CN105894498A - Optical coherent image segmentation method for retina - Google Patents

Abstract

The invention discloses an optical coherent image segmentation method for a retina. Filtering is carried out on a retina macula lutea image line by line by means of Gaussian filter; an initial local contour of the image is obtained by using a multi-resolution-ratio method; and then a middle contour of the retina macula lutea image is obtained rapidly by using a level set method to obtain a final image segmentation result. According to the invention, the edge detection result is accurate and the calculation speed is fast. Retina segmentation can be completed without an initial seed point. The method can be applied to the clinic practice conveniently; and the practicability is high.

Description

A kind of retina optics coherent image dividing method

Technical field

The present invention relates to a kind of retina optics coherent image dividing method.

Background technology

At medical domain, macula retinae thickness can be used to quantify the disease such as diabetic macular edema and age-related macular degeneration, the most generally uses the image technology of optical coherence tomography to obtain macula lutea image.But existing macula lutea image partition method arithmetic speed is relatively slow, hinders its Clinical practice.OCT retinal image segmentation method in early days is mainly based upon gray threshold and grey scale change, and these methods are to noise-sensitive and time-consuming.Koozekanani et al. proposes a kind of method of markov random file (Markov random field, MRF) and extracts inner retina and external margin, and the robustness of this autoregression model is better than those methods based on gray threshold.It require that initial seed point just can complete segmentation amphiblestroid to pathology reliably.Mujat et al. have employed the method for deformation batten to split retina nerve layer, and it needs deformation batten to be put near initial profile, and generally also ratio is relatively time-consuming.Chiu points out that the most of speed of method for the segmentation of two and three dimensions OCT image reported in document is relatively slow, and this causes their practicality deficiency clinically.

Summary of the invention

The technical problem to be solved is, not enough for prior art, it is provided that a kind of retina optics coherent image dividing method.

For solving above-mentioned technical problem, the technical solution adopted in the present invention is: a kind of retina optics coherent image dividing method, comprises the following steps:

1) Optical coherence tomography method is utilized to obtain original image H0W0；

2) described image H0W0 is carried out one-dimensional gaussian filtering, obtain H0W0_GAU；

3) use mean filter that described H0W0_GAU is filtered, image H0W0_GAU is compressed to the half of H0W0_GAU by the yardstick H1W0 using multiresolution at vertical direction, obtain image H1W0 and then repeat change of scale several times, including follow-up yardstick H2W0, yardstick H3W1 and last yardstick H4W2；

4) along each column image being asked for respectively on vertical direction the maximum vertical coordinate of upwards difference and downward difference on image H4W2, respectively as each column vertical coordinate position at internal limiting membrane (ILM) edge and layer of retina,pigment epithelium (RPE) edge in image H4W2.

5) in image H4W2, each column vertical coordinate position and the abscissa positions at ILM edge are multiplied by 16 and 4 respectively, as each column vertical coordinate at ILM edge in image H0W0_GAU and abscissa initial position；In H4W2, each column vertical coordinate position and the abscissa positions at RPE edge are multiplied by 16 and 4 respectively, as each column vertical coordinate position at RPE edge in image H0W0_GAU and abscissa initial position.

6) Level Set Method is utilized to through described step 5) chatted image H0W0_GAU carries out level-set segmentation, obtain edge contour C_ILM and C_RPE, and make fitting of a polynomial, the C_ILM_R (internal limiting membrane (ILM) edge contour) being in the image H0W0 to be split obtained and C_RPE_R (layer of retina,pigment epithelium (RPE) edge contour) edge.

7), after obtaining edge contour, use the method for mobile polynomial regression that edge contour is carried out fairing processing.Fairing processing method particularly includes: be current point with the certain point on edge contour, before and after current point, respectively take 16 points, 33 data carry out polynomial regression altogether, and polynomial exponent number is 5 rank.

Described original image resolution is 2000 × 2048.

Described mean filter is the mean filter of 9 × 9.

