CN107657612A - Suitable for full-automatic the retinal vessel analysis method and system of intelligent and portable equipment - Google Patents
Suitable for full-automatic the retinal vessel analysis method and system of intelligent and portable equipment Download PDFInfo
- Publication number
- CN107657612A CN107657612A CN201710970423.XA CN201710970423A CN107657612A CN 107657612 A CN107657612 A CN 107657612A CN 201710970423 A CN201710970423 A CN 201710970423A CN 107657612 A CN107657612 A CN 107657612A
- Authority
- CN
- China
- Prior art keywords
- image
- intelligent
- blood vessel
- eye fundus
- portable equipment
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- 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
- G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
-
- 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
-
- 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
- G06T2207/30101—Blood vessel; Artery; Vein; Vascular
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Eye Examination Apparatus (AREA)
Abstract
The present invention provides a kind of full-automatic retinal vessel analysis method and system suitable for intelligent and portable equipment, the present invention improves significant characteristics method by increasing pre-treatment step, the pretreatment includes eye fundus image is carried out uniforming processing and removes blood vessel center light reflection processing, realize the Fast Segmentation of retinal vessel, method has that stability is good, computing is simple, fireballing advantage.The present invention realizes the Accurate Analysis of a variety of retinal vessels, and pass through optimized algorithm complexity and arithmetic speed, realize and the combination of intelligent and portable equipment, computer-aided diagnosis provided strong applicability in retinal image analysis, calculated systemic blood vessels analysis stable, that accuracy is high.
Description
Technical field
The invention belongs to computer-aided diagnosis field, and in particular to suitable for the full-automatic retina of intelligent and portable equipment
Vessels analysis method and system.
Background technology
The whole world has more than 400,000,000 diabetic, to diabetic, even High-risk Group of Diabetes carry out in a organized way,
Periodically, large-scale eye fundus image examination is the method that current American-European countries generally uses.《Type 2 Diabetes In China guideline of prevention and treatment》
Also indicate that, diabetic once makes a definite diagnosis, and should just carry out annual at least 1 time funduscopy；Diabetic once occurs tight
Pay attention to the symptom of retinopathy, then should carry out within every 2~4 months 1 eyeground check.By doctor to coloured image in clinical at present
Examination one by one, this process workload is larger, and efficiency is low, and subjectivity is larger, therefore needs a kind of automatically diabetes badly and regard
Retinopathy screening method.
The change of retinal vessel is an importance of diabetic retinopathy.It is continuous with computer technology
Development is improved, the method for many retinal vessel quantitative analyses occurs.However, due to by shooting condition, individual differences with
And the influence of lesion, eye fundus image, which exists, exposes the problems such as uneven, easily difference is larger between by noise jamming, image, method
Precision it is relatively low.In addition, retinal vessel analysis bag contains blood vessel segmentation, measurement, distinguishes the multinomial work such as artery and vein, but
Prior art is only capable of realizing single content, can not realize the analysis of system, be restricted very much in the application.On the other hand, it is
It is easy to the use in remote districts, basic hospital, family etc., it is necessary to platform being easily obtained, and existing method is most
Using mainframe computer or work station as platform, high is required to computing hardware, and lack friendly, for user interface.
Traditional blood vessel segmentation method based on graph theory model uses three-dimensional graph theory model, requires high to computing, it is difficult to will
It is transplanted to portable equipment, greatly limit the use range of its method.
The content of the invention
It is an object of the invention to provide a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment
It is real and system, the present invention realize a variety of Accurate Analysis of retinal vessel, and by optimized algorithm complexity and arithmetic speed
Showed and the combination of intelligent and portable equipment, make computer-aided diagnosis can be provided in retinal image analysis strong applicability,
Calculate systemic blood vessels analysis stable, that accuracy is high.
To reach above-mentioned purpose, the technical scheme that the present invention takes is：
A kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment, comprise the following steps：
Step 1：Retinal vessel segmentation based on significant characteristics, including pretreatment, multiple dimensioned frequency-domain residual feature carry
Take, morphological feature extraction, Directional feature extraction, self-information feature extraction, Fusion Features and Threshold segmentation；The present invention is to original
Blood vessel segmentation method based on significant characteristics is improved, and the accurate of blood vessel segmentation is improved by the homogenization to image
Rate.
1.1) pre-process：Including image homogenization and remove the reflection of blood vessel center light.Image homogenization is divided into image again
Uniformed between homogenization and image.In being uniformed in image, Homogenization Treatments are carried out to background with high-pass filter, specially made
High-frequency information is removed with large scale Gaussian filter (sigma=3~5), it is poor then to make with original image.In image interior background
After homogenization, the homogenization between image is used between the image of elimination Different Individual (such as different race, sexs, age)
Difference.Specific method is the grey level histogram of tri- passages of R, G, B for the original image for obtaining background uniformity first and divided
Intermediate value is not calculated.In each channel using brightness curve become change commanders intermediate value be moved to desirable intensity value ranges midpoint it is (excellent
Select on 128).
