CN108009592A - A kind of diabetic retinal classification of images method - Google Patents
A kind of diabetic retinal classification of images method Download PDFInfo
- Publication number
- CN108009592A CN108009592A CN201711349578.8A CN201711349578A CN108009592A CN 108009592 A CN108009592 A CN 108009592A CN 201711349578 A CN201711349578 A CN 201711349578A CN 108009592 A CN108009592 A CN 108009592A
- Authority
- CN
- China
- Prior art keywords
- bnnet
- data
- retinal
- classification
- networks
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
A kind of diabetic retinal classification of images method.First, the pretreatment such as denoising, normalization is carried out to image；Then, in the case of sample data deficiency, the strategy extraction strengthened using transfer learning and data is used for the depth characteristic classified；Finally, the feature of extraction is sent into grader and retina pathological image is divided into five classes.The classification accuracy of the present invention has preferable robustness and generalization up to 93%.
Description
Technical field
The present invention is the automatic classification method of diabetic retinal image, suitable for machine learning, pattern-recognition and doctor
Learn image processing technique field.
Background technology
The automatic classification of diabetic retinopathy reason image has important clinical value in retina pathology figure
As in classification problem, extraction is representative, has differentiation meaning to be characterized in realizing that the key factor of good classification effect is current
Based on the sorting technique of artificial pathological image, mainly there is following limitation:(1) retinal map that the quality of image is collected
The quality of picture is easily influenced be subject to other many factors such as the experiences of illumination, camera lens, machinery equipment and Image Acquisition personnel；
(2) the personal experience doctor of doctor usually assesses the lesion degree of judgement retina by visual inspection retinal images, but
Be in retinopathy image eye recognition be out characterized in it is limited, it is right and since the clinical experience of doctor is different
In same secondary retinal images, different doctors may provide different clinical diagnosis results；(3) spy of pathological image itself
Otherness very little between point retinal images different phases, has carried out certain difficulty to feature extraction and sorting work tape；(4)
Lack large-scale publicly available retinal map image set in field of medical images
The content of the invention
The purpose of the present invention aiming at the problems such as retinal images feature extraction is difficult, classification performance is poor, tape label
The problem of retinal images are few, proposes a kind of diabetic retinopathy classification of images method based on transfer learning.
The technical solution adopted by the present invention is：A kind of diabetic retinal classification of images method,
1) Diabetic of the data source in data modeling and data analysis competition platform (kaggle)
Retinopathy Detection compete, and the retinal images in the data set are all high-resolution RGB images, and according to
Retinal images are divided into normally by lesion degree, Minimal change, moderate lesion, severe lesion, five class of proliferative lesion；
2) denoising, histogram equalization, normalization, removal black surround frame, data enhancing and spy are carried out to retinal image data
The pretreatments such as sign analysis；
3) in order to avoid model by the change of data distribution during training cause convergence rate slow the problems such as,
On the basis of AlexNet networks one is obtained to introducing batch normalization layer before each convolutional layer and full articulamentum
BNnet network structure .BNnet networks are deeper more complicated deep neural networks (DCNN)；
4) a depth sorting device is devised, which proposes a kind of classfying frame of new retinopathy image
Frame, the input of the grader is 4096 dimensions of every width sample by BNnet models using the extraction of transfer learning method in data set
Fc7 layers of characteristic set.
The present invention introduces AlexNet to verify the validity of normalization layer, sets two groups of training methods:(1) with increasing
Retinal images after strong are trained AlexNet networks；(2) BNnet networks are carried out with enhanced retinal images
Training, BNnet introduce batch normalization layer on the basis of AlexNet is original before each convolutional layer and full articulamentum.
The present invention uses two kinds of different training methods to verify the validity of transfer learning to BNnet network structures:
(1) BNnet networks are directly trained with retinal images after enhancing；(2) with ILSVRC2012 data sets to BNnet networks
Pre-training is carried out, then the model learnt is moved to and is learnt again on retina pathological image.
The present invention is in order to verify the validity of data enhancing, to using two kinds of differences by the BNnet models of transfer learning
Learning strategy again:(1) model is learnt again using the data before enhancing；(2) enhanced data the set pair analysis model is used
Learnt again.
The present invention by the softmax layers (fc8) of 1000 nodes of top layer in BNnet networks with one newly comprising 5
The softmax layers of node are that new-fc8 is replaced, and the parameters weighting W random initializtions that will be connected to new-fc8 layers.
