CN104102899B - Retinal vessel recognition methods and device - Google Patents

Abstract

The present invention relates to medicine technology field, in particular to retinal vessel recognition methods and device.This method, including：The retina gray-scale map in green path is extracted from the retinal fundus images of rgb format；Multiple contrast yardsticks are set, contrast metrization is carried out under each contrast yardstick from multiple directions to the pixel on the retina gray-scale map, retina binarized contrast's degree figure is obtained, the wherein pixel in retina binarized contrast degree figure is divided into the pixel on retinal vessel and the pixel on non-retinal vessel；Difference fusion is carried out to multiple retina binarized contrast degree figures of acquisition, and removes noise and artifact in the image obtained after fusion, retina fusion figure is obtained；Central retroreflective regions are determined from retina fusion figure, and central retroreflective regions are filled, retinal vessel distribution map is obtained.The method and device that the present invention is provided, improves the accuracy and reliability of retinal vessel identification.

Description

Retinal vessel recognition methods and device

Technical field

The present invention relates to medicine technology field, in particular to retinal vessel recognition methods and device.

Background technology

Retina occupy the internal layer of wall of eyeball, is the film of layer of transparent.Retina is can uniquely to use that optics is non-to invade Enter the central nervous tissue that mode is observed, it is responsible for transmitting the image information in the external world in real time to brain.In order to ensure it is enough its Containing abundant blood vessel network in normal physiological function, retina, any aberrant angiogenesis can cause view in retina The damage of film function.Many general diseases, such as diabetes, hypertension, artery sclerosis, cardiovascular and cerebrovascular accident and apoplexy, and Many eye diseases, such as retinopathy of prematurity, retinal vein obstruction and age-related macular degeneration, and retinal vessel Change it is closely related.To the diameter of retinal vessel, length, Branch Angle, the accurate horizontal quantification evaluation of tortuous degree and Its longitudinal basis weight evaluation changed with time, disease progression improves the accuracy and efficiency of relevant disease diagnosis for oculist There are important meaning, the diagnosis and treatment standard also change based on These parameters of many diseases.

The popularization for exempting from mydriasis eye-ground photography technology solves the problem of retinal vascular images are gathered well, and it is extensive Applied to fields such as clinical ophthalmology, physical examination, eye disease screenings, as a result proving can be well to retina based on eye-ground photography inspection Relevant vascular diseases make Accurate Diagnosis.But eye-ground photography only realizes the collection of retinal vessel, point of correlated results Analysis also relies on manual identified, mark retinal vessel border, and diagnosis is made further combined with clinic diagnosis specification.Due to view Film blood vessel is numerous, diameter, it is out of shape, come in every shape, manual identified, mark retinal vessel be a very time-consuming, laborious work Make, and also there is the difference between larger gauger between gauger.Therefore, clinic be highly desirable to one kind can be to view Film vascular morphology carries out efficient and objective quantitative analysis method, and the retinal vascular images collected are quantized, this is realized The basic premise of one target is that retinal vessel border is delineated.

A series of computers on carrying out automatic identification to the retinal vessel on fundus photograph existing in the last few years are calculated Method, such as line or Edge track method, multi-scale filtering method, model deformation and machine learning method.Its center line or Edge track method are utilized The linear structure of retinal vessel, sets a starting point, gradually extends since starting point, tracks the out of shape of blood vessel first. Multi-scale filtering rule is to utilize the characteristics of retinal vessel is in width, density and directionality, takes progressive manner to set Different angle, luminance threshold, the pixel on retinal vessel is filtered out from background.Machine learning rule is first selected On image a certain feature of each pixel be discriminant criterion, artificially formulate the threshold value of the index, according to whether more than threshold value by its The pixel is divided into vascular tissue or non-vascular tissue, and by constantly carrying out carrying out to comparing threshold value with the result of manual identified Constantly adjust, improve the accuracy of identification, conventional method includes neutral net, SVMs etc..

Current retinal vessel recognizer is based on different characteristics of image, such as color, intensity, shape, gray scale ladder Continuity of degree, contrast and anatomical structure etc., to recognize the border of retinal vessel, it is recognized for small retinal blood vessels Reliability and accuracy still can not meet the actual demand of retinal vessel identification.

The content of the invention

It is an object of the invention to provide retinal vessel recognition methods and device, the problem of to solve above-mentioned.

Retinal vessel recognition methods is provided in an embodiment of the present invention, including：From the retina eyeground of rgb format The retina gray-scale map in green path is extracted in image；Multiple contrast yardsticks are set, from multiple under each contrast yardstick Direction carries out contrast metrization to the pixel on the retina gray-scale map, obtains retina binarized contrast's degree figure, wherein Pixel in the retina binarized contrast degree figure is divided into the pixel on retinal vessel and the pixel on non-retinal vessel； Difference fusion is carried out to multiple retina binarized contrast degree figures of acquisition, and removes making an uproar in the image obtained after fusion Sound and artifact, obtain retina fusion figure；Central retroreflective regions are determined from retina fusion figure, and it is anti-to the center Light region is filled, and obtains retinal vessel distribution map.

Preferably, it is described that multiple contrast yardsticks are set, from multiple directions to the retina under each contrast yardstick Pixel on gray-scale map carries out contrast metrization, obtains retina binarized contrast's degree figure, including：Using current pixel point as circle The heart, the contrast yardstick currently to determine determines the contrast range of current pixel point as radius；Calculate described current from multiple directions The difference of pixel and each linear structure luminance mean value in the contrast range, and selection maximum is made from multiple differences For the contrast value of the current pixel point, wherein the computing formula of the contrast value is：

Wherein, t_θFor the difference, d is the length of the linear structure, I^θFor a pixel in the linear structure Brightness value, I (i, j) be the current pixel point brightness value, k be normal integer, span be 1~d；Utilize above-mentioned calculating The method of current pixel point calculates the contrast value of all pixels point on the retina gray-scale map in current contrast yardstick, profit Retina binarized contrast's degree figure under current contrast yardstick is obtained with the contrast value of all pixels；Utilize this method Obtain retina binarized contrast's degree figure under multiple contrast yardsticks.

