WO2022120739A1 - 一种基于卷积神经网络的医学图像分割方法及装置 - Google Patents

一种基于卷积神经网络的医学图像分割方法及装置 Download PDF

Info

Publication number
WO2022120739A1
WO2022120739A1 PCT/CN2020/135348 CN2020135348W WO2022120739A1 WO 2022120739 A1 WO2022120739 A1 WO 2022120739A1 CN 2020135348 W CN2020135348 W CN 2020135348W WO 2022120739 A1 WO2022120739 A1 WO 2022120739A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
convolutional neural
neural network
edge
enhanced
Prior art date
Application number
PCT/CN2020/135348
Other languages
English (en)
French (fr)
Inventor
张娜
郑海荣
刘新
申帅
胡战利
梁栋
李烨
邹超
贾森
Original Assignee
深圳先进技术研究院
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 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Priority to PCT/CN2020/135348 priority Critical patent/WO2022120739A1/zh
Publication of WO2022120739A1 publication Critical patent/WO2022120739A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis

Definitions

  • the invention relates to the technical field of medical image segmentation, in particular to a medical image segmentation method and device based on a convolutional neural network.
  • Convolutional Neural Networks are deep learning models or multilayer perceptrons similar to artificial neural networks that are commonly used to analyze image data. Convolutional neural networks mainly use convolutional layers and pooling layers for feature extraction and image size compression. The following takes U-net network as an example to illustrate the application of convolutional neural network in the process of medical image segmentation:
  • FCN full convolutional neural network
  • the model of the neural network has two modes: training mode (Train) and evaluation mode (Eval).
  • Training mode the data is put into the network model group by group for training, and the parameters in the model can better reflect the change process from input to output with the training.
  • evaluation mode the parameters of the model are no longer changed. Input unlabeled image data, and the calculation of the model after training can get the fitting result.
  • the present application provides a medical image segmentation method and device based on a convolutional neural network, which is used to overcome the fact that the generalization performance of the network is affected when the amount of training sample data is insufficient in the prior art; Improve the accuracy of the model and alleviate overfitting in less cases.
  • the present application provides a medical image segmentation data enhancement method based on a convolutional neural network, including:
  • the image to be tested is tested by the trained convolutional neural network, so that the image to be tested is segmented according to the type of the labeled data.
  • the amount of data that can be used as a training sample is small and the effective feature information that can be extracted by the network model in the training sample is insufficient, adding a part of artificial features and edge information to the training sample can increase the number of each training sample.
  • the feature information of a data sample can improve the generalization performance of the network and achieve the technical effect of reducing the sample size required for network fitting.
  • the above method also includes:
  • the obtained multi-channel image data is input into the convolutional neural network obtained by training for fitting, and the segmented image is obtained.
  • the image to be tested (unlabeled, the labeled image is automatically generated by the training model) is also subjected to edge enhancement processing, and combined with the image to be tested to generate multiple
  • the channel image data is input into the convolutional neural network trained in the previous step for contour fitting. Since the parameters of the convolutional neural network model are obtained through the training in the above training steps, the types marked in the above training process can be directly obtained.
  • Segment the image to be tested and output the segmented image compared with related technologies, because the generalization performance of the network model is improved during the training process, the recognition performance of the image to be tested is higher, and the difference between the output segmented image and the labeled image is higher. The error is greatly reduced; in addition, the image to be tested is input into the trained model for contour fitting, and the multi-channel edge enhancement image is added, which is conducive to extracting edge information during the model fitting process, thereby improving the fitting speed of the test. and accuracy.
  • the described raw image obtained is carried out edge enhancement processing, and the step of obtaining the edge enhanced image comprises:
  • the step of performing edge enhancement processing on the image to be tested to obtain the image to be tested with edge enhancement includes:
  • Laplacian, Scharr and/or Canny edge enhancement algorithms are respectively used to perform edge enhancement processing on the acquired images to be tested, to obtain multiple edge-enhanced images to be tested.
  • the original image or the image to be tested is respectively subjected to edge enhancement processing, so that clear edge information of the image can be obtained efficiently and quickly, which is beneficial to improve the speed of image segmentation and the input of effective information features.
  • the original image and the image to be tested are both cerebrovascular images or both are cardiovascular images.
  • the convolutional neural network model is a U-net network segmentation model with multiple input channels.
  • the feature extraction and image size compression are mainly performed by the convolutional layer and the pooling layer, so as to achieve excellent performance of medical image segmentation and effectively segment the target image elements.
  • the step of merging the original image and the edge-enhanced original image to generate multi-channel image data includes:
  • V1 train ⁇ x 1 ,x 2 ,x 3 ,...,x n ⁇
  • x i is an image sample of size r ⁇ r, and the number of samples in the original training set is n; i ⁇ [1,n], n is a positive integer; r is the pixel value;
  • the enhanced training set/test set V2 train consisting of m ⁇ n edge-enhanced original image samples x ij is represented as follows:
  • V2 train ⁇ x i1 ,...,x im ⁇
  • x ij is an edge-enhanced original image sample of size r ⁇ r, and the number of samples in the enhanced training set is n ⁇ m; i ⁇ [1,n], j ⁇ [1,m], m is a positive integer;
  • the training set V train formed by combining the original training set V1 train and the enhanced training set V2 train is expressed as follows:
  • V train V1 train +V2 train
  • the step of combining the image to be tested and the image to be tested with edge enhancement to generate multi-channel image data includes:
  • V1 test ⁇ y 1 ,y 2 ,y 3 ,...,y w ⁇
  • y i is an image sample of size r ⁇ r, and the number of samples in the original test set is w; i ⁇ [1,w], w is a positive integer;
  • the image sample y ij the enhanced test set V2 test composed of m ⁇ n edge-enhanced image samples y ij to be tested is expressed as follows:
  • V2 test ⁇ y i1 ,...,y im ⁇
  • y ij is an edge-enhanced image sample of size r ⁇ r to be tested, and the number of samples in the enhanced test set is n ⁇ m; i ⁇ [1,n], j ⁇ [1,m], m is a positive integer ; r is the pixel value;
  • test set V test formed by merging the original test set V1 test and the enhanced test set V2 test is expressed as follows:
  • V test V1 test + V2 test
  • the network model is trained for the target until the loss function converges, and finally the network model parameters are obtained, and then the training model is obtained; in the testing process, the w multi-channel data (x i , x i1 , x i2 .
  • Different channels are input to the training model for fitting, and finally the classification result is obtained; each original image in the original training set is processed through a variety of edge enhancement algorithms to obtain multiple edge-enhanced original images, and the number of samples that can be obtained is the original training set.
  • the number of augmented training sets is multiple times, and the multi-channel image data formed by the sum of all samples in the augmented training set and the original training set is used as the training set to train the convolutional neural network model, compared with only the samples of the original training set as the training set.
  • the convolutional neural network model is trained, and the image edge feature information of multiple channels is added based on each original image sample, so as to overcome the defect that it is difficult to capture enough feature information due to the insufficient number of original training samples, and improve the network performance. Fitting efficiency, and then achieve fast convergence and stable network model using less image data.
  • the obtained multi-channel image data is input into the convolutional neural network model as input data, and the labeled data of the original image is used as the target of fitting the convolutional neural network model, and the convolutional neural network model is trained to obtain
  • the parameters of the convolutional neural network model to realize the step of segmenting the image to be tested according to the type of the labeled data include:
  • the network segmentation model structure includes 3 downsampling layers, 3 upsampling layers and 3 jump link layers, and the number of input channels of the network is m+1; among them, the network Start with two 3 ⁇ 3 convolutional layers; each downsampling layer includes: 1 MaxPooling layer and two 3 ⁇ 3 convolutional layers; each upsampling layer includes: 1 upsampling layer and two 3 ⁇ 3 layers
