CN106780439A - A kind of method for screening eye fundus image - Google Patents

Abstract

A kind of method for screening eye fundus image, according to the tree network structure and gray distribution features of retinal vessel, by Morlet small echos and the partitioning scheme of Gauss matched filtering, from multiple dimensioned discrete Gaussian kernel to vascular skeleton, so as to improve the contrast of tiny blood vessels and background area；In combination with Gabor wavelet algorithm, the characteristic vector constituted using its wavelet character and green channel half-tone information；Learning training is marked to eye features such as intersection, arch bridge, macula lutea, blackspot, the hickies of artery and vein vascular and blood vessel by deep neural network technology；The application can complete the screening and classification to eyeground picture, and the diagnosis speed of acceleration fundus oculi disease solves the problems, such as to expend largely artificial and inefficiencies, and the final effort for the raising overall social general level of the health.

Description

A kind of method for screening eye fundus image

Technical field

The invention belongs to eyeground field of medical technology, specifically a kind of method for screening eye fundus image.

Background technology

Fundus oculi disease is the major reason of human blindness, its Health and Living quality for seriously having threatened people；Blood high Pressure, artery sclerosis, angiemphraxis, diabetes and age-related macular etc. are various universal and cure more difficult disease, obtained society compared with Many concerns, are the important topics of medical field urgent need solution.

In recent years, clinically have confirmed, the generation of many diseases of human body, development, reflection that all can be different degrees of to eye On bottom, therefore many diseases in early detection and can carry out diagnosis and treatment in advance by the analysis to eyeground feature, to control The further development of disease.This cognition causes another important topic of eyeground medical treatment as medical field, if can be right Eyeground feature is more accurately screened, analyzed and is classified, then Community health is integrally brought up to level higher. Eye-ground photography is that it being capable of relatively straightforward record eye ground in recent years clinically using more funduscopy methods Morphological change and development, particularly multispectral fundus imaging technology is after the first Application of field of ophthalmology so that believe on eyeground Breath can be presented to related personnel with more perspective, depth, and ophthalmologist, oculist is understood based on eyeground picture, can obtain patient To the concern and treatment of early stage, and effective disease development anticipation is provided, so as to be controlled to disease as far as possible or make patient Thoroughly recovery from illness.

But the eyeground picture taken pictures out for fundus camera is extracted, screened, classified, and do multi-level comprehensive point Analysis, is still at home at present a severe problem of comparing.With the office increased with training deciphering doctor's quantity of sufferer picture Limit is present, and the picture of different lesions feature how is quickly sorted from existing a large amount of eyeground pictures and bottom of the normal eyes are separated Picture, and then time loss is become very meaningful by mitigation doctor on the picture sifting sort having little significance.

In recent years, extraction and the analytical technology of eyeground Event Characteristics are more concerned with the world, and have ignored eyeground pathological changes Globality, and specific technique analyzes, and can cause whole screening and classify comprising excessive excessively miscellaneous technology, the fusion between technology Property, complementarity substantially reduce, research cost is very high so that the reduction of the feasibility of Clinical practice, can not give health medical treatment band Go more incomes.

The content of the invention

In view of the above-mentioned problems existing in the prior art, the invention provides a kind of method for screening eye fundus image, can be complete The screening and classification of paired eyeground picture, accelerate the diagnosis speed of fundus oculi disease, solve to expend largely artificial and inefficiencies and ask Topic, and the final effort for the raising overall social general level of the health.

To achieve the above object, the technical scheme is that：A kind of method for screening eye fundus image, specific steps are such as Under：

S1：Receive the eyeground picture for shooting；

S2：Using Morlet small wave converting methods to image segmentation, carry out the extraction of vascular skeleton；

S3：Eyeground picture is carried out Gabor wavelet conversion process；

S4：For Gabor wavelet transformation results, blood vessel is carried out using svm classifier model and non-vascular is sorted out, and record blood Pipe coordinates regional；

S5：Using the sorted blood vessel coordinate information of Gabor wavelet, the blood vessel coordinate extracted with Morlet small echos carries out area Domain compares；

S6：According to image type, signature is carried out；And the feature for mark carries out study instruction using neutral net Practice；

S7：Svm classifier is carried out to the feature that neutral net is extracted.

