject of the intelligent auxiliary diagnosis method in this embodiment is a terminal, which includes but is not limited to mobile terminals such as smart phones, tablet computers, and personal digital assistants, and may also include terminals such as desktop computers.
  • This embodiment adds training steps S201-S202 of the classifier model on the basis of the previous embodiment.
  • S203-S205 in this embodiment are completely the same as S101-S103 in the previous embodiment.
  • S101-S103 in the previous embodiment please refer to the relevant description of S101-S103 in the previous embodiment, which will not be repeated here.
  • the intelligent auxiliary diagnosis method shown in Figure 4, in order to improve the accuracy of image classification, S201-S202, are as follows:
  • S201 Based on the preset generator model, the preset discriminator model, and the preset classifier model, train the sample image and the classification category corresponding to the sample image to obtain a ternary generative confrontation network.
  • Figure 5 is a schematic diagram of the structure of the ternary generative confrontation network provided in this application. Now, taking the application scenario of Alzheimer's disease as an example, combined with the structure of the ternary generative confrontation network in Figure 5, the training results are explained. The process of ternary generation against the network.
  • the generator in Figure 5 refers to the preset generator model, and when the training is completed, the corresponding trained generator model is generated;
  • the discriminator in Figure 5 refers to the preset discriminator model, When the training is completed, the corresponding trained discriminator model is generated; in the training process, the Alzheimer's disease classifier in Figure 5 refers to the preset classifier model, and the corresponding trained classifier model is generated after the training is completed .
  • the preset generator model mainly includes 3D deconvolution layer; the preset discriminator model mainly includes 3D convolution layer, 3D densely connected block, transition layer, fully connected layer, etc.; the preset classifier model mainly includes 3D volume Build-up layers, 3D densely connected blocks, second-order pooling modules, etc. It is worth noting that, here is only Alzheimer's disease as an example, the Alzheimer's disease intelligent auxiliary diagnosis model that can be used to classify MRI images is trained; this method can be used to train classification models for other medical images , There is no restriction on this.
  • each network layer in the preset generator model, the preset discriminator model, and the preset classifier model is quantified.
  • the 3D convolutional layer and the fully connected layer in the preset discriminator model and the preset classifier model, as well as the 3D deconvolution layer in the preset generator model are parameterized through the tensor decomposition method; the fully connected layer
  • the weight matrix, the convolution kernel tensor of the deconvolution layer, and the convolution kernel tensor of the convolution layer can all be expressed in the corresponding tensor form:
  • the weight tensor W of the fully connected layer is decomposed according to the above formula, and the tensor of the fully connected layer is expressed as follows:
  • the tensor decomposition steps of the 3D convolutional layer and deconvolutional layer are as follows:
  • a synthetic image with classification annotation is generated, and finally a synthetic image annotation pair is generated; based on the sample image in the training data and the preset classification Detector model to determine the sample image annotation pair corresponding to the sample image; the sample image annotation pair, the preset real image annotation pair and the generated synthetic image annotation are input to the preset discriminator model for discrimination processing, and the sample image annotation pair corresponds to The first discrimination result, the second discrimination result corresponding to the preset real image annotation pair, and the third discrimination result corresponding to the synthetic image annotation pair; based on the first discrimination result, the second discrimination result, and the third discrimination result, the prediction is calculated Suppose the first loss function corresponding to the generator model, the second loss function corresponding to the preset discriminator model, and the third loss function corresponding to the preset classifier model; based on the first loss function, the second loss function, and the third loss function Respectively update the network parameters
  • the trained classifier model from the ternary generative confrontation network.
  • the ternary generative confrontation network obtained by training is the trained classification model.
  • the trained classification model includes the trained generator model, the trained discriminator model, and the trained classifier model. Obtain the trained classifier model from the trained classification model.
  • S201 may include S2011-S2016, specifically as follows:
  • S2011 Generate a synthetic image annotation pair based on a preset classification label, a one-dimensional Gaussian random vector and the preset generator model.
  • a one-dimensional Gaussian random vector and a preset classification label are used as input, and input into the preset generator model; the input preset classification label is cascaded to each sheet after quantization by one-hot encoding (One-Hot Encoding)
  • the network layer is based on the one-dimensional Gaussian random vector to generate the target feature map; the target feature map is enlarged layer by layer based on the quantized network layer, and the target composite image is finally generated; the target composite image is generated based on the target composite image and preset classification and labeling, and the composite is finally generated Image annotation pair.
  • the semi-supervised learning method is introduced into the disease classification task, which can efficiently and comprehensively use the unlabeled medical image information, and at the same time, generate the classified and labeled information through the preset generator model.
  • Synthetic images play the role of data enhancement, and can train high-precision auxiliary diagnostic models even in the case of small samples.
  • Reduce the demand for labeling training samples reduce the workload of the traditional algorithm for the complicated labeling of training data, further shorten the work cycle of disease diagnosis, accelerate the speed of disease diagnosis, and improve the overall disease recognition effectiveness.
  • S2011 includes S20111-S20113, which are as follows:
  • S20111 Concatenate the preset classification label to the quantized network layer, and generate a target feature map based on the one-dimensional Gaussian random vector.
  • the preset generator model contains the quantized network layer.
  • the preset generator model includes a quantized deconvolution layer.
  • the input preset classification labels are cascaded to each quantized deconvolution layer after one-hot encoding; the target feature map is generated based on the quantized deconvolution layer and one-dimensional Gaussian random vector.
  • the activation function of the deconvolution layer adopts a linear rectification function (Rectified Linear Unit, ReLU) and batch normalization (BN).
  • ReLU linear rectification function
  • BN batch normalization
  • the target feature map is the brain anatomical feature map.
  • S20112 Enlarge the target feature map layer by layer based on the quantized network layer to generate a target composite image.
  • the quantized network layer After multi-layer deconvolution, the quantized network layer enlarges the target feature map layer by layer, and the resulting image is the target composite image.
  • the anatomical feature map of the brain is enlarged layer by layer to generate a synthetic image with the same size as the real MRI image.
  • the last layer of the preset generator model adopts the hyperbolic function tanh activation function.
  • S20113 Generate the synthetic image annotation pair based on the target synthetic image and the preset classification annotation.
  • the terminal generates a synthetic image annotation pair based on the target synthetic image and preset classification annotations. For example, the terminal generates an MRI image with classification annotations based on the synthesized image and preset classification annotations.
  • An MRI image with classification annotations can also be called an MRI image annotation pair.
  • S2012 Based on the sample image and the preset classifier model, predict a sample image annotation pair corresponding to the sample image.
  • the terminal predicts the corresponding category of the sample image based on the sample image in the training data and the preset classifier model, and determines the sample image label corresponding to the sample image based on the sample image and the corresponding category.
  • the sample image is input to a preset classifier model, and the preset classifier model predicts the annotation information corresponding to the sample image, and generates a sample image annotation pair corresponding to the sample image based on the sample image and the annotation information.
  • the sample image is the real unlabeled MRI image
  • the real unlabeled MRI image is input into the preset classifier model
  • the preset classifier model predicts the label corresponding to the real unlabeled MRI image Information, based on the real unlabeled MRI image and the predicted annotation information to generate an MRI image annotation pair.
  • S2013 Perform discrimination processing on the sample image annotation pair, the preset real image annotation pair, and the synthetic image annotation input to the preset discriminator model to obtain the first discrimination result corresponding to the sample image annotation pair, The second discrimination result corresponding to the preset real image annotation pair and the third discrimination result corresponding to the synthetic image annotation pair.
  • the sample image is annotated to input the preset discriminator model for discriminating processing
  • the preset discriminator model extracts the feature information of the sample image annotation pair, obtains the sample feature map corresponding to the sample image annotation pair, and performs discrimination processing on the sample feature map based on the preset discriminator model to obtain the first discrimination result.
  • the preset real image annotations in the training data are used to discriminate the input preset discriminator model.
  • the preset discriminator model extracts the feature information of the preset real image annotation pair to obtain the real feature map corresponding to the preset real image annotation pair , Based on the preset discriminator model, the real feature map is discriminated, and the second discriminating result is obtained.
  • the synthetic image annotation is input to the preset discriminator model for discriminating processing, the preset discriminator model extracts the feature information of the synthetic image annotation pair, and the synthetic feature map corresponding to the synthetic image annotation pair is obtained, and the synthetic feature is based on the preset discriminator model
  • the figure performs the discrimination processing, and the third discrimination result is obtained.
  • S2013 may include S20131-S20134, which are specifically as follows:
  • S20131 Extract feature information of the sample image annotation pair based on the quantized dense convolutional neural network, and obtain a sample feature map corresponding to the sample image annotation pair.
  • the preset discriminator model includes a dense convolutional neural network after Zhang quantization.
  • the sample image corresponding to the sample image annotation pair is input into the preset discriminator model in the form of a third-order tensor.
  • the classification annotations in the sample image annotation pair are cascaded to the preset discriminator model as a condition variable through one-hot encoding.
  • the quantized dense convolutional neural network extracts the feature information of the sample image annotation pair, and obtains a feature map with reserved spatial information, that is, obtains the sample feature map corresponding to the sample image annotation pair.
  • the activation function of the convolutional layer adopts ReLU and Batch Normalization.
  • S20132 Extract the feature information of the preset real image annotation pair based on the quantized dense convolutional neural network, and obtain the real feature map corresponding to the preset real sample image annotation pair.
  • the image corresponding to the real image annotation pair is input into the preset discriminator model in the form of a third-order tensor, and the classification annotations in the real image annotation pair are cascaded to each of the preset discriminator models through one-hot encoding as a condition variable.
  • the quantized dense convolutional neural network extracts the feature information of the real image annotation pair, and obtains the feature map with the reserved spatial information, that is, the real feature map corresponding to the real image annotation pair is obtained.
  • the activation function of the convolutional layer adopts ReLU and Batch Normalization.
  • S20133 Extract feature information in the synthetic image annotation pair based on the quantized dense convolutional neural network, and obtain a synthetic feature map corresponding to the synthetic image annotation pair.
  • the synthesized image is input into the preset discriminator model in the form of a third-order tensor, and the classification label corresponding to the synthesized image is cascaded to each network layer in the preset discriminator model through one-hot encoding as a condition variable.
  • the quantized dense convolutional neural network extracts the feature information of the synthetic image annotation pair, and obtains the feature map with the reserved spatial information, that is, the synthetic feature map corresponding to the synthetic image annotation pair is obtained.
  • the activation function of the convolutional layer adopts ReLU and Batch Normalization.
  • S20134 Perform discrimination processing on the sample feature map, the real feature map, and the synthetic feature map based on the preset discriminator model to obtain the first discrimination result, the second discrimination result, and the The third judgment result.
  • the preset discriminator model includes a fully connected layer after Zhang quantization, and the judgment is made based on the sigmoid function of the layer to obtain the first judgment result corresponding to the sample feature map, the second judgment result corresponding to the real feature map, and the composite The third discrimination result corresponding to the feature map.
  • S2014 Based on the first discrimination result, the second discrimination result, and the third discrimination result, calculate a first loss function corresponding to the preset generator model and a second loss function corresponding to the preset discriminator model A loss function and a third loss function corresponding to the preset classifier model.
  • the core of the network layer tensor decomposition in the preset generator model is updated.