Compared with prior art, the had the beneficial effect that edge detection results of the present invention of the present invention is accurate, calculates speed fast, it is not necessary to initial seed point can complete amphiblestroid segmentation, it is simple to uses clinically, practical.

Accompanying drawing explanation

Fig. 1 is the inventive method flow chart；

Fig. 2 is speckle layer Level Set Method segmentation result schematic diagram of the present invention；

Fig. 3 is that macula lutea layer of the present invention splits final result schematic diagram.

Detailed description of the invention

1, Data Source

By Optical coherence tomography method, we acquire 311 width macula lutea image altogether from 34 patients, and the resolution of original image is 2000 (depth direction) * 2048 (width).

2, macula lutea image partition method based on multiresolution and level set

In order to the macular thickness of complementary medicine is measured, need to obtain the clear profile of macula lutea image, the present invention designs a kind of new macula lutea image partition method based on multiresolution and level set, first by one-dimensional gaussian filtering, original image is filtered by row, multiresolution method is used to obtain image initial local configuration again, finally use Level Set Method can obtain final image segmentation result with the intermediate profile of quick obtaining macula lutea image.The processing procedure of the inventive method is as shown in Figure 1.

(1) gaussian filtering

Gaussian filtering is a kind of linear smoothing filtering, it is adaptable to eliminates Gaussian noise, is widely used in the noise abatement process of image procossing.

Being horizontal layer structure in view of macula retinae image appearance, and in order to save operation time, we use the gaussian filtering of horizontal direction, its formula is as follows:

$g\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}{e}^{-0.5\×{\left(\frac{x}{\mathrm{\σ}}\right)}^2}---\left(1\right)$

Wherein, x is image H0W0 pixel level direction coordinate, and σ is the width parameter of kernel function, determines the radial effect scope of function, π and e is respectively pi and the natural truth of a matter.The horizontal direction gaussian kernel function parameter that the present invention uses is: σ=5.0, window size N=31.

The present invention uses gaussian filtering to be filtered the often row of source images.Needing to illustrate that we are only filtered by row, the reason using gaussian filtering the most in a column direction is so to cause the movement of marginal position.Additionally, we use one-dimensional gaussian filtering, this is because 2-d gaussian filters speed is relatively slow, and two one-dimensional gaussian filterings can be decomposed into.

(2) image down sampling

Multiresolution analysis is a kind of localization time frequency analysis algorithm, can be the powerful analyzing non-stationary signal with time domain and combining of frequency domain representing signal.It carries out multiscale analysis by computing multi signal such as flexible, the translations of basic function, can effective extraction information from signal, be one flexibly, higher-dimension signal processing algorithm fast and effectively.

The present invention carries out down-sampling by multiresolution method and obtains the initial profile of macula lutea image, implements step and comprises the following steps.First, use the mean filter of 9 × 9 to the macula lutea image filtering after gaussian filtering, then use the yardstick H1W0 of multiresolution, image is compressed at vertical direction the half of source images.Then change of scale several times is repeated, including follow-up yardstick H2W0, yardstick H3W1 and last yardstick H4W2.

(3) level set algorithm

Level set algorithm is that Sethian and Osher proposed in 1988, for some calculating of low dimensional are risen to the highest one-dimensional, the description of N-dimensional is regarded as a level of N+1 dimension.

Generally, Level Set Method is expressed as continuous function curved surface φ by implicit for planar closed curve (x, y, t) has the same value curve of same functions value.For example, it is possible to represent that (x, y, in t)=0 at zero level collection φ by implicit for aim curve.I.e. t corresponding to zero level collection is

C (p, t)=(x, y) | φ (x, y, t)=0} (2)

Assume for the planar closed curve developed be C (p, t)=((p, t), y (p, t)), wherein p is arbitrary parameterized variables to x, and t is the time.If the interior of curve is N to normal vector, curvature is k, then curve develops can represent by following partial differential equation along the direction of its normal vector

