WO2023143628A1 - 基于遗传模糊树的视网膜糖尿病变深度网络检测方法 - Google Patents

基于遗传模糊树的视网膜糖尿病变深度网络检测方法 Download PDF

Info

Publication number
WO2023143628A1
WO2023143628A1 PCT/CN2023/077768 CN2023077768W WO2023143628A1 WO 2023143628 A1 WO2023143628 A1 WO 2023143628A1 CN 2023077768 W CN2023077768 W CN 2023077768W WO 2023143628 A1 WO2023143628 A1 WO 2023143628A1
Authority
WO
WIPO (PCT)
Prior art keywords
fuzzy
node
retinal
image
tree
Prior art date
Application number
PCT/CN2023/077768
Other languages
English (en)
French (fr)
Inventor
丁卫平
王海鹏
鞠恒荣
刘传升
耿宇
黄嘉爽
程纯
曹金鑫
秦廷桢
沈鑫杰
潘柏儒
Original Assignee
南通大学
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 南通大学 filed Critical 南通大学
Publication of WO2023143628A1 publication Critical patent/WO2023143628A1/zh

Links

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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/12—Computing arrangements based on biological models using genetic models
    • G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/048—Fuzzy inferencing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/90—Dynamic range modification of images or parts thereof
    • G06T5/92—Dynamic range modification of images or parts thereof based on global image properties
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G06—COMPUTING; 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
    • G06T2207/30041—Eye; Retina; Ophthalmic
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Definitions