  • the convolutional layer and the upsampling layer are bilinear interpolation; the jump link splices the feature map in the downsampling process and the feature map of the same resolution in the upsampling process in the channel dimension.
  • the up-sampled and down-sampled images of the same resolution are spliced in the channel dimension through skip links, and a better up-sampling structure is obtained, that is, the image is segmented or reconstructed, and the classification task is added to simultaneously Complete classification problems such as cells and blood vessels.
  • the obtained multi-channel image data is input into the convolutional neural network model as input data, and the labeled data of the original image is used as the target of fitting the convolutional neural network model, and the convolutional neural network model is trained to obtain
  • the steps of training a model with a convolutional neural network include:
  • a set of multi-channel image data is input to the multi-input channel U-net network segmentation model, and the labeled data of the original image is used as the multi-input channel U-net network segmentation model to fit.
  • the target is to train the multi-input channel U-net network segmentation model, and obtain the multi-input channel U-net network segmentation model parameters;
  • a multi-input channel U-net network segmentation training model is obtained according to the constructed multi-input channel U-net network segmentation model and the multi-input channel U-net network segmentation model parameters.
  • the manual calibration of the training template can be reduced, the influence on the generalization performance of the network model can be reduced under the condition of fewer training samples, and the fitting efficiency in the training process of the network model can be improved.
  • the obtained multi-channel image data is input into a convolutional neural network training model for fitting, and the step of obtaining the segmented image includes:
  • each test image and its edge-enhanced image are used as a set of multi-channel data to input the multi-input channel U-net network segmentation training model to fit the segmented images.
  • the present application also provides a convolutional neural network-based medical image segmentation data enhancement device, comprising a processor and a memory, the memory stores a convolutional neural network-based medical image segmentation data enhancement device.
  • a program the processor executes the steps of the above method when running the medical image segmentation data enhancement program based on the convolutional neural network.
  • the above-mentioned medical image segmentation data enhancement method is presented in the form of computer-readable codes and stored in the memory.
  • the processor When the computer readable code in the memory is executed in the system, the above steps of data enhancement for medical image segmentation based on the convolutional neural network are performed to obtain the technical effect of improving the generalization performance of the network.
  • the number of training samples can be reduced, and thus the labeling data can be reduced.
  • the workload of manual labeling is reduced, and on the other hand, the applicability of the data segmentation technology to the number of training samples is greatly improved.
  • Increasing the edge information can improve the fitting speed and reduce the time for the model to converge, so that the model can be trained with less data.
  • Fig. 1 is the hardware operating environment frame diagram of the medical image segmentation data enhancement method based on convolutional neural network of the application;
  • FIG. 2 is a flowchart of a medical image segmentation data enhancement method provided by a convolutional neural network according to an embodiment of the present application
  • FIG. 3 is a schematic structural diagram of a network model adopted in the medical image segmentation data enhancement method provided by a convolutional neural network according to an embodiment of the present application;
  • FIG. 4 is a schematic structural diagram of a convolutional neural network model adopted by the related art
  • Fig. 5 is a group of original images and annotated images input when the medical image segmentation method of convolutional neural network in the related art performs network model training;
  • Fig. 6 is the image to be tested that is input to the network model trained in Fig. 5 and the output image after its fitting;
  • 7 is a group of multi-channel data and labeling data input when the medical image segmentation data enhancement method of the convolutional neural network according to an embodiment of the application performs network model training;
  • Fig. 8 is the image to be tested that is input to the network model trained in Fig. 7 and the output image after its fitting;
  • FIG. 9 is a comparison diagram of errors in the training and fitting process of the related art and the method of the present application.
  • the terms “connected”, “fixed” and the like should be understood in a broad sense, for example, “fixed” may be a fixed connection, a detachable connection, or an integrated; It can be a mechanical connection, an electrical connection, a physical connection or a wireless communication connection; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the internal connection of two elements or the interaction between the two elements. unless otherwise expressly qualified.
  • “fixed” may be a fixed connection, a detachable connection, or an integrated; It can be a mechanical connection, an electrical connection, a physical connection or a wireless communication connection; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the internal connection of two elements or the interaction between the two elements. unless otherwise expressly qualified.
  • the specific meanings of the above terms in the present invention can be understood according to specific situations.
  • the present application provides a convolutional neural network medical image segmentation data enhancement device 100 including a collector 101 , a memory 102 and a processor 103 , the collector 1 is used to collect medical images, and the memory storage 2 has convolutional neural network.
  • the medical image segmentation data enhancement program of the neural network the processor 3 executes the following steps of the medical image segmentation data enhancement method of the convolutional neural network when executing the program, and the implementation of the method scheme, that is, the scheme of Embodiment 2, is implemented below in conjunction with the above-mentioned device.
  • the method scheme that is, the scheme of Embodiment 2
  • the present application provides a medical image segmentation data enhancement method of a convolutional neural network, which specifically includes the following steps:
  • the input data is obtained through the collector 101 in the device.
  • the input data is a medical image, for example, a medical image obtained by projecting a human organ by a medical imaging instrument such as CT or B-ultrasound.
  • Effective human tissue or organ information to assist doctors in judging lesions usually requires a mathematical model installed inside the computer to perform complex calculations and processing on the image to obtain effective information.
  • the input data is collected and transmitted to the processor 103.
  • the collected medical image is an original image before processing.
  • the processor 103 performs edge enhancement processing on the original image x1 collected at time t1 through a specific image processing algorithm to obtain For the edge-enhanced original image x 11 , perform edge enhancement processing on the original image x 2 collected at time t 2 through the same specific image processing algorithm to obtain an edge-enhanced original image x 21 , repeat the above edge enhancement processing, and collect the image at time t n
  • the original image x n of x n is subjected to edge enhancement processing through the same specific image processing algorithm to obtain an edge-enhanced original image x n1 .
  • edge enhancement processing can be performed on the acquired original image through a variety of specific image processing algorithms, so as to obtain multiple edge-enhanced original images x i1 , x i2 for any original image x i , ..., x im , where i is the serial number of the collection moment, m and n are positive integers, where m ⁇ n.
  • the above-mentioned image processing algorithm is any image processing algorithm that can realize edge enhancement.
  • the original image and the enhanced image are all images that are cropped into uniform shapes (eg, squares, circles, rectangles, etc.) and sizes according to the requirements of the network model.
  • S130 Input the obtained multi-channel image data as a training sample into a convolutional neural network model, use the labeled data of the original image as a fitting target of the convolutional neural network model, train the convolutional neural network model, and obtain a convolutional neural network training model;
  • the tissue or organ that needs to be segmented in each image in the original image set is manually calibrated to form a labeled image, and a corresponding relationship is established between each labeled image and its referenced original image, and edge enhancement is performed based on the original image to obtain an edge-enhanced image.
