CN109949302A - Retinal feature Structural Techniques based on pixel - Google Patents
Retinal feature Structural Techniques based on pixel Download PDFInfo
- Publication number
- CN109949302A CN109949302A CN201910235344.3A CN201910235344A CN109949302A CN 109949302 A CN109949302 A CN 109949302A CN 201910235344 A CN201910235344 A CN 201910235344A CN 109949302 A CN109949302 A CN 109949302A
- Authority
- CN
- China
- Prior art keywords
- feature
- retinal
- pixel
- network
- image
- 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
Landscapes
- Eye Examination Apparatus (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of retinal feature Structural Techniques based on pixel, comprising: 1) acquire colored eye ground image;2) feature structure in acquisition image is labeled, marks blood vessel, exudate, optic cup optic disk, blutpunkte and macula lutea respectively;3) five classes are divided into according to health, nonproliferative diabetic retinopathy degree and Proliferative diabetic retinopathy to the image of acquisition, construct data set;4) FOV extraction is carried out to retina color image, determines and extracts threshold value;5) multi-scale division network is designed, four kinds of retinal images size training patterns are inputted;6) the network ROC curve, the sensitivity and specificity between comparing cell are established.The result shows that the accuracy rate of retinal feature segmentation can be improved in this method, and robustness is good.
Description
Technical field
The present invention relates to a kind of retinal feature Structural Techniques based on pixel, to colored eye ground figure
When as segmentation of structures, accuracy rate has a distinct increment than the prior art.
Background technique
Retinopathy is caused by daily a variety of common diseases as a result, such as diabetes, artery sclerosis, leukaemia
Deng.These diseases will lead to retina length and width, angle even results in the hyperplasia of retinal vessel.Clinically regarded frequently by eyeground
Nethike embrane image carries out screening, analysis and diagnosis to disease.It may determine that whether patient has artery sclerosis by retina arteriovenous,
By blutpunkte and exudate number may determine that whether patient has diabetes, patient may determine that by the detection of optic cup optic disk
Whether glaucoma is had.Therefore, in order to carry out quantitative analysis to disease and to the early prevention of disease, segmentation retina each section knot
Structure related work plays the important and pivotal role, and has to the high efficiency diagnosis of doctor expert and the mankind to the prevention of some persistent ailments
Directive significance, be the embodiment to be promoted the well-being of mankind using artificial intelligence.
There is many segmentation eyeground structure very early by whole world Experts ' Attention therewith in the segmentation problem of eye ground
Traditional algorithm, can briefly be divided into unsupervised and have the two major classes dividing method of supervision.In recent years, deep learning algorithm obtains
Key breakthrough forms abstract further further feature by learning shallow-layer feature, and finds the distribution of data accordingly.With biography
System method is compared, and deep learning allows computer autonomous learning from observation data, is settled a dispute by the parties concerned themselves before us according to learning outcome
It is solved the problems, such as with traditional algorithm.And many segmentation eyeground methods are disclosed.Such as Khalaf ties convolutional neural networks CNN
Structure simplification distinguishes big blood vessel, thin vessels and the background in eye fundus image, and is carried out using various sizes of convolution kernel
Adjustment, has obtained good segmentation result in digital retinal images vessel extraction library DRIVE.The it is proposeds such as Fu are based on CNN and item
The retinal images blood vessel segmentation method of part random field CRF, this method are utilized using blood vessel segmentation as border detection issue handling
CNN generates segmentation probability graph, obtains binary segmentation result in conjunction with CRF.After CNN, AlexNet and VGGNet are had also been proposed
It is split.But still there are many drawbacks:
(1) the problems such as data set is too small, data set mark inaccuracy;
(2) for there are the retinal images of large area pathological characters, between lesion region, lesion region and normal region
Between interfere with each other, seriously affect segmentation effect;
(3) small part retinal image quality is influenced more by environment when acquiring, and leads to characteristic details and background contrasts
It is low, it is artificial to be directly manually labeled the technical experience for relying on operator to data set, it is larger by subjective factor, efficiency also compared with
It is low.