  • the goal of the preset generator model is to generate images that can deceive the preset discriminator model by simulating real images.
  • the preset generator model simulates the characteristics of the real MRI brain anatomy to generate a close-to-real MRI image that can deceive the preset discriminator model. Therefore, the loss of the preset generator model consists of two parts. One part is to deceive the preset discriminator model to make it judge that the MRI image generated by the preset generator model is marked as true; the other part is the real MRI image and the generated MRI image
  • the reconstruction loss between; can be expressed as:
  • the loss of the preset discriminator model consists of three parts, as follows:
  • the preset classifier model In the process of training the preset classifier model, according to the loss function G-loss gradient of the preset classifier model in the back propagation process, update the kernel matrix G k of the network layer tensor decomposition in the preset classifier model [i k ,j k ] parameters.
  • the goal of the preset classifier model is to automatically extract the feature information in the preprocessed image and perform classification; for example, automatically extract the MRI brain anatomical structure feature for classification, and divide the MRI image into 3 categories: normal, Alzheimer's disease And mild cognitive impairment.
  • the loss of the preset classifier model consists of two parts, one part is the supervised loss, that is, the cross entropy of the real image and the generated image for classification tasks; the other part is the unsupervised loss, that is, deceiving the preset discriminator model to make it It is judged that the labeling pair of MRI images generated by the preset classifier model on the unlabeled MRI images is true. It can be expressed as:
  • Calculating R L for annotated pairs of real MRI images is equivalent to calculating the KL divergence between the distribution P c (x,y) learned by the preset classifier model and the real data distribution P real (x,y).
  • the preset generator model generates MRI image annotation pairs close to the true distribution, which can improve the classification performance of the classifier model.
  • R p is introduced to calculate the cross entropy of MRI image annotation pairs; minimizing R p is equivalent to minimizing KL dispersion Degree D KL (P g (x,y)
  • the preset classifier model minimizes the KL divergence by indirectly minimizing R p to minimize the KL divergence D KL (P g (x ,y)
  • S2015 Based on the first loss function, the second loss function, and the third loss function, the preset generator model, the preset discriminator model, and the Preset the network parameters corresponding to each of the classifier models.
  • the terminal updates the corresponding network parameters of the preset generator model, the preset discriminator model, and the preset classifier model through the gradient descent of the backpropagation algorithm according to the calculated first loss function, second loss function, and third loss function. .
  • the weight value of each network layer in each model of the preset generator model, the preset discriminator model, and the preset classifier model is updated according to the first loss function, the second loss function, and the third loss function.
  • continue training based on the preset generator model, preset discriminator model, and preset classifier model after updating the parameters. That is, the sample image and the classification category corresponding to the sample image are continuously trained based on each model after updating the parameters.
  • the preset generator model, the preset discriminator model, and the preset classifier model are cooperatively trained, such as repeated training in the "generation-discrimination-classification" cooperative mode.
  • the terminal detects that the first loss function, the second loss function, and the third loss function are all converged during the repeated "generation-discrimination-classification" cooperative mode training process, the training stops, and the trained ternary generation confrontation network is obtained at this time , That is, the trained classification model. It can also set the number of iterations in advance. After the training of the number of iterations is performed, the training is considered to be completed, and a trained ternary generation confrontation network is obtained at this time.
  • the medical image to be classified is acquired through the terminal; the medical image to be classified is preprocessed to obtain the preprocessed image; the preprocessed image is classified based on the trained classification model to obtain the corresponding classification result.
  • the trained classification model contains the network layer after tensor decomposition and the second-order pooling module, when medical image processing is based on this classification model, the internal structure information and internal correlation of the medical image are retained, and the second-order pooling The module uses the dependency relationship between different regions of the medical image and the correlation information between different channels of high-level features. Under the action of the self-attention mechanism, the weight of important feature channels is large, and the weight of unimportant channels is small, so as to extract the correlation information related to the disease.
  • the image is classified based on the ternary generative confrontation network, and the classifier network model is added to the generator and discriminator of the traditional binary generative confrontation network, and the loss function of the compatibility is designed to alleviate the traditional two-dimensional confrontation network.
  • the problem of instability of meta-generative confrontation network training is solved, and the problem of inconsistent target convergence points for the dual-generative confrontation network discriminator to complete classification and distinguish true and false at the same time, so that the generator model and the classifier model can be replaced by collaborative training
  • the traditional binary generative confrontation network is a training method for generating confrontation, so that the generator model and the classifier model can reach the optimal at the same time, accelerate the convergence, and make the generated confrontation network easier to reach the Nash equilibrium.
  • each network layer in the ternary generation confrontation network is compressed by the Zhang quantization method instead of the traditional vectorization method, which reduces the parameters and has a regularization effect on the network model, which solves the problem of high
  • the problem of excessive parameter and over-fitting in the classification and recognition of resolution images and the internal spatial structure information of the image and the internal correlation between different voxels can be maintained through the Zhang quantization method, which solves the image of the vectorized network layer
  • second-order pooling is used to replace traditional first-order pooling (maximum pooling or average pooling), making full use of the second-order information of the overall image to automatically extract more information under the action of the self-attention mechanism.
  • the discriminative features improve the accuracy of the classification of the classifier model; the ternary generative confrontation network in this application uses a semi-supervised learning method to cooperate with the training through the preset generator model, the preset discriminator model, and the preset classifier model. It reduces the network model’s need for image labeling information, and makes full use of unlabeled data to realize an intelligent network model with high accuracy and robustness; it also reduces the number of parameters, improves computing efficiency, and helps reduce terminal costs. Performance requirements, thereby improving the efficiency of intelligent auxiliary diagnosis.
  • FIG. 6 is a schematic diagram of a medical image-based intelligent auxiliary diagnosis terminal provided by an embodiment of the present application.
  • the units included in the terminal are used to execute the steps in the embodiments corresponding to FIG. 1 and FIG. 4.
  • FIG. 6 please refer to the relevant descriptions in the respective embodiments of FIG. 1 and FIG. 4.
  • Figure 6 only the parts related to this embodiment are shown. See Figure 6, including:
  • the obtaining unit 310 is used to obtain medical images to be classified
  • the preprocessing unit 320 is configured to preprocess the medical image to be classified to obtain a preprocessed image
  • the classification unit 330 is configured to input the preprocessed image into a trained classification model for classification processing to obtain a classification category corresponding to the preprocessed image; wherein, the classification model includes a network layer after quantization and a second-order Pooling module; the classification model is based on a preset generator model, a preset discriminator model, and a preset classifier model, a ternary generative confrontation network obtained by training the sample image and the classification category corresponding to the sample image.
  • the trained classification model includes a trained classifier model.
  • classification unit 330 includes:
  • a processing unit configured to use the classifier model to normalize the preprocessed image to obtain a target image
  • An extraction unit configured to use the classifier model to extract key features in the target image to obtain a global high-order feature map
  • the classification category obtaining unit is configured to use the classifier model to obtain the classification category corresponding to the global high-order feature map.
  • extraction unit is specifically configured to:
  • the first feature map is weighted based on the weight vector to obtain the global high-order feature map.
  • the terminal also includes:
  • the training unit is used to train the sample image and the classification category corresponding to the sample image based on the preset generator model, the preset discriminator model, and the preset classifier model to obtain a ternary generative confrontation network;
  • the model acquisition unit is configured to acquire the trained classifier model from the ternary generative confrontation network.
  • the training unit includes:
  • a generating unit configured to generate a synthetic image annotation pair based on a preset classification label, a one-dimensional Gaussian random vector, and the preset generator model;
  • a determining unit configured to predict a sample image annotation pair corresponding to the sample image based on the sample image and the preset classifier model
  • the discrimination unit is configured to perform discrimination processing on the sample image tagging pair, the preset real image tagging pair, and the synthetic image tagging pair input to the preset discriminator model to obtain the first corresponding to the sample image tagging pair A judgment result, a second judgment result corresponding to the preset real image annotation pair, and a third judgment result corresponding to the synthetic image annotation pair;
  • the calculation unit is configured to calculate the first loss function corresponding to the preset generator model and the corresponding to the preset discriminator model based on the first discrimination result, the second discrimination result, and the third discrimination result The second loss function of and the third loss function corresponding to the preset classifier model;
  • An update unit configured to update the preset generator model and the preset discriminator model by using a backpropagation algorithm gradient descent based on the first loss function, the second loss function, and the third loss function And network parameters corresponding to each of the preset classifier models;
  • the network generation unit is configured to stop training when the first loss function, the second loss function, and the third loss function all converge to obtain the ternary generation confrontation network.
  • the preset generator model includes the quantized network layer.
  • the generating unit is specifically configured to:
  • the preset discriminator model includes a dense convolutional neural network after quantization.
  • discrimination unit is specifically used for:
  • the sample feature map, the real feature map, and the synthetic feature map are respectively subjected to discrimination processing to obtain the first discrimination result, the second discrimination result, and the third discrimination result. Determine the result.
  • FIG. 7 is a schematic diagram of an intelligent auxiliary diagnosis terminal based on medical images according to another embodiment of the present application.
  • the terminal 4 of this embodiment includes a processor 40, a memory 41, and computer-readable instructions 42 stored in the memory 41 and running on the processor 40.
  • the processor 40 executes the computer readable instructions 42, the steps in the above embodiments of the medical image-based intelligent assisted diagnosis method for each terminal are implemented, such as S101 to S103 shown in FIG. 1.
  • the processor 40 implements the functions of the units in the foregoing embodiments when the processor 40 executes the computer-readable instructions 42, for example, the functions of the units 310 to 330 shown in FIG. 6.
  • the computer-readable instruction 42 may be divided into one or more units, and the one or more units are stored in the memory 41 and executed by the processor 40 to complete the present application .
  • the one or more units may be a series of computer-readable instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer-readable instructions 42 in the terminal 4.
  • the computer-readable instructions 42 may be acquired by the acquisition unit, the preprocessing unit, and the classification unit, and the specific functions of each unit are as described above.
  • the terminal may include, but is not limited to, a processor 40 and a memory 41.
  • FIG. 7 is only an example of the terminal 4, and does not constitute a limitation on the terminal 4. It may include more or less components than those shown in the figure, or a combination of certain components, or different components, such as
  • the terminal may also include input and output terminals, network access terminals, buses, and so on.
  • the so-called processor 40 may be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory 41 may be an internal storage unit of the terminal 4, such as a hard disk or a memory of the terminal 4.
  • the memory 41 may also be an external storage terminal of the terminal 4, such as a plug-in hard disk equipped on the terminal 4, a smart memory card (Smart Media Card, SMC), or a Secure Digital (SD) card, Flash Card, etc. Further, the memory 41 may also include both an internal storage unit of the terminal 4 and an external storage terminal.
  • the memory 41 is used to store the computer readable instructions and other programs and data required by the terminal.
  • the memory 41 can also be used to temporarily store data that has been output or will be output.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Medical Informatics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Databases & Information Systems (AREA)
  • Public Health (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Quality & Reliability (AREA)
  • Pathology (AREA)
  • Image Analysis (AREA)