Grader of the present invention includes two full articulamentum ip1 and ip2, in order to avoid over-fitting, in each full articulamentum
Add Dropout layers respectively afterwards, and linear amending unit (relu1 and relu2) is introduced after two full linking layers, with solution
Certainly gradient disperse problem.
The present invention effect be：(1) a kind of taxonomy model of new retinopathy image is proposed, which has good
Classification performance mainly have benefited from two important parts:Depth characteristic expression is obtained by transfer learning and one is based on
The classification policy of deeper network introduces AlexNet networks on batch specification layer (batch normalization) and using migration
The strategy of study obtains the more powerful depth characteristic expression needed for image classification, and when classifying to retina to grader
More accurately classifying quality can be obtained by having introduced the taxonomic structure that full articulamentum and Dropout layer form deeper；(2) compare
In traditional medical image sorting technique, the depth characteristic obtained by transfer learning represents have to prediction retina pathological image
There is more preferable robustness to be asked in order to avoid model causes convergence rate slow etc. during training by the change of data distribution
Topic, on the basis of AlexNet networks to each convolutional layer and full articulamentum before introduce batch normalization layer and obtain one
A BNnet network structures .BNnet networks are deeper more complicated deep neural networks (DCNN)；(3) present invention is also set
A depth characteristic grader is counted, lesion retinal images are divided into normally, Minimal change, moderate lesion, severe lesion, is increased
5 classifications of natural disposition lesion.
Brief description of the drawings
Fig. 1 is image preprocessing of the present invention:(a) to delete meaningless image, (b) to remove black surround frame, (c) is straight for image
Side's figure equalization, (d) normalizes for view data；
Depth characteristic extracting method network frame figure of Fig. 2 present invention based on transfer learning；
Fig. 3 depth sorting device structure charts of the present invention.
Specific implementation method
See Fig. 1, Fig. 2, Fig. 3, a kind of diabetic retinal classification of images method, the invention is characterised in that：
1) Diabetic of the data source in data modeling and data analysis competition platform (kaggle)
Retinopathy Detection compete, and the retinal images in the data set are all high-resolution RGB images, and according to
Retinal images are divided into normally by lesion degree, Minimal change, moderate lesion, severe lesion, five class of proliferative lesion；
2) denoising, histogram equalization, normalization, removal black surround frame, data enhancing and spy are carried out to retinal image data
The pretreatments such as sign analysis；
3) in order to avoid model by the change of data distribution during training cause convergence rate slow the problems such as,
On the basis of AlexNet networks one is obtained to introducing batch normalization layer before each convolutional layer and full articulamentum
BNnet network structure .BNnet networks are deeper more complicated deep neural networks (DCNN)；
4) a depth sorting device is devised, which proposes a kind of classfying frame of new retinopathy image
Frame, the input of the grader is 4096 dimensions of every width sample by BNnet models using the extraction of transfer learning method in data set
Fc7 layers of characteristic set.
The present invention introduces AlexNet to verify the validity of normalization layer, sets two groups of training methods:(1) with increasing
Retinal images after strong are trained AlexNet networks；(2) BNnet networks are carried out with enhanced retinal images
Training, BNnet introduce batch normalization layer on the basis of AlexNet is original before each convolutional layer and full articulamentum.
The present invention uses two kinds of different training methods to verify the validity of transfer learning to BNnet network structures:
(1) BNnet networks are directly trained with retinal images after enhancing；(2) with ILSVRC2012 data sets to BNnet networks
Pre-training is carried out, then the model learnt is moved to and is learnt again on retina pathological image.
The present invention is in order to verify the validity of data enhancing, to using two kinds of differences by the BNnet models of transfer learning
Learning strategy again:(1) model is learnt again using the data before enhancing；(2) enhanced data the set pair analysis model is used
Learnt again.
The present invention by the softmax layers (fc8) of 1000 nodes of top layer in BNnet networks with one newly comprising 5
The softmax layers of node are that new-fc8 is replaced, and the parameters weighting W random initializtions that will be connected to new-fc8 layers.
Grader of the present invention includes two full articulamentum ip1 and ip2, in order to avoid over-fitting, in each full articulamentum
Add Dropout layers respectively afterwards, and linear amending unit (relu1 and relu2) is introduced after two full linking layers, with solution
Certainly gradient disperse problem.
