WO2021077522A1 - 全息微波乳房肿块识别方法及识别系统 - Google Patents

全息微波乳房肿块识别方法及识别系统 Download PDF

Info

Publication number
WO2021077522A1
WO2021077522A1 PCT/CN2019/119952 CN2019119952W WO2021077522A1 WO 2021077522 A1 WO2021077522 A1 WO 2021077522A1 CN 2019119952 W CN2019119952 W CN 2019119952W WO 2021077522 A1 WO2021077522 A1 WO 2021077522A1
Authority
WO
WIPO (PCT)
Prior art keywords
breast
neural network
convolutional neural
deep convolutional
network model
Prior art date
Application number
PCT/CN2019/119952
Other languages
English (en)
French (fr)
Inventor
王露露
Original Assignee
深圳技术大学
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 深圳技术大学 filed Critical 深圳技术大学
Publication of WO2021077522A1 publication Critical patent/WO2021077522A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30068—Mammography; Breast
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30096—Tumor; Lesion

Definitions

  • the application belongs to the field of microwave imaging technology, and specifically relates to a holographic microwave breast mass recognition method and recognition system.
  • Microwave imaging is a new biomedical imaging method.
  • HM holographic microwave
  • HM imaging still has many shortcomings, such as long imaging scan time, high computational cost, low image resolution, and noise interference.
  • the cost of acquiring three-dimensional image data by scanning directly is relatively high. Reconstructing three-dimensional images from two-dimensional images is a common method, but the image quality is not guaranteed and often cannot meet people's needs.
  • CNN Convolutional Neural Network
  • the CNN architecture requires a large number of training data sets, which makes it more difficult to classify medical images because it takes a lot of time and manpower to create professionally labeled training data sets.
  • CNN may over-adapt and challenge learning the best image features.
  • the superficial CNN is too general to capture the nuances between these images; while the deep neural network (DNN) may become highly sensitive to nuances, but cannot capture the overall similarity between these images.
  • the present application provides a holographic microwave breast mass recognition method and recognition system.
  • the present application provides a holographic microwave breast mass recognition method, which includes the following steps:
  • Adjust the structural parameters of the deep convolutional neural network model use the training set to train the deep convolutional neural network model of each structural parameter, and obtain the deep convolutional neural network model with the required breast mass recognition accuracy;
  • the above-mentioned holographic microwave breast mass recognition method further includes the following steps:
  • the specific process of amplifying the HM color sample images without breast masses and breast masses, and using the amplified images to construct the training set and the test set is as follows:
  • a deep convolutional neural network model is designed; among them, the deep convolutional neural network model includes convolutional layer, pooling layer and fully connected layer.
  • the deep convolutional neural network-based recognition model for breast-free masses and breast masses includes an input module, a feature learning module, an image classification module, and an output module;
  • the feature learning module includes a three-layer convolution unit.
  • the first and second layer convolution units each include a convolution layer, a batch normalization layer, an excitation layer, and a pooling layer.
  • the third layer convolution unit includes a convolution layer, Batch standardization layer and incentive layer. Among them, the excitation layer uses the ReLU function.
  • the image classification module includes a fully connected layer and SoftMax classification function
  • the convolution layer performs a convolution operation on the input breast HM image through different numbers and sizes of convolution kernels, and extracts a feature map; in the convolution process, the two-dimensional breast HM image is used as input data, and the convolution kernel is moved Generate the final image on the entire two-dimensional breast HM image;
  • the convolution operation process is:
  • C(x,y) is the element in the output matrix of the convolution layer
  • A(x,y) is the element in the input matrix of the convolution layer
  • B(i,j) is the element in the convolution kernel
  • y is the yth column in the matrix
  • i is the ith row in the convolution kernel
  • j is the jth column in the convolution kernel
  • M is the size of the input matrix
  • N is the convolution The size of the nucleus
  • the extracted feature map is:
  • W s represents the kernel
  • * represents the convolution operator
  • X r is the input value of the r-th feature map
  • r is a natural number
  • b s is the bias term
  • the pooling process of the pooling layer is:
  • U(x', y') is the element in the output matrix of the pooling layer
  • m, n are integers in [0, ⁇ I]
  • ⁇ I is the step size of downsampling, which is a finite positive integer.
  • the fully connected layer processes the output of the pooling layer, and discards elements in the fully connected layer with a probability of 0.3-0.5.
  • the specific process of using the training set to train the deep convolutional neural network model of each structural parameter to obtain the required breast mass recognition accuracy is:
  • the deep convolutional neural network model with different structural parameters is trained through the training set, and the deep convolutional neural network model with the required breast mass recognition accuracy is obtained.
  • the present application also provides a holographic microwave breast mass recognition system, which includes an image acquisition module, an image amplification module, a model construction module, a training module, and a recognition module;
  • the image acquisition module is used to acquire HM color sample images without breast masses and HM color sample images with breast masses;
  • the image amplification module is used to amplify HM color sample images without breast masses and HM color sample images with breast masses, and use the amplified images to construct a training set and a test set;
  • the model building module is used to build a deep convolutional neural network model
  • the training module uses the training set to train the deep convolutional neural network model of each structural parameter to obtain the deep convolutional neural network model with the required breast mass recognition accuracy;
  • the recognition module uses the deep convolutional neural network model with the required breast mass recognition accuracy to perform a breast mass recognition test on the test set, and obtain an HM image with a breast mass.
  • the above-mentioned holographic microwave breast mass recognition system further includes a storage module and a display module.
  • the storage module is used to store HM images without breast masses and HM images with breast masses
  • the display module is used to display HM images without breast masses, There are HM images of breast masses and the diagnostic accuracy of breast masses.
  • the training module includes an adjustment unit, a combination unit and a training unit;
  • the adjustment unit is used to adjust the structural parameters of the deep convolutional neural network model in a preset area according to a decreasing law according to the size of the convolution kernel and a law of doubling the number of convolution kernels;
  • the combination unit is used to combine different structural parameters of the deep convolutional neural network model according to different sizes and numbers of convolution kernels, so as to construct a deep convolutional neural network model with different structural parameters;
  • the training unit uses the training set to train deep convolutional neural network models with different structural parameters to select a deep convolutional neural network model with a required breast mass recognition accuracy.
  • the present application also provides a computer storage medium on which a computer program is stored.
  • the computer program is executed by a processor, the steps of any one of the above-mentioned methods for identifying holographic microwave breast masses are realized .