  • the invention relates to the technical field of medical information intelligent processing, in particular to a method for detecting retinal diabetes variable depth network based on genetic fuzzy tree.
  • Retinal images store vascular information that is closely related to ophthalmological blindness diseases.
  • the health of retinal blood vessels is of great significance for doctors to diagnose diabetic cardiovascular and cerebrovascular diseases and various ophthalmic diseases in an early stage.
  • clinical manual segmentation of retinal blood vessels is not only a huge workload but also requires a high level of experience and skills for medical personnel.
  • different medical personnel may have different segmentation results on the same retinal image, and manual segmentation can no longer meet clinical needs.
  • the purpose of the present invention is to solve the above problems, and proposes a method for detecting retinal diabetes variable depth network based on genetic fuzzy tree.
  • the deep network detection method of retinal diabetes based on genetic fuzzy tree comprises the following steps:
  • the network model U-net is divided into a compression path and an expansion path. There are 4 times of downsampling and 4 times of upsampling. Before each sampling, two convolutions and one maximum pooling are performed. In the path, 4 times of downsampling are used to compress the features of the retinal image, and the extended path performs 4 times of upsampling on the effective feature layer obtained by the last downsampling, and at the same time connect the corresponding feature layers in the downsampling, and finally use 1*1 volume
  • the product normalizes the retinal feature map Feature Map, and uses the enhanced image data to train the constructed model to obtain an image segmentation model;
  • S3 Perform fuzzy processing on the blood vessel images segmented by the model, calculate the membership degree of the attributes in the image and the fuzzy information gain, and conduct training with the real diagnosis results to obtain the decision rules of the branch nodes of the fuzzy decision tree and the result set of the leaf nodes. Make classification predictions;
  • S4 Encode each node of the decision tree, construct a fitness function at the same time, measure the pros and cons of the fuzzy tree model from the two aspects of accuracy and complexity, use the accurate function E to represent the accuracy of the model, the smaller the E, the higher the accuracy High, the complexity is reflected by the number M of leaf nodes of the tree. The smaller M is, the lower the complexity of the model is.
  • the fitness function s(T) is defined as follows:
  • step S2 the specific steps of the step S2 are as follows:
  • Step S2.1 Divide the retinal lesion data set into training set and verification set according to 9:1 and input them into the training network;
  • Step S2.2 Build the compression path of the network model U-net.
  • the compression path performs 4 downsampling on the retinal image to obtain 5 preliminary effective feature layers.
  • Each preliminary effective feature layer is a stack of convolution and maximum pooling.
  • Each convolution discards the edge information of the image, and performs 2*2 maximum pooling on the 561*580 retinal image obtained by convolution.
  • the channel number of the retinal feature map Feature Map is doubled for each downsampling;
  • Step S2.3 Build the extension path of the network model U-net.
  • the extension path includes 4 times of upsampling, and each upsampling reduces the number of retinal feature map Feature Map channels in the previous layer to half through 2*2 deconvolution , the length and width of the image are doubled, and the corresponding feature layers in the downsampling will be connected during upsampling. Since the edge information of the image is lost during convolution, appropriate cropping is used during connection to ensure that the image size after connection is consistent.
  • U- The net network finally uses 1*1 convolution to normalize the retinal feature map Feature Map;
  • Step S2.4 The loss function uses cross entropy and SoftMax. For the category of each pixel of the retinal lesion image, the predicted probability of belonging to the lesion area and normal area is p and 1-p respectively.
  • the SoftMax function of the pixel form is:
  • ak(x) represents the activation value of the kth layer of the pixel point x in the feature map
  • K is the number of classes
  • p k (x) is the classification result of the class k for the pixel point x
  • the cross-entropy loss function E is defined as:
  • ⁇ ⁇ 1,...,K ⁇
  • l(x) is the real label of each pixel x
  • p l(x) (x) is the classification result of the real label
  • w(x) is the pixel
  • the weight map of x distinguishes the weight of each pixel, and its weight calculation formula is as follows:
  • w c (x) is a weight map used to balance a certain type of frequency, sorted according to the distance from pixel point x to the border of retinal lesions from near to far, d 1 (x) represents the distance of the first ranking, d 2 ( x) represents the distance of the second ranking, w 0 is the initial value of the weight, its value is 10, and the standard deviation ⁇ is set to 5;
  • Step S2.5 Use the stochastic gradient descent of the convolutional neural network framework Caffe to train and optimize the network, and train the built model with the goal of minimizing the loss function and maximizing the prediction accuracy.
  • step S3 the specific steps of the step S3 are as follows:
  • Step S3.1 Fuzzy the lesion area of the retinal blood vessel, especially for the edge area of the lesion, to obtain the membership degree of the continuous value attribute of the image.
  • the attribute value is a membership degree of a [0,1] interval, which is more natural and Reasonably describe the imprecise information of the edge of the lesion area;
  • c is the mean value of the normal distribution
  • ⁇ is the standard deviation of the normal distribution
  • Step S3.5 Construct child nodes according to the attribute value of the extended attribute of the node, and recursively process each child nodes.
  • step S4 the specific steps of the step S4 are as follows:
  • Step S4.2 Measure the pros and cons of the fuzzy tree model from the two aspects of the accuracy and complexity of predicting retinopathy, use the accuracy function E to represent the accuracy of the model, the smaller the E, the higher the accuracy, that is, the accuracy of predicting retinopathy The higher the rate, the complexity is reflected by the number M of leaf nodes of the tree. The smaller M is, the lower the complexity of the model is. Therefore, the fitness function s(T) is:
  • the fitness of each retinal lesion fuzzy tree is calculated, that is, the population is initialized;
  • Step S4.3 According to the fitness of each fuzzy tree, use the roulette method to select a pair of parent individuals.
  • the probability of each fuzzy tree being selected is proportional to its fitness value. Let the total number of individuals be N, some
  • the fitness value of an individual x i is expressed as f(xi )
  • the probability of the individual being selected is p(xi )
  • the cumulative probability is q(xi )
  • the corresponding calculation formula is as follows:
  • the cumulative probability q( xi ) represents the sum of the selection probabilities of all individuals before the individual, which is equivalent to the range turned on the roulette wheel, the larger the range, the easier it is to select;
  • Step S4.4 Randomly select an intersection point k for the parent individual, perform an intersection operation with probability P c to generate a new individual, and randomly generate a mutation probability P on [0,1] for each gene of the new individual m , to mutate;
  • Step S4.5 Recalculate the fitness of the new individual in the environment, compare it with the optimal value, and update the population.
  • the generation of individuals with the highest fitness is obtained, that is, the optimal fuzzy decision tree is obtained.
  • the present invention considers the characteristics of medical images, integrates the genetic fuzzy tree in the deep learning network, and more naturally and reasonably describes the inaccurate information of the edge of the retinal vascular lesion area, so that the accuracy and interpretability of the model results are enhanced , even when the data scale is limited, better results can be obtained.
  • the accuracy rate index is introduced, and the penalty item in the loss function is dynamically adjusted according to the distance between the sample category and the true value to further improve the classification accuracy.
  • a method for detecting retinal diabetes variable depth network based on genetic fuzzy tree of the present invention can accurately segment the blood vessel image in the retina, more accurately identify the end of the blood vessel, and analyze and classify the extracted retinal features through the fuzzy decision tree , can improve the accuracy of detection, and can strengthen the interpretability of diagnostic results, effectively improving the reliability of diagnostic results, using genetic algorithms to optimize the structure of fuzzy trees, further improving the accuracy of detection, and effectively Help doctors diagnose retinal diabetes and allow patients to obtain the best treatment period.
  • Fig. 1 shows the flow chart of a kind of retinal diabetes deep network detection method based on genetic fuzzy tree of the present invention
  • Fig. 2 shows the flowchart of generating fuzzy decision tree for predicting the presence of lesions in the segmented retinal image according to the present invention