Image after image homogenization processing is coloured image CI, coloured image CIOpened through Morphological Grayscale in operation removal
Heart light reflects, and obtains coloured image CII；
1.2) multiple dimensioned frequency-domain residual feature extraction：Coloured image C first after extraction pretreatmentIIGreen channel
(G) multiple dimensioned gaussian pyramid, is generated, Fourier's change is carried out using the image on each yardstick of gaussian pyramid as input
Change, obtain phase and amplitude；Obtained composing residual error in a frequency domain according to amplitude, residual error knot by phase and is composed by inversefouriertransform
Conjunction obtains strength characteristic image, after obtaining multiple dimensioned strength characteristic, the strength characteristic on different scale is merged to obtain most
Whole frequency-domain residual characteristic image；
1.3) morphological feature extracts：To coloured image CIIGreen channel (G) carry out bottom cap conversion extraction blood vessel, specifically
The practice for first pass through closed operation delete blood vessel, then by with original color image CIIGreen channel (G) make the difference, obtain morphology spy
Levy image；
1.4) Directional feature extraction：To coloured image CIIGreen channel (G) carry out two-dimensional Gabor filtering, at interval of one
Determine one Gabor operator image of angle calculation, respectively with coloured image CIIGreen channel (G) do convolution and obtain different directions
Blood vessel enhancing figure, then the blood vessel enhancing figure of different directions is subjected to image co-registration, obtain final direction character image；
1.5) self-information feature extraction：Coloured image C is obtained firstIIRed channel (R), green channel (G) and blueness
The average gray image of passage (B), then on R channel images, G channel images, channel B image, average gray image respectively
Do grey level histogram and calculate self-information feature, R, G, channel B image and self-information feature corresponding to average gray image are passed through
Fusion obtains final self-information characteristic image；
1.6) Fusion Features and Threshold segmentation：Obtain image (i.e. frequency-domain residual characteristic image, the form of four feature enhancings
Learn characteristic image, direction character image, self-information characteristic image) after Weighted Fusion, then employ triangle it is global from
Adapt to threshold segmentation method and carry out vessel extraction.
Step 2：The accurate measurement of retinal blood vessels caliber, include foundation, the model optimization of two-dimentional graph theory model, and blood
Pipe diameter measures.
2.1) based on vessel segmentation, vascular skeleton is obtained by binaryzation, blood vessel is extracted according to vascular skeleton
Center line, remove the branch point on center line and crosspoint；Then based on vessel centerline, established along blood vessel bearing of trend
The two-dimentional Mathematical Model of Graph Theory of blood vessel；Cost function extracts the shade of gray of blood vessel using Steerable filter device；
2.2) by the solution of the maximum flow problem to two-dimentional Mathematical Model of Graph Theory, obtained in two-dimentional graph theory model optimal
Cut-off rule, the exact boundary of blood vessel is just obtained, the distance between border is calculated along the vertical direction of blood vessel bearing of trend, so that it may
To obtain blood vessel diameter.
Step 3：Based on blood vessel segmentation, using convolutional neural networks method, point of progress artery and vein
Class, including collection training image, training pattern, test three parts.
3.1) training image is gathered：In the training stage, the hand labeled artery of 20~40 eye fundus images (is labeled as A
Class) and vein (being labeled as B classes).Then (present invention is with area vasculosa for angiosomes corresponding to artery and vein on eye fundus image
Domain is ROI) generate training image with sliding window (window size is 7~20 × 7~20) traversal；
3.2) convolutional neural networks model training：The present invention uses AlexNet network structures, and network depth is 8~10 layers
(preferably 8 layers), calculating consumption can be significantly reduced, and because characteristics of image is relatively single, single training set size is 300~
500, Learning Step is 0.0001~0.0002, and training is not less than 40 cycles of training；
3.3) test：Blood vessel segmentation (i.e. step 1) is obtained first with improved significant characteristics method.Then in ROI
The test image of interior structure and training image formed objects centered on each pixel, the input of all test images is trained
Model obtain test image label (probability for A classes and be B classes probability), using this label as foundation, it is determined that test
The half-tone information of image medium vessels, rebuild the artery and vein classification chart of blood vessel.
A kind of full-automatic retinal vessel analysis system suitable for intelligent and portable equipment, including intelligent and portable equipment and
The full-automatic retinal vessel analysis App (Application, mobile applications) built in the equipment, is implanted into App
Blood vessel segmentation and diameter measuring method (step 1 and step 2)；Remote transmission function is set, blood can be achieved at remote computation end
Pipe segmentation, classification (step 3) and diameter measurement；Operated for the ease of user, this App includes account management, getImage, figure
The functional modules such as picture is analyzed, result shows, stored, transmitting.
The present invention has the advantages that compared to prior art：
The present invention realizes the systematic analytic method for the retinal vessel that may migrate to intelligent and portable equipment, to conventional method
Largely simplified, ensure the stability and accuracy of computing while improving arithmetic speed.The present invention passes through
Using the method (i.e. increase pretreatment) of improved significant characteristics, the Fast Segmentation of retinal vessel is realized, method has
Stability is good, computing is simple, fireballing advantage.
Further, present invention improves over the design of graph theory model, threedimensional model is decomposed into two two dimensional models, protected
Computational complexity is greatly reduced while staying accuracy, and successfully this method is transplanted on portable equipment.
Further, arteria retina and vein are classified using convolutional neural networks method, significantly improved point
The accuracy of class.The present invention greatly reduces the computing of model by the improvement to network structure while accuracy is ensured
Complexity, success implementation method is from work station to the transfer of removable portable equipment.