By each network of comparison we have found that PixelNet model is that the pixel based on characteristic pattern carries out study segmentation,
It is suitble to retinal fundus images segmentation.The model extracts convolution feature using VGG-16 convolution, the pixel sampled to one, from
Corresponding feature is extracted on multiple convolution characteristic patterns, hypercolumn descriptor is established, is then input to this feature
One MLP multilayer perceptron, recently enters classification results.In addition, the network main thought is originally sampling plan when training
Slightly, accelerate training.It proposes stratified sampling pixel-based, sparse sampling, finds from training image
The seldom pixel of middle sampling can be obtained by good result.The present invention introduce first how to establish the data set for improving oneself with
More extensive training collection is provided.Then it on the basis of PixelNet network portion structure, is replaced originally with more perfect VGG-19
VGG-16 to increase the width of network.Meanwhile adjusting each layer of num_output parameter of network structure, that is, it is defeated to adjust each layer
Characteristic pattern number out, maximum compression model size is under conditions of model accuracy rate guarantees to improve test rate.In addition,
Three layers of perceptron part reasonably change to improve testing efficiency the number of plies according to the complexity of eyeground different characteristic.
Finally, we, which are reasonably allocated to network model in multiple GPU, is split test to test set image, realize each in image
Individual segmentation between structure.
Summary of the invention
In view of this, the invention proposes a kind of retinal feature Structural Techniques based on pixel, this method can
To improve the accuracy rate of retinal feature segmentation, and robustness is good.
In order to achieve the above objectives, first aspect according to the invention provides the retinal feature knot based on pixel
Structure dividing method, specific technical solution, including the following steps:
Step 1: acquiring colored eye ground image;
Step 2: the feature structure in acquisition image being labeled, blood vessel is marked respectively, exudate, optic cup optic disk, goes out
Blood point and macula lutea;
Step 3: to the image of acquisition according to health, nonproliferative diabetic retinopathy degree and proliferative diabetic
Retinopathy is divided into five classes, constructs data set;
Step 4: FOV extraction being carried out to retina color image, determines and extracts threshold value;
Step 5: design multi-scale division network inputs four kinds of retinal images size training patterns;
Step 6: establishing the network ROC curve, the sensitivity and specificity between comparing cell.
Compared with prior art, the beneficial effects of the present invention are:
Firstly, the present invention use oneself last 2 years building data set, using transfer learning make retina photograph no matter from
It is all more comprehensive than existing dividing method on quality and quantity.Then, data set is assigned under multiple GPU by improved more
Scale network is trained and tests, and substantially increases segmentation efficiency and accuracy rate.Finally, segmentation result is passed through the interface MFC
It shows.This method comprehensively considers timeliness and accuracy rate, and blood vessel accuracy rate is 0.927, and optic disk optic cup accuracy rate is
0.997, exudate accuracy rate is 0.939, and macula lutea accuracy rate is 0.997, and blutpunkte accuracy rate is 0.904, and this method is suitable
Wider retina eye fundus picture is closed, and robustness is good.
Detailed description of the invention
Fig. 1 overall framework schematic diagram, i.e. Figure of abstract;
The multiple dimensioned network structure of Fig. 2;
Distribution of each stage eye fundus image of Fig. 3 diabetic in data set;
The original retinal images of Fig. 4 (a-c) and the FOV (d-f) being accordingly partitioned into;
The ROC curve of the multiple dimensioned network of Fig. 5;
Fig. 6 retinal structure segmentation effect figure;
Fig. 7 network training parameter algorithm flow chart;
Fig. 8 compares the ROC curve of heterogeneous networks.
Specific embodiment
The present invention is described in further detail With reference to embodiment.