  • Fig. 4 shows the retinal blood vessel diagram that the network model U-net of the present invention divides, and the left picture is the original The original image, the label mask in the middle, and the segmented image on the right;
  • Fig. 5 shows that the present invention sets up one of branch of fuzzy decision tree
  • Fig. 6 is a flow chart showing the optimization of the generated fuzzy decision tree using genetic algorithm in the present invention.
  • the present invention discloses a method for detecting retinal diabetes mellitus in depth network based on a genetic fuzzy tree, which belongs to the field of intelligent processing of medical information.
  • the present invention comprises the following steps:
  • the network model U-net is divided into a compression path and an expansion path. There are 4 times of downsampling and 4 times of upsampling. Before each sampling, two convolutions and one maximum pooling are performed. In the path, 4 times of downsampling are used to compress the features of the retinal image, and the extended path performs 4 times of upsampling on the effective feature layer obtained by the last downsampling, and at the same time connect the corresponding feature layers in the downsampling, and finally use 1*1 volume Normalize the retinal feature map Feature Map, train the built model with the enhanced image data, and obtain the image segmentation model;
  • S3 Perform fuzzy processing on the vascular image segmented by the model, calculate the membership degree and fuzzy information gain of the attributes in the fuzzy vascular image, train with the real diagnosis results, and obtain the decision rules and leaves of the branch nodes of the fuzzy decision tree The result set of the node is used for classification prediction;
  • S4 Encode each node of the decision tree, construct a fitness function at the same time, measure the pros and cons of the fuzzy tree model from the two aspects of accuracy and complexity, use the accurate function E to represent the accuracy of the model, the smaller the E, the higher the accuracy High, the complexity is reflected by the number M of leaf nodes of the tree. The smaller M is, the lower the complexity of the model is.
  • the fitness function s(T) is defined as follows:
  • step S2 The specific steps of step S2 are as follows:
  • Step S2.1 Divide the retinal lesion data set into training set and verification set according to 9:1 and input them into the training network;
  • Step S2.2 Build the compression path of the network model U-net.
  • the compression path performs 4 downsampling on the retinal image to obtain 5 preliminary effective feature layers.
  • Each preliminary effective feature layer is a stack of convolution and maximum pooling.
  • Each convolution discards the edge information of the image, and performs 2*2 maximum pooling on the 561*580 retinal image obtained by convolution.
  • the channel number of the retinal feature map Feature Map is doubled for each downsampling;
  • Step S2.3 Build the extension path of the network model U-net.
  • the extension path includes 4 times of upsampling, and each upsampling reduces the number of retinal feature map Feature Map channels in the previous layer to half through 2*2 deconvolution , the length and width of the image are doubled, and the corresponding feature layers in the downsampling will be connected during upsampling. Since the edge information of the image is lost during convolution, appropriate cropping is used during connection to ensure that the image size after connection is consistent.
  • U- The net network finally uses 1*1 convolution to normalize the retinal feature map Feature Map;
  • Step S2.4 The loss function uses cross entropy and SoftMax. For the category of each pixel of the retinal lesion image, the predicted probability of belonging to the lesion area and normal area is p and 1-p respectively.
  • the SoftMax function of the pixel form is:
  • ak(x) represents the activation value of the kth layer of the pixel point x in the feature map
  • K is the number of classes
  • p k (x) is the classification result of the class k for the pixel point x
  • the cross-entropy loss function E is defined as:
  • ⁇ ⁇ 1,...,K ⁇
  • l(x) is the real label of each pixel x
  • p l(x) (x) is the classification result of the real label
  • w(x) is the pixel
  • the weight map of x distinguishes the weight of each pixel, and its weight calculation formula is as follows:
  • w c (x) is a weight map used to balance a certain type of frequency, sorted according to the distance from pixel point x to the border of retinal lesions from near to far, d 1 (x) represents the distance of the first ranking, d 2 ( x) represents the distance of the second ranking, w 0 is the initial value of the weight, its value is 10, and the standard deviation ⁇ is set to 5;
  • Step S2.5 Use the stochastic gradient descent of the convolutional neural network framework Caffe to train and optimize the network, and train the built model with the goal of minimizing the loss function and maximizing the prediction accuracy.
  • step S3 The specific steps of step S3 are as follows:
  • Step S3.1 Fuzzy the lesion area of the retinal blood vessel, especially for the edge area of the lesion, to obtain the membership degree of the continuous value attribute of the image.
  • the attribute value is a membership degree of a [0,1] interval, which is more natural and Reasonably describe the imprecise information of the edge of the lesion area;
  • c is the mean value of the normal distribution
  • ⁇ is the standard deviation of the normal distribution
  • Step S3.3 According to the attributes of the parent node and the training set corresponding to the parent node, and the section Click the attribute value on the attribute of the parent node to construct the fuzzy subclass set A m corresponding to the node.
  • Step S3.5 Construct child nodes according to the attribute value of the extended attribute of the node, and recursively process each child node.
  • step S4 The concrete steps of step S4 are as follows:
  • Step S4.2 Measure the pros and cons of the fuzzy tree model from the two aspects of the accuracy and complexity of predicting retinopathy, use the accuracy function E to represent the accuracy of the model, the smaller the E, the higher the accuracy, that is, the accuracy of predicting retinopathy The higher the rate, the complexity is reflected by the number M of leaf nodes of the tree. The smaller M is, the lower the complexity of the model is. Therefore, the fitness function s(T) is:
  • the fitness of each retinal lesion fuzzy tree is calculated, that is, the population is initialized;
  • Step S4.3 According to the fitness of each fuzzy tree, use the roulette method to select a pair of parent individuals.
  • the probability of each fuzzy tree being selected is proportional to its fitness value. Let the total number of individuals be N, some
  • the fitness value of an individual x i is expressed as f(xi )
  • the probability of the individual being selected is p(xi )
  • the cumulative probability is q(xi )
  • the corresponding calculation formula is as follows:
  • the cumulative probability q( xi ) represents the sum of the selection probabilities of all individuals before the individual, which is equivalent to the range turned on the roulette wheel, the larger the range, the easier it is to select;
  • Step S4.4 Randomly select an intersection point k for the parent individual, perform an intersection operation with probability P c to generate a new individual, and randomly generate a mutation probability P on [0,1] for each gene of the new individual m , to mutate;
  • Step S4.5 Recalculate the fitness of the new individual in the environment, compare it with the optimal value, and update the population.
  • the maximum evolutionary number T 150 or the fitness of the optimal individual and the fitness of the population do not increase for 10 consecutive generations
  • the generation of individuals with the highest fitness is obtained, that is, the optimal fuzzy decision tree is obtained.
  • the present invention considers the characteristics of medical images, integrates the genetic fuzzy tree in the deep learning network, and more naturally and reasonably describes the inaccurate information of the edge of the retinal vascular lesion area, making the model structure
  • the accuracy and interpretability of the results are enhanced, and better results can be obtained even when the data scale is limited.
  • the accuracy index is introduced, and the penalty item in the loss function is dynamically adjusted according to the distance between the sample category and the true value, and further Improve classification accuracy.
  • a method for detecting retinal diabetes variable depth network based on genetic fuzzy tree of the present invention can accurately segment the blood vessel image in the retina, identify the end of the blood vessel more accurately, analyze and classify the extracted retinal features through the fuzzy decision tree, and can Improve the accuracy of detection, and strengthen the interpretability of diagnostic results, effectively improve the reliability of diagnostic results, use genetic algorithm to optimize the structure of the fuzzy tree, further improve the accuracy of detection, and can effectively help doctors Diagnosis of retinal diabetes allows patients to obtain the best treatment period.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Pathology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Automation & Control Theory (AREA)
  • Fuzzy Systems (AREA)
  • Quality & Reliability (AREA)
  • Probability & Statistics with Applications (AREA)
  • Databases & Information Systems (AREA)
  • Physiology (AREA)
  • Genetics & Genomics (AREA)
  • Image Analysis (AREA)