Abstract

本申请适用于计算机技术领域，提供了一种基于医学图像的智能辅助诊断方法及终端，包括：获取待分类医学图像；对待分类医学图像进行预处理，得到预处理图像；将预处理图像输入已训练的分类模型进行分类处理，得到预处理图像对应的分类类别；其中，该分类模型包含经过张量化后的网络层以及二阶池化模块。由于已训练的分类模型中包含经过张量分解后的网络层以及二阶池化模块，基于该分类模型对图像处理时，能提取与病变相关的更有判别性的特征，进而提升医学图像分类的准确率。

Description

一种基于医学图像的智能辅助诊断方法及终端 技术领域
本申请属于计算机技术领域，尤其涉及一种基于医学图像的智能辅助诊断方法及终端。
背景技术
目前，随着社会的发展，基于深度网络模型对医学图像分类进行智能辅助诊断的应用越来越广泛。然而，传统的深度网络模型对医学图像处理时，丢失了组织结构的内部信息和不同区域之间的相关性信息，这种空间结构信息的丢失最终将导致分类结果不准确，严重影响疾病智能辅助诊断模型的性能。
技术问题
本申请实施例的目的之一在于：提供了一种基于医学图像的智能辅助诊断方法及终端，以解决传统的深度网络模型对医学图像处理时，丢失了大量图像的内部结构信息和内部相关性信息，导致分类结果不准确的问题。
技术解决方案
为解决上述技术问题，本申请实施例采用的技术方案是：
第一方面，提供了一种基于医学图像的智能辅助诊断方法，包括：
获取待分类医学图像；
对所述待分类医学图像进行预处理，得到预处理图像；
将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
在一个实施例中，为了使提取到的全局高阶特征图保留更多的重要特征信息，进而提高分类模型分类的准确率，将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别包括：采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像；
采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图；
采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
在一个实施例中，为了提取到包含更多特征信息的全局高阶特征图，采用所述分类器 模型提取所述目标图像中的关键特征，得到全局高阶特征图包括：通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第一特征图；
通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维，得到降维后的第二特征图；
计算所述第二特征图对应的权重向量；
基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
在一个实施例中，为了提高图像分类的准确率，本申请还包括：
基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络；
从所述三元生成对抗网络中获取所述已训练的分类器模型。
在一个实施例中，为了提高图像分类的准确率以及分类的速度，基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络包括：基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对；
基于所述样本图像以及所述预设分类器模型，预测所述样本图像对应的样本图像标注对；
将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果；
基于所述第一判别结果、所述第二判别结果以及所述第三判别结果，计算所述预设生成器模型对应的第一损失函数、所述预设判别器模型对应的第二损失函数以及所述预设分类器模型对应的第三损失函数；
基于所述第一损失函数、所述第二损失函数以及所述第三损失函数通过反向传播算法梯度下降分别更新所述预设生成器模型、所述预设判别器模型以及所述预设分类器模型各自对应的网络参数；
当所述第一损失函数、所述第二损失函数以及所述第三损失函数均收敛时，停止训练，得到所述三元生成对抗网络。
在一个实施例中，为了减少训练网络模型过程中对医学图像的需求，可以通过预设生成器模型生成带有分类标注的合成图像，基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对包括：将所述预设分类标注级联至所述张量化后的网络层，并基于所述一维高斯随机向量生成目标特征图；
基于所述张量化后的网络层将所述目标特征图逐层放大，生成目标合成图像；
基于所述目标合成图像以及所述预设分类标注，生成所述带有分类标注的合成图像。
在一个实施例中，为了提高分类模型的分类性能，进而提高图像分类的准确率，将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果包括：
基于所述张量化后的密集型卷积神经网络提取所述样本图像标注对的特征信息，得到所述样本图像标注对所对应的样本特征图；
基于所述张量化后的密集型卷积神经网络提取所述预设真实图像标注对的特征信息，得到所述预设真实样本图像标注对所对应的真实特征图；
基于所述张量化后的密集型卷积神经网络提取所述合成图像标注对中的特征信息，得到所述合成图像标注对所对应的合成特征图；
基于所述预设判别器模型对所述样本特征图、所述真实特征图以及所述合成特征图分别进行判别处理，得到所述第一判别结果、所述第二判别结果以及所述第三判别结果。
第二方面，提供了一种基于医学图像的智能辅助诊断终端，该终端包括：
获取单元，用于获取待分类医学图像；
预处理单元，用于对所述待分类医学图像进行预处理，得到预处理图像；
分类单元，用于将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
所述已训练的分类模型包含已训练的分类器模型。
进一步地，所述分类单元包括：
处理单元，用于采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像；
提取单元，用于采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图；
分类类别获取单元，用于采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
进一步地，所述提取单元具体用于：
通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第 一特征图；
通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维，得到降维后的第二特征图；
计算所述第二特征图对应的权重向量；
基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
进一步地，终端还包括：
训练单元，用于基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络；
模型获取单元，用于从所述三元生成对抗网络中获取所述已训练的分类器模型。
进一步地，所述训练单元包括：
生成单元，用于基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对；
确定单元，用于基于所述样本图像以及所述预设分类器模型，预测所述样本图像对应的样本图像标注对；
判别单元，用于将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果；
计算单元，用于基于所述第一判别结果、所述第二判别结果以及所述第三判别结果，计算所述预设生成器模型对应的第一损失函数、所述预设判别器模型对应的第二损失函数以及所述预设分类器模型对应的第三损失函数；
更新单元，用于基于所述第一损失函数、所述第二损失函数以及所述第三损失函数通过反向传播算法梯度下降分别更新所述预设生成器模型、所述预设判别器模型以及所述预设分类器模型各自对应的网络参数；
网络生成单元，用于当所述第一损失函数、所述第二损失函数以及所述第三损失函数均收敛时，停止训练，得到所述三元生成对抗网络。
所述预设生成器模型包含所述张量化后的网络层。
进一步地，所述生成单元具体用于：
将所述预设分类标注级联至所述张量化后的网络层，并基于所述一维高斯随机向量生成目标特征图；
基于所述张量化后的网络层将所述目标特征图逐层放大，生成目标合成图像；
基于所述目标合成图像以及所述预设分类标注，生成所述合成图像标注对。
所述预设判别器模型包含张量化后的密集型卷积神经网络。
进一步地，所述判别单元具体用于：
基于所述张量化后的密集型卷积神经网络提取所述样本图像标注对的特征信息，得到所述样本图像标注对所对应的样本特征图；
基于所述张量化后的密集型卷积神经网络提取所述预设真实图像标注对的特征信息，得到所述预设真实样本图像标注对所对应的真实特征图；
基于所述张量化后的密集型卷积神经网络提取所述合成图像标注对中的特征信息，得到所述合成图像标注对所对应的合成特征图；
基于所述预设判别器模型对所述样本特征图、所述真实特征图以及所述合成特征图分别进行判别处理，得到所述第一判别结果、所述第二判别结果以及所述第三判别结果。
第三方面，提供了另一种终端，包括处理器、输入设备、输出设备和存储器，所述处理器、输入设备、输出设备和存储器相互连接，其中，所述存储器用于存储支持终端执行上述方法的计算机程序，所述计算机程序包括程序指令，所述处理器被配置用于调用所述程序指令，执行以下步骤：
获取待分类医学图像；
对所述待分类医学图像进行预处理，得到预处理图像；
将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
第四方面，提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现以下步骤：
获取待分类医学图像；
对所述待分类医学图像进行预处理，得到预处理图像；
将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
有益效果
本申请实施例，通过终端获取待分类医学图像；对待分类医学图像进行预处理，得到预处理图像；基于已训练的分类模型对预处理图像进行分类处理，得到对应的分类结果。由于该已训练的分类模型中包含经过张量分解后的网络层以及二阶池化模块，基于该分类模型对医学图像处理时，保留了医学图像的内部结构信息和内部相关性，二阶池化模块通过医学图像不同区域的依赖关系和高阶特征不同通道间的相关性信息，在自注意力机制的作用下，使重要的特征通道权重大，不重要的通道权重小，从而提取与病变相关的更有判别性的特征，提高疾病智能辅助诊断的准确率。进一步地，本申请中基于三元生成对抗网络对图像进行分类，在传统二元生成对抗网络的生成器和判别器的基础上加入了分类器网络模型，通过设计兼容性的损失函数缓解传统二元生成对抗网络训练不稳定的问题，解决了二元生成对抗网络判别器同时完成分类和判别真假两个任务目标收敛点不一致的问题，使生成器模型和分类器模型通过协作训练的方式替代传统二元生成对抗网络的生成对抗的训练方式，从而使生成器模型和分类器模型能够同时达到最优，加快收敛，使生成对抗网络更容易达到纳什均衡。本申请中对医学图像进行分类时，通过张量化方法替代传统的向量化方法压缩三元生成对抗网络中的每个网络层，在减少参数的同时对网络模型起到正则化效果，解决了高分辨率图像在分类和识别时参数量过大、过拟合的问题；且通过张量化方法能够保持内部空间结构信息和不同体素之间的内部相关性，解决了向量化网络层的内部结构信息损失问题；本申请中通过二阶池化替代传统的一阶池化(最大池化或者平均池化)，充分利用整体图像的二阶信息在自注意力机制的作用下自动提取更有判别性的特征，提高分类器模型分类的准确率；本申请中的三元生成对抗网络通过预设生成器模型、预设判别器模型和预设分类器模型采用半监督学习方式共同配合训练，减弱了网络模型对图像标注信息的需求，充分利用无标注数据，进而实现具有高精度、高鲁棒性的智能网络模型；且减少了参数数量，提高了计算效率，有助于降低对终端的性能要求，从而加快患者的疾病诊断过程。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或示范性技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1是本申请一实施例提供的一种基于医学图像的智能辅助诊断方法的实现流程图；
图2是本申请提供的已训练的分类器模型的结构示意图；
图3是本申请提供的二阶池化模块的结构示意图；