Universe network block schematic illustration of the invention is as shown in Figure 2.Here four branching networks extract feature respectively independently,
Each branch parameter is not shared, therefore inputs section start in data, and four branches are respectively to initial data training, in four channels
End the feature vector that the last one convolutional layer in each channel extracts is connected, keep characteristic information more abundant.By four
The feature vector in a channel is spliced according to 1: 1: 1: 1 ratio, is finally output to next full articulamentum, makes to connect entirely
Layer sufficiently combines the respective feature of four channel networks, has taken into account the general type generic features and target data set of pre-training model
Data special characteristic.It is consistent in addition to patch dimensional parameters per network pre-training structure all the way with PixelNet network.
The specific implementation process of technical solution of the present invention is illustrated With reference to embodiment.
1. multi-scale division network structure
Here four branching networks extract feature respectively independently, and each branch parameter is not shared, therefore is inputted in data
At beginning, four branches respectively to initial data training, four channels end by the last one convolutional layer in each channel
The feature vector of extraction is connected, and keeps characteristic information more abundant.By the feature vector in four channels according to 1: 1: 1: 1 ratio into
Row splicing, is finally output to next full articulamentum, full articulamentum is made sufficiently to combine the respective feature of four channel networks,
The general type generic features of pre-training model and the data special characteristic of target data set are taken into account.Per network pre-training knot all the way
Structure is consistent in addition to patch dimensional parameters with PixelNet network.
2. experiment
2.1 data set
It is 2124 × 2056 that the data set that this experiment uses, which is 1444 × 1444 and 940 pixels comprising 560 pixels,
1500 retinal fundus images of total 283 patients, age bracket male differ, differ within women 24-75 years old for 25-83 years old, wrap
Multiple pathological grades are included, and all marks and approves by hospital ophthalmology expert.For more comprehensive data set, we are also adopted
The blood vessel of DRIVE data set marks, and has chosen the feature mark of 20 sheet photo of STARE data set, amounts to 1520 retinal maps
Picture.Wherein 1320 are used for training pattern, remaining is for verifying model.All images all first carry out adaptive thresholding before the input
It is worth processing method to divide the boundary FOV, threshold value 23 from the background of retinal fundus images.According to the multiple dimensioned input of network
The retinal images that 1320 are partitioned into the boundary FOV are divided into 224 × 224,448 × 448,896 × 896 patch by feature,
Input along with complete image as network.
2.2 segmentation standard
In order to more objectively compare the similarities and differences of segmentation result and artificial calibration result, four kinds of Statistical indicators are introduced:
(1) true positives refer to the pixel that really blood vessel is also accurately identified by model as blood vessel;(2) false negative refer to really blood vessel but by
Model is identified as the pixel of non-vascular;(3) true negative refers to the picture that really non-vascular is also accurately identified by model as non-vascular
Vegetarian refreshments;(4) false positive refers to that really non-vascular is but identified as the pixel of blood vessel by model.Pass through definition spirit again on this basis
The superiority and inferiority of the trained model of the parameter evaluations network such as sensitivity, specificity, accuracy rate.According to the variation relation between Se and 1-Sp
Receiver operating curve's ROC curve is drawn, the area AUC under ROC curve reflects the performance of dividing method, and AUC is 1
Classifier is exactly perfect classifier.
2.3 training
The each structure of retina is different, but it is seen that each structure can be divided into two according to our network structures
Class.We term it concentrated structures for one kind, and such as optic cup optic disk and macula lutea, another kind of we term it cover type structures, such as blood
Pipe, exudate, blutpunkte.Concentrated is structurally characterized in that such retinal structure often only occurs in a certain piece of region of image, institute
Accounting example is smaller relative to general image, and structural integrity is smaller.Cover type is structurally characterized in that such retinal structure generally goes out
Each place of present image, proportion is larger relative to whole figure image, and structure is caused to have certain globality.
For cover type structure, we use the multiple dimensioned input network shown described in a upper section such as Fig. 2, Shu Ru 224
× 224,448 × 448,896 × 896 and complete image, while inputting details, do not forget the entirety for paying attention to cover type structure
Property, and the update that can carry out weight is derived based on back-propagation algorithm, training parameter algorithm flow block diagram such as Fig. 7 shows.