The traditional method using medical image training network from the beginning of present invention substitution, using ILSVRC2012 data
Set pair BNnet networks carry out pre-training, then obtained model is moved to and is learnt again specifically on retinal images,
The depth characteristic extraction frame that the present invention uses is to use the softmax layers (fc8) of 1000 nodes of top layer in BNnet networks
One new softmax layer comprising 5 nodes i.e. new-fc8 is replaced, and the parameters weighting W that will be connected to new-fc8 layers
Random initializtion devises a depth sorting device and classifies to retina pathological image, and the input of the grader is data
Concentrate every width sample by BNnet models using the characteristic set of 4096 fc7 layers of dimensions of the transfer learning method extraction graders
Including two full articulamentum ip1 and ip2, in order to avoid over-fitting, Dropout is added respectively after each full articulamentum
Layer, and linear amending unit (relu1 and relu2) is introduced after two full linking layers, to solve the problems, such as gradient disperse.
Claims (6)
- A kind of 1. diabetic retinal classification of images method, it is characterised in that：1) Diabetic Retinopathy of the data source in data modeling and data analysis competition platform kaggle Detection competes, and the retinal images in the data set are all high-resolution RGB images, and will be regarded according to lesion degree Nethike embrane image is divided into normally, Minimal change, moderate lesion, severe lesion, five class of proliferative lesion；2) denoising, histogram equalization, normalization, removal black surround frame, data enhancing and feature point are carried out to retinal image data The pretreatment such as analysis；3) in order to avoid model by the change of data distribution during training cause convergence rate slow the problems such as, On the basis of AlexNet networks one is obtained to introducing batch normalization layer before each convolutional layer and full articulamentum BNnet network structure .BNnet networks are deeper more complicated deep neural network DCNN；4) a depth sorting device is devised, which proposes a kind of taxonomy model of new retinopathy image, should The input of grader is 4096 dimensions fc7 layer of every width sample by BNnet models using the extraction of transfer learning method in data set Characteristic set.
- 2. a kind of diabetic retinal classification of images method according to claim 1, it is characterised in that in order to test The validity that normalization layer is introduced to AlexNet is demonstrate,proved, two groups of training methods are set:(1) with enhanced retinal images pair AlexNet networks are trained；(2) BNnet networks are trained with enhanced retinal images, BNnet is in AlexNet On the basis of original, batch normalization layer is introduced before each convolutional layer and full articulamentum.
- 3. a kind of diabetic retinal classification of images method according to claim 1, it is characterised in that in order to test The validity of transfer learning is demonstrate,proved, two kinds of different training methods are used to BNnet network structures:(1) with retinal map after enhancing As being directly trained to BNnet networks；(2) pre-training is carried out to BNnet networks with ILSVRC2012 data sets, then will study To model move to and learn again on retina pathological image.
- 4. a kind of diabetic retinal classification of images method according to claim 1, it is characterised in that in order to test The validity of data enhancing is demonstrate,proved, to using two kinds of different learning strategies again by the BNnet models of transfer learning:(1) use Data before enhancing learn model again；(2) learnt again using enhanced data the set pair analysis model.
- A kind of 5. diabetic retinal classification of images method according to claim 1, it is characterised in that the present invention It is by one new softmax layer comprising 5 nodes of the softmax layer fc8 of 1000 nodes of top layer in BNnet networks New-fc8 is replaced, and the parameters weighting W random initializtions that will be connected to new-fc8 layers.