Abstract

本发明涉及医学信息智能处理技术领域，具体涉及基于遗传模糊树的视网膜糖尿病变深度网络检测方法。首先对视网膜图像进行增强处理，将病变区域展宽，对正常区域进行压缩；然后搭建网络模型U-net，准确分割出视网膜血管以及血管末梢图像；接着将模型分割出的血管图像与真实诊断结果进行训练，构建出可解释的模糊决策树；其次对决策树权值编码并且构造适应度函数，基于遗传算法对多棵决策树进行组合优化；最后引入准确率指标动态调整损失函数中的惩罚项。本发明的有益效果是可精确地识别出视网膜糖尿病变血管末梢，提高检测分类准确度，更有效地帮助医生诊断视网膜糖尿病变，让患者获得最佳治疗时期。

Description

基于遗传模糊树的视网膜糖尿病变深度网络检测方法
本申请要求于2022年01月26日提交中国专利局、申请号为202210094089.7、发明名称为“基于遗传模糊树的视网膜糖尿病变深度网络检测方法”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及医学信息智能处理技术领域，尤其涉及基于遗传模糊树的视网膜糖尿病变深度网络检测方法。
背景技术
视网膜图像中保存着和眼科致盲疾病有重要联系的血管信息，视网膜血管的健康状况对于医生及早诊断糖尿病心脑血管疾病及多种眼科疾病具有重要意义。但是由于视网膜血管自身结构复杂，同时易受采集环境中光照因素的影响，使得临床上手动分割视网膜血管不仅工作量巨大而且对医疗人员的经验和技能要求颇高。另外不同的医疗人员对同一幅视网膜图像的分割结果可能存在差异，手动分割已不能满足临床的需要。
随着计算机技术的不断发展，利用人工智能技术对视网膜电子病历血管图像进行自动分割，以对眼科疾病进行辅助诊断和决策，成为了国内外学者关注的研究热点。深度学习凭借其在识别应用中超高的预测准确率，在图像处理领域获得了极大关注，深度学习中的卷积神经网络模型凭借其局部感知、参数共享的特殊结构在图像处理方面有着独特的优越性。本专利从深度学习、模糊决策树两个角度对视网膜图像数据进行分析和处理，对视网膜病变图像进行分割和预测。
发明内容
本发明的目的是解决上述问题，提出了基于遗传模糊树的视网膜糖尿病变深度网络检测方法。
为了实现上述目的，本发明采用了如下技术方案：
基于遗传模糊树的视网膜糖尿病变深度网络检测方法，包括以下步骤：
S1：将视网膜图像进行增强处理，视网膜图像的病变区域与周围正常区域会呈现出视觉上较为明显的不同特征，形成为不同图像区域，采用图像增强Gamma校正方法，对感兴趣的图像病变区域进行展宽，对不感兴趣的背景区域进行压缩；
S2：搭建网络模型U-net，该网络模型U-net分为压缩路径和扩展路径，共有4次下采样和4次上采样，每次采样前进行两次卷积和一次最大池化，压缩路径中用4次下采样对视网膜图像进行特征压缩，扩展路径对最后一次下采样得到的有效特征层进行4次上采样，同时把下采样中对应的特征层进行连接，最后使用1*1卷积将视网膜特征图Feature Map进行归一化，用增强后的图像数据对构建的模型进行训练，得到图像分割模型；
S3：将模型分割出来的血管图像进行模糊化处理，计算图像中属性的隶属度以及模糊信息增益，与真实诊断结果进行训练，得到模糊决策树分支节点的决策规则和叶子结点的结果集合，进行分类预测；
S4：对决策树各个节点进行编码，同时构造适应度函数，从准确度和复杂度两个方面来衡量模糊树模型的优劣，用准确函数E来表示模型的精度，E越小则精度越高，用树的叶节点个数M来反映复杂度，M越小则模型的复杂度越低，适应度函数s(T)定义如下：
其中：WE、WM分别为精度E和叶节点个数M的权值，且WE+WM＝1，s(T)表示树T的适应度；
基于遗传算法对多棵决策树进行组合优化；
S5：最后引入准确率指标，根据样本类别与真实值的距离，动态调整损失函数中的惩罚项，进一步提升分类准确率。
作为本发明的优选技术方案：所述步骤S2的具体步骤如下：
步骤S2.1：将视网膜病变数据集按9∶1划分为训练集和验证集输入到训练网络中；
步骤S2.2：搭建网络模型U-net的压缩路径，压缩路径对视网膜图像进行4次下采样，得到5个初步有效特征层，每个初步有效特征层为卷积和最大池化的堆叠，使用3*3的卷积核对输入大小为565*584的视网膜图像进行两次卷积操作，每次卷积舍弃图像的边缘信息，将卷积得到的561*580视网膜图像进行2*2最大池化，每次下采样将视网膜特征图Feature Map的通道数增加为原来的两倍；
步骤S2.3：搭建网络模型U-net的扩展路径，扩展路径中包含4次上采样，每次上采样通过2*2反卷积将上一层中视网膜特征图Feature Map通道数缩小为一半，图像的长宽翻倍，同时上采样时会将下采样中对应的特征层进行连接，由于卷积时丢失图像边缘信息，因此连接时采用适当的裁剪来保证连接后图像大小一致，U-net网络最后使用1*1卷积来将视网膜特征图Feature Map进行归一化；
步骤S2.4：损失函数使用交叉熵和SoftMax，对于视网膜病变图像每个像素点的类别，预测其属于病变区域和正常区域的概率分别为p和1-p，像素点形式SoftMax函数为：
其中：ak(x)表示像素点x在特征图中的第k层的激活值，K是类的数量，pk(x)是类k对像素点x的分类结果；
交叉熵损失函数E定义为：
其中：Ω＝{1,...,K}，l(x)是每个像素点x的真实标签，pl(x)(x)是真实标签的分类结果，w(x)是像素点x的权重图，区分每个像素点的权重，其权重计算公式如下：
其中：wc(x)是用来平衡某一类频率的权重图，根据像素点x到视网膜病变边界距离由近到远进行排序，d1(x)表示排序第一的距离，d2(x)表示排序第二的距离，w0为权重初始值，其值为10，标准差σ设置为5；
步骤S2.5：用卷积神经网络框架Caffe的随机梯度下降对网络进行训练优化，以最小化损失函数和最大化预测准确率为目标，训练所搭建的模型。
作为本发明的优选技术方案：所述步骤S3的具体步骤如下：
步骤S3.1：对视网膜血管的病变区域进行模糊化，尤其针对病变边缘区域，得到图像连续值属性的隶属度，模糊化后属性值是一个[0,1]区间的隶属度，更自然、合理的描述病变区域边缘的不精确信息；
步骤S3.2：计算视网膜病变区域属性的模糊信息增益，设A＝{(ui,μA(ui)),ui∈U}是属性集U上隶属函数为μA(ui)的模糊属性集合，高斯隶属函数μA(ui)计算方式如式5，U＝{u1,u2,...,ui,...,um}是属性的离散集合，m为属性的个数，且第i个属性的模糊度μi＝μA(ui)，则模糊集合A的模糊性度量E(A)为：

其中：c为正态分布的均值，σ为正态分布的标准差；
选取具有最高模糊信息增益的属性作为根节点的属性；
步骤S3.3：根据父节点的属性和父节点所对应的训练集，以及该节点在父节点属性上的属性值，构造该节点所对应的模糊子类集Am，在模糊子类集Am上，依据要划分的目标类C＝{c1,c2,...,cm}，计算每个模糊子集的模糊信息增益；
步骤S3.4：计算节点Node中某目标类ci的置信度i＝1,2,3,...,m，根据规定的最大置信水平β和最小置信水平α，判断是否生成叶子结点：
其中：An是模糊子类集Am中未使用过的属性集合，且n＜m，aj是属性集An中的第j个属性，j＝1,2,...,n，LeafNode表示叶子结点；
步骤S3.5：根据该节点扩展属性的属性值构造子节点，递归处理各 个子节点。
作为本发明的优选技术方案：所述步骤S4的具体步骤如下：
步骤S4.1：对上一步骤中训练的模糊决策树进行编码，转化为遗传算法可解的个体形式：规定根节点编号N0为1，当非根节点为左子节点时编号Nl＝2×Np，其中Np为父节点的编号，当非根节点为右子节点时编号Nr＝2×Np+1，得到各节点编号后，以节点自身的编号、左右子节点以及父节点的编号按顺序构造出一个四元组，作为该节点的编码Ncode，若无父节点或者无子节点则对应位置上码值为0，将树中各节点的编码进行连接，得到整棵树的矩阵编码；
步骤S4.2：从预测视网膜病变准确度和复杂度两个方面来衡量模糊树模型的优劣，用准确函数E来表示模型的精度，E越小则精度越高，即预测视网膜病变的准确率越高，复杂度由树的叶节点个数M来反映，M越小则模型的复杂度越低，因此适应度函数s(T)为：
其中：WE、WM分别为精度E和叶节点个数M的权值，且WE+WM＝1，s(T)表示树T的适应度；
根据构造的适应度函数，求出每个视网膜病变模糊树的适应度，即初始化种群群体；
步骤S4.3：根据每棵模糊树的适应度，利用轮盘赌方法，选出一对父代个体，每棵模糊树被选中的概率与其适应度值成比例，设个体总数为N，某一个体xi的适应度值表示为f(xi)，该个体被选中的概率为p(xi)，累积概率为q(xi)，对应的计算公式如下：