  • the label image of the edge-enhanced image is the same as the label image of the original image referenced before processing, and the label images of all the original images form label data; the processor 103 converts the above m+1 channel image data.
  • the convolutional neural network model is input as a training sample, and the labeled data of the original image is used as the fitting target of the convolutional neural network model based on the above correspondence, and the convolutional neural network model is trained, and the volume is obtained when the loss function of the network model converges.
  • the weight and bias of the neural network model parameter neuron, and the network parameters are substituted into the convolutional neural network model to obtain the training model.
  • the testing step includes: testing the image to be tested through a convolutional neural network training model, so as to realize the segmentation of the image to be tested according to the type of the labeled data.
  • the uncalibrated image to be tested is input into the above training model, and tested by the training model, so as to realize the fitting of the segmentation target contour in the image to be tested, and output the calibrated image, that is, the segmented image.
  • the processor 102 performs edge enhancement processing on the original image to obtain samples that are many times larger than the original image, and obtains a sufficient number of training sample sets.
  • the parameters in the model can better reflect the change process from input to output with the training, which can effectively promote the convergence of the evaluation function of the network model, and Quickly obtain stable training results; and in the evaluation mode, the training model formed according to the parameters determined in the training process is tested on the image to be tested, and the calculation can be fitted to obtain results similar to those of the annotation.
  • S200 includes:
  • the specifications of the image to be tested here include parameters such as shape and size, which are the same as those of the original image.
  • S230 Input the obtained multi-channel image data into a convolutional neural network training model for fitting to obtain a segmented image.
  • the m+1 channel data formed based on the image to be tested and its multiple edge-enhanced images is input into the training model for fitting to obtain a segmented image.
  • the step S110 of generating the edge-enhanced image in the training step includes:
  • the above-mentioned image features of the training samples can be expanded to four times the original size.
  • other well-known edge enhancement algorithms can be used to perform edge enhancement processing on the original image to obtain edge-enhanced images.
  • edge enhancement processing on the original image to obtain edge-enhanced images.
  • more edge features can be extracted as Feature information, by adding channels to enhance edge information, improves the generalization performance of the network model when the amount of sample data is insufficient.
  • the convolutional neural network model in S100 is a U-net network segmentation model, and before step S100, S010 is included to construct a U-net network segmentation model with multiple input channels and multiple output channels; see FIG.
  • the network segmentation model structure includes 3 downsampling layers, 3 upsampling layers and 3 jump link layers, and the number of input channels of the network is N; wherein, the network starts with two 3 ⁇ 3 convolutional layers; each Each downsampling layer includes: 1 MaxPooling layer and two 3 ⁇ 3 convolutional layers; each upsampling layer includes: 1 upsampling layer and two 3 ⁇ 3 convolutional layers, and the upsampling layer is bilinear interpolation ; The jump link splices the feature map in the downsampling process with the feature map of the same resolution in the upsampling process in the channel dimension.
  • the step S120 of generating multi-channel image data in the training step includes:
  • V1 train ⁇ x 1 ,x 2 ,x 3 ,...,x n ⁇
  • x i is an image sample of size r ⁇ r, and the number of samples in the original training set is n; i ⁇ [1,n], n is a positive integer; r is the pixel value;
  • V2 train ⁇ x i1 ,...,x im ⁇
  • x ij is an edge-enhanced original image sample of size r ⁇ r, and the number of samples in the enhanced training set is n ⁇ m; i ⁇ [1,n], j ⁇ [1,m], m is a positive integer;
  • the training set V train formed by combining the original training set V1 train and the enhanced training set V2 train is expressed as follows:
  • V train V1 train +V2 train
  • the above scheme does not change the number of samples in the training set of the related art, but generates multiple enhanced images based on each original image of the original training set, and realizes the input of multi-channel image data to make up for the insufficient number of original images. technical defects of the image information.
  • S130 includes:
  • the parameters of the convolutional neural network model include the weight and bias of neurons. In order to fit these parameters, a large amount of sample information is required for network training.
  • S132 Obtain a multi-input channel U-net network segmentation training model according to the constructed multi-input channel U-net network segmentation model and the multi-input channel U-net network segmentation model parameters.
  • the convolutional neural network model is trained through multiple edge-enhanced samples.
  • the edge feature information can provide effective information for the model to classify images and improve the fitting speed.
  • the step S210 of generating an edge-enhanced image in a similar test step includes:
  • the step S220 of generating multi-channel image data in the testing step includes:
  • V1 test ⁇ y 1 ,y 2 ,y 3 ,...,y w ⁇
  • y i is an image sample of size r ⁇ r, and the number of samples in the original test set is w; i ⁇ [1,w], w is a positive integer;
  • V2 test ⁇ y i1 ,...,y im ⁇
  • y ij is an edge-enhanced image sample of size r ⁇ r to be tested, and the number of samples in the enhanced test set is n ⁇ m; i ⁇ [1,n], j ⁇ [1,m], m is a positive integer ; r is the pixel value; ;
  • test set V test formed by merging the original test set V1 test and the enhanced test set V2 test is expressed as follows:
  • V test V1 test + V2 test
  • test image In order to maintain the consistency of incoming and outgoing data and network training, the test image also uses the above edge enhancement algorithm to generate the same number of channel image data, so as to obtain segmented images by fitting the above training model.
  • step S230 includes:
  • each test image and its edge-enhanced image are used as a set of multi-channel data to input a multi-input channel U-net network segmentation training model to perform fitting to obtain a segmented image.
  • FIG. 4 See Figure 4 for the single-input channel cerebrovascular segmentation U-net network model of the architecture, and see Figure 5 for a set of samples in the original training set, including the input original image and the labeled image to be fitted by the network.
  • a set of U-net network sample data the left side is the original image of 512 ⁇ 512 pixels, and the right side is the 512 ⁇ 512 pixel annotated image formed by manually calibrating the blood vessels in the left image, only through the original training
  • the training model is obtained by training the network model in Figure 4. See Figure 6.
  • the image of 512 ⁇ 512 pixels on the left is used as the test image and input to the above training model.
  • the segmentation image obtained by fitting is shown in Figure 6.
  • the four-input channel cerebral blood vessel segmentation U-net network model shown in Figure 3 is a four-input network model.
  • the network model in Figure 4 is a single input channel. Referring to Figure 7, it is a set of samples in the training set, from left to right, the original image of 512 ⁇ 512 pixels in Figure 5, and the Laplacian, Scharr, and Canny edge enhancement algorithms are used to perform edge enhancement processing on the original image in Figure 5.
  • FIG 9 from left to right are the training error fitting speed simulation graph, the test error fitting speed simulation graph, and the training error-test error fitting speed simulation graph, in which the dotted line is the original training set and the original test set respectively.
  • Figure 4 shows the fitting curve of the network model training for training and testing. The solid line is the comparison of the multi-channel image data combined with the original training set and the enhanced training set and the multi-channel image data combined with the original test set and the enhanced test set.
  • Figure 4 It can be seen from the training error simulation diagram that the loss function decline speed of the two networks is not significantly different during training; from the test error fitting speed simulation diagram, it can be seen that the edge enhancement network is performed.
  • the fitting speed is faster; from the training error-test error fitting speed simulation graph, it can be seen that the edge-enhanced network is more stable and has better generalization performance.
  • the edge-enhanced information can reduce the time for model convergence, thus using less data for training out the network model; and the latter has a smaller overall fitting error than the former.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