Preferably, multiple retina binarized contrast degree figures of described pair of acquisition carry out difference fusion, and removal is melted Noise and artifact in the image obtained after conjunction, obtain retina fusion figure, including：

Step A：Contrast yardstick d is determined from multiple contrast yardsticks₀, will be in the contrast yardstick d₀Lower acquisition is regarded Nethike embrane binarized contrast's degree image is defined as with reference to comparison diagram R₀, wherein the contrast yardstick d₀In multiple contrast yardsticks It is non-minimum；

Step B：By the contrast yardstick d₀Reduce interval delta d and obtain contrast yardstick d_i, will be in the contrast yardstick d_iUnder obtain The retina binarized contrast's degree image taken is defined as comparison diagram R_i；

Step C：Using described with reference to comparison diagram R₀And the comparison diagram R_iSame position on contrast value be worth To fused images U_i；Using described with reference to comparison diagram R₀And the comparison diagram R_iSame position on the difference of contrast value obtain To difference image D_i, wherein when described with reference to comparison diagram R₀And the comparison diagram R_iSame position on contrast value it is identical when, D_iValue on the position is 1, otherwise D_iValue on the position is 0；

Step D：Remove the fused images U_iAnd the difference image D_iOn noise and artifact, and using removing noise And the fused images U of pseudo- movie queen_iAnd the difference image D_iObtain fusion figure；

Step E：Replace described with reference to comparison diagram R using the fusion figure₀, and the contrast yardstick d of the fusion figure_iReplace For the contrast yardstick d₀, repeat step B~D is until obtained d₀Equal to 1；

Step F：From the reference comparison diagram R finally obtained₀Middle removal noise and artifact, obtain the retina fusion figure.

Preferably, it is described to remove the fused images U_iAnd the difference image D_iOn noise and artifact, including：Utilize FormulaCalculate the fused images U_iAnd the difference image D_iOn target image flexibility, wherein S is institute The most major diameter of target image is stated, V is the volume of the target image；Judge whether the flexibility is less than the flexibility threshold of setting Value, if it is, the target image is determined as small tufted structure or non-elongated shape structure, and removes the target image.

Preferably, it is described to determine contrast yardstick d from multiple contrast yardsticks₀, including：According to the retina eyeground The Breadth Maximum of retinal vessel in image determines the contrast yardstick d₀。

Preferably, this method also includes：The interval delta d is determined according to the resolution ratio of the retinal fundus images.

Preferably, central retroreflective regions are determined in the fusion figure from the retina, including：From horizontal direction and vertically Retina fusion figure, obtains elongate structure non-vascular region described in scanning direction；By the elongate structure non-vascular region It is defined as central retroreflective regions.

The embodiment of the present invention additionally provides a kind of retinal vessel identifying device, including：Image zooming-out module, for from The retina gray-scale map in green path is extracted in the retinal fundus images of rgb format；Multiple dimensioned multi-direction quantization modules, are used In setting multiple contrast yardsticks, the pixel on the retina gray-scale map is clicked through from multiple directions under each contrast yardstick Row contrast metrization, obtains retina binarized contrast's degree figure, wherein the pixel in the retina binarized contrast degree figure point For the pixel in the pixel on retinal vessel and non-retinal vessel；Difference Fusion Module, for the multiple described of acquisition Retina binarized contrast's degree figure carries out difference fusion, and removes noise and artifact in the image obtained after fusion, depending on Nethike embrane fusion figure；Module is filled, for determining central retroreflective regions from retina fusion figure, and it is reflective to the center Region is filled, and obtains retinal vessel distribution map.

Preferably, the multiple dimensioned multi-direction quantization modules, including：Contrast range determination sub-module, for current picture Vegetarian refreshments is the center of circle, and the contrast yardstick currently to determine determines the contrast range of current pixel point as radius；To ratio submodule Block, for calculating the current pixel point and the difference of each linear structure luminance mean value in the contrast range from multiple directions Value, and contrast value of the maximum as the current pixel point is chosen from multiple differences, wherein the meter of the contrast value Calculating formula is：

Wherein, t_θFor the difference, D is the length of the linear structure, I^θFor the brightness value of a pixel in the linear structure, I (i, j) is described current The brightness value of pixel, k is normal integer, and span is 1~d；Calculate current using the method for above-mentioned calculating current pixel point The contrast value of all pixels point in yardstick on the retina gray-scale map is contrasted, the contrast of all pixels is utilized It is worth to retina binarized contrast's degree figure under current contrast yardstick；The view under multiple contrast yardsticks is obtained using this method Film binarized contrast's degree figure.

Retinal vessel recognition methods provided in an embodiment of the present invention and device, first the retina eyeground from rgb format The image in green path is extracted in image, because retinal vessel has highest contrast with retina background in the path Degree, is converted into binarized contrast's degree image using multiple dimensioned multi-direction contrast quantification method by the image of green path afterwards, The value of each pixel represents whether it has difference compared with ambient background in binarized contrast's degree image, uses this method It can avoid, due to retinal vessel diameters, density, the different influences to identification in position, improving the accuracy of blood vessel identification. Further, the image of different directions is carried out into difference fusion in multiple yardsticks using difference fusion method to remove on image simultaneously Noise.Finally, central retroreflective regions are determined in filling retina fusion figure, it is possible thereby to remove because retinal vessel center is reflective And the dark space caused in blood vessel recognition result, the reliability and accuracy of retinal vessel identification are further increased, thus The accuracy and reliability of retinal vessel identification are improved using the method and device of the embodiment of the present invention, also can more meet and regard The actual demand of retinal vasculature identification.