In the present invention, above three prediction is integrated, and is set as the objective function of multitask network training.
Herein, by LossK1, LossK2, LossK3, LossK4It is expressed as the loss of one to four branches with Loss Bin, and comes
It is lost from the binary system of final output.The objective function of whole network can be expressed as Loss=Loss K1+Loss K2+Loss
K3+Loss K4+Loss Bin.In each training step, loss function can be propagated from branch one to four.It therefore, can be with base
Network is updated in five kinds of different classification tasks.The design considers five predictions in generalized framework, it is more suitable for me
Multistage deep neural network backpropagation.In short, loss can be with is defined as:For
Each detection layers m, there are training samples.For an image, object and the distribution without object is very uneven, therefore use is adopted
Sample eliminates this imbalance.It can be defined as
2.4 compared with other networks
The network is compared with tri- kinds of networks of PiexlNet, Unet and DeepLabv3+ of open source are split effect, and four
A network is all based on the data set of we and processing, is compared from three accuracy rate, sensitivity and specificity indexs.Pay attention to
It is the present invention accurate segmentation effect of each feature structure of eye fundus image in order to obtain herein, it will be respectively to each feature knot
Structure carry out three accuracy rate, sensitivity and specificity indexs judge, be respectively blood vessel, exudate, blutpunkte, optic cup optic disk and
Macula lutea.The ROC curve of final four networks output is as shown in Figure 7.Multi-scale model method has higher than single channel network
Accuracy.Compared with art methods, the method achieve higher precision, it is small to solve eyeground data scale, mark essence
Spend the inaccurate problem undesirable so as to cause eyeground segmentation of structures effect, at the same realize detection optical fundus blood vessel, exudate,
Optic cup optic disk, blutpunkte whole feature structure.
The foregoing is only a preferred embodiment of the present invention, is not intended to limit the scope of the present invention, should
Understand, the present invention is not limited to existing scheme as described herein, the purpose of these implementations description is to help in this field
Technical staff practice the present invention.Any those of skill in the art are easy in the feelings for not departing from spirit and scope of the invention
It is further improved under condition and perfect, therefore the present invention is only limited by the content and range of the claims in the present invention,
It includes alternative in the spirit and scope of the invention being defined by the appended claims and equivalent that it, which is intended to cover all,
Scheme.
Claims (5)
1. a kind of retinal feature Structural Techniques based on pixel, including the following steps:
Step 1: acquiring colored eye ground image;
Step 2: the feature structure in acquisition image being labeled, marks blood vessel, exudate, optic cup optic disk, blutpunkte respectively
And macula lutea;
Step 3: to the image of acquisition according to health, nonproliferative diabetic retinopathy degree and proliferative diabetic view
Film lesion is divided into five classes, constructs data set;
Step 4: FOV extraction being carried out to retina color image, determines and extracts threshold value;
Step 5: design multi-scale division network inputs four kinds of retinal images size training patterns;
Step 6: establishing the network ROC curve, the sensitivity and specificity between comparing cell.
2. the retinal feature Structural Techniques according to claim 1 based on pixel, which is characterized in that step 3
In, this method constructs the data set of oneself, and different from disclosed eyeground data set of having increased income, our data set is to difference
Age, different sexes, the different state of an illness are marked and have been illustrated respectively, including each pathological grade, and all pass through hospital ophthalmology
Expert, which marks, to be approved, wherein most eye ground photos are one group data of patient during one section, it is more square in this way
Just illness analysis is done after.
3. the retinal feature Structural Techniques according to claim 1 based on pixel, which is characterized in that step 4
In, in order to which retinal images are split from background, we will be with like attribute by using k means clustering method
Image pixel grouping, in retinal fundus images reduce different colours quantity, execute adaptive thresholding method with from
Divide the boundary FOV in the background of retinal fundus images, since purpose is to distinguish FOV and background, threshold value must be big
Threshold value is selected by changing threshold value to 30 (corresponding to black) come paired observation from 0 in the intensity value of background pixel.