Further, the above method, also includes：

S8：Based on the characteristic model of training, incoming eyeground picture is matched；By the marker characteristic trained, to eyeground figure Piece carries out automatic screening and classification.

Further, pretreatment separation is carried out, and use self-adapting histogram to the eyeground picture for receiving in step S1 Balance optimizing image.

Further, after vascular skeleton extraction is carried out, using small in multiple dimensioned Gauss matched filtering optimization image Vasculature part, and binary conversion treatment is carried out to the image after Gauss matched filtering.

Further, Gabor wavelet conversion process, its core is Gaussian function, by adjust the parameter of Gaussian function come Approach the gray-scale intensity distribution of vessel cross-sections.

Further, according to image type, manually blood vessel, arch bridge, optic disk, optic cup, macula lutea, blackspot, hickie etc. are entered Row signature.

Further, the Morlet small wave converting methods, specially：

Wavelet conversion coefficient T ω (b, θ, a) reflect the similarity degree of gradation of image distribution curve and wavelet sequence function, Wherein * represents conjugation；C ω represent normalization constant；ω represents the small echo to be analyzed；B is displacement vector；θ is the anglec of rotation；a It is scale factor；R is wave filter.

Further, Gabor wavelet transformation for mula is：

Wherein,It is Gaussian function, referred to as window function, wherein a>b>0, g_a(t-b) it is one " window function " of individual time localization, wherein parameter b is used to move in parallel window, and e is natural Exponents.

Used as further, reversely the method for solving of renewal weight is when the neutral net is finely tuned：L (w) is launched It is that derivation is carried out to W after the function on weight W, weight is updated by gradient and the last linear combination for updating weight V；Institute L (w) is stated to be specially：

Wherein, w is weight, and L (w) is the loser after network is once finely tuned, and N is the size for once finely tuning input data, l_w (xⁱ) it is the distance between label value and predicted value difference, λ r (w) is weight constant.

The present invention can obtain following technique effect due to using above technical method：By to shooting the eye for completing The algorithm combination mode that bottom photo is used using the application, can overcome the noise error brought during picture uneven illumination, reach The purpose of effective reliable extraction is carried out to optical fundus blood vessel；It is trained with reference to the corresponding image feature information of eye various disease conditions Practise, further up to the purpose for carrying out sifting sort by characteristics of image to eyeground picture, oculist only needs on this basis Photograph image to filtering out carries out confirmation analysis, so as to save the quality time of doctor, greatly improves operating efficiency.

Brief description of the drawings

For clearer explanation embodiments of the invention or the technical scheme of prior art, below will be to embodiment or existing The accompanying drawing to be used needed for having technology description does one and simply introduces, it should be apparent that, drawings in the following description are only Some embodiments of the present invention, for those of ordinary skill in the art, on the premise of not paying creative work, may be used also Other accompanying drawings are obtained with according to these accompanying drawings.

Fig. 1 is original fundus photograph G channel images and its histogram in embodiment；

Fig. 2 is image and its histogram after being equalized in embodiment, and the unit is mainly using the directionality of Morlet small echos The direction character of signal Analysis improves the contrast of eye fundus image tiny blood vessels and background；

Fig. 3 is the characteristic function figure of retinal vessel in embodiment, it is therefore an objective to carry out the skeletal extraction of optical fundus blood vessel, its pin Each pixel to image carries out many-valued selection, and wavelet conversion coefficient mould maximum is defeated as the characteristic function of current point Go out；

Fig. 4 is the process of convolution figure of neutral net CNN principles in embodiment, the figure by the way of many convolution kernels to marking Piece carries out convolution feature extraction；

Fig. 5 is the multilayer convolution feature overall situationization schematic diagram of neutral net CNN principles in embodiment, and it can be original signal Feature strengthens, and reduces noise；

Fig. 6 carries out eye disease picture sifting sort method for a kind of wavelet algorithm combination deep neural network of the application Overall structure schematic block diagram.