  • this application builds a deep convolutional neural network model, and uses the training set to train the deep convolutional neural network model of each structural parameter to obtain the desired breast mass
  • the deep convolutional neural network model with recognition accuracy is used to perform the breast mass recognition test on the test set to obtain HM images with breast masses
  • this application can significantly reduce labor
  • the recognition error rate of breast image feature extraction and background selection can realize rapid classification of HM images without breast lumps and HM images with breast lumps, and accurate recognition.
  • the deep convolutional neural network model constructed by this application is robust.
  • This application applies the method based on deep convolutional network to the specific problem of breast mass detection, which can effectively improve the sensitivity and accuracy of breast mass detection, and realize the detection of HM images without breast masses and HM images with breast masses. auto recognition.
  • FIG. 1 is a flowchart of a method for identifying a holographic microwave breast mass according to an embodiment of the application.
  • Fig. 2(a) is a normal breast image without a breast mass in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • Figure 2(b) is a high-density normal breast image without breast masses in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • Figure 2(c) is an abnormal breast image with a breast mass in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • FIG. 3 is a schematic structural diagram of a recognition model for breast masses without breasts and breast masses based on a deep convolutional neural network in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • FIG. 4 is a schematic structural diagram of a deep convolutional neural network model in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • FIG. 5 is a schematic diagram of training results of a deep convolutional neural network model in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • FIG. 6 is a breast image classification diagram based on a deep convolutional neural network model in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • FIG. 7 is a holographic microwave breast mass recognition diagram based on a deep convolutional neural network in a holographic microwave breast mass recognition method provided by an embodiment of the application.
  • FIG. 8 is a structural block diagram of a holographic microwave breast mass recognition system provided by an embodiment of the application.
  • Image acquisition module 2. Image amplification module; 3. Model building module; 4. Training module; 5. Recognition module.
  • this application provides a holographic microwave breast mass recognition method, which includes the following steps:
  • HM color sample images without breast masses and HM color sample images with breast masses by performing high-speed scanning on the HM imaging system platform.
  • the HM color sample image without breast masses and the HM color sample image with breast masses can be amplified by image rotation. 75% of the amplified images are used to construct the training set, and 25% of the images are used to construct the test set.
  • the deep convolutional neural network model uses the training set to train the deep convolutional neural network model of each structural parameter, and obtain the deep convolutional neural network model with the required breast mass recognition accuracy.
  • the deep convolutional neural network model that requires breast mass recognition accuracy is usually the deep convolutional neural network model with the highest breast mass recognition accuracy in the training results.
  • the holographic microwave breast lump recognition method of this application also includes the following steps:
  • step S2 the specific process of amplifying the HM color sample images without breast lumps and breast lumps, and using the amplified images to construct the training set and the test set is as follows:
  • step S23 Amplify the HM grayscale images without breast masses and breast masses that have been preprocessed in step S22, and use the amplified images to construct a training set and a test set.
  • step S3 the specific process of constructing the deep convolutional neural network model is:
  • the feature learning module includes three layers of convolutional units.
  • the first and second layers of convolutional units both include convolutional layers, batch normalization layers, excitation layers, and pooling layers.
  • the third layer of convolutional units includes convolutional layers and batch normalization. Layer and incentive layer. Among them, the excitation layer uses the ReLU function.
  • the image classification module includes a fully connected layer and SoftMax classification function.
  • the convolution layer mainly performs convolution operations on the input breast HM image through different numbers and sizes of convolution kernels, and extracts feature maps. Among them, the convolution operation process is:
  • C(x,y) is the element in the output matrix of the convolution layer
  • A(x,y) is the element in the input matrix of the convolution layer
  • B(i,j) is the element in the convolution kernel Element
  • x is the xth row in the matrix
  • y is the yth column in the matrix
  • i is the ith row in the convolution kernel
  • j is the jth column in the convolution kernel
  • M is the size of the input matrix
  • the extracted feature map can be expressed as:
  • W s represents the kernel
  • * represents the convolution operator
  • X r is the input value of the r-th feature map
  • r is a natural number
  • b s is the bias term.
  • the two-dimensional breast HM image is used as input data, and the convolution kernel is moved to the entire two-dimensional breast HM image to generate the final image.
  • the batch normalization layer uses the following normalization methods to forcibly pull the input value distribution of any neuron in each layer of neural network back to a standard normal distribution with a mean of 0 and a variance of 1, so that the activation input value falls on a non-linear function to compare the input
  • the batch standardization layer can choose a relatively large initial learning rate, which greatly improves the training speed and eliminates the problem of parameter selection.
  • the specific process is:
  • the output of neuron type is the mean value Output value for the kth classification result, the standard deviation of the neuron output value is Among them, ⁇ is a small constant, the purpose is to prevent Approaching 0, the purpose of batch normalization is to adjust the input data of each layer of the neural network to a standard normal distribution with a mean value of zero and a variance of 1.
  • the pooling layer performs down-sampling operations, which are mainly used for feature dimensionality reduction, compressing the number of data and parameters, reducing overfitting, and improving the fault tolerance of the model.
  • the pooling process of the pooling layer is:
  • U(x', y') is the element in the output matrix of the pooling layer
  • m, n are integers in [0, ⁇ I]
  • ⁇ I is the step size of downsampling, which is a finite positive integer
  • the fully connected layer performs information integration on the entire image patch and provides the final classification; the fully connected layer processes the output of the pooling layer and discards the elements in the fully connected layer with a probability of 0.3-0.5.
  • the deep convolutional neural network model includes a convolutional layer and a pooling layer. And three layers of fully connected layer.
  • C represents the convolution kernel
  • the number on the left of C represents the size of the convolution kernel
  • the number on the right of C represents the number of convolution kernels.
  • 9C16 indicates that the convolution layer is 16 9 ⁇ 9 convolution kernels
  • S indicates the pooling layer
  • S2 indicates that the pooling layer template is 2 ⁇ 2.
  • step S4 the training set is used to train the deep convolutional neural network model with different structural parameters, and the deep convolutional neural network model with the recognition accuracy of breast lumps is optimized.
  • the specific process is:
  • the selection range of the size of the convolution kernel can be [9, 7, 5, 3, 1] and the selection range of the number of convolution kernels can be [16, 32, 64, 128, 256].
  • a deep convolutional neural network model in which the convolutional layer is three layers is used for description.
  • model Network structure 1 9C16-S2-7C32-S2-5C64 2 7C16-S2-5C32-S2-3C64 3 5C16-S2-3C32-S2-1C64
  • the deep convolutional neural network model with different structural parameters through the training set, and optimize the deep convolutional neural network model to obtain the required breast mass recognition accuracy.
  • the deep convolutional neural network model with the required breast mass recognition accuracy is usually the deep convolutional neural network model with the highest breast mass recognition accuracy in the obtained training results.
  • the deep convolutional neural network model with different structure parameters is trained one by one while the other layer structure parameters of the deep convolutional neural network model remain unchanged.
  • the training result of the deep convolutional neural network model is shown in Figure 5.
  • the first deep convolutional neural network model in Table 2 has the highest training accuracy, that is, the highest recognition rate. This model is selected as the optimized deep convolutional neural network model, and the optimized deep convolutional neural network model is used for The follow-up HM breast mass identification is in progress.
  • the breast image based on the deep convolutional neural network model is divided into muscle type, fat type and tumor type.
  • the tumor-type breast image as shown in FIG. 7 is recognized.
  • the present application also provides a holographic microwave breast mass recognition system, which includes an image acquisition module 1, an image amplification module 2, a model construction module 3, a training module 4, and a recognition module 5.
  • the image acquisition module 1 is used to acquire HM color sample images without breast masses and HM color sample images with breast masses.
  • HM color sample images without breast masses and HM color sample images with breast masses have corresponding category labels.
  • the image amplification module 2 is used to amplify HM color sample images without breast masses and HM color sample images with breast masses, and use the amplified images to construct a training set and a test set.
  • Model building module 3 is used to build a deep convolutional neural network model.
  • the training module 4 uses the training set to train the deep convolutional neural network model of each structural parameter to obtain the deep convolutional neural network model with the required breast mass recognition accuracy.
  • the recognition module 5 uses the deep convolutional neural network model with the required breast mass recognition accuracy to perform a breast mass recognition test on the test set, and obtain an HM image with a breast mass.
  • the holographic microwave breast mass recognition system of the present application also includes a storage module and a display module.
  • the storage module is used to store HM images without breast lumps and HM images with breast lumps.
  • the display module is used to display HM images without breast lumps, HM images with breast lumps, and diagnostic accuracy of breast lumps.
  • the training module 4 includes an adjustment unit, a combination unit, and a training unit, wherein the adjustment unit is used to perform a decrease in the size of the convolution kernel and the number of convolution kernels in a predetermined area to increase the depth of the convolutional nerve.
  • the structural parameters of the network model are adjusted.
  • the combination unit is used to combine different structural parameters of the deep convolutional neural network model according to different sizes and numbers of convolution kernels to construct a deep convolutional neural network model with different structural parameters.
  • the training unit uses the training set to train deep convolutional neural network models with different structural parameters to select the deep convolutional neural network model with the required breast mass recognition accuracy.
  • the deep convolutional neural network model with the required breast mass recognition accuracy is usually the deep convolutional neural network model with the highest breast mass recognition accuracy in the training results.
  • the holographic microwave breast lump recognition system provided in the above embodiment only uses the division of the above program modules for illustration. In practical applications, the above processing can be allocated to different program modules as needed, that is, the holographic microwave
  • the internal structure of the breast mass recognition system is divided into different program modules to complete all or part of the processing described above.
  • the holographic microwave breast mass recognition system provided in the above-mentioned embodiment and the embodiment of the holographic microwave breast mass recognition method belong to the same concept. For the specific implementation process, please refer to the method embodiment, which will not be repeated here.
  • This application builds a deep convolutional neural network model, and uses the training set to train the deep convolutional neural network model of each structural parameter to obtain the deep convolutional neural network model with the required breast mass recognition accuracy; use the required breast mass
  • the deep convolutional neural network model with recognition accuracy performs breast lump recognition test on the test set to obtain HM images with breast lump; this application can significantly reduce the recognition error rate of artificial breast image feature extraction and background selection, deep convolutional neural
  • the network model is robust, and can quickly classify and accurately identify HM images without breast lumps and HM images with breast lumps; this application applies the method based on deep convolutional networks to HM detection of breast lumps This specific problem can effectively improve the sensitivity and accuracy of breast mass detection, and realize automatic recognition of HM images without breast masses and HM images with breast masses.
  • an embodiment of the present application also provides a holographic microwave breast mass recognition device, which includes: processing And a memory for storing computer programs that can run on the processor. When the processor is used to run the computer program, the following steps are executed:
  • Adjust the structural parameters of the deep convolutional neural network model use the training set to train the deep convolutional neural network model of each structural parameter, and obtain the deep convolutional neural network model with the required breast mass recognition accuracy;
  • the embodiment of the present application also provides a computer storage medium, which is a computer-readable storage medium, for example, a memory including a computer program, which can be executed by a processor in a consensus device to complete the foregoing The steps in the holographic microwave breast lump recognition method.
  • a computer storage medium which is a computer-readable storage medium, for example, a memory including a computer program, which can be executed by a processor in a consensus device to complete the foregoing The steps in the holographic microwave breast lump recognition method.
  • the computer-readable storage medium may be a magnetic random access memory (FRAM, ferromagnetic random access memory), a read-only memory (ROM, Read Only Memory), a programmable read-only memory (PROM, Programmable Read-Only Memory), and an erasable Programmable Read-Only Memory (EPROM, Erasable Programmable Read-Only Memory), Electrically Erasable Programmable Read-Only Memory (EEPROM, Electrically Erasable Programmable Read-Only Memory), Flash Memory, Magnetic Surface Memory, Optical Disk , Or CD-ROM (Compact Disc Read-Only Memory) and other storage.
  • FRAM magnetic random access memory
  • ROM read-only memory
  • PROM programmable read-only memory
  • EPROM erasable Programmable Read-Only Memory
  • EPROM Erasable Programmable Read-Only Memory
  • EEPROM Electrically Erasable Programmable Read-Only Memory
  • Flash Memory Magnetic Surface Memory, Optical Disk , Or CD-ROM (Compact