Brief description of the drawings
Fig. 1 is the workflow diagram of present system.
Fig. 2 is the blood vessel segmentation flow chart based on significance analysis.
Fig. 3 is the blood vessel diameter measurement procedure figure based on graph theory.
Fig. 4 is eye fundus image vascular arteriovenous classification process figure.
Fig. 5 is mobile phone app interface schematic diagrams.
Embodiment
Below in conjunction with the accompanying drawings and embodiment is described in further details to the present invention.
As shown in figure 1, suitable for the full-automatic retinal vessel analysis method of intelligent and portable equipment, comprise the following steps：
Step 1：Original image is pre-processed, extracts a variety of significant characteristics of image, carries out retinal vessel point
Cut；
Step 2：Vascular skeleton is extracted from vessel segmentation, and combines original image information, establishes retinal vessel
Graph theory model, measure the diameter of blood vessel；
Step 3：The deep learning sorting technique of arteria retina and vein is established, blood vessel point is being carried out to original image
On the basis of cutting, the classification of artery and vein is realized using deep learning sorting technique.
In the first step, the method that the present invention proposes the improved significant characteristics for eye fundus image blood vessel segmentation.
Because the reasons such as age, disease, race have background difference between the image of eye fundus image, additionally due to its image-forming principle can also be made
Into the uneven illumination one in image.Existing blood vessel segmentation method can not all solve this problem well.Significant characteristics
Method has the advantages of computational complexity is low, arithmetic speed is fast, but is influenceed by above-mentioned picture quality, existing conspicuousness
Characterization method accuracy is not high.The present invention has carried out homogenization processing before significant characteristics are extracted to image, so as to significantly
Improve the efficiency of blood vessel segmentation.Blood vessel segmentation method of the present invention specifically includes following steps：Pretreatment, feature extraction, information
Fusion and four steps of Threshold segmentation.
1.1：Pretreatment includes image homogenization and removes two parts of blood vessel center light reflection
Original image (eye fundus image for being is marked in Fig. 2) is carried on the back with a high-pass filter first
The Homogenization Treatments of scape.Convolution specially is done with large-sized Gaussian filter (sigma=5) and original image, eliminates height
Frequency information, only retain low-frequency information, it is poor then to make with original image.Formula is as follows：
In formula, (x, y) is pixel point coordinates,It is Gaussian filter, I (x, y) is original image, and R (x, y) is that background is equal
Image after homogenizing.Normalization inside image between homogenization processing and then progress image, it is therefore an objective to eliminate what is newly inputted
Difference between original image and the original image analyzed, for example, Different Individual (different ethnic groups, age, sex etc.)
Difference between image.Specific method is that red, green, the ash of blue three passages are obtained on the image of background uniformity
Histogram is spent, and calculates intermediate value respectively.Then brightness curve change is utilized respectively in each channel to change commanders under respective channel
Intermediate value be moved on unified midpoint (gray value 128).In this conversion, point of the gray value less than 0 is arranged to 0, gray scale
Point of the value more than 255 is arranged to 255.
On eye fundus image, blood vessel center is commonly present a bright striped, and referred to as center light reflects, and it is to blood vessel segmentation
There is certain influence, should be removed.For after the homogenization processing between image interior background homogenization and image
Eye fundus image, blood vessel is dark compared with background, therefore the result corroded make it that blood vessel is thicker, and center light reflection is eliminated；Again
Blood vessel is set to return to original width by expansion so that blood vessel returns to original width.Expansion and the radius of erosion operator
It is 3~5 (preferably 3) pixels.
1.2：Significant characteristics extraction includes frequency-domain residual feature, direction character, morphological feature and self-information feature and carried
Take four parts (Fig. 2)
Frequency-domain residual feature extraction：The green channel (G) of original image after pretreatment is generated into multiple dimensioned Gauss gold
Word tower Iδ(x, y), δ=0,1,2 ... n, yardstick is represented, whereinL represents the short side length of image, and n represents multiple dimensioned
Yardstick number, it is specific generate gaussian pyramid method it is as follows：
In formulaFor Gaussian filter.
Obtain composing residual error in each yardstick of gaussian pyramid, concrete operations are as follows：
Rδ(u, v)=ln (Aδ(u,v))-ln(Aδ(u,v))*h
In formula, i is imaginary symbols, and (u, v) is the pixel point coordinates under frequency domain, Fδ(u, v) is Fourier transformation result
Amplitude, Aδ(u, v) is phase, and h is mean filter, Rδ(u, v) is the spectrum residual error on δ yardsticks.
Phase and spectrum residual error are combined and obtain strength characteristic image by inversefouriertransform：
A series of (corresponding δ) strength characteristic image f are finally givenδ(x,y).Obtain fδAfter (x, y), by different scale
Image carries out being upsampled to original scale, and the image of the different scale after up-sampling then is carried out into gray scale is averaged, so as to obtain
Final frequency-domain residual characteristic image (feature enhancing image E1)。
Morphological feature extracts：The green channel (G) of original image after pretreatment is become using bottom cap and brings extraction
Go out blood vessel (feature enhancing image E2).Specific formula is as follows：
In formula, f is the green channel (G) of the original image after pretreatment, and b is structural detail," Θ " difference
Refer to expansion and etching operation.