累积概率q(xi)表示个体之前所有个体的选择概率之和，它相当于轮盘上转过的范围，范围越大越容易选到；
步骤S4.4：对父代个体随机选择一个交叉点k，以概率Pc进行交叉操作，生成新的个体，对新个体的每个基因，随机在[0,1]上产生一个变异概率Pm，进行变异；
步骤S4.5：重新计算新个体在环境中的适应度，与最优值进行比较，更新种群，当最大进化代数T＝150或者最优个体的适应度和群体适应度连续10代不再上升时，得到适应度最高的一代个体，即得到了最优的模糊决策树。
本发明有益效果：
1、本发明考虑了医学图像的特征，在深度学习网络中融合了遗传模糊树，更自然、合理的描述视网膜血管病变区域边缘的不精确信息，使得模型结果的精确性和可解释性得到加强，在数据规模有限的情况下也能收获较好的效果，引入准确率指标，根据样本类别与真实值的距离，动态调整损失函数中的惩罚项，进一步提升分类准确率。
2、本发明的一种基于遗传模糊树的视网膜糖尿病变深度网络检测方法，可以准确分割出视网膜中的血管图像，更精确地识别出血管末梢，通过模糊决策树对提取的视网膜特征进行分析分类，能够提高检测的准确度，并且能够使诊断结果的可解释性得到加强，有效提高了诊断结果的可靠性，使用遗传算法优化模糊树的结构，更进一步提高了检测的准确度，能够有效的帮助医生诊断视网膜糖尿病变，让患者获得最佳的治疗时期。
说明书附图
下面结合附图对本发明作进一步说明：
图1所示为本发明一种基于遗传模糊树的视网膜糖尿病变深度网络检测方法的流程图；
图2所示为本发明预测分割后视网膜图像存在病变生成模糊决策树的流程图；
图3所示为本发明图像预处理时Gamma图像增强对比图，左图为原始图像的黑白图，右图为gamma＝1.2时的增强效果图；
图4所示为本发明网络模型U-net分割出的视网膜血管图，左图为原 始图像，中间为标签mask，右图为分割后图像；
图5所示为本发明建立模糊决策树的其中一段分支；
图6所示为本发明对生成的模糊决策树使用遗传算法进行优化的流程图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合实施例，对本发明进行进一步详细说明。当然，通过参考附图描述的示例仅用于解释本发明，而不能解释为对本发明的限制。
如图1所示，本发明公开了基于遗传模糊树的视网膜糖尿病变深度网络检测方法，属于医学信息智能处理领域。本发明包括以下步骤：
S1：将视网膜图像进行增强处理，视网膜图像的病变区域与周围正常区域会呈现出视觉上较为明显的不同特征，形成为不同图像区域，采用图像增强Gamma校正方法，对感兴趣的图像病变区域进行展宽，对不感兴趣的背景区域进行压缩；
S2：搭建网络模型U-net，该网络模型U-net分为压缩路径和扩展路径，共有4次下采样和4次上采样，每次采样前进行两次卷积和一次最大池化，压缩路径中用4次下采样对视网膜图像进行特征压缩，扩展路径对最后一次下采样得到的有效特征层进行4次上采样，同时把下采样中对应的特征层进行连接，最后使用1*1卷积将视网膜特征图Feature Map进行归一化，用增强后的图像数据对构建的模型进行训练，得到图像分割模型；
S3：将模型分割出来的血管图像进行模糊化处理，计算模糊化处理后的血管图像中属性的隶属度以及模糊信息增益，与真实诊断结果进行训练，得到模糊决策树分支节点的决策规则和叶子结点的结果集合，进行分类预测；
S4：对决策树各个节点进行编码，同时构造适应度函数，从准确度和复杂度两个方面来衡量模糊树模型的优劣，用准确函数E来表示模型的精度，E越小则精度越高，用树的叶节点个数M来反映复杂度，M越小则模型的复杂度越低，适应度函数s(T)定义如下：
其中：WE、WM分别为精度E和叶节点个数M的权值，且WE+WM＝1，s(T)表示树T的适应度；
基于遗传算法对多棵决策树进行组合优化；
S5：最后引入准确率指标，根据样本类别与真实值的距离，动态调整损失函数中的惩罚项，进一步提升分类准确率。
步骤S2的具体步骤如下：
步骤S2.1：将视网膜病变数据集按9∶1划分为训练集和验证集输入到训练网络中；
步骤S2.2：搭建网络模型U-net的压缩路径，压缩路径对视网膜图像进行4次下采样，得到5个初步有效特征层，每个初步有效特征层为卷积和最大池化的堆叠，使用3*3的卷积核对输入大小为565*584的视网膜图像进行两次卷积操作，每次卷积舍弃图像的边缘信息，将卷积得到的561*580视网膜图像进行2*2最大池化，每次下采样将视网膜特征图Feature Map的通道数增加为原来的两倍；
步骤S2.3：搭建网络模型U-net的扩展路径，扩展路径中包含4次上采样，每次上采样通过2*2反卷积将上一层中视网膜特征图Feature Map通道数缩小为一半，图像的长宽翻倍，同时上采样时会将下采样中对应的特征层进行连接，由于卷积时丢失图像边缘信息，因此连接时采用适当的裁剪来保证连接后图像大小一致，U-net网络最后使用1*1卷积来将视网膜特征图Feature Map进行归一化；
步骤S2.4：损失函数使用交叉熵和SoftMax，对于视网膜病变图像每个像素点的类别，预测其属于病变区域和正常区域的概率分别为p和1-p，像素点形式SoftMax函数为：
其中：ak(x)表示像素点x在特征图中的第k层的激活值，K是类的数量，pk(x)是类k对像素点x的分类结果；
交叉熵损失函数E定义为：
其中：Ω＝{1,...,K}，l(x)是每个像素点x的真实标签，pl(x)(x)是真实标签的分类结果，w(x)是像素点x的权重图，区分每个像素点的权重，其权重计算公式如下：
其中：wc(x)是用来平衡某一类频率的权重图，根据像素点x到视网膜病变边界距离由近到远进行排序，d1(x)表示排序第一的距离，d2(x)表示排序第二的距离，w0为权重初始值，其值为10，标准差σ设置为5；
步骤S2.5：用卷积神经网络框架Caffe的随机梯度下降对网络进行训练优化，以最小化损失函数和最大化预测准确率为目标，训练所搭建的模型。
步骤S3的具体步骤如下：
步骤S3.1：对视网膜血管的病变区域进行模糊化，尤其针对病变边缘区域，得到图像连续值属性的隶属度，模糊化后属性值是一个[0,1]区间的隶属度，更自然、合理的描述病变区域边缘的不精确信息；
步骤S3.2：计算视网膜病变区域属性的模糊信息增益，设A＝{(ui,μA(ui)),ui∈U}是属性集U上隶属函数为μA(ui)的模糊属性集合，高斯隶属函数μA(ui)计算方式如式5，U＝{u1,u2,...,ui,...,um}是属性的离散集合，m为属性的个数，且第i个属性ui的模糊度μi＝μA(ui)，则模糊集合A的模糊性度量E(A)为：