图4是本申请另一实施例提供的一种基于医学图像的智能辅助诊断方法的实现流程图；
图5是本申请提供的三元生成对抗网络的结构示意图；
图6是本申请一实施例提供的一种基于医学图像的智能辅助诊断终端的示意图；
图7是本申请另一实施例提供的一种终端的示意图。
本发明的实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本申请，并不用于限定本申请。
需说明的是，当部件被称为“固定于”或“设置于”另一个部件，它可以直接在另一个部件上或者间接在该另一个部件上。当一个部件被称为是“连接于”另一个部件，它可以是直接或者间接连接至该另一个部件上。术语“上”、“下”、“左”、“右”等指示的方位或位置关系为基于附图所示的方位或位置关系，仅是为了便于描述，而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作，因此不能理解为对本申请的限制，对于本领域的普通技术人员而言，可以根据具体情况理解上述术语的具体含义。术语“第一”、“第二”仅用于便于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明技术特征的数量。“多个”的含义是两个或两个以上，除非另有明确具体的限定。
为了说明本申请所述的技术方案，以下结合具体附图及实施例进行详细说明。
本申请实施例提供了一种基于医学图像的智能辅助诊断方法，可应用于对医学图像进行分类。本发明能够高效处理高维输入数据的同时保证最优的分类性能，具有较强的实用性和推广性。适用于所有能用医学图像进行诊断的疾病分类任务，为方便描述，本专利以阿尔茨海默症为例。可通过该方法对脑部核磁共振成像(Magnetic Resonance Imaging，MRI)进行分类，根据分类结果如正常老人、轻度认知障碍和阿尔茨海默症，对阿尔茨海默症(Alzheimer’s disease，AD)进行智能辅助诊断。具体地，基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络，训练好的三元生成对抗网络中包含已训练的生成器模型、已训练的判别器模型、已训练的分类器模型，基于已训练的分类器模型对MRI图像进行分类，得到对应的分类结果。其中，训练好的三元生成对抗网络中包含经过张量分解后的网络层以及二阶 池化模块。
本申请中基于三元生成对抗网络对图像进行分类，相比传统的二元生成对抗网络来说，缓解了传统生成对抗网络训练不稳定的问题，解决了传统二元生成对抗网络中判别器同时判别真假和分类两个目标任务的收敛点不一致的问题；本申请中对图像进行分类时，通过张量化方法替代传统的向量化方法压缩三元生成对抗网络中的每个网络层，在减少参数的同时对网络模型起到正则化效果，解决了高分辨率图像在分类和识别时参数量过大、过拟合的问题；且通过张量化方法能够保持内部空间结构信息和不同体素之间的内部相关性，解决了向量化网络层的内部结构损失问题。本申请中通过二阶池化替代传统的一阶池化(最大池化或者平均池化)，充分利用整体图像的二阶信息在自注意力机制的作用下自动提取更有判别性的特征，提高分类器分类的准确率。本申请中的三元生成对抗网络将预设生成器模型、预设判别器模型、预设分类器模型采用半监督学习方式共同配合训练，减弱了网络模型对图像标注信息的需求，充分利用无标注数据，进而实现具有高精度、高鲁棒性的智能网络模型；且减少了参数数量，提高了计算效率，有助于降低对终端的性能要求，加快了医学图像分类的速度，提高疾病诊断效率。
请参见图1，图1是本申请实施例提供的一种基于医学图像的智能辅助诊断方法的示意流程图。本实施例中智能辅助诊断方法的执行主体为终端，终端包括但不限于智能手机、平板电脑、个人数字助理(Personal Digital Assistant，PDA)等移动终端，还可以包括台式电脑等终端。如图1所示的智能辅助诊断方法可包括：
S101：获取待分类医学图像。
终端在检测到医学图像分类指令时，获取待分类医学图像。医学图像分类指令是用于指示终端进行医学图像分类的指令。图像分类指令可以由用户触发，如医生点击终端中的图像分类选项。获取待分类医学图像可以是用户上传至终端的待分类医学图像，也可以是终端根据图像分类指令中包含的文件标识，获取该文件标识对应的文本文件，并提取文本文件中的待分类医学图像。
S102：对所述待分类医学图像进行预处理，得到预处理图像。
终端对待分类医学图像进行预处理，得到预处理图像。具体地，终端将待分类医学图像处理为单颜色通道图像，将单颜色通道图像进行拼接，得到预处理图像。单颜色通道图像为由一种颜色元素的信息组成的颜色通道图像。保存图像颜色信息的通道称为颜色通道,每个颜色通道都存放着图像中颜色元素的信息。例如，在RGB色彩模式(RGB color mode，RGB)中，R表示一个红色通道，G表示一个绿色通道，B表示一个蓝色通道。终端可以通过调用预设函数将待分类医学图像的通道模式转换为多个单颜色通道图像；通过调用的 预设函数将多个单颜色通道图像拼接，得到预处理图像。
S103：将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
终端将预处理图像输入已训练的分类模型进行分类处理，得到与预处理图像对应的分类类别。其中，已训练的分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络；该已训练的分类模型中包含经过张量分解后的网络层以及二阶池化模块。
在训练过程中，预设生成器模型、预设判别器模型以及预设分类器模型三者相互配合、共同协作对训练数据进行训练，得到训练好的三元生成对抗网络。其中，训练数据可以包括样本图像、样本图像对应的分类类别(即预设真实图像标注)和无标注的样本图像；具体地，终端基于预设分类标注、一维高斯随机向量以及预设生成器模型，生成带有分类标注的合成图像，最终生成合成图像标注对；基于样本图像以及预设分类器模型，预测得到样本图像对应的样本图像标注对；将样本图像标注对、预设真实图像标注对以及合成图像标注对输入预设判别器模型进行判别处理，得到样本图像标注对所对应的第一判别结果、预设真实图像标注对所对应的第二判别结果以及合成图像标注对所对应的第三判别结果；基于第一判别结果、第二判别结果以及第三判别结果，计算预设生成器模型对应的第一损失函数、预设判别器模型对应的第二损失函数以及预设分类器模型对应的第三损失函数；基于第一损失函数、第二损失函数以及第三损失函数分别通过反向传播算法梯度下降更新预设生成器模型、预设判别器模型以及预设分类器模型各自对应的网络参数；当第一损失函数、第二损失函数以及第三损失函数均收敛时，停止训练，得到三元生成对抗网络，即得到已训练的分类模型；已训练的分类模型中包含已训练的生成器模型、已训练的判别器模型和已训练的分类器模型。将预处理图像输入已训练的分类器模型，已训练的分类器模型对预处理图像进行归一化处理，得到目标图像；采用已训练的分类器模型提取目标图像中的关键特征，得到全局高阶特征图；已训练的分类器模型获取全局高阶特征图对应的分类类别并输出该分类类别，即得到预处理图像对应的分类类别。
进一步地，为了使提取到的全局高阶特征图保留更多的重要特征信息，进而提高分类模型分类的准确率，S103可以包括S1031-S1033，具体如下：
S1031：采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像。
已训练的分类模型中包含已训练的生成器模型、已训练的判别器模型和已训练的分类 器模型，可通过已训练的分类器模型对预处理图像进行分类处理。具体地，将预处理图像输入已训练的分类器模型中，对预处理图像进行归一化处理，得到目标图像。例如，获取预处理图像对应的数据，对该数据进行线性变化，使得预处理图像对应的体素值在[-1，1]之间。
S1032：采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图。
终端采用已训练的分类器模型提取目标图像中的关键特征，得到全局高阶特征图；已训练的分类器模型中包含经过张量分解后的网络层以及二阶池化模块。请参阅图2，图2是本申请提供的已训练的分类器模型的结构示意图，如图2所示，已训练的分类器模型中包含3D卷积层、3D平均池化层、3D密集连接块1、二阶池化模块1、过渡层1、3D密集连接块2、二阶池化模块2、过渡层2、3D密集连接块3、二阶池化模块3、全连接层。
具体地，在训练分类模型的过程中，对预设生成器模型、预设判别器模型以及预设分类器模型中的每个网络层进行张量化。以分类器模型为例，在训练过程中，对预设分类器模型中3D卷积层、3D平均池化层、全连接层等网络层进行张量化。终端基于张量分解后的网络层以及已训练的分类器模型中二阶池化模块提取目标图像中的关键特征，得到全局高阶特征图。
本实施例中，对分类器模型中的每个网络层都进行张量化，通过张量化方法替代传统的向量化方法压缩分类器模型中的每个网络层，在减少参数的同时对网络模型起到正则化效果，解决了高分辨率图像在分类和识别时参数量过大、过拟合的问题；本实施例中，通过二阶池化替代传统的一阶池化，充分利用输入的预处理图像的二阶信息，在自注意力机制的作用下自动提取与病变相关的更有判别性的特征，进而提高分类器分类的准确率；且分类器模型采用3D卷积层替代2D卷积层，使得输入的图像可以不做任何降维以张量的形式做为输入，保留了图像的空间信息，减少了2D网络层造成的空间信息损失。
进一步地，为了提取到包含更多特征信息的全局高阶特征图，S1032可以包括S10321-S10324，具体如下：
S10321：通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第一特征图。
终端通过已训练的分类器模型中的张量化后的网络层提取目标图像中的特征，得到第一特征图。例如，通过已训练的分类器模型中的3D卷积层、3D平均池化层等提取目标图像中的特征，得到第一特征图。
S10322：通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维， 得到降维后的第二特征图。
终端通过已训练的分类器模型中的二阶池化模块对第一特征图进行通道降维，得到降维后的第二特征图。请参阅图3，图3是本申请提供的二阶池化模块的结构示意图，二阶池化模块包含预缩模块以及校准模块。具体地，通过1×1×1卷积对输入的4维特征图(即第一特征图)进行通道降维，得到降维后的第二特征图。
S10323：计算所述第二特征图对应的权重向量。
终端通过已训练的分类器模型计算第二特征图对应的权重向量。请参阅图3，具体地，计算降维后的第二特征图中不同通道之间两两通道的协方差信息，得到协方差矩阵；根据协方差矩阵通过分组卷积和1×1×1卷积得到与4维特征图通道数相同的权重向量。
S10324：基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
终端基于计算得到的权重向量对第一特征图进行加权，使第一特征图中重要的通道权重大，不重要的通道权重小，得到更有代表性的全局高阶特征图。具体地，已训练的分类器模型通过反向传播算法使第一特征图中重要的通道权重大，不重要的通道权重小，以此提取更有代表性的特征信息，得到全局高阶特征图。
S1033：采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
已训练的分类器模型获取全局高阶特征图对应的分类类别并输出该分类类别，即得到预处理图像对应的分类类别。进一步地，该分类类别可用于辅助疾病诊断本申请实施例，通过终端获取待分类医学图像；对待分类医学图像进行预处理，得到预处理图像；基于已训练的分类模型对预处理图像进行分类处理，得到对应的分类结果。由于该已训练的分类模型中包含经过张量分解后的网络层以及二阶池化模块，基于该分类模型对医学图像处理时，保留了医学图像的内部结构信息和内部相关性，二阶池化模块通过医学图像不同区域的依赖关系和高阶特征不同通道间的相关性信息，在自注意力机制的作用下，使重要的特征通道权重大，不重要的通道权重小，从而提取与病变相关的更有判别性的特征，提高疾病智能辅助诊断的准确率。进一步地，本申请中基于三元生成对抗网络对图像进行分类，在传统二元生成对抗网络的生成器和判别器的基础上加入了分类器网络模型，通过设计兼容性的损失函数缓解传统二元生成对抗网络训练不稳定的问题，解决了二元生成对抗网络判别器同时完成分类和判别真假两个任务目标收敛点不一致的问题，使生成器模型和分类器模型通过协作训练的方式替代传统二元生成对抗网络的生成对抗的训练方式，从而使生成器模型和分类器模型能够同时达到最优，加快收敛，使生成对抗网络更容易达到纳什均衡。本申请中对医学图像进行分类时，通过张量化方法替代传统的向量化方法压缩三元生 成对抗网络中的每个网络层，在减少参数的同时对网络模型起到正则化效果，解决了高分辨率图像在分类和识别时参数量过大、过拟合的问题；且通过张量化方法能够保持内部空间结构信息和不同体素之间的内部相关性，解决了向量化网络层的内部结构损失问题；本申请中通过二阶池化替代传统的一阶池化(最大池化或者平均池化)，充分利用整体图像的二阶信息在自注意力机制的作用下自动提取更有判别性的特征，提高分类器模型分类的准确率；本申请中的三元生成对抗网络通过预设生成器模型、预设判别器模型和预设分类器模型采用半监督学习方式共同配合训练，减弱了网络模型对图像标注信息的需求，充分利用无标注数据，进而实现具有高精度、高鲁棒性的智能网络模型；且减少了参数数量，提高了计算效率，有助于降低对终端的性能要求，从而加快辅助诊断流程。