Landscapes

  • Engineering & Computer Science (AREA)
  • Quality & Reliability (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)
  • Ultra Sonic Daignosis Equipment (AREA)

Abstract

一种全息微波乳房肿块识别方法及识别系统，识别方法包括以下步骤：分别获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像(S1)；对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并构建训练集和测试集(S2)；构建深度卷积神经网络模型(S3)；调节深度卷积神经网络模型的结构参数，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需病灶识别准确率的深度卷积神经网络模型(S4)；使用所需病灶识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像(S5)，能够有效地提高乳房肿块检测的灵敏度和准确率。

Description

全息微波乳房肿块识别方法及识别系统 技术领域
本申请属于微波成像技术领域，具体涉及一种全息微波乳房肿块识别方法及识别系统。
背景技术
微波成像是一种新的生物医学成像方法。研究表明，全息微波(holographic microwave，HM)具有肿瘤检测灵敏度高的优点，为乳腺癌的早期诊断提供了可能。随着HM技术在生物影像领域的推广应用，人们对高清晰HM图像和快速成像的需求日益增长。但因受制于算法和成像系统设计的缺陷，HM成像依然存在诸多不足，如成像扫描时间长、计算成本高、图像分辨率低、噪声干扰等。直接扫描获取三维图像数据的成本较高，从二维图像重构三维立体图像是常用的方法，但图像质量没有保障，经常不能满足人们的需求。
深度学习是生物医学成像领域的前沿技术，已成功应用于生物医学图像分类。卷积神经网络(CNN)是深度学习的一种，可用于生物医学图像分类。CNN体系结构需要大量的训练数据集，这使得对医疗图像进行分类变得更加困难，因为创建专业标记的训练数据集需要花费大量的时间和人力。当只涉及到小的训练数据集时，CNN可能会过度适应和挑战学习最佳的图像特征。肤浅的CNN过于笼统，无法捕捉到这些图像之间的细微差别；而深度神经网络(DNN)可能对细微差别变得高度敏感，但无法捕捉到这些图像之间的整体相似性。
发明内容
为至少在一定程度上克服相关技术中存在的问题，本申请提供了一种全息微波乳房肿块识别方法及识别系统。
根据本申请实施例的第一方面，本申请提供了一种全息微波乳房肿块识 别方法，其包括以下步骤：
分别获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像；
对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集；
构建深度卷积神经网络模型；
调节深度卷积神经网络模型的结构参数，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型；
使用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像。
上述全息微波乳房肿块识别方法还包括以下步骤：
对带有乳房肿块的HM图像中的乳房肿块进行二次分类筛查。
上述全息微波乳房肿块识别方法中，所述对无乳房肿块和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集的具体过程为：
分别获取无乳房肿块和有乳房肿块的HM彩色样本图像的病人信息以及图像的长、宽、高和像素信息；
将获取的无乳房肿块和有乳房肿块的HM彩色样本图像转化为灰度图像，并对灰度图像进行图像归一化预处理，提取特征；
对预处理完成的无乳房肿块和有乳房肿块的HM灰度图像进行扩增，并利用扩增后的图像构建训练集和测试集。
上述全息微波乳房肿块识别方法中，所述构建深度卷积神经网络模型的具体过程为；
构建基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型；
根据基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型，设计深度卷积神经网络模型；其中，深度卷积神经网络模型包含卷积层、池化层 和全连接层。
进一步地，所述基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型包括输入模块、特征学习模块、图像分类模块和输出模块；
所述特征学习模块包括三层卷积单元，第一层和第二层卷积单元均包括卷积层、批量标准化层、激励层和池化层，第三层卷积单元包括卷积层、批量标准化层和激励层。其中，激励层使用ReLU函数。
图像分类模块包括全连接层和SoftMax分类函数；
所述卷积层通过不同数量和大小的卷积核对输入的乳房HM图像进行卷积操作，并提取特征图；在卷积过程中，以二维乳房HM图像作为输入数据，将卷积核移到整个二维乳房HM图像上，生成最终图像；
卷积操作过程为：
Figure PCTCN2019119952-appb-000001
式中，C(x，y)为卷积层输出矩阵中的元素，A(x，y)为卷积层输入矩阵中的元素，B(i，j)为卷积核中的元素，x为矩阵中的第x行，y为矩阵中的第y列，i为卷积核中的第i行，j为卷积核中的第j列，M为输入矩阵的大小，N为卷积核的大小；
提取的特征图为：
O s＝∑ rW s*X r+b s，
式中，W s表示内核，*表示卷积运算符，X r为第r个特征图的输入值，r为自然数，b s是偏压项；
所述池化层的池化过程为：
U(x′，y′)＝max(R(x+m，y+n))，
式中，U(x′，y′)为池化层输出矩阵中的元素，m，n为[0，ΔI]中的整数，ΔI是下采样的步长，为有限的正整数，在池化层后构建归一化层，将U(x′，y′)规范得到归一化层输出矩阵中的元素，
Figure PCTCN2019119952-appb-000002
式中，V(x，y)为归一化层输出矩阵中的元素；σ为缩放常数，σ＝0.0001；μ为指数常数，μ＝0.75；M为输入矩阵的通道数；
所述全连接层处理池化层的输出，以0.3-0.5的概率舍弃全连接层中的元素。
上述全息微波乳房肿块识别方法中，所述利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型的具体过程为：
在给定区域内按照卷积核大小呈递减规律、卷积核数量成倍递增规律对深度卷积神经网络模型的结构参数进行调节；
根据不同的卷积核大小和数量，组合获得不同的深度卷积神经网络模型的结构参数，并构建出不同结构参数的深度卷积神经网络模型；
通过训练集对不同结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型。
根据本申请实施例的第二方面，本申请还提供了一种全息微波乳房肿块识别系统，其包括图像获取模块、图像扩增模块、模型构建模块、训练模块和识别模块；
所述图像获取模块，用于获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像；
所述图像扩增模块，用于对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集；
所述模型构建模块，用于构建深度卷积神经网络模型；
所述训练模块，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，以获得所需乳房肿块识别准确率的深度卷积神经网络模型；
所述识别模块，利用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像。
上述全息微波乳房肿块识别系统还包括存储模块和显示模块，所述存储模块用于存储无乳房肿块的HM图像和有乳房肿块的HM图像，所述显示模 块用于显示无乳房肿块的HM图像、有乳房肿块的HM图像以及乳房肿块诊断准确率。
上述全息微波乳房肿块识别系统中，所述训练模块包括调节单元、组合单元和训练单元；
所述调节单元用于在预设区域内按照按卷积核大小呈递减规律、卷积核数量成倍递增规律对深度卷积神经网络模型的结构参数进行调节；
所述组合单元用于根据不同的卷积核大小和数量，组合获得深度卷积神经网络模型的不同的结构参数，以构建出不同结构参数的深度卷积神经网络模型；
所述训练单元利用训练集对不同结构参数的深度卷积神经网络模型进行训练，以选出所需乳房肿块识别准确率的深度卷积神经网络模型。
根据本申请实施例的第三方面，本申请还提供了一种计算机存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现上述任一项全息微波乳房肿块识别方法的步骤。