Brief description of the drawings

Fig. 1 shows the flow chart of retinal vessel recognition methods in the embodiment of the present invention；

Fig. 2-1~2-3 shows multiple dimensioned and direction explanation schematic diagram in the embodiment of the present invention；

Fig. 3-1~3-2 shows the retina binarized contrast obtained in the embodiment of the present invention under different contrast yardsticks Degree figure；

Fig. 4-1 shows the partial enlarged drawing of retina binarized contrast's degree figure in Fig. 3-1；

Fig. 4-2 shows the partial enlarged drawing of retina binarized contrast's degree figure in Fig. 3-2；

Fig. 5 shows the flow chart of difference fusion method in the embodiment of the present invention；

Fig. 6 shows the effect diagram of difference fusion method in the embodiment of the present invention；

Fig. 7 shows the effect diagram of central reflective areas fill method in the embodiment of the present invention；

Fig. 8 shows the structural representation of retinal vessel identifying device in the embodiment of the present invention；

Fig. 9 shows the verification the verifying results figure of the retinal vessel recognition methods to the embodiment of the present invention.

Embodiment

The present invention is described in further detail below by specific embodiment and with reference to accompanying drawing.

The embodiments of the invention provide a kind of retinal vessel recognition methods, as shown in figure 1, main processing steps include：

Step S11：The retina gray-scale map in green path is extracted from the retinal fundus images of rgb format；

Step S12：Multiple contrast yardsticks are set, from multiple directions on retina gray-scale map under each contrast yardstick Pixel carry out contrast metrization, retina binarized contrast's degree figure is obtained, wherein in retina binarized contrast degree figure Pixel is divided into the pixel on retinal vessel and the pixel on non-retinal vessel；

Step S13：Difference fusion is carried out to multiple retina binarized contrast degree figures of acquisition, and removes acquisition after fusion Image in noise and artifact, obtain retina fusion figure；

Step S14：Central retroreflective regions are determined from retina fusion figure, and central retroreflective regions are filled, are obtained To retinal vessel distribution map.

Retinal vessel recognition methods provided in an embodiment of the present invention, first from the retinal fundus images of rgb format The image in green path is extracted, because retinal vessel has highest contrast with retina background in the path, afterwards The image of green path is converted into by binarized contrast's degree image using multiple dimensioned multi-direction contrast quantification method, in binaryzation pair Whether value than each pixel in degree image represents it compared with ambient background with difference, can be avoided using this method Due to retinal vessel diameters, density, the different influences to identification in position, the accuracy of blood vessel identification is improved.Further, The image of different directions is carried out while removing the noise on image by difference fusion in multiple yardsticks using difference fusion method.Most Afterwards, central retroreflective regions are determined in filling retina fusion figure, it is possible thereby to remove because retinal vessel center is reflective and in blood The dark space caused in pipe recognition result, further increases the reliability and accuracy of retinal vessel identification, thus utilizes this The method of inventive embodiments improves the accuracy and reliability of retinal vessel identification, also can more meet retinal vessel identification Actual demand.

Because posterior pole of eyeball portion is in arc-shaped, the illumination light of fundus camera illumination intensity and heterogeneity in imaging, accordingly Fundus photograph on identical institutional framework the existing inhomogeneity of brightness, central brightness is higher, and periphery brightness is relatively low.Except position Put outer, the thickness of blood vessel, image quality can cause the heterogeneity of identical institutional framework brightness.Drawn to overcome by a variety of causes In the retinal fundus images risen the problem of identical institutional framework brightness disproportionation one, the linear character based on retinal vessel, this Retinal vessel is identified using multiple dimensioned direction contrast quantification method in inventive embodiments.

Multiple dimensioned direction contrast quantification method refers to contrast the pixel in image with two, direction angle from multiple dimensioned Metrization, carries out binaryzation by the pixel in retina gray level image, i.e., is divided into the pixel in retina gray level image and regarding The pixel in pixel and non-retinal vessel on retinal vasculature.

Specifically use the method that retinal vessel is identified in multiple dimensioned direction for：Multiple contrast yardsticks are set, Contrast metrization is carried out from multiple directions to the pixel on retina gray-scale map under each contrast yardstick, retina two is obtained Value contrast figure, including：Using current pixel point as the center of circle, the contrast yardstick currently to determine determines current pixel as radius The contrast range of point；

From multiple directions calculating current pixel point and the difference of each linear structure luminance mean value in contrast range, and from Contrast value of the maximum as current pixel point is chosen in multiple differences, the computing formula of wherein contrast value is：

Wherein, t_θFor difference, d is the length of linear structure, I^θFor the brightness value of a pixel in linear structure, I (i, j) is the brightness value of current pixel point, and k is normal integer, and span is 1~d；

All pictures in current contrast yardstick on retina gray-scale map are calculated using the method for above-mentioned calculating current pixel point The contrast value of vegetarian refreshments, retina binarized contrast's degree under current contrast yardstick is obtained using the contrast value of all pixels point Figure；Retina binarized contrast's degree figure under multiple contrast yardsticks is obtained using this method.

When calculating the contrast value of pixel according to the computational methods of above-mentioned contrast value, if the brightness of current pixel point Higher than the brightness of its background, then T (i, j) value is negative, if the value of current pixel point is less than the brightness of its background, T (i, j) value For just.

Because the brightness value of retinal vessel is generally low compared with retina background, therefore T (i, j) is maximum just Value is defined as the contrast value of the pixel.