其中：c为正态分布的均值，σ为正态分布的标准差；
选取具有最高模糊信息增益的属性作为根节点的属性；
步骤S3.3：根据父节点的属性和父节点所对应的训练集，以及该节 点在父节点属性上的属性值，构造该节点所对应的模糊子类集Am，在模糊子类集Am上，依据要划分的目标类C＝{c1,c2,...,cm}，计算每个模糊子集的模糊信息增益；
步骤S3.4：计算节点Node中某目标类ci的置信度i＝1,2,3,...,m，根据规定的最大置信水平β和最小置信水平α，判断是否生成叶子结点：
其中：An是模糊子类集Am中未使用过的属性集合，且n＜m，aj是属性集An中的第j个属性，j＝1,2,...,n，LeafNode表示叶子结点；
步骤S3.5：根据该节点扩展属性的属性值构造子节点，递归处理各个子节点。
步骤S4的具体步骤如下：
步骤S4.1：对上一步骤中训练的模糊决策树进行编码，转化为遗传算法可解的个体形式：规定根节点编号N0为1，当非根节点为左子节点时编号Nl＝2×Np，其中Np为父节点的编号，当非根节点为右子节点时编号Nr＝2×Np+1，得到各节点编号后，以节点自身的编号、左右子节点以及父节点的编号按顺序构造出一个四元组，作为该节点的编码Ncode，若无父节点或者无子节点则对应位置上码值为0，将树中各节点的编码进行连接，得到整棵树的矩阵编码；
步骤S4.2：从预测视网膜病变准确度和复杂度两个方面来衡量模糊树模型的优劣，用准确函数E来表示模型的精度，E越小则精度越高，即预测视网膜病变的准确率越高，复杂度由树的叶节点个数M来反映，M越小则模型的复杂度越低，因此适应度函数s(T)为：
其中：WE、WM分别为精度E和叶节点个数M的权值，且WE+WM＝1，s(T)表示树T的适应度；
根据构造的适应度函数，求出每个视网膜病变模糊树的适应度，即初始化种群群体；
步骤S4.3：根据每棵模糊树的适应度，利用轮盘赌方法，选出一对父代个体，每棵模糊树被选中的概率与其适应度值成比例，设个体总数为N，某一个体xi的适应度值表示为f(xi)，该个体被选中的概率为p(xi)，累积概率为q(xi)，对应的计算公式如下：

累积概率q(xi)表示个体之前所有个体的选择概率之和，它相当于轮盘上转过的范围，范围越大越容易选到；
步骤S4.4：对父代个体随机选择一个交叉点k，以概率Pc进行交叉操作，生成新的个体，对新个体的每个基因，随机在[0,1]上产生一个变异概率Pm，进行变异；
步骤S4.5：重新计算新个体在环境中的适应度，与最优值进行比较，更新种群，当最大进化代数T＝150或者最优个体的适应度和群体适应度连续10代不再上升时，得到适应度最高的一代个体，即得到了最优的模糊决策树。
本发明用图像增强Gamma方法对视网膜图像数据集进行预处理，如图3所示，左边为原始图像，右边为Gamma＝1.2时的增强效果图；然后将数据集划按9:1划分为训练集与验证集，训练分割网络U-net，当损失函数达到最小时，分割结果如图4所示，左图为原始图像，中间是标签mask，右图为U-net分割结果；选择视网膜血管图像中血管宽度、血管弯曲度、分形维度等特征作为决策属性，按照图2所示方法建立模糊决策树，图5为模糊树的部分分支；将生成的模糊决策树进行矩阵编码，使用图6所示遗传算法来优化模糊决策树；最后计算预测的准确率，根据样本类别与真实值的距离，评价模型的准确率。
本发明考虑了医学图像的特征，在深度学习网络中融合了遗传模糊树，更自然、合理的描述视网膜血管病变区域边缘的不精确信息，使得模型结 果的精确性和可解释性得到加强，在数据规模有限的情况下也能收获较好的效果，引入准确率指标，根据样本类别与真实值的距离，动态调整损失函数中的惩罚项，进一步提升分类准确率。
本发明的一种基于遗传模糊树的视网膜糖尿病变深度网络检测方法，可以准确分割出视网膜中的血管图像，更精确地识别出血管末梢，通过模糊决策树对提取的视网膜特征进行分析分类，能够提高检测的准确度，并且能够使诊断结果的可解释性得到加强，有效提高了诊断结果的可靠性，使用遗传算法优化模糊树的结构，更进一步提高了检测的准确度，能够有效的帮助医生诊断视网膜糖尿病变，让患者获得最佳的治疗时期。
以上所述的具体实施方案，对本发明的目的、技术方案和有益效果进行了进一步的详细说明，所应理解的是，以上所述仅为本发明的具体实施方案而已，并非用以限定本发明的范围，任何本领域的技术人员，在不脱离本发明的构思和原则的前提下所做出的等同变化与修改，均应属于本发明保护的范围。

Claims (4)

  1. 基于遗传模糊树的视网膜糖尿病变深度网络检测方法，其特征在于，包括以下步骤：
    S1：将视网膜图像进行增强处理，视网膜图像的病变区域与周围正常区域会呈现出视觉上较为明显的不同特征，形成为不同图像区域，采用图像增强Gamma校正方法，对感兴趣的图像病变区域进行展宽，对不感兴趣的背景区域进行压缩；
    S2：搭建网络模型U-net，该网络模型U-net分为压缩路径和扩展路径，共有4次下采样和4次上采样，每次采样前进行两次卷积和一次最大池化，压缩路径中用4次下采样对视网膜图像进行特征压缩，扩展路径对最后一次下采样得到的有效特征层进行4次上采样，同时把下采样中对应的特征层进行连接，最后使用1*1卷积将视网膜特征图Feature Map进行归一化，用增强后的图像数据对构建的模型进行训练，得到图像分割模型；
    S3：将模型分割出来的血管图像进行模糊化处理，计算图像中属性的隶属度以及模糊信息增益，与真实诊断结果进行训练，得到模糊决策树分支节点的决策规则和叶子结点的结果集合，进行分类预测；
    S4：对决策树各个节点进行编码，同时构造适应度函数，从准确度和复杂度两个方面来衡量模糊树模型的优劣，用准确函数E来表示模型的精度，E越小则精度越高，用树的叶节点个数M来反映复杂度，M越小则模型的复杂度越低，适应度函数s(T)定义如下：
    其中：WE、WM分别为精度E和叶节点个数M的权值，且WE+WM＝1，s(T)表示树T的适应度；
    基于遗传算法对多棵决策树进行组合优化；
    S5：最后引入准确率指标，根据样本类别与真实值的距离，动态调整损失函数中的惩罚项，进一步提升分类准确率。
  2. 根据权利要求1所述的基于遗传模糊树的视网膜糖尿病变深度网络检测方法，其特征在于，所述步骤S2的具体步骤如下：
    步骤S2.1：将视网膜病变数据集按9∶1划分为训练集和验证集输入 到训练网络中；
    步骤S2.2：搭建网络模型U-net的压缩路径，压缩路径对视网膜图像进行4次下采样，得到5个初步有效特征层，每个初步有效特征层为卷积和最大池化的堆叠，使用3*3的卷积核对输入大小为565*584的视网膜图像进行两次卷积操作，每次卷积舍弃图像的边缘信息，将卷积得到的561*580视网膜图像进行2*2最大池化，每次下采样将视网膜特征图Feature Map的通道数增加为原来的两倍；
    步骤S2.3：搭建网络模型U-net的扩展路径，扩展路径中包含4次上采样，每次上采样通过2*2反卷积将上一层中视网膜特征图Feature Map通道数缩小为一半，图像的长宽翻倍，同时上采样时会将下采样中对应的特征层进行连接，由于卷积时丢失图像边缘信息，因此连接时采用适当的裁剪来保证连接后图像大小一致，U-net网络最后使用1*1卷积来将视网膜特征图Feature Map进行归一化；
    步骤S2.4：损失函数使用交叉熵和SoftMax，对于视网膜病变图像每个像素点的类别，预测其属于病变区域和正常区域的概率分别为p和1-p，像素点形式SoftMax函数为：
    其中：ak(x)表示像素点x在特征图中的第k层的激活值，K是类的数量，pk(x)是类k对像素点x的分类结果；
    交叉熵损失函数E定义为：
    其中：Ω＝{1,...,K}，l(x)是每个像素点x的真实标签，pl(x)(x)是真实标签的分类结果，w(x)是像素点x的权重图，区分每个像素点的权重，其权重计算公式如下：
    其中：wc(x)是用来平衡某一类频率的权重图，根据像素点x到视网膜 病变边界距离由近到远进行排序，d1(x)表示排序第一的距离，d2(x)表示排序第二的距离，w0为权重初始值，其值为10，标准差σ设置为5；
    步骤S2.5：用卷积神经网络框架Caffe的随机梯度下降对网络进行训练优化，以最小化损失函数和最大化预测准确率为目标，训练所搭建的模型。
  3. 根据权利要求1所述的基于遗传模糊树的视网膜糖尿病变深度网络检测方法，其特征在于，所述步骤S3的具体步骤如下：
    步骤S3.1：对视网膜血管的病变区域进行模糊化，尤其针对病变边缘区域，得到图像连续值属性的隶属度，模糊化后属性值是一个[0,1]区间的隶属度，更自然、合理的描述病变区域边缘的不精确信息；
    步骤S3.2：计算视网膜病变区域属性的模糊信息增益，设A＝{(ui,μA(ui)),ui∈U}是属性集U上隶属函数为μA(ui)的模糊属性集合，高斯隶属函数μA(ui)计算方式如式5，U＝{u1,u2,...,ui,...,um}是属性的离散集合，m为属性的个数，且第i个属性的模糊度μi＝μA(ui)，则模糊集合A的模糊性度量E(A)为：