Specific embodiment

To make the purpose, technical scheme and advantage of embodiments of the invention clearer, with reference to the embodiment of the present invention In accompanying drawing, clearly complete description is carried out to the technical scheme in the embodiment of the present invention：

Embodiment 1

The present embodiment provides a kind of method for screening eye fundus image, comprises the following steps that：

1. the eyeground picture of shooting is received；

2. the G channel images for 24 RGB retinal images of rgb format carry out pretreatment separation；

3. neutralization contrast is excessively collected using illumination inequality present in self-adapting histogram equilibrium optimization image, intensity profile The low problem of degree；

4. using Morlet small echos formula to image segmentation, carry out the extraction of vascular skeleton；

5. using tiny blood vessels part in multiple dimensioned Gauss matched filtering optimization image；

6. the binaryzation based on delayed threshold values method carries out binary conversion treatment to the image after Gauss matched filtering, so as to arrange Except most of non-vascular pixel；

7. G passages picture is carried out Gabor wavelet conversion process, because its core is Gaussian function, by adjusting Gauss The parameter of function come approach vessel cross-sections gray-scale intensity be distributed；

8. blood vessel is carried out to pixel using two class svm classifier models and non-vascular is sorted out；

9. the sorted blood vessel coordinate information of Gabor wavelet is used, the blood vessel coordinate extracted with Morlet small echos carries out area Domain compares；

10. with reference to the feature of image type, spy manually is carried out to blood vessel, arch bridge, optic disk, optic cup, macula lutea, blackspot, hickie etc. Levy mark；

Feature for mark carries out learning training using deep neural network, and the feature that neutral net is extracted is entered Row svm classifier；

Based on the characteristic model of training, incoming eyeground picture is matched；

By the marker characteristic trained, automatic screening and classification are carried out to eyeground picture；

According to the tree network structure and gray distribution features of retinal vessel, filter is matched by Morlet small echos and Gauss The partitioning scheme of ripple, from multiple dimensioned discrete Gaussian kernel to vascular skeleton, so as to improve the contrast of tiny blood vessels and background area Degree；In combination with Gabor wavelet method, the characteristic vector constituted using its wavelet character and green channel half-tone information, next gram The separation interference brought to optical fundus blood vessel in the presence of uneven illumination situation is taken, and carries out backwards calculation correction, so as to realize to eye Bottom blood vessel and the reliable extraction in optic disk region.

On the basis of wavelet algorithm is separated to blood vessel, optic disk, by deep neural network technology to artery and vein vascular and The eye features such as intersection, arch bridge, macula lutea, blackspot, the hickie of blood vessel are marked learning training, with reference to hypertension, artery sclerosis, The disease image features such as angiemphraxis, diabetes, age-related macular and uvea seepage are compared test such that it is able to quick sieve Choosing has pair eyeground picture that should be understood that characteristics of lesion.

Embodiment 2

The present embodiment to be related in embodiment 1 Morlet wavelet transformations, Gabor wavelet fusion feature vector and to image The nerual network technique that feature carries out deep learning is specifically illustrated.

(1) Morlet small wave converting methods, specially

Wherein * represents conjugation；C ω represent normalization constant；ω represents the small echo to be analyzed；B is displacement vector；θ is rotation Gyration；A is scale factor, and r is wave filter；Wavelet conversion coefficient T ω (b, θ, a) reflect gradation of image distribution curve with it is small The similarity degree of wave train array function, when the local frequencies of image and the close wavelet function frequency of oscillation of corresponding scale, T ω (b, θ, modulus value a) are accordingly larger, and along yardstick direction of principal axis, the position wire definition of wavelet conversion coefficient mould maximum is small " ridge " of wave conversion, by its maximum as the characteristic function value output of current point, has thus obtained the spy of retinal vessel Levy function.

(2) the optical fundus blood vessel optimization based on Gabor wavelet and SVM is extracted and counterpropagation network correction；

Gabor wavelet transformation for mula is：

Wherein,It is Gaussian function, referred to as window function, wherein a>b>0, g_a(t-b) it is one " window function " of individual time localization, e is natural Exponents, and wherein parameter b is used to move in parallel window, during to cover whole Domain.

(3) represent feature to carry out the depth nerve net of learning training and Classification and Identification in retina based on eyeground picture Network technology；

Reversely the formula of renewal weight is when its god of net's channels and collaterals is finely tuned：

The solution mode of gradient is to carry out derivation to W after L (w) is expanded into the function on weight W, by gradient and upper The linear combination for once updating weight V updates weight.

Wavelet transformation is the partial transformation of space and frequency, thus information can be effectively extracted from signal, and by stretching The calculation function such as contracting and translation carries out multiple dimensioned refinement analysis to function and signal, and other feature extraction algorithms that compare come Say, the ageing and accuracy of the method is all very outstanding, its extraction to foundation characteristics such as optical fundus blood vessel, optic disks has very much Effect.