It is many in the conventional method to judge brightness with the presence or absence of " change of ladder sample " by comparing the difference of consecutive points brightness, and " change of ladder sample " whether there is by the brightness of comparison object pixel and surrounding linear structure in the present invention.

In view of retinal vessel is generally slim-lined construction, it is mainly used in recognizing view using the method in above-mentioned multi-angle direction The brightness in the horizontal of film blood vessel occurs the position of " change of ladder sample ", that is, determining the border of blood vessel, and such as Fig. 2-1 shows blood vessel side The schematic diagram on boundary.

Further to being illustrated to " multiple dimensioned " and the concept in " direction " in Fig. 2-1~2-3.

The schematic diagram of vessel borders is shown in wherein Fig. 2-1, Fig. 2-2 is shown can be from 8 for current pixel point Distribution arrangement compares its difference with linear structure brightness value around.Fig. 2-3 is shown using current pixel point Pi as the center of circle, to set Fixed contrast yardstick d determines the contrast range of current pixel point for radius, and linear structure is scanned for out of contrast range, utilizes The linear structure scanned carries out brightness ratio pair with current pixel point, wherein contrast yardstick d value can be 1~50 picture Element, contrast yardstick d is relevant with the resolution ratio of retinal fundus images, when the resolution ratio of retinal fundus images is different, view Corresponding change can also occur for the diameter of film blood vessel, therefore contrast yardstick d needs to be entered according to the resolution ratio of retinal fundus images Row adjustment.

The retinal vessel in retina gray-scale map is recognized using multiple dimensioned direction quantization method in the embodiment of the present invention, its Middle multiscale contrast quantifies to be the direction contrast in order to solve retinal vessel density between individual, the problem of width is different Metrization is to solve the problems, such as the identification of different directions retinal vessel.It can be removed due to imaging illumination not by this method Influence of the blood vessel brightness disproportionation to identification, can not only strengthen the image of blood vessel, moreover it is possible to improve whole image brightness caused by Homogeneity.

Retina gray-scale map can be converted into retina binaryzation by contrasting quantization method by above-mentioned multiple dimensioned direction Contrast figure.

The low-resolution image that resolution ratio is 565x584 pixels is shown on the left of Fig. 3-1, right side a takes 1 for contrast yardstick d Retina binarized contrast's degree figure after being converted during pixel；Retina binaryzation when b takes 3 pixel for contrast yardstick d after conversion Contrast figure；Retina binarized contrast's degree figure when c takes 5 pixel for contrast yardstick d after conversion；D takes 7 pictures for contrast yardstick d Retina binarized contrast's degree figure after being converted when plain.

The high-definition picture that resolution ratio is 1924x1556 pixels is shown on the left of Fig. 3-2, right side a takes for contrast yardstick d Retina binarized contrast's degree figure after being converted during 5 pixel；Retina two-value when b takes 10 pixel for contrast yardstick d after conversion Change contrast figure；Retina binarized contrast's degree figure when c takes 20 pixel for contrast yardstick d after conversion；D takes for contrast yardstick d Retina binarized contrast's degree figure after being converted during 30 pixel.

Black picture element represents being averaged for the projecting linear structure of brightness of the pixel wherein in Fig. 3-1~3-2 Brightness, can be seen that small retinal blood vessels are strengthened from Fig. 3-1~3-2.

Fig. 4 is that a in Fig. 3 partial enlarged drawing, wherein Fig. 4-1 is the retina that contrast yardstick d takes 1 pixel in Fig. 3-1 B in the enlarged drawing of binarized contrast's degree figure, Fig. 4-1 is the retina binarized contrast that contrast yardstick d takes 7 pixels in Fig. 3-1 Spend the enlarged drawing of figure；C in Fig. 4-2 is the amplification of retina binarized contrast's degree figure that contrast yardstick d takes 5 pixels in Fig. 3-2 Figure, the d in Fig. 4-2 is the enlarged drawing of retina binarized contrast's degree figure that contrast yardstick d takes 30 pixels in Fig. 3-2.

It can be seen that when contrast yardstick d is smaller, small retinal blood vessels obtain a certain degree of increasing from Fig. 4-1~4-2 It is strong but noise is larger, conversely when contrast yardstick d it is larger when, noise is smaller but can lose a part of thin vessels.

Because the position that the noise and artifact on obtained retina binarized contrast's degree figure occur is random, this method It is middle that the retina binarized contrast's degree figure obtained under different contrast yardstick d is merged by the way of gradual, to go Except the noise and artifact on contrast figure.

The method merged in the embodiment of the present invention using difference is entered to multiple retina binarized contrast degree figures to acquisition Row fusion, and noise and artifact in the image obtained after fusion are removed, retina fusion figure is obtained, as shown in Figure 5 main place Reason step includes：

Step A：Contrast yardstick d is determined from multiple contrast yardsticks₀, will be in contrast yardstick d₀The retina two-value of lower acquisition Change contrast image to be defined as with reference to comparison diagram R₀, wherein contrast yardstick d₀It is non-minimum in multiple contrast yardsticks.

In order to illustrate that embodiment of the present invention difference fusion method removes the effect of noise and artifact, further combined with specifically showing The effect that example is merged to difference is illustrated.