    其中：c为正态分布的均值，σ为正态分布的标准差；
    选取具有最高模糊信息增益的属性作为根节点的属性；
    步骤S3.3：根据父节点的属性和父节点所对应的训练集，以及该节点在父节点属性上的属性值，构造该节点所对应的模糊子类集Am，在模糊子类集Am上，依据要划分的目标类C＝{c1,c2,...,cm}，计算每个模糊子集的模糊信息增益；
    步骤S3.4：计算节点Node中某目标类ci的置信度根据规定的最大置信水平β和最小置信水平α，判断是否生成叶子结点：
    其中：An是模糊子类集Am中未使用过的属性集合，且n＜m，aj是属性集An中的第j个属性，j＝1,2,...,n，LeafNode表示叶子结点；
    步骤S3.5：根据该节点扩展属性的属性值构造子节点，递归处理各个子节点。
  4. 根据权利要求1所述的基于遗传模糊树的视网膜糖尿病变深度网络检测方法，其特征在于，所述步骤S4的具体步骤如下：
    步骤S4.1：对上一步骤中训练的模糊决策树进行编码，转化为遗传算法可解的个体形式：规定根节点编号N0为1，当非根节点为左子节点时编号Nl＝2×Np，其中Np为父节点的编号，当非根节点为右子节点时编号Nr＝2×Np+1，得到各节点编号后，以节点自身的编号、左右子节点以及父节点的编号按顺序构造出一个四元组，作为该节点的编码Ncode，若无父节点或者无子节点则对应位置上码值为0，将树中各节点的编码进行连接，得到整棵树的矩阵编码；
    步骤S4.2：从预测视网膜病变准确度和复杂度两个方面来衡量模糊树模型的优劣，用准确函数E来表示模型的精度，E越小则精度越高，即预测视网膜病变的准确率越高，复杂度由树的叶节点个数M来反映，M越小则模型的复杂度越低，因此适应度函数s(T)为：
    其中：WE、WM分别为精度E和叶节点个数M的权值，且WE+WM＝1，s(T)表示树T的适应度；
    根据构造的适应度函数，求出每个视网膜病变模糊树的适应度，即初始化种群群体；
    步骤S4.3：根据每棵模糊树的适应度，利用轮盘赌方法，选出一对父代个体，每棵模糊树被选中的概率与其适应度值成比例，设个体总数为N，某一个体xi的适应度值表示为f(xi)，该个体被选中的概率为p(xi)，累积概率为q(xi)，对应的计算公式如下：

    累积概率q(xi)表示个体之前所有个体的选择概率之和，它相当于轮盘上转过的范围，范围越大越容易选到；
    步骤S4.4：对父代个体随机选择一个交叉点k，以概率Pc进行交叉操作，生成新的个体，对新个体的每个基因，随机在[0,1]上产生一个变异概率Pm，进行变异；
    步骤S4.5：重新计算新个体在环境中的适应度，与最优值进行比较，更新种群，当最大进化代数T＝150或者最优个体的适应度和群体适应度连续10代不再上升时，得到适应度最高的一代个体，即得到了最优的模糊决策树。
PCT/CN2023/077768 2022-01-26 2023-02-23 基于遗传模糊树的视网膜糖尿病变深度网络检测方法 WO2023143628A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210094089.7 2022-01-26
CN202210094089.7A CN114494196B (zh) 2022-01-26 2022-01-26 基于遗传模糊树的视网膜糖尿病变深度网络检测方法

Publications (1)

Publication Number Publication Date
WO2023143628A1 true WO2023143628A1 (zh) 2023-08-03

Family

ID=81476976

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/077768 WO2023143628A1 (zh) 2022-01-26 2023-02-23 基于遗传模糊树的视网膜糖尿病变深度网络检测方法

Country Status (2)

Country Link
CN (1) CN114494196B (zh)
WO (1) WO2023143628A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117871792A (zh) * 2024-03-13 2024-04-12 河北省建筑科学研究院有限公司 用于对园区绿植碳汇的动态监测方法及系统
CN118096769A (zh) * 2024-04-29 2024-05-28 中国科学院宁波材料技术与工程研究所 一种视网膜oct图像的分析方法及装置