一种基于卷积神经网络的医学图像分割数据增强方法及装置，该方法包括：对获取的原始图像进行边缘增强处理，获得边缘增强的原始图像（S10）；将原始图像和边缘增强的原始图像合并生成多通道图像数据（S20）；将获得的多通道图像数据作为训练样本输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得卷积神经网络训练模型（S30），通过卷积神经网络训练模型对待测图像进行测试以实现对待测图像按照所述标注数据的类型进行分割（S40）。用于解决相关技术中由于训练样本数量不足导致的网络泛化性能降低，从而导致网络模型测试误差较大甚至失真的问题。

Description

一种基于卷积神经网络的医学图像分割方法及装置 技术领域
本发明涉及医学图像分割技术领域，具体是一种基于卷积神经网络的医学图像分割方法及装置。
背景技术
卷积神经网络(Convolutional Neural Networks)是一种深度学习模型或类似于人工神经网络的多层感知器，常用来分析图像数据。卷积神经网络主使用由卷积层和池化层进行特征提取和图像大小的压缩。下面以U－net网络为例对卷积神经网络在医学图像分割过程中的应用进行说明：
是一种卷积神经网络；是在全卷积神经网络(FCN)的基础上发展出来的；能够有效分割目标图像要素的网络。U－net网络在拟合的过程中，需要计算出每一个神经元的权重(weight)和偏置(bias)两个参数。
为了拟合这些参数，需要大量的样本进行网络的训练。其中每一组样本均包含了输入的原始图像和网络要拟合的标注数据。神经网络的模型具有两种模式：训练模式(Train)和评估模式(Eval)。在训练模式中，将数据一组一组的放入网络模型进行训练，模型中的参数随着训练能越来越好的反映出输入到输出的变化过程。在样本量足够的情况下，网络模型的评价函数(Loss)逐渐收敛时，通常代表网络获得了稳定的训练结果。这时模型的训练就完成了。在评估模式中，模型的参数不再变动。输入未经标注的图像数据，经过训练完成的模型的计算可以得到拟合的结果。
卷积神经网络的训练依赖于大量的图像数据和标注数据，而在数据量不足的情况下，网络的泛化性能会受到影响。
发明内容
本申请提供一种基于卷积神经网络的医学图像分割方法及装置，用于克服现有技术中在训练样本数据量不足的情况下，导致网络的泛化性能会受到 影响；实现在样本量较少的情况下提升模型的准确率并缓解过拟合的产生。
为实现上述目的，第一方面，本申请提供一种基于卷积神经网络的医学图像分割数据增强方法，包括：
对获取的原始图像进行边缘增强处理，获得边缘增强的原始图像；
将原始图像和边缘增强的原始图像合并生成多通道图像数据；
将获得的多通道图像数据作为训练样本输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得训练的卷积神经网络；
通过训练的卷积神经网络对待测图像进行测试以实现对待测图像按照所述标注数据的类型进行分割。
通过采用上述的技术方案，在能作为训练样本的数据量较少、训练样本中能被网络模型提取的有效特征信息不足的情况下，通过在训练样本中增加一部分人工特征及边缘信息可以增加每一个数据样本的特征信息，进而提升网络的泛化性能，实现降低网络拟合需要的样本量的技术效果。
优选的，上述方法还包括：
将待测图像进行边缘增强处理，获得边缘增强的待测图像；
将待测图像和边缘增强的待测图像合并生成多通道图像数据；
将获得的多通道图像数据输入训练获得的卷积神经网络中进行拟合，获得分割图像。
通过采用上述的技术方案，基于要保持出入数据和网络训练时的一致性，将待测图像(未标注，通过训练模型自动生成标注图像)同样进行边缘增强处理，并与待测图像合并生成多通道图像数据输入上一步骤中训练完成的卷积神经网络中进行轮廓拟合，由于卷积神经网络模型的各参数均通过上述训练步骤训练获得，由此可直接按照上述训练过程中标注的类型对待测图像进行分割并输出分割完成的图像；相对于相关技术，由于训练过程中提高了网络模型的泛化性能，对于待测图像的识别性能更高，输出的分割图像与标注图像之间的误差大大减小；此外相对单独将待测图像输入训练的模型中进行轮廓拟合，由于增加了多通道边缘增强图像，有利于模型拟合过程中提取边缘信息，从而能提升测试的拟合速度和准确性。
优选的，所述对获取的原始图像进行边缘增强处理，获得边缘增强图像 的步骤包括：
分别采用Laplacian、Scharr和/或Canny边缘增强算法对获取的原始图像进行边缘增强处理，获得多个边缘增强的原始图像；
所述将待测图像进行边缘增强处理，获得边缘增强的待测图像的步骤包括：
分别采用Laplacian、Scharr和/或Canny边缘增强算法对获取的待测图像进行边缘增强处理，获得多个边缘增强的待测图像。
通过采用上述的技术方案，分别对原始图像或待测图像进行边缘增强处理，能够高效、快速地获得图像清晰的边缘信息，有利于提升图像分割的速度和有效信息特征的输入。
优选的，所述原始图像及待测图像均为脑血管图像或均为心血管图像。
通过采用上述的技术方案，通过实验模拟获得的网络模型泛化性能更为优异，对不同输入数据的稳定性能较佳。
优选的，所述卷积神经网络模型为多输入通道的U－net网络分割模型。
通过采用上述的技术方案，主使用由卷积层和池化层进行特征提取和图像大小的压缩，实现了对医学图像分割的卓越性能，有效分割目标图像要素。
优选的，所述将原始图像和边缘增强的原始图像合并生成多通道图像数据的步骤包括：
获取n个尺寸和形状均相同的原始图像样本x i，n个原始图像样本x i构成的原始训练集V1 train表示如下：
V1 train＝{x 1,x 2,x 3,…,x n}
x i为一个大小为r×r的图像样本，原始训练集的样本个数为n；i∈[1,n]，n为正整数；r为像素值；
对任一原始图像样本x i进行第j次边缘增强获得边缘增强的原始图像样本x ij，对n个原始图像样本x i分别进行m次边缘增强获得的m×n个边缘增强的原始图像样本x ij，m×n个边缘增强的原始图像样本x ij构成的增强训练集/测试集V2 train表示如下：
V2 train＝{x i1,…,x im}
x ij为一个大小为r×r的边缘增强的原始图像样本，增强训练集的样本个数为n×m；i∈[1,n]，j∈[1,m]，m为正整数；
原始训练集V1 train与增强训练集V2 train合并形成的训练集V train表示如下：
V train＝V1 train+V2 train
＝{x 1,x 2,x 3,…,x n}+{x 1,x 2,x 3,…,x n} 1+…+{x 1,x 2,x 3,…,x n} m
＝{(x 1,x 11,x 12,…,x 1m),(x 2,x 21,x 22,…,x 2m),(x 3,x 31,x 32,…,x 3m),…,(x n,x n1,x n2,…,x nm)}
(x i,x i1,x i2…,x im)形成一个m+1通道图像样本，训练集V traint的样本个数为n；训练集V train生成m+1通道图像数据；
所述将待测图像和边缘增强的待测图像合并生成多通道图像数据的步骤包括：
获取n个尺寸和形状均相同的待测图像样本y i，n个待测图像样本y i构成的原始测试集V1 test表示如下：
V1 test＝{y 1,y 2,y 3,…,y w}
y i为一个大小为r×r的图像样本，原始测试集的样本个数为w；i∈[1,w]，w为正整数；
对任一待测样本y i进行第j次边缘增强获得边缘增强的待测图像样本y ij，对w个待测样本y i分别进行m次边缘增强获得的m×n个边缘增强的待测图像样本y ij，m×n个边缘增强的待测图像样本y ij构成的增强测试集V2 test表示如下：
V2 test＝{y i1,…,y im}
y ij为一个大小为r×r的边缘增强的待测图像样本，增强测试集的样本个数为n×m；i∈[1,n]，j∈[1,m]，m为正整数；r为像素值；
原始测试集V1 test与增强测试集V2 test合并形成的测试集V test表示如下：
V test＝V1 test+V2 test
＝{y 1,y 2,y 3,…,y n}+{y 1,y 2,y 3,…,y n} 1+…+{y 1,y 2,y 3,…,y n} m
＝{(y 1,y 11,y 12,…,y 1m),(y 2,y 21,y 22,…,y 2m),(y 3,y 31,y 32,…,y 3m),…,(y n,y n1,y n2,…,y nm)}
(y i,y i1,y i2…,y im)形成一个m+1通道图像样本，测试集V test的样本个数为n；训练集/测试集V test生成m+1通道图像数据。
通过采用上述的技术方案，训练过程中分别将n个多通道数据(x i,x i1,x i2…,x im)各自经不同的通道输入网络模型，并以原始图像样本x i的标注数据 为目标对网络模型进行训练，直到损失函数收敛，最终获得网络模型参数，进而获得训练模型；测试过程中分别将w个多通道数据(x i,x i1,x i2…,x im)各自经不同的通道输入训练模型进行拟合，最终获得分类结果；通过多种边缘增强算法对原始训练集中的每幅原始图像进行处理获得多个边缘增强的原始图像，能够获得样本数量是原始训练集样本数量多倍的增强训练集，将增强训练集与原始训练集中所有样本的总和形成的多通道图像数据作为训练集对卷积神经网络模型进行训练，相对于仅采用原始训练集的样本作为训练集对卷积神经网络模型进行训练，基于每个原始图像样本增加了多个通道的图像边缘特征信息，从而克服在原始训练样本数量不足导致难于抓取到足够的特征信息的缺陷，提升了网络的拟合效率，进而实现使用更少的图像数据快速获得收敛、稳定的网络模型。
优选的，在所述将获得的多通道图像数据作为输入数据输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得卷积神经网络模型参数，以实现对待测图像按照所述标注数据的类型进行分割的步骤之前包括：