根据本申请的上述具体实施方式可知，至少具有以下有益效果：本申请通过构建深度卷积神经网络模型，并利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型；利用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像；本申请能够显著降低人工乳房图像特征提取和背景选择的识别错误率，能够实现对无乳房肿块的HM图像和有乳房肿块的HM图像进行快速分类，并准确地进行识别。
本申请构建的深度卷积神经网络模型的鲁棒性强。
本申请将基于深度卷积网络的方法应用到乳房肿块HM检测这一具体问题，能够有效地提高乳房肿块检测的灵敏度和准确率，实现对无乳房肿块的HM图像和有乳房肿块的HM图像的自动识别。
应了解的是，上述一般描述及以下具体实施方式仅为示例性及阐释性的，其并不能限制本申请所欲主张的范围。
附图说明
下面的所附附图是本申请的说明书的一部分，其示出了本申请的实施例，所附附图与说明书的描述一起用来说明本申请的原理。
图1为本申请实施例提供的一种全息微波乳房肿块识别方法的流程图。
图2(a)为本申请实施例提供的一种全息微波乳房肿块识别方法中无乳房肿块的正常乳房图像。
图2(b)为本申请实施例提供的一种全息微波乳房肿块识别方法中无乳房肿块的高密度正常乳房图像。
图2(c)为本申请实施例提供的一种全息微波乳房肿块识别方法中有乳房肿块的异常乳房图像。
图3为本申请实施例提供的一种全息微波乳房肿块识别方法中基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型的结构示意图。
图4为本申请实施例提供的一种全息微波乳房肿块识别方法中深度卷积神经网络模型的结构示意图。
图5为本申请实施例提供的一种全息微波乳房肿块识别方法中深度卷积神经网络模型的训练结果示意图。
图6为本申请实施例提供的一种全息微波乳房肿块识别方法中基于深度卷积神经网络模型的乳房图像分类图。
图7为本申请实施例提供的一种全息微波乳房肿块识别方法中的基于深度卷积神经网络的全息微波乳房肿块识别图。
图8为本申请实施例提供的一种全息微波乳房肿块识别系统的结构框图。
附图标记说明：
1、图像获取模块；2、图像扩增模块；3、模型构建模块；4、训练模块；5、识别模块。
具体实施方式
为使本申请实施例的目的、技术方案和优点更加清楚明白，下面将以附图及详细叙述清楚说明本申请所揭示内容的精神，任何所属技术领域技术人 员在了解本申请内容的实施例后，当可由本申请内容所教示的技术，加以改变及修饰，其并不脱离本申请内容的精神与范围。
本申请的示意性实施例及其说明用于解释本申请，但并不作为对本申请的限定。另外，在附图及实施方式中所使用相同或类似标号的元件/构件是用来代表相同或类似部分。
关于本文中所使用的“第一”、“第二”、…等，并非特别指称次序或顺位的意思，也非用以限定本申请，其仅为了区别以相同技术用语描述的元件或操作。
关于本文中所使用的方向用语，例如：上、下、左、右、前或后等，仅是参考附图的方向。因此，使用的方向用语是用来说明并非用来限制本创作。
关于本文中所使用的“包含”、“包括”、“具有”、“含有”等等，均为开放性的用语，即意指包含但不限于。
关于本文中所使用的“及/或”，包括所述事物的任一或全部组合。
关于本文中的“多个”包括“两个”及“两个以上”；关于本文中的“多组”包括“两组”及“两组以上”。
关于本文中所使用的用语“大致”、“约”等，用以修饰任何可以细微变化的数量或误差，但这些微变化或误差并不会改变其本质。一般而言，此类用语所修饰的细微变化或误差的范围在部分实施例中可为20％，在部分实施例中可为10％，在部分实施例中可为5％或是其他数值。本领域技术人员应当了解，前述提及的数值可依实际需求而调整，并不以此为限。
某些用以描述本申请的用词将于下或在此说明书的别处讨论，以提供本领域技术人员在有关本申请的描述上额外的引导。
如图1所示，本申请提供了一种全息微波乳房肿块识别方法，其包括以下步骤：
S1、如图2所示，分别获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像，其中，无乳房肿块的HM彩色样本图像和有乳房肿块HM彩色样本图像均带有相应的种类标签。
具体地，可以通过在HM成像系统平台上进行高速扫描，获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像。
S2、对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集。
具体地，可以通过图像旋转对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增。扩增后的图像中75％的图像用于构建训练集，25％的图像用于构建测试集。
S3、构建深度卷积神经网络模型，以用于识别无乳房肿块的HM图像和有乳房肿块的HM图像。
S4、调节深度卷积神经网络模型的结构参数，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型。其中，所需乳房肿块识别准确率的深度卷积神经网络模型通常为训练结果中乳房肿块识别准确率最高的深度卷积神经网络模型。
S5、使用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像。
本申请全息微波乳房肿块识别方法还包括以下步骤：
S6、对带有乳房肿块的HM图像中的乳房肿块进行二次分类筛查，以降低假阳性的误诊率。
上述步骤S2中，对无乳房肿块和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集的具体过程为：
S21、分别获取无乳房肿块和有乳房肿块的HM彩色样本图像的病人信息以及图像的长、宽、高和像素信息。
S22、将获取的无乳房肿块和有乳房肿块的HM彩色样本图像转化为灰度图像，并对灰度图像进行图像归一化预处理，提取特征以减小图像尺寸。
S23、对步骤S22预处理完成的无乳房肿块和有乳房肿块的HM灰度图像进行扩增，并利用扩增后的图像构建训练集和测试集。
上述步骤S3中，构建深度卷积神经网络模型的具体过程为：
S31、构建基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型，其中，如图3所示，基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型包括输入模块、特征学习模块、图像分类模块和输出模块。
特征学习模块包括三层卷积单元，第一层和第二层卷积单元均包括卷积层、批量标准化层、激励层和池化层，第三层卷积单元包括卷积层、批量标准化层和激励层。其中，激励层使用ReLU函数。
图像分类模块包括全连接层和SoftMax分类函数。
卷积层主要通过不同数量和大小的卷积核对输入的乳房HM图像进行卷积操作，并提取特征图。其中，卷积操作过程为：
Figure PCTCN2019119952-appb-000003
式(1)中，C(x，y)为卷积层输出矩阵中的元素，A(x，y)为卷积层输入矩阵中的元素，B(i，j)为卷积核中的元素，x为矩阵中的第x行，y为矩阵中的第y列，i为卷积核中的第i行，j为卷积核中的第j列，M为输入矩阵的大小，N为卷积核的大小。
提取的特征图可以表示为：
Figure PCTCN2019119952-appb-000004
式(2)中，W s表示内核，*表示卷积运算符，X r为第r个特征图的输入值，r为自然数，b s是偏压项。
在卷积过程中，以二维乳房HM图像作为输入数据，将卷积核移到整个二维乳房HM图像上，生成最终图像。
批量标准化层通过以下规范化手段，把每层神经网络任意神经元这个输入值的分布强行拉回到均值为0，方差为1的标准正态分布，使得激活输入值落在非线性函数对输入比较敏感的区域，批量标准化层可以选择比较大的初始学习率，极大的提高训练速度，省去参数选择的问题，具体过程为：
Figure PCTCN2019119952-appb-000005
式(3)中，神经元型态的输出为均值
Figure PCTCN2019119952-appb-000006
为第k个分类结果输出值，神经元输出值的标准差为
Figure PCTCN2019119952-appb-000007
其中，ε是很小的常数，目的是防止
Figure PCTCN2019119952-appb-000008
趋近于0，批量归一化的目的就是把神经网络每一层的输入数据都调整到均值为零，方差为1的标准正态分布。
激励层使用的ReLU函数具体为：
Figure PCTCN2019119952-appb-000009
池化层进行下采样操作，主要用于特征降维，压缩数据和参数的数量，减小过拟合，同时提高模型的容错性。池化层的池化过程为：
U(x′，y′)＝max(R(x+m，y+n))     (5)
式(5)中，U(x′，y′)为池化层输出矩阵中的元素，m，n为[0，ΔI]中的整数，ΔI是下采样的步长，为有限的正整数，在池化层后构建归一化层，将U(x′，y′)规范得到归一化层输出矩阵中的元素，