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114494196B (zh) * 2022-01-26 2023-11-17 南通大学 基于遗传模糊树的视网膜糖尿病变深度网络检测方法
CN114972279A (zh) * 2022-06-06 2022-08-30 南通大学 用于视网膜血管不确定边界精准分割的三支U-Net方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111325725A (zh) * 2020-02-19 2020-06-23 京东方科技集团股份有限公司 视网膜图像识别方法及装置、电子设备、存储介质
US20200234445A1 (en) * 2018-04-13 2020-07-23 Bozhon Precision Industry Technology Co., Ltd. Method and system for classifying diabetic retina images based on deep learning
CN112883962A (zh) * 2021-01-29 2021-06-01 北京百度网讯科技有限公司 眼底图像识别方法、装置、设备、存储介质以及程序产品
CN112950555A (zh) * 2021-02-05 2021-06-11 广州中医药大学第一附属医院 一种基于深度学习的2型糖尿病心血管病变图像分类方法
CN113012163A (zh) * 2021-03-11 2021-06-22 山东大学 一种基于多尺度注意力网络的视网膜血管分割方法、设备及存储介质
CN114494196A (zh) * 2022-01-26 2022-05-13 南通大学 基于遗传模糊树的视网膜糖尿病变深度网络检测方法

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9778021B2 (en) * 2013-08-29 2017-10-03 Carl Zeiss Meditec, Inc. Evaluation of optical coherence tomographic data prior to segmentation
CN104463215A (zh) * 2014-12-10 2015-03-25 东北大学 基于视网膜图像处理的微小动脉瘤发生风险预测系统
CN109523524B (zh) * 2018-11-07 2020-07-03 电子科技大学 一种基于集成学习的眼底图像硬性渗出检测方法
CN109635862B (zh) * 2018-12-05 2021-08-24 合肥奥比斯科技有限公司 早产儿视网膜病plus病变分类方法
CN110459267B (zh) * 2019-08-12 2023-05-26 岭南师范学院 一种基于改进自适应遗传算法的人体体成分预测方法
CN111240353B (zh) * 2020-01-07 2021-06-15 南京航空航天大学 基于遗传模糊树的无人机协同空战决策方法
CN111815574B (zh) * 2020-06-18 2022-08-12 南通大学 一种基于粗糙集神经网络的眼底视网膜血管图像分割方法
CN111931816A (zh) * 2020-07-09 2020-11-13 河南工业大学 一种视网膜图像平行处理方法及装置
CN112185523B (zh) * 2020-09-30 2023-09-08 南京大学 基于多尺度卷积神经网络的糖尿病视网膜病变分类方法
CN112258486B (zh) * 2020-10-28 2023-04-07 汕头大学 基于进化神经架构搜索的眼底图像视网膜血管分割方法
CN113160119A (zh) * 2021-02-04 2021-07-23 广州中医药大学第一附属医院 一种基于深度学习的糖尿病视网膜病变图像分类方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200234445A1 (en) * 2018-04-13 2020-07-23 Bozhon Precision Industry Technology Co., Ltd. Method and system for classifying diabetic retina images based on deep learning
CN111325725A (zh) * 2020-02-19 2020-06-23 京东方科技集团股份有限公司 视网膜图像识别方法及装置、电子设备、存储介质
CN112883962A (zh) * 2021-01-29 2021-06-01 北京百度网讯科技有限公司 眼底图像识别方法、装置、设备、存储介质以及程序产品
CN112950555A (zh) * 2021-02-05 2021-06-11 广州中医药大学第一附属医院 一种基于深度学习的2型糖尿病心血管病变图像分类方法
CN113012163A (zh) * 2021-03-11 2021-06-22 山东大学 一种基于多尺度注意力网络的视网膜血管分割方法、设备及存储介质
CN114494196A (zh) * 2022-01-26 2022-05-13 南通大学 基于遗传模糊树的视网膜糖尿病变深度网络检测方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
WAN-LI ZHOU, QIU HONG-ZE, YIN ZUO-HAI: "Optimization of fuzzy decision tree with multiattribute based on genetic algorithm", COMPUTER ENGINEERING AND APPLICATIONS, HUABEI JISUAN JISHU YANJIUSUO, CN, vol. 45, no. 3, 21 January 2009 (2009-01-21), CN , pages 148 - 151, XP093081641, ISSN: 1002-8331, DOI: 10.3778/j.issn.1002-8331.2009.03.044 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117871792A (zh) * 2024-03-13 2024-04-12 河北省建筑科学研究院有限公司 用于对园区绿植碳汇的动态监测方法及系统
CN117871792B (zh) * 2024-03-13 2024-05-14 河北省建筑科学研究院有限公司 用于对园区绿植碳汇的动态监测方法及系统
CN118096769A (zh) * 2024-04-29 2024-05-28 中国科学院宁波材料技术与工程研究所 一种视网膜oct图像的分析方法及装置

Also Published As

Publication number Publication date
CN114494196B (zh) 2023-11-17
CN114494196A (zh) 2022-05-13

Similar Documents

Publication Publication Date Title
WO2023143628A1 (zh) 基于遗传模糊树的视网膜糖尿病变深度网络检测方法
US10706333B2 (en) Medical image analysis method, medical image analysis system and storage medium
CN108364006B (zh) 基于多模式深度学习的医学图像分类装置及其构建方法
CN109376636B (zh) 基于胶囊网络的眼底视网膜图像分类方法
US20220148191A1 (en) Image segmentation method and apparatus and storage medium
WO2016192612A1 (zh) 基于深度学习对医疗数据进行分析的方法及其智能分析仪
Kumar et al. Breast cancer classification of image using convolutional neural network
CN109345538A (zh) 一种基于卷积神经网络的视网膜血管分割方法
CN110246577B (zh) 一种基于人工智能辅助妊娠期糖尿病遗传风险预测的方法
CN110459328A (zh) 一种评估心脏骤停的临床决策支持系统
Mahapatra et al. A novel framework for retinal vessel segmentation using optimal improved frangi filter and adaptive weighted spatial FCM
CN113113130A (zh) 一种肿瘤个体化诊疗方案推荐方法
CN111080643A (zh) 基于眼底图像的糖尿病及相关疾病的分类方法及设备
Brintha et al. A food recommendation system for predictive diabetic patients using ANN and CNN
CN115985503B (zh) 基于集成学习的癌症预测系统
CN111028232A (zh) 基于眼底图像的糖尿病分类方法及设备
CN115601346A (zh) 一种基于深度学习的多模态mri对膝关节软骨损伤的多层次分类方法
CN111047590A (zh) 基于眼底图像的高血压分类方法及设备
Han et al. DM-CNN: Dynamic Multi-scale Convolutional Neural Network with uncertainty quantification for medical image classification
CN112712895B (zh) 针对2型糖尿病并发症的多模态大数据的数据分析方法
CN113421250A (zh) 一种基于无病变影像训练的眼底疾病智能诊断方法
CN115937590A (zh) 一种并联融合CNN和Transformer的皮肤病图像分类方法
Mathina Kani et al. Classification of skin lesion images using modified Inception V3 model with transfer learning and augmentation techniques
CN112419256A (zh) 基于模糊图神经网络的糖网病眼底图像分级方法
Sundaramoorthy et al. Hybrid Optimization with Recurrent Neural Network-based Medical Image Processing for Predicting Interstitial Lung Disease

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: 23746504

Country of ref document: EP

Kind code of ref document: A1