Directional feature extraction：One prescription is generated to different from the green channel (G) of the original image after pretreatment
Gabor operator images, the expression formula of Gabor filtering are as follows：
Wherein, (x, y) is pixel point coordinates, and i is imaginary symbols,For the frequency of sinusoidal signal, σxAnd σyRespectively
The standard deviation of Gabor filter in the x and y direction.
20 ° of generations, one Gabor operator image is often rotated from 0 ° to 160 °, totally nine, the size of image is maximum blood vessel
1.3~1.5 times of diameter, the present embodiment value are 16 × 16 pixels.After the Gabor operators image of generation and pretreatment
Green channel (G) convolution of original image, obtains the blood vessel enhancing figure of different directions：
In formula,For direction character image (i.e. blood vessel enhancing figure), G (x, y) is coloured image CIIGreen channel (G),
γ is Gabor operator images.After the blood vessel enhancing figure of different directions is obtained, image is merged, completely regarded
Retinal vasculature.The specific practice is, for each pixel (coordinate is (x, y)) in the original image after pretreatment, selection
The maximum gray value of the point is as fused images (feature enhancing image E on nine different directions characteristic images3) gray value.
Self-information feature：The red channel (R), green channel (G), blueness for obtaining pretreated original image first are logical
Road (B) and average gray level image AVE=(R+G+B)/3, the altogether grey level histogram of four passages, and calculate each gray scale respectively
The frequency that level occurs：
In formula, nkRepresent the frequency that gray level k occurs, rkK-th of gray level is represented, M and N are the images represented with pixel
Length and width.
After obtaining gray scale frequency, the self-information of each grey level is calculated, computational methods are as follows：
L(rk)=- ln (p (rk))
So, the less grey level of probability of occurrence can be endowed higher gray value；Likewise, probability of occurrence is larger
Grey level is endowed relatively low gray value.Finally, by the gray value of each pixel in original image possessed by the gray value
Self-information L (rk) replace.Due in eye fundus image, the number of the pixel that the grey level where some thin vessels is included compared with
Few, after above-mentioned processing, some thin vessels can obtain a certain degree of enhancing.Red channel is obtained in the same way
(R), the self-information image of green channel (G), blue channel (B) and average gray level image AVE, respectively SR,SG,SB,SAVE, and
Fusion is weighted, obtains final feature enhancing image E4。
E4=w1SR+w2SG+w3SB+w4SAVE
In formula, h1,h2,h3, and h4Respectively image SR,SG,SB,SAVETotal gray value.
1.3：The image of four feature enhancings is weighted fusion
Fusion method is following (Fig. 2)：Calculate the gray value sum of all pixels point on every image respectively first, i.e., total ash
Angle value；Then total gray value of each image ratio shared in four images is calculated；By the negative of the counted ratio of previous step
Weight w of the logarithm value as the figure, utilizes formulaObtain Saliency maps picture.
1.4：Threshold segmentation
After Saliency maps picture is generated, the blood vessel on eye fundus image has obtained effective enhancing, and noise information obtains
Suppress, then can be so that retinal vessel to be split by the method for Threshold segmentation.Triangle is employed in the present invention
Global adaptive threshold fuzziness method, enters after row threshold division, obtains blood vessel segmentation figure (output image).
In second step, blood vessel diameter measuring method is referring to Fig. 3.After row threshold division is entered, as a result may there is two
Individual problem：One is that the border of angiosomes has a circle non-vascular to be detected as blood vessel；Another is present on image
Noise does not remove completely.AND operation, which is done, for the two problems present invention with mask images and blood-vessel image first removes trimming
Boundary's (binaryzation), then carries out denoising to bianry image, and specific method is less than dividing for 10 pixels to remove area on bianry image
Cut region.After obtaining final blood vessel segmentation, the figure undercutting row mathematical modeling in Graph-theoretical Approach is employed.Specific practice is,
Vessel centerline is extracted to blood vessel segmentation figure first.In the blood vessel on heart line image, branch point and crosspoint are eliminated.Remove this
After a little points, whole vascular tree becomes some Vessel sections, and the breakpoint of these fragments only has an adjacent pixel.From one
End points sets out, and tracking blood vessel is marked to another end points, and to each Vessel sections.
Then the two-dimentional Mathematical Model of Graph Theory of blood vessel is established based on vessel centerline.For the blood vessel of each mark
Fragment, the extension of each centerline pixels is calculated with PCA using every side m adjacent pixels of centerline pixels
Direction.Parameter m value is decided by the size of image, about the 0.005~0.007 of the image length of side times.Along bearing of trend
Two two-dimentional graph theory models are respectively created in the opposite direction of normal direction and normal direction.Specific method is with each centerline pixels
For starting point, establishing model node along the opposite direction of normal direction or normal direction, (node separation is 0.1~0.2 pixel, node
Number is 150~200).Connected between the model node that adjacent center pixel is established by unidirectional side.
In the selection of cost function, the present invention extracts the shade of gray of blood vessel using Steerable filter device on the original image
Information.The cost function of sub-pix node is realized by linear difference.By the adjustment to model parameter, can further improve
The smooth degree of vascular wall.In order to find the gradient in the particular blood vessel of different places different directions, with an one-dimensional Gauss
Function does convolution algorithm.Gaussian function is as follows：
Wherein, σ is the standard deviation of Gaussian function, by the solution of the maximum flow problem to graph theory model, is tied according to solving
Fruit can obtain the exact boundary of blood vessel and the bearing of trend of blood vessel.By finding blood along the vertical direction of blood vessel bearing of trend
Tube edge circle, the accurate diameter of blood vessel can be obtained.