- A kind of 6. diabetic retinal classification of images method according to claim 1, it is characterised in that grader Including two full articulamentum ip1 and ip2, in order to avoid over-fitting, Dropout is added respectively after each full articulamentum Layer, and linear amending unit relu1 and relu2 is introduced after two full linking layers, to solve the problems, such as gradient disperse.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711349578.8A CN108009592A (en) | 2017-12-15 | 2017-12-15 | A kind of diabetic retinal classification of images method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711349578.8A CN108009592A (en) | 2017-12-15 | 2017-12-15 | A kind of diabetic retinal classification of images method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108009592A true CN108009592A (en) | 2018-05-08 |
Family
ID=62059263
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711349578.8A Pending CN108009592A (en) | 2017-12-15 | 2017-12-15 | A kind of diabetic retinal classification of images method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108009592A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108665457A (en) * | 2018-05-16 | 2018-10-16 | 腾讯科技（深圳）有限公司 | Image-recognizing method, device, storage medium and computer equipment |
| CN108765422A (en) * | 2018-06-13 | 2018-11-06 | 云南大学 | A kind of retinal images blood vessel automatic division method |
| CN108876775A (en) * | 2018-06-12 | 2018-11-23 | 广州图灵人工智能技术有限公司 | The rapid detection method of diabetic retinopathy |
| CN109508673A (en) * | 2018-11-13 | 2019-03-22 | 大连理工大学 | It is a kind of based on the traffic scene obstacle detection of rodlike pixel and recognition methods |
| CN109543749A (en) * | 2018-11-22 | 2019-03-29 | 云南大学 | Drawing sentiment analysis method based on deep learning |
| CN110084809A (en) * | 2019-05-06 | 2019-08-02 | 成都医云科技有限公司 | Diabetic retinopathy data processing method, device and electronic equipment |
| CN110210570A (en) * | 2019-06-10 | 2019-09-06 | 上海延华大数据科技有限公司 | The more classification methods of diabetic retinopathy image based on deep learning |
| CN111028232A (en) * | 2019-12-31 | 2020-04-17 | 上海鹰瞳医疗科技有限公司 | Diabetes classification method and equipment based on fundus images |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105868785A (en) * | 2016-03-30 | 2016-08-17 | 乐视控股（北京）有限公司 | Image identification method based on convolutional neural network and image identification system thereof |
| CN106055576A (en) * | 2016-05-20 | 2016-10-26 | 大连理工大学 | Rapid and effective image retrieval method under large-scale data background |
| CN106127725A (en) * | 2016-05-16 | 2016-11-16 | 北京工业大学 | A kind of millimetre-wave radar cloud atlas dividing method based on multiresolution CNN |
| CN106485251A (en) * | 2016-10-08 | 2017-03-08 | 天津工业大学 | Egg embryo classification based on deep learning |
| CN106570141A (en) * | 2016-11-04 | 2017-04-19 | 中国科学院自动化研究所 | Method for detecting approximately repeated image |
| CN106934798A (en) * | 2017-02-20 | 2017-07-07 | 苏州体素信息科技有限公司 | Diabetic retinopathy classification stage division based on deep learning |
| CN107169956A (en) * | 2017-04-28 | 2017-09-15 | 西安工程大学 | Yarn dyed fabric defect detection method based on convolutional neural networks |
| CN107203134A (en) * | 2017-06-02 | 2017-09-26 | 浙江零跑科技有限公司 | A kind of front truck follower method based on depth convolutional neural networks |
| CN107330449A (en) * | 2017-06-13 | 2017-11-07 | 瑞达昇科技(大连)有限公司 | A kind of BDR sign detection method and device |
-
2017
- 2017-12-15 CN CN201711349578.8A patent/CN108009592A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105868785A (en) * | 2016-03-30 | 2016-08-17 | 乐视控股（北京）有限公司 | Image identification method based on convolutional neural network and image identification system thereof |
| CN106127725A (en) * | 2016-05-16 | 2016-11-16 | 北京工业大学 | A kind of millimetre-wave radar cloud atlas dividing method based on multiresolution CNN |
| CN106055576A (en) * | 2016-05-20 | 2016-10-26 | 大连理工大学 | Rapid and effective image retrieval method under large-scale data background |
| CN106485251A (en) * | 2016-10-08 | 2017-03-08 | 天津工业大学 | Egg embryo classification based on deep learning |
| CN106570141A (en) * | 2016-11-04 | 2017-04-19 | 中国科学院自动化研究所 | Method for detecting approximately repeated image |
| CN106934798A (en) * | 2017-02-20 | 2017-07-07 | 苏州体素信息科技有限公司 | Diabetic retinopathy classification stage division based on deep learning |
| CN107169956A (en) * | 2017-04-28 | 2017-09-15 | 西安工程大学 | Yarn dyed fabric defect detection method based on convolutional neural networks |
| CN107203134A (en) * | 2017-06-02 | 2017-09-26 | 浙江零跑科技有限公司 | A kind of front truck follower method based on depth convolutional neural networks |