Figure PCTCN2019119952-appb-000010
式(6)中，V(x，y)为归一化层输出矩阵中的元素；σ为缩放常数，σ＝0.0001；μ为指数常数，μ＝0.75；M为输入矩阵的通道数。
全连接层对整个图像补丁进行信息集成，并提供最终分类；全连接层处理池化层的输出，以0.3-0.5的概率舍弃全连接层中的元素。
S32、根据基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型，设计深度卷积神经网络模型，其中，如图4所示，深度卷积神经网络模型包含卷积层、池化层和全连接层三层。
如表1所示，设计出卷积为三层的深度卷积神经网络模型。其中，C表示卷积核，C左边的数字表示卷积核大小，C右边的数字表示卷积核数。例如，9C16表示卷积层为16个9×9的卷积核；S表示池化层，S2表示池化层 模板为2×2。
表1深度卷积神经网络模型的训练准确率
深度 网络结构 训练准确率(％)
3 9C16-S2-7C32-S2-5C64 100％
上述步骤S4中，利用训练集对不同结构参数的深度卷积神经网络模型进行训练，优选出乳房肿块识别准确率的深度卷积神经网络模型，其具体过程为：
S41、对于深度卷积神经网络模型，在给定区域内按照卷积核大小呈递减规律、卷积核数量成倍递增规律对深度卷积神经网络模型的结构参数进行调节。
具体地，卷积核大小的选择范围可以为[9，7，5，3，1]，卷积核数量的选择范围可以为[16，32，64，128，256]。
S42、根据不同的卷积核大小和数量，组合获得不同的深度卷积神经网络模型的结构参数，从而构建出不同结构参数的深度卷积神经网络模型。
例如，以卷积层为三层的深度卷积神经网络模型进行说明。
在[16，32，64，128，256]区域内选择[16，32，64]、[32，64，128]和[64，128，256]三种不同数量卷积核的结构参数组合，在[9，7，5，3，1]区域内选择[9，7，5]、[7，5，3]和[5，3，1]的三种卷积核大小结构参数的组合，在卷积层为三层的深度卷积神经网络模型下，可以根据以上结构参数组合构建出9种深度卷积神经网络模型，9种深度卷积神经网络模型的结构参数如表2所示。
表2不同卷积核参数的深度卷积神经网络模型
模型 网络结构
1 9C16-S2-7C32-S2-5C64
2 7C16-S2-5C32-S2-3C64
3 5C16-S2-3C32-S2-1C64
4 9C32-S2-7C64-S2-5C128
5 7C32-S2-5C64-S2-3C128
6 5C32-S2-3C64-S2-1C128
7 9C64-S2-7C128-S2-5C256
8 7C64-S2-5C128-S2-3C256
9 5C64-S2-3C128-S2-1C256
S43、通过训练集对不同结构参数的深度卷积神经网络模型进行训练，优化获得所需乳房肿块识别准确率的深度卷积神经网络模型。其中，所需乳房肿块识别准确率的深度卷积神经网络模型，通常为所得训练结果中乳房肿块识别准确率最高的深度卷积神经网络模型。
以卷积层为三层的深度卷积神经网络模型为例，在深度卷积神经网络模型其他各层结构参数保持不变的情况下，逐一对不同结构参数的深度卷积神经网络模型进行训练，深度卷积神经网络模型的训练结果如图5所示。
表2中第一种深度卷积神经网络模型的训练准确率最高，即识别率最高，选择该模型作为优化后的深度卷积神经网络模型，将该优化后的深度卷积神经网络模型用于后续的HM乳房肿块识别中。
如图6所示，将基于深度卷积神经网络模型的乳房图像分为肌肉型、脂肪型和肿瘤型。采用本申请全息微波乳房肿块识别方法，识别出如图7所示的肿瘤型乳房图像。
如图8所示，本申请还提供了一种全息微波乳房肿块识别系统，其包括图像获取模块1、图像扩增模块2、模型构建模块3、训练模块4和识别模块5。
其中，图像获取模块1，用于获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像。无乳房肿块的HM彩色样本图像和有乳房肿块HM彩色样本图像带有相应的种类标签。
图像扩增模块2，用于对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集。
模型构建模块3，用于构建深度卷积神经网络模型。
训练模块4，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，以获得所需乳房肿块识别准确率的深度卷积神经网络模型。
识别模块5，利用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像。
本申请全息微波乳房肿块识别系统还包括存储模块和显示模块。其中，存储模块用于存储无乳房肿块的HM图像和有乳房肿块的HM图像。显示模块用于显示无乳房肿块的HM图像、有乳房肿块的HM图像以及乳房肿块诊断准确率等结果。
具体地，训练模块4包括调节单元、组合单元和训练单元，其中，调节单元用于在预设区域内按照按卷积核大小呈递减规律、卷积核数量成倍递增规律对深度卷积神经网络模型的结构参数进行调节。
组合单元，用于根据不同的卷积核大小和数量，组合获得深度卷积神经网络模型的不同的结构参数，以构建出不同结构参数的深度卷积神经网络模型。
训练单元，利用训练集对不同结构参数的深度卷积神经网络模型进行训练，以选出所需乳房肿块识别准确率的深度卷积神经网络模型。其中，所需乳房肿块识别准确率的深度卷积神经网络模型，通常为训练结果中乳房肿块识别准确率最高的深度卷积神经网络模型。
需要说明的是：上述实施例提供的全息微波乳房肿块识别系统仅以上述各程序模块的划分进行举例说明，实际应用中，可以根据需要而将上述处理分配由不同的程序模块完成，即将全息微波乳房肿块识别系统的内部结构划分成不同的程序模块，以完成以上描述的全部或者部分处理。另外，上述实施例提供的全息微波乳房肿块识别系统与全息微波乳房肿块识别方法实施例属于同一构思，其具体实现过程详见方法实施例，这里不再赘述。
本申请通过构建深度卷积神经网络模型，并利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷 积神经网络模型；利用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像；本申请能够显著降低人工乳房图像特征提取和背景选择的识别错误率，深度卷积神经网络模型的鲁棒性强，能够实现对无乳房肿块的HM图像和有乳房肿块的HM图像进行快速分类，并准确地进行识别；本申请将基于深度卷积网络的方法应用到乳房肿块HM检测这一具体问题，能够有效地提高乳房肿块检测的灵敏度和准确率，实现对无乳房肿块的HM图像和有乳房肿块的HM图像的自动识别。
基于上述全息微波乳房肿块识别系统中各模块的硬件实现，为了实现本申请实施例提供的全息微波乳房肿块识别方法，本申请实施例还提供了一种全息微波乳房肿块识别装置，其包括：处理器和用于存储能够在处理器上运行的计算机程序的存储器。其中所述处理器用于运行所述计算机程序时，执行如下步骤：
分别获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像；
对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并构建训练集和测试集；
构建深度卷积神经网络模型，以用于识别无乳房肿块的HM图像和有乳房肿块的HM图像；
调节深度卷积神经网络模型的结构参数，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型；
使用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像。
在示例性实施例中，本申请实施例还提供了一种计算机存储介质，是计算机可读存储介质，例如，包括计算机程序的存储器，上述计算机程序可由共识装置中的处理器执行，以完成前述全息微波乳房肿块识别方法中的所述 步骤。
计算机可读存储介质可以是磁性随机存取存储器(FRAM，ferromagnetic random access memory)、只读存储器(ROM，Read Only Memory)、可编程只读存储器(PROM，Programmable Read-Only Memory)、可擦除可编程只读存储器(EPROM，Erasable Programmable Read-Only Memory)、电可擦除可编程只读存储器(EEPROM，Electrically Erasable Programmable Read-Only Memory)、快闪存储器(Flash Memory)、磁表面存储器、光盘、或只读光盘(CD-ROM，Compact Disc Read-Only Memory)等存储器。
以上所述仅为本申请示意性的具体实施方式，在不脱离本申请的构思和原则的前提下，任何本领域的技术人员所做出的等同变化与修改，均应属于本申请保护的范围。