For this step A, a subgraphs in such as Fig. 6 show contrast yardstick d₀Reference comparison diagram R during for 30 pixel₀。

Step B：Will contrast yardstick d₀Reduce interval delta d and obtain contrast yardstick d_i, will be in contrast yardstick d_iThe view of lower acquisition Film binarized contrast's degree image is defined as comparison diagram R_i。

In order to embody comparison diagram R_iWith comparison diagram R₀Difference, in step B example, comparison diagram R_iContrast yardstick d_i The b subgraphs being taken as in 5 pixels, such as Fig. 6 show contrast yardstick d_iIt is taken as comparison diagram R during 5 pixels_i。

Step C：Using with reference to comparison diagram R₀And comparison diagram R_iSame position on contrast value be worth to fusion figure As U_i；Using with reference to comparison diagram R₀And comparison diagram R_iSame position on the difference of contrast value obtain difference image D_i, wherein When with reference to comparison diagram R₀And comparison diagram R_iSame position on contrast value it is identical when, D_iValue on the position is 1, otherwise D_iValue on the position is 0.

Carry out acquisition after difference fusion using the b subgraphs in a subgraphs and Fig. 6 in Fig. 6 as the c subgraphs in Fig. 6 are shown Difference image D_i, the d subgraphs in Fig. 6 show the fused images U of acquisition_i。

Step D：Remove fused images U_iAnd difference image D_iOn noise and artifact, and using removing noise and pseudo- movie queen Fused images U_iAnd difference image D_iObtain fusion figure；

According to above-mentioned example, the e subgraphs in Fig. 6 show the fusion figure acquired.

Step E：Replaced using fusion figure with reference to comparison diagram R₀, and the contrast yardstick d of fusion figure_iReplace with contrast yardstick d₀, Repeat step B~D is until obtained d₀Equal to 1；

Step F：From the reference comparison diagram R finally obtained₀Middle removal noise and artifact, obtain retina fusion figure.

F subgraphs in Fig. 6 are only to be carried out with a subgraphs (d=30 pixels) in Fig. 6 with the b subgraphs (d=5 pixels) in Fig. 6 Difference fusion method abate the noise after effect.

G subgraphs in Fig. 6 are gradually to be decreased to d values after the abating the noise of 5 pixels by 30 pixels using difference fusion method Effect.

Fused images U is removed in the embodiment of the present invention_iAnd difference image D_iOn noise and artifact, including：Utilize formulaCalculate fused images U_iAnd difference image D_iOn target image flexibility, wherein S is most long for target image Footpath, V is the volume of target image；

Judge whether flexibility is less than the flexibility threshold value of setting, if it is, target image is determined as small tufted structure Or non-elongated shape structure, and remove target image.

Contrast yardstick d is determined from multiple contrast yardsticks₀, including：Retinal vessel in retinal fundus images Breadth Maximum determine contrast yardstick d₀。

This method also includes：Interval delta d is determined according to the resolution ratio of retinal fundus images.

Take the blood vessel in above-mentioned difference fusion method difference image to be gradually fused noise simultaneously and (side is used in such as Fig. 6 The macular region that collimation mark goes out) also it is removed.As shown in f subgraphs in Fig. 6, when Δ d values larger (such as taking 25 pixels), some are made an uproar Sound may be identified as blood vessel.And (for example taking 5 pixels) when Δ d values are smaller, noise can seldom be identified as blood vessel, and adopt With the gradual difference fusion method of the embodiment of the present invention can efficiently against it is above-mentioned the problem of, can effectively remove noise simultaneously It can be prevented effectively from that blood vessel is identified as noise or noise is identified as blood vessel.

The center of the thicker retina arteriovenous I and II branch of diameter often has obvious reflective existing in fundus photograph As, this cause the relatively low brightness originally of blood vessel middle position occur it is false increase, influence the identification of retinal vessel, show as " dark space " occurs in blood vessel center on contrast figure, and the arrow in such as Fig. 7 b subgraphs is signified.It is reflective in order to solve blood vessel center The dark space that region occurs, routinely needs to perform " closing " operation to fill dark space.But when the kernel setting for performing padding When incorrect, it is possible to can cause blood vessel mistake adjacent around and existing blood vessel being merged.

A kind of method for filling central reflective areas is provided in the embodiment of the present invention, is specifically included

From horizontal direction and vertical scan direction retina fusion figure, obtain elongate structure non-vascular region；Will be elongated Shape structure non-vascular region is defined as central retroreflective regions.

A subgraphs in such as Fig. 7 show the high brightness reflective tape in retinal vessel center, and b subgraphs are shown in level and vertical Nogata is to scanning for small-sized non-vascular region (i.e. elongate structure non-vascular region) and filled, and the effect after filling is such as As shown in the c subgraphs in Fig. 7.

It is therefore seen that can effectively remove retinal blood using the fill method of the central reflective areas of the embodiment of the present invention The dark space in pipe center, prevents from blood vessel mistake adjacent around and existing blood vessel being merged, and improves retinal vessel and knows Other accuracy.

The embodiment of the present invention additionally provides a kind of retinal vessel identifying device, includes as shown in Figure 8：

Image zooming-out module 81, for extracting the retina in green path from the retinal fundus images of rgb format Gray-scale map；

Multiple dimensioned multi-direction quantization modules 82, for setting multiple contrast yardsticks, from multiple under each contrast yardstick Direction carries out contrast metrization to the pixel on retina gray-scale map, obtains retina binarized contrast's degree figure, wherein view Pixel in film binarized contrast's degree figure is divided into the pixel on retinal vessel and the pixel on non-retinal vessel；

Difference Fusion Module 83, carries out difference fusion, and go for multiple retina binarized contrast degree figures to acquisition Except the noise and artifact in the image obtained after fusion, retina fusion figure is obtained；

Module 84 is filled, for determining central retroreflective regions from retina fusion figure, and central retroreflective regions are carried out Filling, obtains retinal vessel distribution map.