, it is necessary to be analyzed to feature, trained, screened and classified after foundation characteristic is extracted, so can be only achieved final Purpose.Neural network algorithm is the best quasi-mode matching algorithm of classifying quality, and it carries out spy by a small amount of mark data Levy extraction, training is finely adjusted to different network layers, and network characterization is polymerized to process the classification of large data sets And regression problem, the network can select bioengineering network, artificial neural network etc., more authentic and valid simulation clinical practice Environment, by the screening to eyeground picture feature, classification, so as to realize the assistant analysis to clinical treatment.

The above, the only present invention preferably specific embodiment, but protection scope of the present invention is not limited thereto, Any one skilled in the art in the technical scope of present disclosure, technology according to the present invention scheme and its Inventive concept is subject to equivalent or change, should all be included within the scope of the present invention.

Claims (8)

1. it is a kind of screen eye fundus image method, it is characterised in that comprise the following steps that：

S1：Receive the eyeground picture for shooting；

S2：Using Morlet small wave converting methods to image segmentation, carry out the extraction of vascular skeleton；

S3：Eyeground picture is carried out Gabor wavelet conversion process；

S4：For Gabor wavelet transformation results, blood vessel is carried out using svm classifier model and non-vascular is sorted out, and record blood vessel seat Mark region；

S5：Using the sorted blood vessel coordinate information of Gabor wavelet, the blood vessel coordinate extracted with Morlet small echos carries out region ratio It is right；

S6：According to image type, signature is carried out；And the feature for mark carries out learning training using neutral net；

S7：Svm classifier is carried out to the feature that neutral net is extracted.

2. according to claim 1 it is a kind of screen eye fundus image method, it is characterised in that the above method, also include：

S8：Based on the characteristic model of training, incoming eyeground picture is matched；By the marker characteristic trained, eyeground picture is entered Row automatic screening and classification.

3. it is according to claim 1 or claim 2 it is a kind of screen eye fundus image method, it is characterised in that to receiving in step S1 Eyeground picture, carry out pretreatment separation, and image is optimized using self-adapting histogram equilibrium.

4. according to claim 3 it is a kind of screen eye fundus image method, it is characterised in that extract it vascular skeleton is carried out Afterwards, using tiny blood vessels part in multiple dimensioned Gauss matched filtering optimization image, and the image after Gauss matched filtering is carried out Binary conversion treatment.

5. according to claim 3 it is a kind of screen eye fundus image method, it is characterised in that Gabor wavelet conversion process, its Core is Gaussian function, and the gray-scale intensity distribution of vessel cross-sections is approached by adjusting the parameter of Gaussian function.

6. a kind of method for screening eye fundus image according to claim 3, it is characterised in that artificial right according to image type Blood vessel, arch bridge, optic disk, optic cup, macula lutea, blackspot, hickie etc. carry out signature.

7. according to claim 3 it is a kind of screen eye fundus image method, it is characterised in that the Morlet wavelet transformation sides Method, specially：

$T\mathrm{\ω}(b,\mathrm{\θ},a)={\mathrm{C\ω}}^{-1/2}\frac{1}{a}\∫{\mathrm{\ω}}^{*}(a^{-1}r-\mathrm{\θ}(x-b\left)\right)/\left(x\right)d^{2}x;$

Wavelet conversion coefficient T ω (b, θ, a) reflect the similarity degree of gradation of image distribution curve and wavelet sequence function, its Middle * represents conjugation；C ω represent normalization constant；ω represents the small echo to be analyzed；B is displacement vector；θ is the anglec of rotation；A is Scale factor, r is wave filter.

8. a kind of method for screening eye fundus image according to claim 3, it is characterised in that Gabor wavelet transformation for mula is：

$G_f(a,b,\mathrm{\ω})={\∫}_{-\mathrm{\∞}}^{\mathrm{\∞}}f\left(t\right)g_a(t-b)e^{-t\mathrm{\ω}t}dt$

Wherein,It is Gaussian function, referred to as window function, wherein a>b>0, g_a(t-b) when being one Between localize " window function ", wherein parameter b be used for move in parallel window, e is natural Exponents.