In 3rd step, the flow of arteriovenous classification is referring to Fig. 4, and first, the method manually split is opened one's eyes base map 20
The classification true value of artery (being labeled as A classes) and vein (being labeled as B classes) is established as on.In the training stage, adaptive threshold is utilized
The method of segmentation finds image ROI region.In region institute is traveled through using 9 × 9 sliding window a little, and using each point in
The heart creates 9 × 9 training image, and the true value of training image obtains from corresponding manual classification image, obtains 3960494 altogether
A class training images, 577649 B class training images.Present invention employs the Tensorflow1.0 versions of Google and
Modularization deep learning storehouse TFLearn transparent TensorFlow.Computer core equipment is Intel Duo i7 6700k CPU
And the GPU (8GB) of NVIDIA GeForce GTX 1080.
For solving the problems, such as that data volume is less (only 20 eye fundus images), present invention employs the side of transfer learning
Method, by transfer learning natural image data set ImageNet, fine setting learns last classification layer parameter, improves the accuracy of classification
(other classification layers are learnt using natural image data set, and layer of finally classifying is learnt using training image).
In network selection, AlexNet networks are have selected, the network only has 8 layers, can significantly reduce calculating consumption, and because image is special
Sign is relatively single, and the network can ensure accuracy.Preferable single training set is 400, Learning Step 0.0001, training 40
Individual cycle of training, the accuracy of training is 91%.
In test phase, the ROI region of image after blood vessel segmentation is found using adaptive threshold method, with ROI region
Point centered on, establish 9 × 9 test image, input the deep learning model established, obtain label, and rebuild final
Test result, specific practice be, vascular group is judged with the magnitude relationship of the respective probability of A classes in this label and B classes, then with ash
Angle value rebuilds the artery and vein classification chart of blood vessel, for example, the gray value of figure medium sized vein is 1, the gray value of artery is 2.
Graphic user interface ginsengs of the full-automatic retinal vessel analysis App on portable terminal device, such as Android smartphone
See Fig. 5.This app includes following module.1) account management module：System allows multiple users to make on same mobile phone
With each user can obtain account and password by registering (Register), carry out user's login before analysis first
(Login).2) it is loaded into eye fundus image module：Image (Download an image) or logical can be loaded into from mobile phone picture library
Cross to take pictures and obtain eye fundus image.3) automated image analysis module：This system can complete automatic graphical analysis, including blood vessel point
Cut and measured with blood vessel diameter.4) image transmitting and remote analysis module：For arteriovenous classification, or for the accuracy of mobile phone compared with
It is low to realize offline blood vessel segmentation and the situation of diameter measurement, it can select to transmit original image to remote work station,
Realize graphical analysis on work station, including it is blood vessel segmentation (VesselSegmention), diameter measurement (WidthMeasure), dynamic
Vein is classified, and result can be back into portable terminal device.5) display of result and memory module：The result of image procossing can be
Show and store on mobile phone, and can be sent by way of Email.6) operation instruction module：Provided in this module in detail
Thin operating instruction and the explanation to result.
Claims (10)
- A kind of 1. full-automatic retinal vessel analysis method suitable for intelligent and portable equipment, it is characterised in that：The analysis method Comprise the following steps：1) significant characteristics extraction is carried out after being pre-processed to eye fundus image, then passes sequentially through Fusion Features and threshold value point Cut, obtain blood vessel segmentation image；The pretreatment includes eye fundus image is carried out uniforming processing and removal blood vessel center light is anti- Penetrate processing；2) according to blood vessel segmentation image, retinal blood vessels caliber is measured using two-dimentional graph theory model.
- 2. a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment according to claim 1, it is special Sign is：In the step 1), homogenization processing comprises the following steps：First, eye fundus image is carried out uniforming in image, made Eye fundus image background uniformity, uniformed then carrying out image to the eye fundus image of background uniformity, return the eye fundus image One changes.
- 3. a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment according to claim 2, it is special Sign is：Homogenization specifically includes following steps in described image：Removed using the large scale Gaussian filter that sigma is 3~5 The high-frequency information of original image, it is poor then to make with original image, and the original image is the eye fundus image without pretreatment；Institute Homogenization specifically includes following steps between stating image：Tri- passages of R, G, B of the original image of background uniformity are obtained first Grey level histogram simultaneously calculates intermediate value respectively, and then becoming intermediate value of changing commanders using brightness curve in each channel is moved to desirable ash On the midpoint of angle value scope.
- 4. a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment according to claim 1, it is special Sign is：In the step 1), remove blood vessel center light reflection processing and comprise the following steps：By the eyeground by homogenization processing Image carries out Morphological Grayscale and opens operation.
- 5. a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment according to claim 1, it is special Sign is：In the step 1), significant characteristics extraction includes frequency-domain residual feature, direction character, morphological feature and confidence Cease four parts of feature extraction.