Wherein multiple dimensioned multi-direction quantization modules 82, including：Contrast range determination sub-module, for using current pixel point as The center of circle, the contrast yardstick currently to determine determines the contrast range of current pixel point as radius；To ratio submodule, it is used for Calculate the difference of each linear structure luminance mean value in current pixel point and contrast range from multiple directions, and from multiple differences The middle contrast value for choosing maximum as current pixel point, the computing formula of wherein contrast value is：

Wherein, t_θFor difference, d is The length of linear structure, I^θFor the brightness value of a pixel in linear structure, I (i, j) is the brightness value of current pixel point, K is normal integer, and span is 1~d；

All pictures in current contrast yardstick on retina gray-scale map are calculated using the method for above-mentioned calculating current pixel point The contrast value of vegetarian refreshments, retina binarized contrast's degree under current contrast yardstick is obtained using the contrast value of all pixels point Figure；Retina binarized contrast's degree figure under multiple contrast yardsticks is obtained using this method.

The computer of view hymenology blood vessel recognizes that its essence is that each pixel on fundus photograph is divided into two classes：On blood vessel Pixel or non-vascular (tissues surrounding vascular) on pixel.Therefore, the result of identification may have four kinds of possible (two kinds of correct knots Fruit and two kinds of error results)：(1) true positives：Pixel on blood vessel is identified as blood vessel pixel；(2) true negative：Will be not in blood vessel On pixel be identified as non-vascular pixel；(3) false positive：Pixel not on blood vessel is identified as blood vessel pixel；(4) it is false cloudy Property：Pixel on blood vessel is identified as non-vascular pixel.True positives represent the ability that algorithm correctly recognizes vascular tissue.

With reference to the method for other retina recognizer research performance verifications, DRIVE (Digital Retinal are used Images for Vessel Extaction) and two public affairs of STARE (STructured Analysis of the Retina) Database is evaluated and tested to the recognition performance of the retinal vessel recognition methods of the embodiment of the present invention altogether.

DRIVE databases are by 40 colored fundus photographs, wherein 20 are used to verify, 20 are used to train in addition.Use Canon CR-5 exempts from the shooting of mydriasis fundus camera, and visual field size is 550 × 550 pixels.The figure that 20 are used to verify in database Piece is manually delineated the border of retinal vessel by two diagosis persons.

STARE opens one's eyes negative film including 20, is shot using TopCon TRV-50, and 20 have 10 pictures to have lesion figure Piece.Visual field size is 650 × 550 pixels, retinal vessel is delineated by hand by two diagosis persons with DRIVE class databases seemingly Border.

Because the resolution ratio of picture contained by DRIVE and STARE databases is relatively low, therefore also use the third image resolution ratio Higher database HRF (High-Resolution Fundus) verifies the performance of recognizer.

In HRF databases comprising 15 normal persons, 15 glaucomas, 15 diabetics eye fundus image, use Canon CR-1 are shot.

Result using manual identified in above-mentioned three databases is normative reference, from accuracy, sensitiveness, accuracy three Aspect evaluates this method to retinal blood length of tube, the performance of volume identification.Accuracy criticizes pixel (including the blood really recognized Pipe and non-vascular) with whole visual field in all pixels point ratio.When evaluating retinal vessel skeleton recognition performance Accuracy is not calculated, because retinal vessel skeleton is the linear structure of only one pixel, good accuracy is had.It is quick Perception criticizes the ratio of the pixel and manual identified pixel really recognized, reflects that how many pixel is correctly identified as blood vessel group Knit.How many pixel the ratio of what the accuracy pixel that correctly recognizes of finger counting method and algorithm were recognized a little, react by mistake It is identified as vascular tissue.

In view of manual segmentation may not 100% accurate segmentation retinal vessel border, set different permissions Error range, i.e., in certain location of pixels, if result and the difference of the result of manual identified that algorithm is recognized are in error range (pixel), then it is assumed that two methods are consistent in the recognition result of the point.List different error range algorithms respectively in the result Performance, in DRIVE and STARE databases, error range is 0 to 3 pixels, is 0-10 pixels in HFR databases.

Add a retinal vessel skeleton recognizer outside traditional evaluating method, auxiliary compare manual identified with Computer recognizes the uniformity in retinal vessel skeleton (center line) identification.In order to avoid local noise is to skeleton (center Line) extraction, equally set error range.

In method, the skeletal extraction algorithm researched and developed using Cornea et al..Fig. 9 is the retina extracted using the algorithm Vascular skeleton, the branch of different stage is marked using different colours.

DRIVE database test results

The result of the identification obtained using the retinal vessel recognition methods of the embodiment of the present invention and the one of DRIVE databases Cause property is as shown in table 1.

Compared with the result of diagosis person one and the manual identified of diagosis person two capacity of blood vessel and length sensitiveness 90% To 70%, accuracy is 99%, and the accuracy of volume and length is 94%.

STARE database test results

It is as shown in table 2 in the recognition performance of STARE databases using the retinal vessel recognition methods of the embodiment of the present invention. When allowable error takes 3 pixel, compared with the manual identified result of two diagosis persons, the volume of blood vessel, length sensitiveness difference For 90% and 70%.Accuracy is 97%.It is about 85% in the accuracy of the capacity of blood vessel of diagosis person one and length, in diagosis person two The accuracy of capacity of blood vessel and length is about 93%.

HRF database test results

It is as shown in table 3 with HRF database result of the comparison.When allowable error is 6 pixels, in Healthy People capacity of blood vessel And the sensitiveness of length is 92% and 83%.It is 92% and 75% in the sensitiveness of glaucoma patient medium vessels volume and length. It is 91% and 73% in the sensitiveness of diabetic retina patient's medium vessels volume and length.