| CN107330449A (en) * | 2017-06-13 | 2017-11-07 | 瑞达昇科技(大连)有限公司 | A kind of BDR sign detection method and device |
Non-Patent Citations (4)
| Title |
|---|
| BIN ZHOU等: ""SeatBelt Detection Using Convolutional Neural Network BN-AlexNet"", 《INTERNATIONAL CONFERENCE ON INTELLIGENT COMPUTING》 * |
| SERGEY IOFFE等: ""Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift"", 《JMLR.ORG》 * |
| 丁蓬莉等: ""糖尿病性视网膜图像的深度神经网络分类方法"", 《计算机应用》 * |
| 姜枫等: ""基于随机Dropout卷积神经网络的人体行为识别方法研究"", 《测试技术学报》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108665457A (en) * | 2018-05-16 | 2018-10-16 | 腾讯科技（深圳）有限公司 | Image-recognizing method, device, storage medium and computer equipment |
| CN110335269A (en) * | 2018-05-16 | 2019-10-15 | 腾讯医疗健康(深圳)有限公司 | The classification recognition methods of eye fundus image and device |
| CN108665457B (en) * | 2018-05-16 | 2023-12-19 | 腾讯医疗健康(深圳)有限公司 | Image recognition method, device, storage medium and computer equipment |
| CN108876775A (en) * | 2018-06-12 | 2018-11-23 | 广州图灵人工智能技术有限公司 | The rapid detection method of diabetic retinopathy |
| CN108765422A (en) * | 2018-06-13 | 2018-11-06 | 云南大学 | A kind of retinal images blood vessel automatic division method |
| CN109508673A (en) * | 2018-11-13 | 2019-03-22 | 大连理工大学 | It is a kind of based on the traffic scene obstacle detection of rodlike pixel and recognition methods |
| CN109543749A (en) * | 2018-11-22 | 2019-03-29 | 云南大学 | Drawing sentiment analysis method based on deep learning |
| CN110084809A (en) * | 2019-05-06 | 2019-08-02 | 成都医云科技有限公司 | Diabetic retinopathy data processing method, device and electronic equipment |
| CN110084809B (en) * | 2019-05-06 | 2021-03-16 | 成都医云科技有限公司 | Diabetic retinopathy data processing method and device and electronic equipment |
| CN110210570A (en) * | 2019-06-10 | 2019-09-06 | 上海延华大数据科技有限公司 | The more classification methods of diabetic retinopathy image based on deep learning |
| CN111028232A (en) * | 2019-12-31 | 2020-04-17 | 上海鹰瞳医疗科技有限公司 | Diabetes classification method and equipment based on fundus images |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108009592A (en) | A kind of diabetic retinal classification of images method | |
| Kwasigroch et al. | Deep CNN based decision support system for detection and assessing the stage of diabetic retinopathy | |
| CN110909780B (en) | Image recognition model training and image recognition method, device and system | |
| Shan et al. | A deep learning method for microaneurysm detection in fundus images | |
| Khan et al. | Cataract detection using convolutional neural network with VGG-19 model | |
| CN107330889A (en) | A kind of traditional Chinese medical science tongue color coating colour automatic analysis method based on convolutional neural networks | |
| CN106530295A (en) | Fundus image classification method and device of retinopathy | |
| CN107330393A (en) | A kind of neonatal pain expression recognition method based on video analysis | |
| Chen et al. | Cell nuclei detection and segmentation for computational pathology using deep learning | |
| CN110490242A (en) | Training method, eye fundus image classification method and the relevant device of image classification network | |
| CN102567734A (en) | Specific value based retina thin blood vessel segmentation method | |
| Tu et al. | SUNet: A lesion regularized model for simultaneous diabetic retinopathy and diabetic macular edema grading | |
| Xu et al. | Fully deep learning for slit-lamp photo based nuclear cataract grading | |
| Niu et al. | Automatic localization of optic disc based on deep learning in fundus images | |
| CN111046978A (en) | Skin mirror image identification method based on StyleGANs and decision fusion | |
| Nandy et al. | An incremental feature extraction framework for referable diabetic retinopathy detection | |
| CN110197483A (en) | Deep basal pit crack detection method based on vision signal | |
| Syed et al. | Detection of tumor in MRI images using artificial neural networks | |
| Lee et al. | Effective computer-assisted automatic cervical vertebrae extraction with rehabilitative ultrasound imaging by using K-means clustering | |
| CN110084796A (en) | A kind of analysis method of complex texture CT image | |
| Pan et al. | Automated drawing psychoanalysis via house-tree-person test | |
| Karthik et al. | Design and Implementation of Multi-Retinal Disease Classification Using Deep Neural Network | |
| CN109255794A (en) | A kind of full convolution edge feature detection method of standard component depth | |
| Zhao et al. | A Tongue Color Classification Method in TCM Based on Transfer Learning | |
| Guefrachi et al. | Automated diabetic retinopathy screening using deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180508 |