请参见图4，图4是本申请另一实施例提供的一种基于医学图像的智能辅助诊断方法的示意流程图。本实施例中智能辅助诊断方法的执行主体为终端，终端包括但不限于智能手机、平板电脑、个人数字助理等移动终端，还可以包括台式电脑等终端。
本实施例在上一实施例的基础上增加了分类器模型的训练步骤S201-S202。本实施例中S203-S205与上一实施例中的S101-S103完全相同，具体请参阅上一实施例中S101-S103的相关描述，此处不赘述。如图4所示的智能辅助诊断方法，为了提高图像分类的准确率，S201-S202，具体如下：
S201：基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络。
请参考图5，图5为本申请提供的三元生成对抗网络的结构示意图，现以阿尔茨海默症这一应用场景为例结合图5中的三元生成对抗网络的结构，说明训练得到三元生成对抗网络的过程。在训练过程中，图5中的生成器指预设生成器模型，当训练完成后生成对应的已训练的生成器模型；在训练过程中，图5中的判别器指预设判别器模型，当训练完成后生成对应的已训练的判别器模型；在训练过程中，图5中的阿尔茨海默症分类器指预设分类器模型，当训练完成后生成对应的已训练的分类器模型。预设生成器模型中主要包括3D反卷积层；预设判别器模型中主要包括3D卷积层、3D密集连接块、过渡层、全连接层等；预设分类器模型中主要包括3D卷积层、3D密集连接块、二阶池化模块等。值得说明的是，此处仅是以阿尔茨海默症为例，训练得到可用于分类MRI图像的阿尔兹海默症智能辅助诊断模型；可通过该方法训练得到用于其它医学图像的分类模型，对此不做限定。
在训练过程中，对预设生成器模型、预设判别器模型以及预设分类器模型中的每个网络层进行张量化。具体地，预设判别器模型和预设分类器模型中3D卷积层和全连接层，以及预设生成器模型中的3D反卷积层通过张量分解方法进行参数压缩；全连接层的权重 矩阵、反卷积层的卷积核张量、卷积层的卷积核张量均可以表示为对应的张量形式：
W((i 1,j 1),...,(i d,j d))＝G 1[i 1,j 1]G 2[i 2,j 2]...G d[i d,j d]
将全连接层的权重张量W按照上述公式进行张量分解，得到全连接层的张量化表示如下：
Figure PCTCN2019119491-appb-000001
3D卷积层和反卷积层的张量分解步骤如下：
网络输入层：
Figure PCTCN2019119491-appb-000002
网络输出层：
Figure PCTCN2019119491-appb-000003
卷积核：
Figure PCTCN2019119491-appb-000004
3D卷积层张量化后得到：
Figure PCTCN2019119491-appb-000005
其中，
Figure PCTCN2019119491-appb-000006
在训练过程中，基于预设分类标注、一维高斯随机向量以及预设生成器模型，生成带有分类标注的合成图像，最终生成合成图像标注对；基于训练数据中的样本图像以及预设分类器模型，确定该样本图像对应的样本图像标注对；将样本图像标注对、预设真实图像标注对以及生成的合成图像标注对输入预设判别器模型进行判别处理，得到样本图像标注对所对应的第一判别结果、预设真实图像标注对所对应的第二判别结果以及合成图像标注对所对应的第三判别结果；基于第一判别结果、第二判别结果以及第三判别结果，计算预设生成器模型对应的第一损失函数、预设判别器模型对应的第二损失函数以及预设分类器模型对应的第三损失函数；基于第一损失函数、第二损失函数以及第三损失函数分别通过反向传播算法梯度下降更新预设生成器模型、预设判别器模型以及预设分类器模型各自对应的网络参数；当第一损失函数、第二损失函数以及第三损失函数均收敛时，停止训练，得到三元生成对抗网络，即得到已训练的分类模型。其中，已训练的分类模型中包含已训练的生成器模型、已训练的判别器模型和已训练的分类器模型。
S202：从所述三元生成对抗网络中获取所述已训练的分类器模型。
从三元生成对抗网络中获取已训练的分类器模型。具体地，训练得到的三元生成对抗网络即为已训练的分类模型，已训练的分类模型中包含已训练的生成器模型、已训练的判 别器模型和已训练的分类器模型，终端从已训练的分类模型中获取已训练的分类器模型。
进一步地，为了提高图像分类的准确率以及分类的速度，在另一实施方式中，S201可包括S2011-S2016，具体如下：
S2011：基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对。
具体地，将一维高斯随机向量以及预设分类标注作为输入，输入至预设生成器模型中；输入的预设分类标注经过独热编码(One-Hot Encoding)级联至每个张量化后的网络层，并基于一维高斯随机向量生成目标特征图；基于张量化后的网络层将目标特征图逐层放大，最终生成目标合成图像；基于目标合成图像以及预设分类标注，最终生成合成图像标注对。
进一步地，为了减少网络模型对医学图像的需求，将半监督学习方法引入到疾病分类任务中，能高效全面地利用无标注的医学图像信息，同时通过预设生成器模型生成带有分类标注的合成图像，起到数据增强的作用，在较小样本情况下也能训练高精度的辅助诊断模型。减小了对标注训练样本的需求量，缩减了传统算法中对训练数据繁杂的标注工作的工作量，进一步地缩短了疾病诊断的工作周期，加快了疾病诊断的速度，提升了疾病识别的整体效率。S2011包括S20111-S20113，具体如下：
S20111：将所述预设分类标注级联至所述张量化后的网络层，并基于所述一维高斯随机向量生成目标特征图。
预设生成器模型中包含张量化后的网络层。例如，预设生成器模型中包含张量化后的反卷积层。输入的预设分类标注经过独热编码级联至每个张量化后的反卷积层；基于张量化后的反卷积层以及一维高斯随机向量生成目标特征图。其中，反卷积层激活函数采用线性整流函数(Rectified Linear Unit，ReLU)、批量标准化(Batch Normalization，BN)。在阿尔茨海默症的辅助诊断识别中，目标特征图为脑部解剖特征图。
S20112：基于所述张量化后的网络层将所述目标特征图逐层放大，生成目标合成图像。
经过多层反卷积后，张量化后的网络层将目标特征图逐层放大，得到的图像为目标合成图像。例如，将脑部解剖特征图逐层放大，生成与真实MRI图像尺寸大小相同的合成图像。其中，预设生成器模型的最后一层采用双曲函数tanh激活函数。
S20113：基于所述目标合成图像以及所述预设分类标注，生成所述合成图像标注对。
终端基于目标合成图像以及预设分类标注，生成合成图像标注对。例如，终端基于合成图像以及预设分类标注生成带有分类标注的MRI图像。也可将带有分类标注的MRI图像称为MRI图像标注对。
S2012：基于所述样本图像以及所述预设分类器模型，预测所述样本图像对应的样本 图像标注对。
终端基于训练数据中的样本图像以及预设分类器模型，预测该样本图像的相应类别，基于样本图像以及相应类别确定样本图像对应的样本图像标注。具体地，将样本图像输入至预设分类器模型，该预设分类器模型预测样本图像对应的标注信息，基于样本图像以及标注信息生成样本图像对应的样本图像标注对。以阿尔茨海默症为例，样本图像即为真实无标注的MRI图像，将真实无标注的MRI图像输入预设分类器模型，该预设分类器模型预测真实无标注的MRI图像对应的标注信息，基于真实无标注的MRI图像以及预测的标注信息生成MRI图像标注对。
S2013：将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果。
具体地，将样本图像标注对输入预设判别器模型进行判别处理，
预设判别器模型提取样本图像标注对的特征信息，得到样本图像标注对所对应的样本特征图，基于预设判别器模型对样本特征图进行判别处理，得到第一判别结果。将训练数据中的预设真实图像标注对输入预设判别器模型进行判别处理，预设判别器模型提取预设真实图像标注对的特征信息，得到预设真实图像标注对所对应的真实特征图，基于预设判别器模型对真实特征图进行判别处理，得到第二判别结果。将合成图像标注对输入预设判别器模型进行判别处理，预设判别器模型提取合成图像标注对的特征信息，得到合成图像标注对所对应的合成特征图，基于预设判别器模型对合成特征图进行判别处理，得到第三判别结果。
进一步地，为了提高分类模型的分类性能，进而提高图像分类的准确率，S2013可以包括S20131-S20134，具体如下：
S20131：基于所述张量化后的密集型卷积神经网络提取所述样本图像标注对的特征信息，得到所述样本图像标注对所对应的样本特征图。
预设判别器模型中包含张量化后的密集型卷积神经网络。样本图像标注对所对应的样本图像以三阶张量的形式输入至预设判别器模型中，样本图像标注对中的分类标注经过独热编码作为条件变量级联至预设判别器模型中的每个网络层中。具体地，张量化后的密集型卷积神经网络提取样本图像标注对的特征信息，得到保留空间信息的特征图，即得到样本图像标注对所对应的样本特征图。其中，卷积层激活函数采用ReLU以及Batch Normalization。
S20132：基于所述张量化后的密集型卷积神经网络提取所述预设真实图像标注对的特征信息，得到所述预设真实样本图像标注对所对应的真实特征图。
真实图像标注对所对应的图像以三阶张量的形式输入至预设判别器模型中，真实图像标注对中的分类标注经过独热编码作为条件变量级联至预设判别器模型中的每个网络层中。具体地，张量化后的密集型卷积神经网络提取真实图像标注对的特征信息，得到保留空间信息的特征图，即得到真实图像标注对所对应的真实特征图。其中，卷积层激活函数采用ReLU以及Batch Normalization。
S20133：基于所述张量化后的密集型卷积神经网络提取所述合成图像标注对中的特征信息，得到所述合成图像标注对所对应的合成特征图。
合成图像以三阶张量的形式输入至预设判别器模型中，合成图像对应的分类标注经过独热编码作为条件变量级联至预设判别器模型中的每个网络层中。具体地，张量化后的密集型卷积神经网络提取合成图像标注对中的特征信息，得到保留空间信息的特征图，即得到合成图像标注对所对应的合成特征图。其中，卷积层激活函数采用ReLU以及Batch Normalization。
S20134：基于所述预设判别器模型对所述样本特征图、所述真实特征图以及所述合成特征图分别进行判别处理，得到所述第一判别结果、所述第二判别结果以及所述第三判别结果。
具体地，预设判别器模型包含张量化后的全连接层，基于该层的sigmoid函数进行判断，得到与样本特征图对应的第一判别结果、与真实特征图对应的第二判别结果、合成特征图对应的第三判别结果。
S2014：基于所述第一判别结果、所述第二判别结果以及所述第三判别结果，计算所述预设生成器模型对应的第一损失函数、所述预设判别器模型对应的第二损失函数以及所述预设分类器模型对应的第三损失函数。
具体地，在对预设生成器模型训练的过程中，在反向传播过程中根据预设生成器模型的损失函数G-loss梯度下降，更新预设生成器模型中网络层张量分解的核矩阵G k[i k,j k]参数。预设生成器模型的目标是通过模拟真实的图像，生成能够欺骗预设判别器模型的图像。例如，预设生成器模型通过模拟真实的MRI脑部解剖结构特征，生成能够欺骗预设判别器模型的接近真实的MRI图像。因此，预设生成器模型其损失包含两部分，一部分为欺骗预设判别器模型，使其判别预设生成器模型生成的MRI图像标注对为真；另一部分为真实MRI图像与生成的MRI图像之间的重建损失；可表示为：
Figure PCTCN2019119491-appb-000007
在对预设判别器模型训练的过程中，在反向传播过程中根据预设判别器模型的损失函数G-loss梯度下降，更新预设判别器模型中网络层张量分解的核矩阵G k[i k,j k]参数。预设判别器模型的目标是判定真实的图像标注对为真，判定由预设生成器模型生成的带有分类标注的合成图像标注对以及预设分类器模型预测得到的图像标注对为假。例如，预设判别器模型的目标是判定真实的MRI图像标注对为真，判定由预设生成器模型生成的MRI图像标注对以及预设分类器模型预测得到的MRI图像标注对为假。因此，预设判别器模型的损失包含三部分，具体如下：