构建多输入通道U－net网络分割模型；所述网络分割模型结构包括3个降采样层和3个上采样层以及3个跳链接层，该网络的输入通道数为m+1；其中，网络开始为两个3×3卷积层；每个降采样层包括：1个MaxPooling层和两个3×3卷积层；每个上采样层包括：1个上采样层和两个3×3卷积层，上采样层为双线性插值；跳链接将下采样过程中的特征图与上采样同分辨率的特征图在通道维度拼接。
通过采用上述的技术方案，通过跳链接，将上采样和下采样的同分辨率图像在通道维度拼接，得到了更好的上采样结构，即分割或重构图像，并通过增加分类任务来同时完成细胞、血管等分类问题。
优选的，在所述将获得的多通道图像数据作为输入数据输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得卷积神经网络训练模型的步骤包括：
在训练集中按照原始图像及其边缘增强的原始图像为一组多通道图像数据输入多输入通道U－net网络分割模型，以该原始图像的标注数据作为多输入通道U－net网络分割模型拟合的目标对多输入通道U－net网络分割模型进 行训练，获得多输入通道U－net网络分割模型参数；
根据构建的所述多输入通道U－net网络分割模型和多输入通道U－net网络分割模型参数获得多输入通道U－net网络分割训练模型。
通过采用上述的技术方案，可减少训练样板的人工标定，并且能够在训练样本较少的情况下降低其对网络模型泛化性能的影响，并提升网络模型训练过程中的拟合效率。
优选的，所述将获得的多通道图像数据输入获得卷积神经网络训练模型中进行拟合，获得分割图像的步骤包括：
在测试集中按照每个测试图像及其边缘增强图像做为一组多通道数据输入多输入通道U－net网络分割训练模型进行拟合获得分割图像。
通过采用上述的技术方案，由于将边缘增强图像与待测图像均输入网络训练模型进行测试，有利于训练模型提取更多的特征信息，从而能够提升训练模型拟合拟合的准确性。
为实现上述目的，第二方面，本申请还提供一种基于卷积神经网络的医学图像分割数据增强装置，包括处理器和存储器，所述存储器存储有基于卷积神经网络的医学图像分割数据增强程序；所述处理器在运行所述基于卷积神经网络的医学图像分割数据增强程序时执行上述方法的步骤。
通过采用上述的技术方案，将上述医学图像分割数据增强方法以计算机可读代码的形式呈现并存储于存储器内，基于卷积神经网络的医学图像分割数据增强装置在安装上述计算机程序后，处理器在系统内运行存储器内的计算机可读代码时，执行上述基于卷积神经网络的医学图像分割数据增强的步骤获得提升网络泛化性能的技术效果。
本申请提供的传输方法、采集卡及数据传输系统具有如下综合技术效果：
在医学图像分割过程中，训练样本数量较少，不足以提取更多的特征信息进行网络模型训练时，通过本方案的数据增强方法能获得多通道数据，以弥补上述不足，实现提高网络模型的泛化性能。
通过本方案的数据增强方法，能够降低训练样本的数量，从而也降低了标注数据，一方面降低了人工标注工作量，另一方面大大提高了数据分割技术对训练样本数量的适用性。
通过本方案的数据增强方法，尤其在心脑血管图像的分割中，泛化性能提升显著，训练误差和测试误差降低到5％以下。
增加边缘信息可以提升拟合速度，减少模型收敛的时间，从而使用更少的数据训练出模型。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单的介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，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示出的结构获得其他的附图。
图1为本申请基于卷积神经网络的医学图像分割数据增强方法的硬件运行环境框架图；
图2为本申请一实施例提供的卷积神经网络的医学图像分割数据增强方法的流程图；
图3为本申请一实施例提供的卷积神经网络的医学图像分割数据增强方法中采用的网络模型结构示意图；
图4为相关技术采用的卷积神经网络模型结构示意图；
图5为相关技术中卷积神经网络的医学图像分割方法进行网络模型训练时输入的一组原始图像和标注图像；
图6为向图5训练好的网络模型中输入的待测图像及经其拟合后的输出图像；
图7为本申请一实施例的卷积神经网络的医学图像分割数据增强方法进行网络模型训练时输入的一组多通道数据和标注数据；
图8为向图7训练好的网络模型中输入的待测图像及经其拟合后的输出图像；
图9为相关技术与本申请的方法训练和拟合过程的误差对比图。
本发明目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明的一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
需要说明，本发明实施例中所有方向性指示(诸如上、下、左、右、前、后……)仅用于解释在某一特定姿态(如附图所示)下各部件之间的相对位置关系、运动情况等，如果该特定姿态发生改变时，则该方向性指示也相应地随之改变。
另外，在本发明中如涉及“第一”、“第二”等的描述仅用于描述目的，而不能理解为指示或暗示其相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。在本发明的描述中，“多个”的含义是至少两个，例如两个，三个等，除非另有明确具体的限定。
在本发明中，除非另有明确的规定和限定，术语“连接”、“固定”等应做广义理解，例如，“固定”可以是固定连接，也可以是可拆卸连接，或成一体；可以是机械连接，也可以是电连接，还可以是物理连接或无线通信连接；可以是直接相连，也可以通过中间媒介间接相连，可以是两个元件内部的连通或两个元件的相互作用关系，除非另有明确的限定。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在本发明中的具体含义。
另外，本发明各个实施例之间的技术方案可以相互结合，但是必须是以本领域普通技术人员能够实现为基础，当技术方案的结合出现相互矛盾或无法实现时应当认为这种技术方案的结合不存在，也不在本发明要求的保护范围之内。
实施例一
如图1所示，本申请提供一种卷积神经网络的医学图像分割数据增强装置100包括采集器101、存储器102和处理器103，采集器1用于采集医学图像，存储器存储2有卷积神经网络的医学图像分割数据增强程序，处理器3在执行该程序时执行以下卷积神经网络的医学图像分割数据增强方法的步骤，下面结合上述装置对该方法方案即实施例二的方案的实施进行详细说明。
实施例二
如图2所示，本申请提供一种卷积神经网络的医学图像分割数据增强方法，具体包括以下步骤：
S100，训练步骤，包括：
S110，对获取的原始图像进行边缘增强处理，获得边缘增强的原始图像；
通过装置中的采集器101获取输入数据，本方案中输入数据为医学图像，例如：CT或B超等医学影像仪器对人体器官进行投影获得的医学图像，为了从医学图像中获取更为清楚和有效的人体组织或器官的信息，以协助医生判断病灶，通常需要安装在计算机内部的数学模型对图像进行复杂的计算和处理，以获得有效信息，本方案中采集器101以设定的频率对输入数据进行采集并传输给处理器103，采集的医学图像在没有进行处理前为原始图像，处理器103对t 1时刻采集的原始图像x 1通过一具体的图像处理算法进行边缘增强处理，获得边缘增强的原始图像x 11，对t 2时刻采集的原始图像x 2通过同一具体的图像处理算法进行边缘增强处理，获得边缘增强的原始图像x 21，重复上述边缘增强处理，对t n时刻采集的原始图像x n通过同一具体的图像处理算法进行边缘增强处理，获得边缘增强的原始图像x n1。在本方案的其他实施例中，可通过多种具体的图像处理算法对采集的原始图像进行边缘增强处理，从而针对任一原始图像x i获得多个边缘增强的原始图像x i1，x i2，…，x im，其中i为采集时刻的序号，m，n分别为正整数，其中m＜n。上述图像处理算法为任意可实现边缘增强的图像处理算法。需要说明的是，这里的原始图像及增强图像均为按照网络模型需求裁剪为统一形状(例如正方形、圆形、长方形等)和尺寸的图像。
S120，将原始图像和边缘增强的原始图像合并生成多通道图像数据；
将上述原始图像的集合{x 1,x 2,x 3,…,x n}与边缘增强的原始图像的集合﹛{x 1,x 2,x 3,…,x n} 1，{x 1,x 2,x 3,…,x n} 2，…，{x 1,x 2,x 3,…,x n} m﹜合并，作为训练集输入卷积神经网络模型，(x i,x i1,x i2…,x im)形成一个m+1通道图像样本经过m+1个输入通道卷积神经网络模型。
S130，将获得的多通道图像数据作为训练样本输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得卷积神经网络训练模型；
预先对原始图像集合中每个图像中需要分割的组织或器官进行人工标定形成标注图像，每个标注图像与其参照的原始图像之间建立有对应关系，其中基于原始图像进行边缘增强获得边缘增强图像与该原始图像之间也具有对应关系，边缘增强图像的标注图像与其处理前参照的原始图像的标注图像相同，所有原始图像的标注图像形成标注数据；处理器103将上述m+1通道图像数据作为训练样本输入卷积神经网络模型，基于上述对应关系将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，在网络模型的损失函数收敛时获得卷积神经网络模型参数神经元的权重(weight)和偏置(bias)，将网络参数代入卷积神经网络模型进而获得训练模型。
S200，测试步骤，包括：通过卷积神经网络训练模型对待测图像进行测试以实现对待测图像按照所述标注数据的类型进行分割。
将没有标定的待测图像输入上述训练模型，经训练模型测试，以实现对待测图像中的分割目标轮廓进行拟合，输出标定图像即分割图像。
上述方案中，通过处理器102对原始图像进行边缘增强处理获得比原始图像数量多数倍的样本，获得充足数量的训练样本集，将上述数据以原始图像及其标注图像为单位、以边缘增强图像及其标注图像为单位一组一组的输入网络模型进行训练，模型中的参数随着训练能越来越好的反应出输入到输出的变化过程，能够有效促进网络模型的评价函数收敛，并快速获得稳定的训练结果；并且在评估模式中，按照训练过程确定的参数形成的训练模型对待测图像进行测试，经其计算可以拟合获得与标注相似的结果。