Claims (10)

  1. 一种全息微波乳房肿块识别方法，其特征在于，包括以下步骤：
    分别获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像；
    对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集；
    构建深度卷积神经网络模型；
    调节深度卷积神经网络模型的结构参数，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型；
    使用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像。
  2. 根据权利要求1所述的全息微波乳房肿块识别方法，其特征在于，还包括以下步骤：
    对带有乳房肿块的HM图像中的乳房肿块进行二次分类筛查。
  3. 根据权利要求1或2所述的全息微波乳房肿块识别方法，其特征在于，所述对无乳房肿块和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集的具体过程为：
    分别获取无乳房肿块和有乳房肿块的HM彩色样本图像的病人信息以及图像的长、宽、高和像素信息；
    将获取的无乳房肿块和有乳房肿块的HM彩色样本图像转化为灰度图像，并对灰度图像进行图像归一化预处理，提取特征；
    对预处理完成的无乳房肿块和有乳房肿块的HM灰度图像进行扩增，并利用扩增后的图像构建训练集和测试集。
  4. 根据权利要求1或2所述的全息微波乳房肿块识别方法，其特征在于，所述构建深度卷积神经网络模型的具体过程为；
    构建基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型；
    根据基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型，设计 深度卷积神经网络模型；其中，深度卷积神经网络模型包含卷积层、池化层和全连接层。
  5. 根据权利要求4所述的全息微波乳房肿块识别方法，其特征在于，所述基于深度卷积神经网络的无乳房肿块和有乳房肿块识别模型包括输入模块、特征学习模块、图像分类模块和输出模块；
    所述特征学习模块包括三层卷积单元，第一层和第二层卷积单元均包括卷积层、批量标准化层、激励层和池化层，第三层卷积单元包括卷积层、批量标准化层和激励层。其中，激励层使用ReLU函数。
    图像分类模块包括全连接层和SoftMax分类函数；
    所述卷积层通过不同数量和大小的卷积核对输入的乳房HM图像进行卷积操作，并提取特征图；在卷积过程中，以二维乳房HM图像作为输入数据，将卷积核移到整个二维乳房HM图像上，生成最终图像；
    卷积操作过程为：
    Figure PCTCN2019119952-appb-100001
    式中，C(x，y)为卷积层输出矩阵中的元素，A(x，y)为卷积层输入矩阵中的元素，B(i，j)为卷积核中的元素，x为矩阵中的第x行，y为矩阵中的第y列，i为卷积核中的第i行，j为卷积核中的第j列，M为输入矩阵的大小，N为卷积核的大小；
    提取的特征图为：
    O s＝∑ rW s*X r+b s，
    式中，W s表示内核，*表示卷积运算符，X r为第r个特征图的输入值，r为自然数，b s是偏压项；
    所述池化层的池化过程为：
    U(x′，y′)＝max(R(x+m，y+n))，
    式中，U(x′，y′)为池化层输出矩阵中的元素，m，n为[0，ΔI]中的整数，ΔI是下采样的步长，为有限的正整数，在池化层后构建归一化层，将U(x′，y′) 规范得到归一化层输出矩阵中的元素，
    Figure PCTCN2019119952-appb-100002
    式中，V(x，y)为归一化层输出矩阵中的元素；σ为缩放常数，σ＝0.0001；μ为指数常数，μ＝0.75；M为输入矩阵的通道数；
    所述全连接层处理池化层的输出，以0.3-0.5的概率舍弃全连接层中的元素。
  6. 根据权利要求1或2所述的全息微波乳房肿块识别方法，其特征在于，所述利用训练集对各个结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型的具体过程为：
    在给定区域内按照卷积核大小呈递减规律、卷积核数量成倍递增规律对深度卷积神经网络模型的结构参数进行调节；
    根据不同的卷积核大小和数量，组合获得不同的深度卷积神经网络模型的结构参数，并构建出不同结构参数的深度卷积神经网络模型；
    通过训练集对不同结构参数的深度卷积神经网络模型进行训练，获得所需乳房肿块识别准确率的深度卷积神经网络模型。
  7. 一种全息微波乳房肿块识别系统，其特征在于，包括图像获取模块、图像扩增模块、模型构建模块、训练模块和识别模块；
    所述图像获取模块，用于获取无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像；
    所述图像扩增模块，用于对无乳房肿块的HM彩色样本图像和有乳房肿块的HM彩色样本图像进行扩增，并利用扩增后的图像构建训练集和测试集；
    所述模型构建模块，用于构建深度卷积神经网络模型；
    所述训练模块，利用训练集对各个结构参数的深度卷积神经网络模型进行训练，以获得所需乳房肿块识别准确率的深度卷积神经网络模型；
    所述识别模块，利用所需乳房肿块识别准确率的深度卷积神经网络模型对测试集进行乳房肿块识别测试，获取带有乳房肿块的HM图像。
  8. 根据权利要求7所述的全息微波乳房肿块识别系统，其特征在于，还包括存储模块和显示模块，所述存储模块用于存储无乳房肿块的HM图像和有乳房肿块的HM图像，所述显示模块用于显示无乳房肿块的HM图像、有乳房肿块的HM图像以及乳房肿块诊断准确率。
  9. 根据权利要求7或8所述的全息微波乳房肿块识别系统，其特征在于，所述训练模块包括调节单元、组合单元和训练单元；
    所述调节单元用于在预设区域内按照按卷积核大小呈递减规律、卷积核数量成倍递增规律对深度卷积神经网络模型的结构参数进行调节；
    所述组合单元用于根据不同的卷积核大小和数量，组合获得深度卷积神经网络模型的不同的结构参数，以构建出不同结构参数的深度卷积神经网络模型；
    所述训练单元利用训练集对不同结构参数的深度卷积神经网络模型进行训练，以选出所需乳房肿块识别准确率的深度卷积神经网络模型。
  10. 一种计算机存储介质，其特征在于，其上存储有计算机程序，所述计算机程序被处理器执行时实现权利要求1至6任一项所述全息微波乳房肿块识别方法的步骤。
PCT/CN2019/119952 2019-10-25 2019-11-21 全息微波乳房肿块识别方法及识别系统 WO2021077522A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911021127.0 2019-10-25
CN201911021127.0A CN110782444A (zh) 2019-10-25 2019-10-25 全息微波乳房肿块识别方法及识别系统