Figure PCTCN2019119491-appb-000008
在对预设分类器模型训练的过程中，在反向传播过程中根据预设分类器模型的损失函数G-loss梯度下降，更新预设分类器模型中网络层张量分解的核矩阵G k[i k,j k]参数。预设分类器模型的目标是自动提取预处理图像中的特征信息并进行分类；例如，自动提取MRI脑部解剖结构特征进行分类，将MRI图像分为3个类别：正常、阿尔茨海默症和轻度认知障碍。因此，预设分类器模型的损失包含两部分，一部分为有监督的损失，即真实图像和生成图像进行分类任务的交叉熵；另一部分为无监督的损失，即欺骗预设判别器模型使其判别预设分类器模型对无标注的MRI图像生成的MRI图像标注对为真。具体可表示为：
Figure PCTCN2019119491-appb-000009
L supervised＝R L+α pR p
Figure PCTCN2019119491-appb-000010
Figure PCTCN2019119491-appb-000011
Figure PCTCN2019119491-appb-000012
对真实MRI图像标注对计算R L，等价于计算预设分类器模型学到的分布P c(x,y)与真实数据分布P real(x,y)之间的KL散度。预设生成器模型生成接近于真实分布的MRI图像标注对，可提高分类器模型的分类性能，因而引入R p计算生成MRI图像标注对的交叉熵；最 小化R p等价于最小化KL散度D KL(P g(x,y)||P c(x,y))；由于D KL(P g(x,y)||P c(x,y))无法直接计算，故KL散度D KL(P g(x,y)||P c(x,y))也无法直接计算，预设分类器模型通过间接最小化R p达到最小化KL散度D KL(P g(x,y)||P c(x,y))的目的。
S2015：基于所述第一损失函数、所述第二损失函数以及所述第三损失函数通过反向传播算法梯度下降分别更新所述预设生成器模型、所述预设判别器模型以及所述预设分类器模型各自对应的网络参数。
终端根据计算得到的第一损失函数、第二损失函数以及第三损失函数通过反向传播算法梯度下降分别更新预设生成器模型、预设判别器模型以及预设分类器模型各自对应的网络参数。例如，根据第一损失函数、第二损失函数以及第三损失函数更新预设生成器模型、预设判别器模型以及预设分类器模型各个模型中每一个网络层的权重值等。然后基于更新参数后的预设生成器模型、预设判别器模型以及预设分类器模型继续训练。即基于更新参数后的各个模型继续对样本图像以及样本图像对应的分类类别进行训练。
S2016：当所述第一损失函数、所述第二损失函数以及所述第三损失函数均收敛时，停止训练，得到所述三元生成对抗网络。
预设生成器模型、预设判别器模型以及预设分类器模型三者合作训练，如反复以“生成-判别-分类”合作模式训练。当终端检测到第一损失函数、第二损失函数以及第三损失函数在反复“生成-判别-分类”合作模式训练过程中均收敛时，停止训练，此时得到训练好的三元生成对抗网络，即已训练的分类模型。也可以是预先设置迭代次数，当执行该迭代次数的训练后，认为训练完成，此时得到训练好的三元生成对抗网络。
本申请实施例，通过终端获取待分类医学图像；对待分类医学图像进行预处理，得到预处理图像；基于已训练的分类模型对预处理图像进行分类处理，得到对应的分类结果。由于已训练的分类模型中包含经过张量分解后的网络层以及二阶池化模块，基于该分类模型对医学图像处理时，保留了医学图像的内部结构信息和内部相关性，二阶池化模块通过医学图像不同区域的依赖关系和高阶特征不同通道间的相关性信息，在自注意力机制的作用下，使重要的特征通道权重大，不重要的通道权重小，从而提取与病变相关的更有判别性的特征，提高疾病智能辅助诊断的准确率。进一步地，本申请中基于三元生成对抗网络对图像进行分类，在传统二元生成对抗网络的生成器和判别器的基础上加入了分类器网络模型，通过设计兼容性的损失函数缓解传统二元生成对抗网络训练不稳定的问题，解决了二元生成对抗网络判别器同时完成分类和判别真假两个任务目标收敛点不一致的问题，使生成器模型和分类器模型通过协作训练的方式替代传统二元生成对抗网络的生成对抗的训 练方式，从而使生成器模型和分类器模型能够同时达到最优，加快收敛，使生成的对抗网络更容易达到纳什均衡。本申请中对医学图像进行分类时，通过张量化方法替代传统的向量化方法压缩三元生成对抗网络中的每个网络层，在减少参数的同时对网络模型起到正则化效果，解决了高分辨率图像在分类和识别时参数量过大、过拟合的问题；且通过张量化方法能够保持图像内部空间结构信息和不同体素之间的内部相关性，解决了向量化网络层的图像内部结构损失问题；本申请中通过二阶池化替代传统的一阶池化(最大池化或者平均池化)，充分利用整体图像的二阶信息在自注意力机制的作用下自动提取更有判别性的特征，提高分类器模型分类的准确率；本申请中的三元生成对抗网络通过预设生成器模型、预设判别器模型和预设分类器模型采用半监督学习方式共同配合训练，减弱了网络模型对图像标注信息的需求，充分利用无标注数据，进而实现具有高精度、高鲁棒性的智能网络模型；且减少了参数数量，提高了计算效率，有助于降低对终端的性能要求，从而提高智能辅助诊断效率。
请参见图6，图6是本申请一实施例提供的一种基于医学图像的智能辅助诊断终端的示意图。该终端包括的各单元用于执行图1、图4对应的实施例中的各步骤。具体请参阅图1、图4各自对应的实施例中的相关描述。为了便于说明，仅示出了与本实施例相关的部分。参见图6，包括：
获取单元310，用于获取待分类医学图像；
预处理单元320，用于对所述待分类医学图像进行预处理，得到预处理图像；
分类单元330，用于将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
所述已训练的分类模型包含已训练的分类器模型。
进一步地，所述分类单元330包括：
处理单元，用于采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像；
提取单元，用于采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图；
分类类别获取单元，用于采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
进一步地，所述提取单元具体用于：
通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第一特征图；
通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维，得到降维后的第二特征图；
计算所述第二特征图对应的权重向量；
基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
进一步地，终端还包括：
训练单元，用于基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络；
模型获取单元，用于从所述三元生成对抗网络中获取所述已训练的分类器模型。
进一步地，所述训练单元包括：
生成单元，用于基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对；
确定单元，用于基于所述样本图像以及所述预设分类器模型，预测所述样本图像对应的样本图像标注对；
判别单元，用于将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果；
计算单元，用于基于所述第一判别结果、所述第二判别结果以及所述第三判别结果，计算所述预设生成器模型对应的第一损失函数、所述预设判别器模型对应的第二损失函数以及所述预设分类器模型对应的第三损失函数；
更新单元，用于基于所述第一损失函数、所述第二损失函数以及所述第三损失函数通过反向传播算法梯度下降分别更新所述预设生成器模型、所述预设判别器模型以及所述预设分类器模型各自对应的网络参数；
网络生成单元，用于当所述第一损失函数、所述第二损失函数以及所述第三损失函数均收敛时，停止训练，得到所述三元生成对抗网络。
所述预设生成器模型包含所述张量化后的网络层。
进一步地，所述生成单元具体用于：
将所述预设分类标注级联至所述张量化后的网络层，并基于所述一维高斯随机向量生成目标特征图；
基于所述张量化后的网络层将所述目标特征图逐层放大，生成目标合成图像；
基于所述目标合成图像以及所述预设分类标注，生成所述合成图像标注对。
所述预设判别器模型包含张量化后的密集型卷积神经网络。
进一步地，所述判别单元具体用于：
基于所述张量化后的密集型卷积神经网络提取所述样本图像标注对的特征信息，得到所述样本图像标注对所对应的样本特征图；
基于所述张量化后的密集型卷积神经网络提取所述预设真实图像标注对的特征信息，得到所述预设真实样本图像标注对所对应的真实特征图；
基于所述张量化后的密集型卷积神经网络提取所述合成图像标注对中的特征信息，得到所述合成图像标注对所对应的合成特征图；
基于所述预设判别器模型对所述样本特征图、所述真实特征图以及所述合成特征图分别进行判别处理，得到所述第一判别结果、所述第二判别结果以及所述第三判别结果。
请参见图7，图7是本申请另一实施例提供的一种基于医学图像的智能辅助诊断终端的示意图。如图7所示，该实施例的终端4包括：处理器40、存储器41以及存储在所述存储器41中并可在所述处理器40上运行的计算机可读指令42。所述处理器40执行所述计算机可读指令42时实现上述各个终端在基于医学图像的智能辅助诊断方法实施例中的步骤，例如图1所示的S101至S103。或者，所述处理器40执行所述计算机可读指令42时实现上述各实施例中各单元的功能，例如图6所示单元310至330功能。
示例性的，所述计算机可读指令42可以被分割成一个或多个单元，所述一个或者多个单元被存储在所述存储器41中，并由所述处理器40执行，以完成本申请。所述一个或多个单元可以是能够完成特定功能的一系列计算机可读指令段，该指令段用于描述所述计算机可读指令42在所述终端4中的执行过程。例如，所述计算机可读指令42可以被获取单元、预处理单元以及分类单元，各单元具体功能如上所述。
所述终端可包括，但不仅限于，处理器40、存储器41。本领域技术人员可以理解，图7仅仅是终端4的示例，并不构成对终端4的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如所述终端还可以包括输入输出终端、网络接入终端、总线等。
所称处理器40可以是中央处理单元(Central Processing Unit，CPU)，还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处 理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器41可以是所述终端4的内部存储单元，例如终端4的硬盘或内存。所述存储器41也可以是所述终端4的外部存储终端，例如所述终端4上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器41还可以既包括所述终端4的内部存储单元也包括外部存储终端。所述存储器41用于存储所述计算机可读指令以及所述终端所需的其他程序和数据。所述存储器41还可以用于暂时地存储已经输出或者将要输出的数据。以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种基于医学图像的智能辅助诊断方法，其特征在于，包括：
    获取待分类医学图像；
    对所述待分类医学图像进行预处理，得到预处理图像；
    将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
  2. 如权利要求1所述的智能辅助诊断方法，其特征在于，所述已训练的分类模型包含已训练的分类器模型，所述将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别包括：
    采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像；
    采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图；
    采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