作为实施例二的一个优选方式，S200包括：
S210，将待测图像进行边缘增强处理，获得边缘增强的待测图像；
与上述S110相似，需要说明的是，这里的待测图像的规格包括形状和尺寸等参数与原始图像相同。
S220，将待测图像和边缘增强的待测图像合并生成多通道图像数据；
与上述S120相似，需要说明的是，这里对待测图像进行边缘增强处理的算法与对原始图像进行边缘增强处理的算法相同。
S230，将获得的多通道图像数据输入卷积神经网络训练模型中进行拟合，获得分割图像。
与上述S120相似，基于待测图像及其多个边缘增强图像形成的m+1通道数据输入训练模型中进行拟合，获得分割图像。
作为实施例二的另一个优选方式，训练步骤中边缘增强图像生成的步骤S110包括：
S111，采用Laplacian边缘增强算法对获取的原始图像进行边缘增强处理，获得第一边缘增强的原始图像；
S112，采用Scharr边缘增强算法对获取的原始图像进行边缘增强处理，获得第二边缘增强的原始图像；
S113，采用Canny边缘增强算法对获取的原始图像进行边缘增强处理，获得第三边缘增强的原始图像。
上述可将训练样本的图像特征扩充至原来的四倍，此外还可以采用其他公知的边缘增强算法对原始图像进行边缘增强处理，获得边缘增强图像，在训练过程中可提取更多的边缘特征作为特征信息，通过增加增强边缘信息的通道，提升了在样本数据量不足时网络模型的泛化性能。
作为实施例二的又一个优选方式，S100中的卷积神经网络模型为U－net网络分割模型，在步骤S100之前包括S010，构建多输入通道多输出通道U－net网络分割模型；参见图3，所述网络分割模型结构包括3个降采样层和3个上采样层以及3个跳链接层，该网络的输入通道数为N；其中，网络开始为两个3×3卷积层；每个降采样层包括：1个MaxPooling层和两个3×3卷积层；每个上采样层包括：1个上采样层和两个3×3卷积层，上采样层为双线性插值；跳链接将下采样过程中的特征图与上采样同分辨率的特征图在通道维度拼接。
作为实施例二的再一个优选方式，训练步骤中生成多通道图像数据的步骤S120包括：
S121，生成原始训练集：获取n个尺寸和形状均相同的原始图像样本x i，n个原始图像样本x i构成的原始训练集V1 train表示如下：
V1 train＝{x 1,x 2,x 3,…,x n}
x i为一个大小为r×r的图像样本，原始训练集的样本个数为n；i∈[1,n]，n为正整数；r为像素值；
S122，生成增强训练集：对任一原始图像样本x i进行第j次边缘增强后获 得的边缘增强的原始图像样本x ij，对n个原始图像样本x i对分别进行m次边缘增强后获得的m×n个边缘增强的原始图像样本x ij，m×n个边缘增强的原始图像样本x ij构成的增强训练集V2 train表示如下：
V2 train＝{x i1,…,x im}
x ij为一个大小为r×r的边缘增强的原始图像样本，增强训练集的样本个数为n×m；i∈[1,n]，j∈[1,m]，m为正整数；
S123，生成多通道图像数据：原始训练集V1 train与增强训练集V2 train合并形成的训练集V train表示如下：
V train＝V1 train+V2 train
＝{x 1,x 2,x 3,…,x n}+{x 1,x 2,x 3,…,x n} 1+…+{x 1,x 2,x 3,…,x n} m
＝{(x 1,x 11,x 12,…,x 1m),(x 2,x 21,x 22,…,x 2m),(x 3,x 31,x 32,…,x 3m),…,(x n,x n1,x n2,…,x nm)}
(x i,x i1,x i2…,x im)形成一个m+1通道图像样本，训练集V traint的样本个数为n；训练集V train生成m+1通道图像数据。
上述方案相对于相关技术的训练集的样本数量没有改变，但是基于原始训练集的每个原始图像生成多个增强图像，实现了多通道图像数据的输入，以弥补原始图像数量不足导致不能获取足够的图像信息的技术缺陷。
作为实施例二的再一个优选方式，S130包括：
S131，在训练集中按照原始图像及其边缘增强的原始图像为一组多通道图像数据输入多输入通道U－net网络分割模型，以该原始图像的标注数据作为多输入通道U－net网络分割模型拟合的目标对多输入通道U－net网络分割模型进行训练，获得多输入通道U－net网络分割模型参数；
卷积神经网络模型参数包括神经元的权重(weight)和偏置(bias)，为了拟合这些参数，需要大量的样本信息进行网络的训练。
S132，根据构建的所述多输入通道U－net网络分割模型和多输入通道U－net网络分割模型参数获得多输入通道U－net网络分割训练模型。
上述方案中通过多个边缘增强的样本对卷积神经网络模型进行训练，在下采样过程及上采样过程中，边缘特征信息可以为模型对图像进行分类时提供有效信息从而提升拟合速度
类似的测试步骤中边缘增强图像生成的步骤S210包括：
S211，采用Laplacian边缘增强算法对获取的测试图像进行边缘增强处理，获得第一边缘增强的测试图像；
S212，采用Scharr边缘增强算法对获取的测试图像进行边缘增强处理，获得第二边缘增强的测试图像；
S213，采用Canny边缘增强算法对获取的测试图像进行边缘增强处理，获得第三边缘增强的测试图像。
作为实施例二的再一个优选方式，测试步骤中生成多通道图像数据的步骤S220包括：
S221，生成原始测试集：获取n个尺寸和形状均相同的待测图像样本y i，n个待测图像样本y i构成的原始测试集V1 test表示如下：
V1 test＝{y 1,y 2,y 3,…,y w}
y i为一个大小为r×r的图像样本，原始测试集的样本个数为w；i∈[1,w]，w为正整数；；
S222，生成增强测试集：对任一待测样本y i进行第j次边缘增强后获得的边缘增强的待测图像样本y ij，对w个待测样本y i对分别进行m次边缘增强后获得的m×n个边缘增强的待测图像样本y ij，m×n个边缘增强的待测图像样本y ij构成的增强测试集V2 test表示如下：
V2 test＝{y i1,…,y im}
y ij为一个大小为r×r的边缘增强的待测图像样本，增强测试集的样本个数为n×m；i∈[1,n]，j∈[1,m]，m为正整数；r为像素值；；
S223，生成多通道图像数据：原始测试集V1 test与增强测试集V2 test合并形成的测试集V test表示如下：
V test＝V1 test+V2 test
＝{y 1,y 2,y 3,…,y n}+{y 1,y 2,y 3,…,y n} 1+…+{y 1,y 2,y 3,…,y n} m
＝{(y 1,y 11,y 12,…,y 1m),(y 2,y 21,y 22,…,y 2m),(y 3,y 31,y 32,…,y 3m),…,(y n,y n1,y n2,…,y nm)}
(y i,y i1,y i2…,y im)形成一个m+1通道图像样本，测试集V test的样本个数为n；训练集/测试集V test生成m+1通道图像数据。
基于要保持出入数据和网络训练时的一致性，因此测试图像也采用上述边缘增强算法生成同等数量的通道图像数据，以实现通过上述训练模型进行 拟合获得分割图像。
作为实施例二的再一实施方式，步骤S230包括：
S231，在测试集中按照每个测试图像及其边缘增强图像做为一组多通道数据输入多输入通道U－net网络分割训练模型进行拟合获得分割图像。
参见图4为构架的单输入通道脑血管分割U－net网络模型，参见图5为原始训练集其中一组样本，包含了输入的原始图像和网络要拟合的标注图像，为脑血管分割的U－net网络样本数据中的一组：其中左侧为512×512像素的原始图像，右侧为对左图中的血管进行人工标定后形成的512×512像素的标注图像，仅通过原始训练集对图4的网络模型进行训练获得训练模型，参见图6，四组图像中，将左侧512×512像素的图像作为测试图像输入上述训练模型对进行拟合获得的分割图像参见图6右侧的512×512像素的图像，中间为测试图像的标注图像。
构架图3所示的四输入通道脑血管分割U－net网络模型，需要说明的是图3所示的网络模型与图4所示的网络模型的区别仅在于：图3的网络模型为四输入通道，图4的网络模型为单输入通道。参见图7，为训练集其中一组样本，从左到右依次为图5中512×512像素的原始图像、以及分别采用Laplacian、Scharr、Canny边缘增强算法对图5中原始图像进行边缘增强处理获得三幅512×512像素的边缘增强图像及原始图像的标注图像；通过本方案的方法将原始训练集及边缘增强训练集的总和作为训练集形成四通道图像数据对图3网络模型进行训练获得训练模型，参见图8，四组测试图像中，将左侧图像作为测试图像输入上述训练模型对进行拟合获得的分割图像参见图8右侧512×512像素的图像，中间为测试图像的标注图像。
参见图9，自左至右分别为训练误差拟合速度仿真图、测试误差拟合速度仿真图、训练误差-测试误差拟合速度仿真图，其中虚线为采用原始训练集和原始测试集分别对图4的网络模型训练进行训练和测试的拟合曲线，实线为采用原始训练集与增强训练集合并的多通道图像数据和原始测试集与增强测试集合并的多通道图像数据分别对图4的网络模型进行训练和测试的拟合曲线，由训练误差仿真图可以看出两个网络在训练时损失函数下降速度没有明显区别；由测试误差拟合速度仿真图可以看出进行边缘增强的网络拟合速度更快；由训练误差-测试误差拟合速度仿真图可以看出边缘增强的网络更加稳 定，泛化性能更好，边缘增强信息可减少模型收敛的时间，从而使用更少的数据训练出网络模型；并且后者相对前者在整体上的拟合误差更小。
以上所述仅为本发明的优选实施例，并非因此限制本发明的专利范围，凡是在本发明的构思下，利用本发明说明书及附图内容所作的等效结构变换，或直接/间接运用在其他相关的技术领域均包括在本发明的专利保护范围内。