Publications (1)

Publication Number Publication Date
WO2021077522A1 true WO2021077522A1 (zh) 2021-04-29

Family

ID=69386424

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/119952 WO2021077522A1 (zh) 2019-10-25 2019-11-21 全息微波乳房肿块识别方法及识别系统

Country Status (2)

Country Link
CN (1) CN110782444A (zh)
WO (1) WO2021077522A1 (zh)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3093821A1 (en) * 2015-04-16 2016-11-16 Siemens Aktiengesellschaft Method and system for anatomical object pose detection using marginal space deep neural networks
CN107368859A (zh) * 2017-07-18 2017-11-21 北京华信佳音医疗科技发展有限责任公司 病变识别模型的训练方法、验证方法和病变图像识别装置
CN107886514A (zh) * 2017-11-22 2018-04-06 浙江中医药大学 基于深度残差网络的乳腺钼靶图像肿块语义分割方法
CN109447088A (zh) * 2018-10-16 2019-03-08 杭州依图医疗技术有限公司 一种乳腺影像识别的方法及装置
CN109461144A (zh) * 2018-10-16 2019-03-12 杭州依图医疗技术有限公司 一种乳腺影像识别的方法及装置
CN109635835A (zh) * 2018-11-08 2019-04-16 深圳蓝韵医学影像有限公司 一种基于深度学习与迁移学习的乳腺病变区域检测方法
US20190205766A1 (en) * 2018-01-03 2019-07-04 Siemens Healthcare Gmbh Medical Imaging Diffeomorphic Registration based on Machine Learning

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107680082A (zh) * 2017-09-11 2018-02-09 宁夏医科大学 基于深度卷积神经网络和全局特征的肺部肿瘤识别方法
CN110232396B (zh) * 2019-04-09 2022-07-01 贵州大学 X射线乳腺影像深度学习分类方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3093821A1 (en) * 2015-04-16 2016-11-16 Siemens Aktiengesellschaft Method and system for anatomical object pose detection using marginal space deep neural networks
CN107368859A (zh) * 2017-07-18 2017-11-21 北京华信佳音医疗科技发展有限责任公司 病变识别模型的训练方法、验证方法和病变图像识别装置
CN107886514A (zh) * 2017-11-22 2018-04-06 浙江中医药大学 基于深度残差网络的乳腺钼靶图像肿块语义分割方法
US20190205766A1 (en) * 2018-01-03 2019-07-04 Siemens Healthcare Gmbh Medical Imaging Diffeomorphic Registration based on Machine Learning
CN109447088A (zh) * 2018-10-16 2019-03-08 杭州依图医疗技术有限公司 一种乳腺影像识别的方法及装置
CN109461144A (zh) * 2018-10-16 2019-03-12 杭州依图医疗技术有限公司 一种乳腺影像识别的方法及装置
CN109635835A (zh) * 2018-11-08 2019-04-16 深圳蓝韵医学影像有限公司 一种基于深度学习与迁移学习的乳腺病变区域检测方法

Also Published As

Publication number Publication date
CN110782444A (zh) 2020-02-11

Similar Documents

Publication Publication Date Title
JP7458328B2 (ja) マルチ分解能登録を介したマルチサンプル全体スライド画像処理
US10482603B1 (en) Medical image segmentation using an integrated edge guidance module and object segmentation network
Kaur et al. Plant species identification based on plant leaf using computer vision and machine learning techniques
US10776606B2 (en) Methods for delineating cellular regions and classifying regions of histopathology and microanatomy
CN111524137B (zh) 基于图像识别的细胞识别计数方法、装置和计算机设备
CN111291825B (zh) 病灶分类模型训练方法、装置、计算机设备和存储介质
CN109389584A (zh) 基于cnn的多尺度鼻咽肿瘤分割方法
Kolhe et al. Internet of Thing and Machine Learning Approach for Agricultural Application: A Review
US20210118137A1 (en) Method and apparatus of labeling target in image, and computer recording medium
CN108664986B (zh) 基于lp范数正则化的多任务学习图像分类方法及系统
Assad et al. Deep biomedical image classification using diagonal bilinear interpolation and residual network
Junayed et al. ScarNet: development and validation of a novel deep CNN model for acne scar classification with a new dataset
CN113450305B (zh) 医疗图像的处理方法、系统、设备及可读存储介质
Varghese et al. INFOPLANT: Plant recognition using convolutional neural networks
Rafid et al. An effective ensemble machine learning approach to classify breast cancer based on feature selection and lesion segmentation using preprocessed mammograms
Daniel et al. Iris-based image processing for cholesterol level detection using gray level co-occurrence matrix and support vector machine
CN111127400A (zh) 一种乳腺病变检测方法和装置
Xie et al. Tea-YOLOv8s: A tea bud detection model based on deep learning and computer vision
CN103366183B (zh) 一种局灶性病灶的非参数自动检测方法
WO2021077522A1 (zh) 全息微波乳房肿块识别方法及识别系统
CN115937590A (zh) 一种并联融合CNN和Transformer的皮肤病图像分类方法
CN115880245A (zh) 一种基于自监督的乳腺癌疾病分类方法
Kale et al. Identification of Ayurvedic Leaves using Deep Learning
Chaugule et al. A new method using feature extraction for identifying paddy rice species for quality seed selection
Babu et al. Plant Leaves Disease Detection (Efficient Net)

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: 19949924

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19949924

Country of ref document: EP

Kind code of ref document: A1