  3. 如权利要求2所述的智能辅助诊断方法，其特征在于，所述采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图包括：
    通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第一特征图；
    通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维，得到降维后的第二特征图；
    计算所述第二特征图对应的权重向量；
    基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
  4. 如权利要求1至3任一项所述的智能辅助诊断方法，其特征在于，所述获取待分类医学图像之前，还包括：
    基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络；
    从所述三元生成对抗网络中获取所述已训练的分类器模型。
  5. 如权利要求4所述的智能辅助诊断方法，其特征在于，所述基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络包括：
    基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对；
    基于所述样本图像以及所述预设分类器模型，预测所述样本图像对应的样本图像标注对；
    将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果；
    基于所述第一判别结果、所述第二判别结果以及所述第三判别结果，计算所述预设生成器模型对应的第一损失函数、所述预设判别器模型对应的第二损失函数以及所述预设分类器模型对应的第三损失函数；
    基于所述第一损失函数、所述第二损失函数以及所述第三损失函数分别通过反向传播算法梯度下降更新所述预设生成器模型、所述预设判别器模型以及所述预设分类器模型各自对应的网络参数；
    当所述第一损失函数、所述第二损失函数以及所述第三损失函数均收敛时，停止训练，得到所述三元生成对抗网络。
  6. 如权利要求5所述的智能辅助诊断方法，其特征在于，所述预设生成器模型包含所述张量化后的网络层；所述基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对包括：
    将所述预设分类标注级联至所述张量化后的网络层，并基于所述一维高斯随机向量生成目标特征图；
    基于所述张量化后的网络层将所述目标特征图逐层放大，生成目标合成图像；
    基于所述目标合成图像以及所述预设分类标注，生成所述合成图像标注对。
  7. 如权利要求5所述的智能辅助诊断方法，其特征在于，所述预设判别器模型包含张量化后的密集型卷积神经网络；
    所述将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果包括：
    基于所述张量化后的密集型卷积神经网络提取所述样本图像标注对的特征信息，得到所述样本图像标注对所对应的样本特征图；
    基于所述张量化后的密集型卷积神经网络提取所述预设真实图像标注对的特征信息， 得到所述预设真实样本图像标注对所对应的真实特征图；
    基于所述张量化后的密集型卷积神经网络提取所述合成图像标注对中的特征信息，得到所述合成图像标注对所对应的合成特征图；
    基于所述预设判别器模型对所述样本特征图、所述真实特征图以及所述合成特征图分别进行判别处理，得到所述第一判别结果、所述第二判别结果以及所述第三判别结果。
  8. 一种基于医学图像的智能辅助诊断终端，其特征在于，包括：
    获取单元，用于获取待分类医学图像；
    预处理单元，用于对所述待分类医学图像进行预处理，得到预处理图像；
    分类单元，用于将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
  9. 如权利要求8所述的终端，其特征在于，所述已训练的分类模型包含已训练的分类器模型，所述分类单元包括：
    处理单元，用于采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像；
    提取单元，用于采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图；
    分类类别获取单元，用于采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
  10. 如权利要求9所述的终端，其特征在于，所述提取单元具体用于：
    通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第一特征图；
    通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维，得到降维后的第二特征图；
    计算所述第二特征图对应的权重向量；
    基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
  11. 如权利要求8至10任一项所述的终端，其特征在于，所述终端还包括：
    训练单元，用于基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到三元生成对抗网络；
    模型获取单元，用于从所述三元生成对抗网络中获取所述已训练的分类器模型。
  12. 如权利要求11所述的终端，其特征在于，所述训练单元包括：
    生成单元，用于基于预设分类标注、一维高斯随机向量以及所述预设生成器模型，生成合成图像标注对；
    确定单元，用于基于所述样本图像以及所述预设分类器模型，预测所述样本图像对应的样本图像标注对；
    判别单元，用于将所述样本图像标注对、预设真实图像标注对以及所述合成图像标注对输入所述预设判别器模型进行判别处理，得到所述样本图像标注对所对应的第一判别结果、所述预设真实图像标注对所对应的第二判别结果以及所述合成图像标注对所对应的第三判别结果；
    计算单元，用于基于所述第一判别结果、所述第二判别结果以及所述第三判别结果，计算所述预设生成器模型对应的第一损失函数、所述预设判别器模型对应的第二损失函数以及所述预设分类器模型对应的第三损失函数；
    更新单元，用于基于所述第一损失函数、所述第二损失函数以及所述第三损失函数通过反向传播算法梯度下降分别更新所述预设生成器模型、所述预设判别器模型以及所述预设分类器模型各自对应的网络参数；
    网络生成单元，用于当所述第一损失函数、所述第二损失函数以及所述第三损失函数均收敛时，停止训练，得到所述三元生成对抗网络。
  13. 如权利要求12所述的终端，其特征在于，所述生成单元具体用于：
    将所述预设分类标注级联至所述张量化后的网络层，并基于所述一维高斯随机向量生成目标特征图；
    基于所述张量化后的网络层将所述目标特征图逐层放大，生成目标合成图像；
    基于所述目标合成图像以及所述预设分类标注，生成所述合成图像标注对。
    所述预设判别器模型包含张量化后的密集型卷积神经网络。
  14. 如权利要求12所述的终端，其特征在于，所述判别单元具体用于：
    基于所述张量化后的密集型卷积神经网络提取所述样本图像标注对的特征信息，得到所述样本图像标注对所对应的样本特征图；
    基于所述张量化后的密集型卷积神经网络提取所述预设真实图像标注对的特征信息，得到所述预设真实样本图像标注对所对应的真实特征图；
    基于所述张量化后的密集型卷积神经网络提取所述合成图像标注对中的特征信息，得到所述合成图像标注对所对应的合成特征图；
    基于所述预设判别器模型对所述样本特征图、所述真实特征图以及所述合成特征图分 别进行判别处理，得到所述第一判别结果、所述第二判别结果以及所述第三判别结果。
  15. 一种基于医学图像的智能辅助诊断终端，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时，实现如下步骤：
    获取待分类医学图像；
    对所述待分类医学图像进行预处理，得到预处理图像；
    将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本图像对应的分类类别进行训练得到的三元生成对抗网络。
  16. 如权利要求15所述的终端，其特征在于，所述已训练的分类模型包含已训练的分类器模型，所述将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别包括：
    采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像；
    采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图；
    采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
  17. 如权利要求16所述的终端，其特征在于，所述采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图包括：
    通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第一特征图；
    通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维，得到降维后的第二特征图；
    计算所述第二特征图对应的权重向量；
    基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
  18. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现以下步骤：
    获取待分类医学图像；
    对所述待分类医学图像进行预处理，得到预处理图像；
    将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别；其中，所述分类模型包含经过张量化后的网络层以及二阶池化模块；所述分类模型是基于预设生成器模型、预设判别器模型以及预设分类器模型，对样本图像以及样本 图像对应的分类类别进行训练得到的三元生成对抗网络。
  19. 如权利要求18所述的计算机可读存储介质，其特征在于，所述已训练的分类模型包含已训练的分类器模型，所述将所述预处理图像输入已训练的分类模型进行分类处理，得到所述预处理图像对应的分类类别包括：
    采用所述分类器模型对所述预处理图像进行归一化处理，得到目标图像；
    采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图；
    采用所述分类器模型获取所述全局高阶特征图对应的所述分类类别。
  20. 如权利要求19所述的计算机可读存储介质，其特征在于，所述采用所述分类器模型提取所述目标图像中的关键特征，得到全局高阶特征图包括：
    通过所述分类器模型中的所述张量化后的网络层提取所述目标图像中的特征，得到第一特征图；
    通过所述分类器模型中的所述二阶池化模块对所述第一特征图进行通道降维，得到降维后的第二特征图；
    计算所述第二特征图对应的权重向量；
    基于所述权重向量对所述第一特征图进行加权，得到所述全局高阶特征图。
PCT/CN2019/119491 2019-11-19 2019-11-19 一种基于医学图像的智能辅助诊断方法及终端 Ceased WO2021097675A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
EP19953002.3A EP4064124B1 (en) 2019-11-19 2019-11-19 Smart diagnosis assistance method and terminal based on medical images
PCT/CN2019/119491 WO2021097675A1 (zh) 2019-11-19 2019-11-19 一种基于医学图像的智能辅助诊断方法及终端
US17/763,513 US12254684B2 (en) 2019-11-19 2019-11-19 Smart diagnosis assistance method to solve results of inaccurate classification of image, and terminal based on medical images