Claims (10)

  1. 一种基于卷积神经网络的医学图像分割数据增强方法，其特征在于，包括：
    对获取的原始图像进行边缘增强处理，获得边缘增强的原始图像；
    将原始图像和边缘增强的原始图像合并生成多通道图像数据；
    将获得的多通道图像数据作为训练样本输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得卷积神经网络训练模型；
    通过卷积神经网络训练模型对待测图像进行测试以实现对待测图像按照所述标注数据的类型进行分割。
  2. 根据权利要求1所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，所述通过卷积神经网络训练模型对待测图像进行测试以实现对待测图像按照所述标注数据的类型进行分割的步骤包括：
    将待测图像进行边缘增强处理，获得边缘增强的待测图像；
    将待测图像和边缘增强的待测图像合并生成多通道图像数据；
    将获得的多通道图像数据输入卷积神经网络训练模型中进行拟合，获得分割图像。
  3. 根据权利要求2所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，所述对获取的原始图像进行边缘增强处理，获得边缘增强图像的步骤包括：
    分别采用Laplacian、Scharr和/或Canny边缘增强算法对获取的原始图像进行边缘增强处理，获得多个边缘增强的原始图像；
    所述将待测图像进行边缘增强处理，获得边缘增强的待测图像的步骤包括：
    分别采用Laplacian、Scharr和/或Canny边缘增强算法对获取的待测图像进行边缘增强处理，获得多个边缘增强的待测图像。
  4. 根据权利要求2所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，所述原始图像及待测图像均为脑血管图像或均为心血管图像。
  5. 根据权利要求2所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，所述卷积神经网络模型为多输入通道的U－net网络分割模型。
  6. 根据权利要求5所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，所述将原始图像待测图像和边缘增强原始图像合并生成多通道图像数据的步骤包括：
    获取n个尺寸和形状均相同的原始图像样本x i，n个原始图像样本x i构成的原始训练集V1 train表示如下：
    V1 train＝{x 1,x 2,x 3,…,x n}
    x i为一个大小为r×r的图像样本，原始训练集的样本个数为n；i∈[1,n]，n为正整数；r为像素值；
    对任一原始图像样本x i进行第j次边缘增强获得的边缘增强的原始图像样本x ij，对n个原始图像样本x i分别进行m次边缘增强获得m×n个边缘增强的原始图像样本x ij，m×n个边缘增强的原始图像样本x ij构成的增强训练集V2 train表示如下：
    V2 train＝{x i1,…,x im}
    x ij为一个大小为r×r的边缘增强的原始图像样本，增强训练集的样本个数为n×m；i∈[1,n]，j∈[1,m]，m为正整数；
    原始训练集V1 train与增强训练集V2 train合并形成的训练集V train表示如下：
    V train＝V1 train+V2 train
    ＝{x 1,x 2,x 3,…,x n}+{x 1,x 2,x 3,…,x n} 1+…+{x 1,x 2,x 3,…,x n} m
    ＝{(x 1,x 11,x 12,…,x 1m),(x 2,x 21,x 22,…,x 2m),(x 3,x 31,x 32,…,x 3m),…,(x n,x n1,x n2,…,x nm)}
    (x i,x i1,x i2…,x im)形成一个m+1通道图像样本，训练集V traint的样本个数为n；训练集V train生成m+1通道图像数据；
    所述将待测图像和边缘增强的待测图像合并生成多通道图像数据的步骤包括：
    获取n个尺寸和形状均相同的待测图像样本y i，n个待测图像样本y i构成的原始测试集V1 test表示如下：
    V1 test＝{y 1,y 2,y 3,…,y w}
    y i为一个大小为r×r的图像样本，原始测试集的样本个数为w；i∈[1,w]，w为正整数；
    对任一待测样本y i进行第j次边缘增强获得的边缘增强的待测图像样本y ij，对w个待测样本y i分别进行m次边缘增强获得m×n个边缘增强的待测图像样本y ij，m×n个边缘增强的待测图像样本y ij构成的增强测试集V2 test表示如下：
    V2 test＝{y i1,…,y im}
    y ij为一个大小为r×r的边缘增强的待测图像样本，增强测试集的样本个数为n×m；i∈[1,n]，j∈[1,m]，m为正整数；r为像素值；
    原始测试集V1 test与增强测试集V2 test合并形成的测试集V test表示如下：
    V test＝V1 test+V2 test
    ＝{y 1,y 2,y 3,…,y n}+{y 1,y 2,y 3,…,y n} 1+…+{y 1,y 2,y 3,…,y n} m
    ＝{(y 1,y 11,y 12,…,y 1m),(y 2,y 21,y 22,…,y 2m),(y 3,y 31,y 32,…,y 3m),…,(y n,y n1,y n2,…,y nm)}
    (y i,y i1,y i2…,y im)形成一个m+1通道图像样本，测试集V test的样本个数为n；训练集/测试集V test生成m+1通道图像数据。
  7. 根据权利要求6所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，在所述将获得的多通道图像数据作为输入数据输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得卷积神经网络训练模型的步骤之前包括：
    构建多输入通道U－net网络分割模型；所述网络分割模型结构包括3个降采样层和3个上采样层以及3个跳链接层，该网络的输入通道数为m+1；其中，网络开始为两个3×3卷积层；每个降采样层包括：1个MaxPooling层和两个3×3卷积层；每个上采样层包括：1个上采样层和两个3×3卷积层，上采样层为双线性插值；跳链接将下采样过程中的特征图与上采样同分辨率的特征图在通道维度拼接。
  8. 根据权利要求7所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，在所述将获得的多通道图像数据作为输入数据输入卷积神经网络模型，将原始图像的标注数据作为卷积神经网络模型拟合的目标，对卷积神经网络模型进行训练，获得卷积神经网络训练模型的步骤包括：
    在训练集中按照原始图像及其边缘增强的原始图像为一组多通道图像数据输入多输入通道U－net网络分割模型，以该原始图像的标注数据作为多输入通道U－net网络分割模型拟合的目标对多输入通道U－net网络分割模型进行训练，获得多输入通道U－net网络分割模型参数；
    根据构建的所述多输入通道U－net网络分割模型和多输入通道U－net网络分割模型参数获得多输入通道U－net网络分割训练模型。
  9. 根据权利要求8所述的基于卷积神经网络的医学图像分割数据增强方法，其特征在于，所述将获得的多通道图像数据输入获得卷积神经网络训练模型中进行拟合，获得分割图像的步骤包括：
    在测试集中按照每个测试图像及其边缘增强图像做为一组多通道数据输入多输入通道U－net网络分割训练模型进行拟合获得分割图像。
  10. 一种基于卷积神经网络的医学图像分割数据增强装置，其特征在于，包括：
    存储器，存储有基于卷积神经网络的医学图像分割数据增强程序；
    处理器，在运行所述基于卷积神经网络的医学图像分割数据增强程序时执行权利要求1～9任一项所述方法的步骤。
PCT/CN2020/135348 2020-12-10 2020-12-10 一种基于卷积神经网络的医学图像分割方法及装置 WO2022120739A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/135348 WO2022120739A1 (zh) 2020-12-10 2020-12-10 一种基于卷积神经网络的医学图像分割方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/135348 WO2022120739A1 (zh) 2020-12-10 2020-12-10 一种基于卷积神经网络的医学图像分割方法及装置