It can to sum up draw, the embodiments of the invention provide the retinal vessel quantified based on multiple dimensioned Direction Contrast identification Method.The characteristics of this method had not only considered retinal vessel local low-light level but also when being considered using multiple dimensioned method because taking pictures Diverse location, width, the difference of direction blood vessel brightness in fundus photograph caused by uneven illumination.The innovative point of this method include with It is lower 2 points：(1) accuracy that blood vessel is recognized is improved by the way that original eye fundus image is converted into retina binarized contrast's degree figure； (2) image co-registration after conversion is split retinal vessel by way of multi-scale gradual difference is merged.It is local Eye fundus image intensity is standardized by the process of contrast metrization according to tissue characteristics, eliminates retinal blood in fundus photograph Pipe is because the influence of the factor to blood vessel imaging brightness such as size, position, image resolution ratio.On the one hand the process of difference fusion will be many The image that yardstick quantifies to obtain links into an integrated entity, on the other hand can be continuous using a series of differences to comparing image and space Property significantly reduces the influence that image noise and artifact are split to retinal vessel.In addition, retina provided in an embodiment of the present invention Blood vessel recognition methods is easy to implement, different from the algorithm based on machine learning and neutral net, the retina of the embodiment of the present invention Blood vessel recognition methods need not collect mass data in advance and specify clearly rule to be trained algorithm.Using different In the test that public database is carried out to the recognition performance of algorithm, algorithm in normal person, persons suffering from ocular disorders, use different eyeground phases Machine, the picture of different resolution show good stability and accuracy.

Due to having used public database so that retinal vessel recognition methods that can be by the embodiment of the present invention and other The performance of retinal vessel recognizer does lateral comparison.But most of algorithms is entered using STARE or DRIVE databases in existing method The evaluation of row performance.What is behaved oneself best in these algorithms is the entitled " based on characteristics of image of Lupascu et al. research and development Adaboost graders " (AdaBoost classifier, FABC), in DRIVE databases, its best accuracy is 95.9%, Sensitiveness is 72%.Algorithm in same database accuracy be 99%, sensitiveness is 90% and does not need machine learning Process.Odstrcilik et al. algorithm can obtain 95% accuracy when carrying out performance test using HRF databases And 74%-79% sensitiveness, the accuracy of algorithm is about 98%, and sensitiveness is about 90%.But existing algorithm can still be lost Lose some tiny blood vessels and the situation of false positive can be occurred by the interference of noise and artifact.

From table 3 it is observed that when allowable error is 0, the result that computer is split with manually splitting exists certain Difference.The result of that is more nearly truly still not very clearly, but think because computer can be anti-on earth in two methods The contrast of local maxima is answered so the result of its identification can be more accurate, repeatability can be more preferable.Because image recognition is most Whole purpose is the accurate quantification aided disease diagnosis by changing to retinal vascular morphologies, so splitting vessel borders to algorithm The assessment of accuracy is highly important.But the influence that hand is shaken when due to manually delineating vessel borders, real work In be difficult to set up the goldstandard of blood vessel segmentation accuracy evaluation by manual identified.Such as in STARE databases, diagosis person one will 10.4% pixel divides blood vessel into eye fundus image, and diagosis person one divides in eye fundus image 14.9% pixel into blood vessel, reads The thin vessels quantity that piece person two has found is significantly more than diagosis person one.

Past most method when evaluating the performance of recognizer only calculates the identification accuracy of capacity of blood vessel and quick Perception.Because the volume of thin vessels will be far smaller than big blood vessel, therefore the appraisal procedure based on blood vessel may be not enough to examination and calculate Recognition capability of the method to thin vessels.In order to overcome this not enough, it is believed that should also calculate retinal vessel skeleton or center line simultaneously The accuracy and sensitiveness of identification, as a result display algorithm retinal vessel will be less than to the identification sensitiveness of retinal vessel skeleton The sensitiveness of volume identification.Sensitiveness of the computer in identification retina level Four and following branch vessel is also poor.View The further improvement for being calculated as algorithm of film vascular skeleton identification relevant parameter provides more objective, accurate evaluation criterion.

This method is a kind of method that use computer automatically recognizes retinal vessel border on colored fundus photograph.Know Bao Kuo not three processes：Contrast metrization, image co-registration and central reflective tape filling.In simplicity, reliability, accuracy Better than existing method, it can be very good to remove influence of the image background noise to identification.By containing various disease with three, dividing The contrast verification of the manual identified result recognition performance of algorithm in the public database of resolution eye fundus image.

The preferred embodiments of the present invention are these are only, are not intended to limit the invention, for those skilled in the art For member, the present invention can have various modifications and variations.Any modification within the spirit and principles of the invention, being made, Equivalent, improvement etc., should be included within the scope of the present invention.

Claims (7)

1. retinal vessel recognition methods, it is characterised in that including：

The retina gray-scale map in green path is extracted from the retinal fundus images of rgb format；

Multiple contrast yardsticks are set, from multiple directions to the pixel on the retina gray-scale map under each contrast yardstick Contrast metrization is carried out, retina binarized contrast's degree figure under current contrast yardstick is obtained, wherein the retina binaryzation Pixel in contrast figure is divided into the pixel on retinal vessel and the pixel on non-retinal vessel；

Difference fusion is carried out to multiple retina binarized contrast degree figures of acquisition, and removed in the image obtained after fusion Noise and artifact, obtain retina fusion figure；

Central retroreflective regions are determined from retina fusion figure, and the central retroreflective regions are filled, depending on Retinal vasculature distribution map；

It is described that multiple contrast yardsticks are set, from multiple directions to the picture on the retina gray-scale map under each contrast yardstick Vegetarian refreshments carries out contrast metrization, obtains retina binarized contrast's degree figure under current contrast yardstick, including：

Using current pixel point as the center of circle, the contrast yardstick currently to determine determines the contrast range of current pixel point as radius；