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/119491 WO2021097675A1 (zh) 2019-11-19 2019-11-19 一种基于医学图像的智能辅助诊断方法及终端

Publications (1)

Publication Number Publication Date
WO2021097675A1 true WO2021097675A1 (zh) 2021-05-27

Family

ID=75980083

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/119491 Ceased WO2021097675A1 (zh) 2019-11-19 2019-11-19 一种基于医学图像的智能辅助诊断方法及终端

Country Status (3)

Country Link
US (1) US12254684B2 (zh)
EP (1) EP4064124B1 (zh)
WO (1) WO2021097675A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113935969A (zh) * 2021-10-18 2022-01-14 太原理工大学 一种基于领域知识引导的新冠肺炎特异性病例的诊断系统
CN114760132A (zh) * 2022-04-14 2022-07-15 中国电信股份有限公司 信号发送方身份认证方法、装置、存储介质及电子设备
CN115019100A (zh) * 2022-06-14 2022-09-06 新乡医学院 基于生成对抗网络的生物组织图像的自动分类方法和系统
CN115099855A (zh) * 2022-06-23 2022-09-23 广州华多网络科技有限公司 广告文案创作模型制备方法及其装置、设备、介质、产品

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US12333717B2 (en) * 2019-12-17 2025-06-17 Shenzhen Institutes Of Advanced Technology Infrared image sequence-based sleep quality evaluation system and method comprising performing evaluation by a classifier and counting the sleep quality evaluation results to determine the largest number to judge good or poor sleep quality
US11663840B2 (en) * 2020-03-26 2023-05-30 Bloomberg Finance L.P. Method and system for removing noise in documents for image processing
US20220036239A1 (en) * 2020-07-29 2022-02-03 Mastercard International Incorporated Systems and methods to define the card member value from an issuer perspective
EP4252180B1 (en) * 2020-11-24 2026-03-18 Koninklijke Philips N.V. Image feature classification
US20220237434A1 (en) * 2021-01-25 2022-07-28 Samsung Electronics Co., Ltd. Electronic apparatus for processing multi-modal data, and operation method thereof
US11992322B2 (en) * 2021-03-30 2024-05-28 Ionetworks Inc. Heart rhythm detection method and system using radar sensor
CN113377981B (zh) * 2021-06-29 2022-05-27 山东建筑大学 基于多任务深度哈希学习的大规模物流商品图像检索方法
CN113744251B (zh) * 2021-09-07 2023-08-29 上海桐树生物科技有限公司 基于自注意力机制从病理图片预测微卫星不稳定性的方法
DE102021210920A1 (de) * 2021-09-29 2023-03-30 Siemens Healthcare Gmbh Einrichtung und computerimplementiertes Verfahren zum Trainieren eines Systems für maschinelles Lernen zum Zuordnen einer Scan-Untersuchung zu einem standardisierten Identifizierercode
US12555033B1 (en) 2021-12-21 2026-02-17 United Services Automobile Association (Usaa) Automatic identification of, and responding to, cognition impairment
US12387324B2 (en) 2022-02-18 2025-08-12 Arizona Board Of Regents On Behalf Of Arizona State University Systems, methods, and apparatuses for implementing discriminative, restorative, and adversarial (DiRA) learning for self-supervised medical image analysis
US20240273889A1 (en) * 2023-02-09 2024-08-15 Electronics And Telecommunications Research Institute System and method for classifying novel class objects
CN116344004B (zh) * 2023-05-31 2023-08-08 苏州恒瑞宏远医疗科技有限公司 一种影像样本数据扩增方法及装置
CN116994743B (zh) * 2023-07-11 2024-04-09 南通大学 基于序贯三支的置信图卷积神经网络的自闭症分类方法
CN117766155B (zh) * 2024-02-22 2024-05-10 中国人民解放军海军青岛特勤疗养中心 基于人工智能的动态血压医疗数据处理系统
CN117953349B (zh) * 2024-03-22 2024-06-11 广东海洋大学 一种中药材的病虫害检测方法、装置、设备及存储介质
CN119810557B (zh) * 2025-01-06 2026-02-27 山东浪潮科学研究院有限公司 利用生成对抗网络自动生成高质量标注数据的方法和系统
CN120147180B (zh) * 2025-05-15 2025-08-26 江西省科技基础条件平台中心(江西省计算中心) 基于深度学习的图像增强处理方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170344880A1 (en) * 2016-05-24 2017-11-30 Cavium, Inc. Systems and methods for vectorized fft for multi-dimensional convolution operations
CN107578069A (zh) * 2017-09-18 2018-01-12 北京邮电大学世纪学院 图像多尺度自动标注方法
CN107679572A (zh) * 2017-09-29 2018-02-09 深圳大学 一种图像判别方法、存储设备及移动终端
CN108198179A (zh) * 2018-01-03 2018-06-22 华南理工大学 一种生成对抗网络改进的ct医学图像肺结节检测方法
CN108491874A (zh) * 2018-03-19 2018-09-04 天津大学 一种基于生成式对抗网络的图像单分类方法
CN108961272A (zh) * 2018-07-02 2018-12-07 浙江工业大学 一种基于深度卷积对抗生成网络的皮肤病图像的生成方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7680307B2 (en) * 2005-04-05 2010-03-16 Scimed Life Systems, Inc. Systems and methods for image segmentation with a multi-stage classifier
US10540578B2 (en) 2017-12-21 2020-01-21 International Business Machines Corporation Adapting a generative adversarial network to new data sources for image classification
US10546389B2 (en) * 2018-04-06 2020-01-28 Elekta Ab (Publ) Devices and methods for identifying an object in an image
US10847265B2 (en) * 2018-04-06 2020-11-24 Curai, Inc. Systems and methods for responding to healthcare inquiries
US10891731B2 (en) * 2018-05-07 2021-01-12 Zebra Medical Vision Ltd. Systems and methods for pre-processing anatomical images for feeding into a classification neural network
US11056227B2 (en) * 2018-05-23 2021-07-06 Koninklijke Philips N.V. System and method for generating textual descriptions from medical images

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170344880A1 (en) * 2016-05-24 2017-11-30 Cavium, Inc. Systems and methods for vectorized fft for multi-dimensional convolution operations
CN107578069A (zh) * 2017-09-18 2018-01-12 北京邮电大学世纪学院 图像多尺度自动标注方法
CN107679572A (zh) * 2017-09-29 2018-02-09 深圳大学 一种图像判别方法、存储设备及移动终端
CN108198179A (zh) * 2018-01-03 2018-06-22 华南理工大学 一种生成对抗网络改进的ct医学图像肺结节检测方法
CN108491874A (zh) * 2018-03-19 2018-09-04 天津大学 一种基于生成式对抗网络的图像单分类方法
CN108961272A (zh) * 2018-07-02 2018-12-07 浙江工业大学 一种基于深度卷积对抗生成网络的皮肤病图像的生成方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4064124A4 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113935969A (zh) * 2021-10-18 2022-01-14 太原理工大学 一种基于领域知识引导的新冠肺炎特异性病例的诊断系统
CN113935969B (zh) * 2021-10-18 2024-04-12 太原理工大学 一种基于领域知识引导的新冠肺炎特异性病例的诊断系统
CN114760132A (zh) * 2022-04-14 2022-07-15 中国电信股份有限公司 信号发送方身份认证方法、装置、存储介质及电子设备
CN114760132B (zh) * 2022-04-14 2023-10-31 中国电信股份有限公司 信号发送方身份认证方法、装置、存储介质及电子设备
CN115019100A (zh) * 2022-06-14 2022-09-06 新乡医学院 基于生成对抗网络的生物组织图像的自动分类方法和系统
CN115099855A (zh) * 2022-06-23 2022-09-23 广州华多网络科技有限公司 广告文案创作模型制备方法及其装置、设备、介质、产品

Also Published As

Publication number Publication date
EP4064124A4 (en) 2022-11-23
EP4064124A1 (en) 2022-09-28
US20220343638A1 (en) 2022-10-27
EP4064124B1 (en) 2025-05-07
US12254684B2 (en) 2025-03-18

Similar Documents

Publication Publication Date Title
WO2021097675A1 (zh) 一种基于医学图像的智能辅助诊断方法及终端
CN110993094B (zh) 一种基于医学图像的智能辅助诊断方法及终端
Wang et al. MAGE: Automatic diagnosis of autism spectrum disorders using multi-atlas graph convolutional networks and ensemble learning
Pei et al. Multi-scale attention-based pseudo-3D convolution neural network for Alzheimer’s disease diagnosis using structural MRI
Vankdothu et al. Brain image identification and classification on Internet of Medical Things in healthcare system using support value based deep neural network
CN109409416B (zh) 特征向量降维方法和医学图像识别方法、装置及存储介质
US11449985B2 (en) Computer vision for cancerous tissue recognition
Kallenberg et al. Unsupervised deep learning applied to breast density segmentation and mammographic risk scoring
Pang et al. A novel end-to-end classifier using domain transferred deep convolutional neural networks for biomedical images
CN107748900B (zh) 基于区分性卷积神经网络的乳腺肿瘤分类装置和存储介质
Moujahid et al. Convolutional neural networks for multimodal brain MRI images segmentation: A comparative study
Li et al. Residual attention learning network and SVM for malaria parasite detection
WO2021179189A1 (zh) 大脑成瘾性状评估的可视化方法、装置及介质
CN111938644A (zh) 基于lle的阿尔茨海默病早期诊断方法、装置、系统
Liu et al. Self-paced multi-view positive and unlabeled graph learning with auxiliary information
Adnan et al. Enhancing AD Classification with Deep Learning: A Study of Transfer Learning and Snake Optimization on MRI Data.
Zine-dine et al. A Hybrid Model for Alzheimer's Disease Classification Based on Neural Network Architectures Enhanced by GAN Model.
Karthigeyan et al. QCVNET-Leveraging Quantum Technology in Convolutional Neural Networks for Improved Cardiovascular Disease Prediction
Liu et al. Brain MRI patient identification based on capsule network
Pai et al. Characterization of errors in deep learning-based brain MRI segmentation
Janbi et al. Comparing ML and DL approaches to diagnosis COVID-19 from CCGAN augmented CTX dataset
CN121121315B (zh) 一种超声图像分类方法、装置、设备及可读存储介质
Mustafina et al. Identification of psoriasis by images using convolutional neural networks
Emmanuel et al. A brief survey on different methods employed in each module for diagnosing alzheimer disease in the early stage
Kulal Performance Analysis of CNN Models Using MR Images of Pituitary Tumour

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: 19953002

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2019953002

Country of ref document: EP

Effective date: 20220620

WWG Wipo information: grant in national office

Ref document number: 17763513

Country of ref document: US

WWG Wipo information: grant in national office

Ref document number: 2019953002

Country of ref document: EP