4. the retinal feature Structural Techniques according to claim 1 based on pixel, which is characterized in that step 5
In, multi-scale division network pixel-based is designed, four branching networks of network extract feature respectively independently, each branch parameter
It does not share, therefore inputs section start in data, four branches, will be each in the end in four channels respectively to initial data training
The feature vector that the last one convolutional layer in a channel extracts is connected, and keeps characteristic information more abundant, by the feature in four channels
Vector is connected entirely according to 1: 1: 1: 1 ratio, is finally output to next full articulamentum, keeps full articulamentum sufficiently comprehensive
The respective feature of four channel networks, the data of the general type generic features and target data set of having taken into account pre-training model are specific
The combination of feature, four channels enhances the generalization ability of network.
5. the retinal feature Structural Techniques according to claim 1 based on pixel, which is characterized in that step 5
In, the first paths to the 4th paths, corresponding sequence is that from top to bottom, the size of input picture is 224 respectively in Fig. 2
The patch and full-size image of × 224,448 × 448,896 × 896 totally three kinds of sizes, first three paths are to allow e-learning
The local feature of segmentation object, last paths is to allow the global feature of e-learning segmentation object, pre- per network all the way
Training structure is consistent in addition to patch dimensional parameters with PixelNet network.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910235344.3A CN109949302A (en) | 2019-03-27 | 2019-03-27 | Retinal feature Structural Techniques based on pixel |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910235344.3A CN109949302A (en) | 2019-03-27 | 2019-03-27 | Retinal feature Structural Techniques based on pixel |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109949302A true CN109949302A (en) | 2019-06-28 |
Family
ID=67010752
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910235344.3A Pending CN109949302A (en) | 2019-03-27 | 2019-03-27 | Retinal feature Structural Techniques based on pixel |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109949302A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110599491A (en) * | 2019-09-04 | 2019-12-20 | 腾讯医疗健康(深圳)有限公司 | Priori information-based eye image segmentation method, device, equipment and medium |
CN111652866A (en) * | 2020-05-29 | 2020-09-11 | 天津医科大学眼科医院 | Iris angiography quantitative analysis method |
CN112001928A (en) * | 2020-07-16 | 2020-11-27 | 北京化工大学 | Retinal vessel segmentation method and system |
CN112652394A (en) * | 2021-01-14 | 2021-04-13 | 浙江工商大学 | Multi-focus target detection-based retinopathy of prematurity diagnosis system |
CN112837322A (en) * | 2019-11-22 | 2021-05-25 | 北京深睿博联科技有限责任公司 | Image segmentation method and device, equipment and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108520522A (en) * | 2017-12-31 | 2018-09-11 | 南京航空航天大学 | Retinal fundus images dividing method based on the full convolutional neural networks of depth |
CN109448006A (en) * | 2018-11-01 | 2019-03-08 | 江西理工大学 | A kind of U-shaped intensive connection Segmentation Method of Retinal Blood Vessels of attention mechanism |
-
2019
- 2019-03-27 CN CN201910235344.3A patent/CN109949302A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108520522A (en) * | 2017-12-31 | 2018-09-11 | 南京航空航天大学 | Retinal fundus images dividing method based on the full convolutional neural networks of depth |
CN109448006A (en) * | 2018-11-01 | 2019-03-08 | 江西理工大学 | A kind of U-shaped intensive connection Segmentation Method of Retinal Blood Vessels of attention mechanism |
Non-Patent Citations (2)
Title |
---|
AAYUSH BANSAL: "PixelNet: Representation of the pixels, by the pixels, and for the pixels", 《ARXIV:1702.06506V1》 * |