- 6. a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment according to claim 1, it is special Sign is：The analysis method is further comprising the steps of：The deep learning sorting technique of arteria retina and vein is established, right On the basis of eye fundus image carries out blood vessel segmentation, the classification of artery and vein is realized using deep learning sorting technique.
- 7. a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment according to claim 6, it is special Sign is：The deep learning sorting technique comprises the following steps：Deep learning is established by the training to convolutional neural networks Model, the training image of use are to utilize sliding window by the eye fundus image of the artery to 20~40 hand labeleds and vein Traveled through and generated；After the eye fundus image being analysed to carries out blood vessel segmentation according to step 1), with each picture of angiosomes Structure and the test images of training image formed objects centered on vegetarian refreshments, deep learning model is inputted by test image, is obtained pair Answer the label of test image, the label be the image medium vessels be the probability of artery and be vein probability, made with this label For foundation, the half-tone information of test image medium vessels is determined, the artery of retinal vessel is rebuild according to half-tone information and vein divides Class figure.
- 8. a kind of full-automatic retinal vessel analysis method suitable for intelligent and portable equipment according to claim 7, it is special Sign is：The convolutional neural networks use following parameter in training：Network structure is 8~10 layers, single training set size For 300~500, Learning Step is 0.0001~0.0002, and training is not less than 40 cycles of training.
- A kind of 9. full-automatic retinal vessel analysis system suitable for intelligent and portable equipment, it is characterised in that：Including full-automatic Retinal vessel analysis module, the analysis module include being loaded into eye fundus image submodule, automated image analysis submodule and knot The display of fruit and sub-module stored；Eye fundus image submodule is loaded into be used to be loaded into image or logical from intelligent and portable equipment picture library Cross intelligent and portable equipment camera function and obtain eye fundus image；Automated image analysis submodule be used for intelligent and portable equipment or with this Equipment is carried out on the work station of telecommunication, and retinal vessel segmentation, diameter measurement and arteriovenous point are carried out for eye fundus image Class；As a result display and sub-module stored is used to the result of blood vessel segmentation, diameter measurement and arteriovenous classification is shown and stored In intelligent and portable equipment.
- 10. a kind of full-automatic retinal vessel analysis system suitable for intelligent and portable equipment according to claim 9, it is special Sign is：The blood vessel segmentation comprises the following steps：Significant characteristics extraction is carried out after being pre-processed to eye fundus image, then Fusion Features and Threshold segmentation are passed sequentially through, obtains blood vessel segmentation image；The pretreatment includes eye fundus image is carried out uniforming processing and removes blood vessel center light reflection processing；The diameter is surveyed Amount comprises the following steps：According to blood vessel segmentation image, retinal blood vessels caliber is measured using two-dimentional graph theory model；The arteriovenous classification comprises the following steps：On the basis of blood vessel segmentation is carried out to eye fundus image, deep learning is utilized Sorting technique realizes the classification of artery and vein.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710970423.XA CN107657612A (en) | 2017-10-16 | 2017-10-16 | Suitable for full-automatic the retinal vessel analysis method and system of intelligent and portable equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710970423.XA CN107657612A (en) | 2017-10-16 | 2017-10-16 | Suitable for full-automatic the retinal vessel analysis method and system of intelligent and portable equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107657612A true CN107657612A (en) | 2018-02-02 |
Family
ID=61119124
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710970423.XA Pending CN107657612A (en) | 2017-10-16 | 2017-10-16 | Suitable for full-automatic the retinal vessel analysis method and system of intelligent and portable equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107657612A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108376048A (en) * | 2018-03-05 | 2018-08-07 | 四川和生视界医药技术开发有限公司 | The restorative procedure and prosthetic device of retinal vessel |
| CN108492302A (en) * | 2018-03-26 | 2018-09-04 | 北京市商汤科技开发有限公司 | Nervous layer dividing method and device, electronic equipment, storage medium, program |
| CN108550136A (en) * | 2018-03-14 | 2018-09-18 | 宁波大学 | A kind of eye fundus image blood vessel segmentation method |
| CN109360209A (en) * | 2018-09-30 | 2019-02-19 | 语坤（北京）网络科技有限公司 | A kind of coronary vessel segmentation method and system |