Publications (1)

Publication Number Publication Date
WO2022120739A1 true WO2022120739A1 (zh) 2022-06-16

Family

ID=81972996

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/135348 WO2022120739A1 (zh) 2020-12-10 2020-12-10 一种基于卷积神经网络的医学图像分割方法及装置

Country Status (1)

Country Link
WO (1) WO2022120739A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170109880A1 (en) * 2015-10-16 2017-04-20 General Electric Company System and method for blood vessel analysis and quantification in highly multiplexed fluorescence imaging
WO2017067868A1 (en) * 2015-10-19 2017-04-27 Ventana Medical Systems, Inc. Performing segmentation of cells and nuclei in multichannel images
US20180374209A1 (en) * 2017-06-27 2018-12-27 General Electric Company Material segmentation in image volumes
US10304193B1 (en) * 2018-08-17 2019-05-28 12 Sigma Technologies Image segmentation and object detection using fully convolutional neural network
CN109840913A (zh) * 2019-01-21 2019-06-04 中南民族大学 一种乳腺x线图像中肿块分割的方法和系统
CN110570431A (zh) * 2019-09-18 2019-12-13 东北大学 一种基于改进的卷积神经网络的医学图像分割方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170109880A1 (en) * 2015-10-16 2017-04-20 General Electric Company System and method for blood vessel analysis and quantification in highly multiplexed fluorescence imaging
WO2017067868A1 (en) * 2015-10-19 2017-04-27 Ventana Medical Systems, Inc. Performing segmentation of cells and nuclei in multichannel images
US20180374209A1 (en) * 2017-06-27 2018-12-27 General Electric Company Material segmentation in image volumes
US10304193B1 (en) * 2018-08-17 2019-05-28 12 Sigma Technologies Image segmentation and object detection using fully convolutional neural network
CN109840913A (zh) * 2019-01-21 2019-06-04 中南民族大学 一种乳腺x线图像中肿块分割的方法和系统
CN110570431A (zh) * 2019-09-18 2019-12-13 东北大学 一种基于改进的卷积神经网络的医学图像分割方法

Similar Documents

Publication Publication Date Title
WO2021008328A1 (zh) 图像处理方法、装置、终端及存储介质
CN108898175B (zh) 基于深度学习胃癌病理切片的计算机辅助模型构建方法
CN112132817B (zh) 一种混合注意力机制的眼底图像视网膜血管分割方法
WO2022188489A1 (zh) 多模态多病种长尾分布眼科疾病分类模型训练方法和装置
EP4102455A1 (en) Multi-modal medical image processing
CN109949276B (zh) 一种于改进SegNet分割网络的淋巴结检测方法
CN109409416A (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
CN114581662B (zh) 一种脑肿瘤图像的分割方法、系统、装置及存储介质
CN107766874B (zh) 一种超声容积生物学参数的测量方法及测量系统
CN113888412B (zh) 一种用于糖尿病视网膜病变分类的图像超分辨率重建方法
WO2021136368A1 (zh) 钼靶图像中胸大肌区域自动检测方法及装置
CN113298780B (zh) 一种基于深度学习的儿童骨龄评估方法及系统
WO2024087359A1 (zh) 用于内窥镜的病灶检测方法、装置、电子设备及存储介质
CN112926537B (zh) 图像处理方法、装置、电子设备以及存储介质
CN112348059A (zh) 基于深度学习的多种染色病理图像分类方法及系统
CN115512110A (zh) 一种涉及跨模态注意力机制的医学图像肿瘤分割方法
CN112396605A (zh) 网络训练方法及装置、图像识别方法和电子设备
CN113269774B (zh) 一种mri图像的帕金森病分类及标注病灶区域的方法
CN115601299A (zh) 基于图像的肝硬化状态智能评估系统及其方法
CN112489029A (zh) 一种基于卷积神经网络的医学图像分割方法及装置
Wang et al. SERR-U-Net: squeeze-and-excitation residual and recurrent block-based U-Net for automatic vessel segmentation in retinal image
CN113362360B (zh) 基于流体速度场的超声颈动脉斑块分割方法
WO2022120739A1 (zh) 一种基于卷积神经网络的医学图像分割方法及装置
CN111626379A (zh) 肺炎x光图像检测方法
CN111209946A (zh) 三维图像处理方法、图像处理模型训练方法及介质

Legal Events

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

Ref document number: 20964677

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: 20964677

Country of ref document: EP

Kind code of ref document: A1