MING LI: "Retinal Blood Vessel Segmentation Based on Multi-Scale Deep Learning", 《PROCEEDINGS OF THE FEDERATED CONFERENCE ON COMPUTER SCIENCE AND INFORMATION SYSTEMS》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110599491A (en) * | 2019-09-04 | 2019-12-20 | 腾讯医疗健康(深圳)有限公司 | Priori information-based eye image segmentation method, device, equipment and medium |
CN110599491B (en) * | 2019-09-04 | 2024-04-12 | 腾讯医疗健康(深圳)有限公司 | Priori information-based eye image segmentation method, apparatus, device and medium |
CN112837322A (en) * | 2019-11-22 | 2021-05-25 | 北京深睿博联科技有限责任公司 | Image segmentation method and device, equipment and storage medium |
CN111652866A (en) * | 2020-05-29 | 2020-09-11 | 天津医科大学眼科医院 | Iris angiography quantitative analysis method |
CN111652866B (en) * | 2020-05-29 | 2022-12-27 | 天津医科大学眼科医院 | Iris angiography quantitative analysis method |
CN112001928A (en) * | 2020-07-16 | 2020-11-27 | 北京化工大学 | Retinal vessel segmentation method and system |
CN112001928B (en) * | 2020-07-16 | 2023-12-15 | 北京化工大学 | Retina blood vessel segmentation method and system |
CN112652394A (en) * | 2021-01-14 | 2021-04-13 | 浙江工商大学 | Multi-focus target detection-based retinopathy of prematurity diagnosis system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109949302A (en) | Retinal feature Structural Techniques based on pixel | |
Hacisoftaoglu et al. | Deep learning frameworks for diabetic retinopathy detection with smartphone-based retinal imaging systems | |
Wang et al. | A coarse-to-fine deep learning framework for optic disc segmentation in fundus images | |
CN106408564B (en) | A kind of method for processing fundus images based on deep learning, apparatus and system | |
Gupta et al. | Diabetic retinopathy: Present and past | |
CN110197493A (en) | Eye fundus image blood vessel segmentation method | |
CN112132817B (en) | Retina blood vessel segmentation method for fundus image based on mixed attention mechanism | |
Fang et al. | Adam challenge: Detecting age-related macular degeneration from fundus images | |
CN109325942A (en) | Eye fundus image Structural Techniques based on full convolutional neural networks | |
Yang et al. | Efficacy for differentiating nonglaucomatous versus glaucomatous optic neuropathy using deep learning systems | |
Zong et al. | U-net based method for automatic hard exudates segmentation in fundus images using inception module and residual connection | |
Aatila et al. | Diabetic retinopathy classification using ResNet50 and VGG-16 pretrained networks | |
Ou et al. | BFENet: A two-stream interaction CNN method for multi-label ophthalmic diseases classification with bilateral fundus images | |
Firke et al. | Convolutional neural network for diabetic retinopathy detection | |
CN111028230A (en) | Fundus image optic disc and macula lutea positioning detection algorithm based on YOLO-V3 | |
Lu et al. | Automatic classification of retinal diseases with transfer learning-based lightweight convolutional neural network | |
Wang et al. | Accurate disease detection quantification of iris based retinal images using random implication image classifier technique | |
Yu et al. | Intelligent detection and applied research on diabetic retinopathy based on the residual attention network | |
Hu et al. | A multi-center study of prediction of macular hole status after vitrectomy and internal limiting membrane peeling by a deep learning model | |
Caicho et al. | Diabetic retinopathy: Detection and classification using alexnet, ***net and resnet50 convolutional neural networks | |
Guergueb et al. | A review of deep learning techniques for glaucoma detection | |
Tsai et al. | Diagnosis of polypoidal choroidal vasculopathy from fluorescein angiography using deep learning | |
Zhang et al. | Blood vessel segmentation in fundus images based on improved loss function | |
Thanh et al. | A real-time classification of glaucoma from retinal fundus images using AI technology | |
Maji et al. | Automatic grading of retinal blood vessel tortuosity using Modified CNN in deep retinal image diagnosis |
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 |
Application publication date: 20190628 |
|
WD01 | Invention patent application deemed withdrawn after publication |