| CN109671094A (en) * | 2018-11-09 | 2019-04-23 | 杭州电子科技大学 | A kind of eye fundus image blood vessel segmentation method based on frequency domain classification |
| CN110276333A (en) * | 2019-06-28 | 2019-09-24 | 上海鹰瞳医疗科技有限公司 | Eyeground identification model training method, eyeground personal identification method and equipment |
| CN110751636A (en) * | 2019-10-12 | 2020-02-04 | 天津工业大学 | Fundus image retinal arteriosclerosis detection method based on improved coding and decoding network |
| CN111000563A (en) * | 2019-11-22 | 2020-04-14 | 北京理工大学 | Automatic measuring method and device for retinal artery and vein diameter ratio |
| CN111222361A (en) * | 2018-11-23 | 2020-06-02 | 福州依影健康科技有限公司 | Method and system for analyzing hypertension retina vascular change characteristic data |
| CN111954485A (en) * | 2018-04-18 | 2020-11-17 | 株式会社尼康 | Image processing method, program, image processing apparatus, and ophthalmologic system |
| CN112075922A (en) * | 2020-10-14 | 2020-12-15 | 中国人民解放军空军军医大学 | Method for measuring fundus image indexes of type 2 diabetes mellitus and analyzing correlation between fundus image indexes and diabetic nephropathy |
| CN112233789A (en) * | 2020-10-12 | 2021-01-15 | 辽宁工程技术大学 | Regional feature fusion type hypertensive retinopathy classification method |
| CN112862804A (en) * | 2021-03-01 | 2021-05-28 | 河南科技大学第一附属医院 | System and method for processing retina blood vessel image |
| CN113269737A (en) * | 2021-05-17 | 2021-08-17 | 西安交通大学 | Method and system for calculating diameter of artery and vein of retina |
| WO2021169128A1 (en) * | 2020-02-29 | 2021-09-02 | 平安科技（深圳）有限公司 | Method and apparatus for recognizing and quantifying fundus retina vessel, and device and storage medium |
| CN113724857A (en) * | 2021-08-27 | 2021-11-30 | 清华大学深圳国际研究生院 | Automatic diagnosis device for eye ground disease based on eye ground image retina blood vessel |
| TWI755141B (en) * | 2020-11-04 | 2022-02-11 | 財團法人國家實驗研究院 | Convolutional neural networks method and system for image analysis |
| CN114708390A (en) * | 2022-05-25 | 2022-07-05 | 深圳科亚医疗科技有限公司 | Image processing method and device for physiological tubular structure and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120236259A1 (en) * | 2011-01-20 | 2012-09-20 | Abramoff Michael D | Automated determination of arteriovenous ratio in images of blood vessels |
| CN104732499A (en) * | 2015-04-01 | 2015-06-24 | 武汉工程大学 | Retina image enhancement algorithm based on multiple scales and multiple directions |
| CN106683080A (en) * | 2016-12-15 | 2017-05-17 | 广西师范大学 | Retinal fundus image preprocessing method |
-
2017
- 2017-10-16 CN CN201710970423.XA patent/CN107657612A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120236259A1 (en) * | 2011-01-20 | 2012-09-20 | Abramoff Michael D | Automated determination of arteriovenous ratio in images of blood vessels |
| CN104732499A (en) * | 2015-04-01 | 2015-06-24 | 武汉工程大学 | Retina image enhancement algorithm based on multiple scales and multiple directions |
| CN106683080A (en) * | 2016-12-15 | 2017-05-17 | 广西师范大学 | Retinal fundus image preprocessing method |
Non-Patent Citations (6)
| Title |
|---|
| XIAYU XU: "An improved arteriovenous classification method for the early diagnostics of various diseases in retinal image", 《COMPUTER METHODS AND PROGRAMS IN BIOMEDICINE》 * |
| XIAYU XU: "Smartphone-Based Accurate Analysis of Retinal Vasculature towards Point-of-Care Diagnostics", 《SCIENTIFIC REPORTS》 * |
| 张大坤: "《软件技术及系统现状与应用前景》", 31 May 2015 * |
| 杨毅: "视网膜血管分割与动静脉分类方法研究", 《中国硕士学位论文全文数据库信息科技辑》 * |
| 赵中堂: "《基于智能移动终端的行为识别方法研究》", 30 April 2015 * |
| 郭思雨: "《面向检测的图像处理技术》", 31 August 2015, 长沙：湖南大学出版社 * |
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108376048B (en) * | 2018-03-05 | 2021-05-28 | 四川和生视界医药技术开发有限公司 | Retinal blood vessel repair method and repair device |
| CN108376048A (en) * | 2018-03-05 | 2018-08-07 | 四川和生视界医药技术开发有限公司 | The restorative procedure and prosthetic device of retinal vessel |
| CN108550136A (en) * | 2018-03-14 | 2018-09-18 | 宁波大学 | A kind of eye fundus image blood vessel segmentation method |
| CN108492302A (en) * | 2018-03-26 | 2018-09-04 | 北京市商汤科技开发有限公司 | Nervous layer dividing method and device, electronic equipment, storage medium, program |
| CN111954485A (en) * | 2018-04-18 | 2020-11-17 | 株式会社尼康 | Image processing method, program, image processing apparatus, and ophthalmologic system |
| CN111954485B (en) * | 2018-04-18 | 2024-08-23 | 株式会社尼康 | Image processing method, computer-readable storage medium, and ophthalmic system |
| CN109360209A (en) * | 2018-09-30 | 2019-02-19 | 语坤（北京）网络科技有限公司 | A kind of coronary vessel segmentation method and system |
| CN109671094A (en) * | 2018-11-09 | 2019-04-23 | 杭州电子科技大学 | A kind of eye fundus image blood vessel segmentation method based on frequency domain classification |
| CN109671094B (en) * | 2018-11-09 | 2023-04-18 | 杭州电子科技大学 | Fundus image blood vessel segmentation method based on frequency domain classification |
| CN111222361A (en) * | 2018-11-23 | 2020-06-02 | 福州依影健康科技有限公司 | Method and system for analyzing hypertension retina vascular change characteristic data |
| CN111222361B (en) * | 2018-11-23 | 2023-12-19 | 福州依影健康科技有限公司 | Method and system for analyzing characteristic data of change of blood vessel of retina in hypertension |