The current pixel point and the difference of each linear structure luminance mean value in the contrast range are calculated from multiple directions, And contrast value of the maximum as the current pixel point is chosen from multiple differences, wherein the calculating of the contrast value is public Formula is：

$T(i,j)=max\{t_{\mathrm{\θ}}=\frac{1}{d}\underset{k=1}{\overset{d}{\Σ}}{I}_k^{\mathrm{\θ}}-I(i,j),\mathrm{\θ}\∈\[0,2\mathrm{\π}\]\};$

Wherein, t_θFor the difference, d is the length of the linear structure, I^θFor the bright of a pixel in the linear structure Angle value, I (i, j) is the brightness value of the current pixel point, and k is normal integer, and span is 1~d；

All pictures in current contrast yardstick on the retina gray-scale map are calculated using the method for above-mentioned calculating current pixel point The contrast value of vegetarian refreshments, the retina binaryzation pair under current contrast yardstick is obtained using the contrast value of all pixels Than degree figure；Retina binarized contrast's degree figure under multiple contrast yardsticks is obtained using this method.

2. according to the method described in claim 1, it is characterised in that multiple retina binarized contrasts of described pair of acquisition Degree figure carries out difference fusion, and removes noise and artifact in the image obtained after fusion, obtains retina fusion figure, including：

Step A：Contrast yardstick d is determined from multiple contrast yardsticks₀, will be in the contrast yardstick d₀The retina of lower acquisition Binarized contrast's degree image is defined as with reference to comparison diagram R₀, wherein the contrast yardstick d₀It is multiple it is described contrast yardsticks in it is non-most It is small；

Step B：By the contrast yardstick d₀Reduce interval delta d and obtain contrast yardstick d_i, will be in the contrast yardstick d_iLower acquisition Retina binarized contrast's degree image is defined as comparison diagram R_i；

Step C：Using described with reference to comparison diagram R₀And the comparison diagram R_iSame position on being worth to for contrast value melt Close image U_i；Using described with reference to comparison diagram R₀And the comparison diagram R_iSame position on the difference of contrast value obtain difference Partial image D_i, wherein when described with reference to comparison diagram R₀And the comparison diagram R_iSame position on contrast value it is identical when, D_i Value on the position is 1, otherwise D_iValue on the position is 0；

Step D：Remove the fused images U_iAnd the difference image D_iOn noise and artifact, and using removing noise and puppet The fused images U of movie queen_iAnd the difference image D_iObtain fusion figure；

Step E：Replace described with reference to comparison diagram R using the fusion figure₀, and the contrast yardstick d of the fusion figure_iReplace with institute State contrast yardstick d₀, repeat step B~D is until obtained d₀Equal to 1；

Step F：From the reference comparison diagram R finally obtained₀Middle removal noise and artifact, obtain the retina fusion figure.

3. method according to claim 2, it is characterised in that the removal fused images U_iAnd the difference image D_iOn noise and artifact, including：

Utilize formulaCalculate the fused images U_iAnd the difference image D_iOn target image flexibility, its Middle S is the most major diameter of the target image, and V is the volume of the target image；

Judge whether the flexibility is less than the flexibility threshold value of setting, if it is, the target image is determined as small tufted Structure or non-elongated shape structure, and remove the target image.

4. method according to claim 2, it is characterised in that described to determine contrast yardstick from multiple contrast yardsticks d₀, including：

The Breadth Maximum of retinal vessel in the retinal fundus images determines the contrast yardstick d₀。

5. method according to claim 2, it is characterised in that this method also includes：According to the retinal fundus images Resolution ratio determine the interval delta d.

6. according to the method described in claim 1, it is characterised in that determine that center is reflective in the fusion figure from the retina Region, including：

Retina fusion figure, obtains elongate structure non-vascular region from horizontal direction and described in vertical scan direction；Will be described Elongate structure non-vascular region is defined as central retroreflective regions.

7. retinal vessel identifying device, it is characterised in that including：

Image zooming-out module, for extracting the retina gray-scale map in green path from the retinal fundus images of rgb format；

Multiple dimensioned multi-direction quantization modules, for setting multiple contrast yardsticks, from multiple directions pair under each contrast yardstick Pixel on the retina gray-scale map carries out contrast metrization, obtains the retina binarized contrast under current contrast yardstick Degree figure, wherein the pixel in the retina binarized contrast degree figure is divided into pixel and non-retinal vessel on retinal vessel On pixel；

Difference Fusion Module, carries out difference fusion, and remove for multiple retina binarized contrast degree figures to acquisition Noise and artifact in the image obtained after fusion, obtain retina fusion figure；

Module is filled, for determining central retroreflective regions from retina fusion figure, and the central retroreflective regions are entered Row filling, obtains retinal vessel distribution map；The multiple dimensioned multi-direction quantization modules, including：

Contrast range determination sub-module, for using current pixel point as the center of circle, using the contrast yardstick that currently determines as radius, it is determined that The contrast range of current pixel point；

To ratio submodule, for calculating the current pixel point and each wire in the contrast range from multiple directions The difference of structure luminance mean value, and contrast value of the maximum as the current pixel point is chosen from multiple differences, wherein The computing formula of the contrast value is： Wherein, t_θFor the difference, d is the length of the linear structure, I^θFor the brightness of a pixel in the linear structure Value, I (i, j) is the brightness value of the current pixel point, and k is normal integer, and span is 1~d；

All pictures in current contrast yardstick on the retina gray-scale map are calculated using the method for above-mentioned calculating current pixel point The contrast value of vegetarian refreshments, the retina binaryzation pair under current contrast yardstick is obtained using the contrast value of all pixels Than degree figure；Retina binarized contrast's degree figure under multiple contrast yardsticks is obtained using this method.