| CN110276333A (en) * | 2019-06-28 | 2019-09-24 | 上海鹰瞳医疗科技有限公司 | Eyeground identification model training method, eyeground personal identification method and equipment |
| CN110751636A (en) * | 2019-10-12 | 2020-02-04 | 天津工业大学 | Fundus image retinal arteriosclerosis detection method based on improved coding and decoding network |
| CN110751636B (en) * | 2019-10-12 | 2023-04-21 | 天津工业大学 | Fundus image retinal arteriosclerosis detection method based on improved coding and decoding network |
| CN111000563A (en) * | 2019-11-22 | 2020-04-14 | 北京理工大学 | Automatic measuring method and device for retinal artery and vein diameter ratio |
| WO2021169128A1 (en) * | 2020-02-29 | 2021-09-02 | 平安科技（深圳）有限公司 | Method and apparatus for recognizing and quantifying fundus retina vessel, and device and storage medium |
| CN112233789A (en) * | 2020-10-12 | 2021-01-15 | 辽宁工程技术大学 | Regional feature fusion type hypertensive retinopathy classification method |
| CN112075922B (en) * | 2020-10-14 | 2024-09-10 | 中国人民解放军空军军医大学 | Method for measuring fundus image index of type 2 diabetes and analyzing correlation with diabetic nephropathy |
| CN112075922A (en) * | 2020-10-14 | 2020-12-15 | 中国人民解放军空军军医大学 | Method for measuring fundus image indexes of type 2 diabetes mellitus and analyzing correlation between fundus image indexes and diabetic nephropathy |
| TWI755141B (en) * | 2020-11-04 | 2022-02-11 | 財團法人國家實驗研究院 | Convolutional neural networks method and system for image analysis |
| CN112862804A (en) * | 2021-03-01 | 2021-05-28 | 河南科技大学第一附属医院 | System and method for processing retina blood vessel image |
| CN112862804B (en) * | 2021-03-01 | 2023-04-07 | 河南科技大学第一附属医院 | System and method for processing retina blood vessel image |
| CN113269737B (en) * | 2021-05-17 | 2024-03-19 | 北京鹰瞳科技发展股份有限公司 | Fundus retina artery and vein vessel diameter calculation method and system |
| CN113269737A (en) * | 2021-05-17 | 2021-08-17 | 西安交通大学 | Method and system for calculating diameter of artery and vein of retina |
| CN113724857A (en) * | 2021-08-27 | 2021-11-30 | 清华大学深圳国际研究生院 | Automatic diagnosis device for eye ground disease based on eye ground image retina blood vessel |
| CN114708390B (en) * | 2022-05-25 | 2022-09-20 | 深圳科亚医疗科技有限公司 | Image processing method and device for physiological tubular structure and storage medium |
| CN114708390A (en) * | 2022-05-25 | 2022-07-05 | 深圳科亚医疗科技有限公司 | Image processing method and device for physiological tubular structure and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107657612A (en) | Suitable for full-automatic the retinal vessel analysis method and system of intelligent and portable equipment | |
| EP3674968B1 (en) | Image classification method, server and computer readable storage medium | |
| US20200250491A1 (en) | Image classification method, computer device, and computer-readable storage medium | |
| CN111815574B (en) | Fundus retina blood vessel image segmentation method based on rough set neural network | |
| CN108765363B (en) | Coronary artery CTA automatic post-processing system based on artificial intelligence | |
| WO2021068523A1 (en) | Method and apparatus for positioning macular center of eye fundus image, electronic device, and storage medium | |
| CN112132817B (en) | Retina blood vessel segmentation method for fundus image based on mixed attention mechanism | |
| CN108764342B (en) | Semantic segmentation method for optic discs and optic cups in fundus image | |
| CN109087302A (en) | A kind of eye fundus image blood vessel segmentation method and apparatus | |
| CN109584209A (en) | Vascular wall patch identifies equipment, system, method and storage medium | |
| CN108764286A (en) | The classifying identification method of characteristic point in a kind of blood-vessel image based on transfer learning | |
| CN111667490B (en) | Fundus picture cup optic disc segmentation method | |
| CN113269737B (en) | Fundus retina artery and vein vessel diameter calculation method and system | |
| Hu et al. | Automatic artery/vein classification using a vessel-constraint network for multicenter fundus images | |
| CN111797901A (en) | Retinal artery and vein classification method and device based on topological structure estimation | |
| CN109658423A (en) | A kind of optic disk optic cup automatic division method of colour eyeground figure | |
| Guo et al. | BSCN: bidirectional symmetric cascade network for retinal vessel segmentation | |
| Wang et al. | Towards an extended EfficientNet-based U-Net framework for joint optic disc and cup segmentation in the fundus image | |
| CN102567734A (en) | Specific value based retina thin blood vessel segmentation method | |
| CN109087310A (en) | Dividing method, system, storage medium and the intelligent terminal of Meibomian gland texture region | |
| CN114972365A (en) | OCT image choroid segmentation model construction method combined with prior mask and application thereof | |
| Dong et al. | Supervised learning-based retinal vascular segmentation by m-unet full convolutional neural network | |
| CN113362360A (en) | Ultrasonic carotid plaque segmentation method based on fluid velocity field | |
| Khan et al. | Multi-feature extraction with ensemble network for tracing chronic retinal disorders | |
| CN109816665A (en) | A kind of fast partition method and device of optical coherence tomographic image |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180202 |