$\frac{\∂C}{\∂t}=V\left(k\right)N---\left(3\right)$

Due to φ (C (t), t)=0, t is carried out total differential, arrange obtain level set and carry out the equation of curve evolvement be

$\frac{\∂\mathrm{\φ}}{\∂t}=-\▿\mathrm{\φ}V\left(k\right)N=\▿\mathrm{\φ}V\left(k\right)\frac{\▿\mathrm{\φ}}{|\▿\mathrm{\φ}|}=V\left(k\right)|\▿\mathrm{\φ}|---\left(4\right)$

Knowable to above-mentioned analysis, Level Set Method realizes active contour model several advantage: first, and evolution curve can change topological structure along with the evolution of φ, can divide, merge formation wedge angle etc.；In fact, owing to φ remains a complete function in evolutionary process, approximate numerical calculation is easily realized；3rd, Level Set Method can expand to the evolution of higher-dimension curved surface, simplifies the complexity of three-dimensional segmentation.Therefore, we can use quick Level Set Method to obtain intermediate profile.

During implementing, we have employed Level Set Method to obtain the edge contour of macula lutea layer, uses Numford-Shah Level Set Models.The energy minimization formula that this model uses is as follows:

E (u, c)=∫ Ω | | u-u₀||²dxdy+μ∫Ω\c|▽u|²dxdy+v×length(c) (5)

Wherein, μ and v is nonnegative constant, takes 0.3 and 0.5 in the present invention respectively.Ω represents image H0W0_GAU region to be split, and c represents the edge contour that needs are split, and u represents and uses the pixel grey scale average of H0W0_GAU after automatic division method segmentation, u₀The gray average of actual macular region, represent the pixel grey scale gradient in image to be split, length represents length.

After carrying out level-set segmentation calculating intermediate profile for H0W0_GAU, the middle segmentation result at the macula lutea layer edge obtained is as shown in Figure 2.

(4) edge-light using mobile polynomial regression is suitable

As can be seen from Figure 2, the edge using Level Set Method to obtain is rough.We can use the method for mobile polynomial regression that boundary curve is carried out SmoothNumerical TechniqueandIts so that it is line smoothing, removes noise.

In the present invention, we use the method for mobile polynomial regression that boundary curve is carried out fairing.Namely centered by current point, respectively taking 16 points, 33 data carry out polynomial regression altogether, and polynomial exponent number is 5 rank.Use the result such as Fig. 3 after the method.

Table 1 is to use manual segmentation method and the measurement result comparison diagram of the inventive method, it is seen that the method measurement result of the present invention is accurate, and precision is high.

Table 1 collects 31 width image macular thickness measurement results to 34 patients

Claims (6)

1. a retina optics coherent image dividing method, it is characterised in that comprise the following steps:

1) Optical coherence tomography method is utilized to obtain original macula retinae image H0W0；

2) described original image is carried out one-dimensional gaussian filtering, obtain image H0W0_GAU；

3) use mean filter that described image H0W0_GAU is filtered, after then being filtered by mean filter Image change on vertical and horizontal yardstick limit of gradually doing, obtain the image H4W2 after change of scale；

4) along each column image being asked for respectively on vertical direction upwards difference and downward difference on image H4W2 Maximum vertical coordinate, respectively as internal limiting membrane edge in image H4W2 and layer of retina,pigment epithelium edge Each column vertical coordinate position；

5) in image H4W2, each column vertical coordinate position and the abscissa positions at internal limiting membrane edge are multiplied by 16 Hes respectively 4, as each column vertical coordinate at internal limiting membrane edge in image H0W0_GAU and abscissa initial position； In H4W2, each column vertical coordinate position and the abscissa positions at layer of retina,pigment epithelium edge are multiplied by 16 respectively With 4, as each column vertical coordinate position at layer of retina,pigment epithelium edge in image H0W0_GAU and horizontal stroke Coordinate initial position；

6) utilize Level Set Method to through described step 5) process after image carry out level-set segmentation, obtain limit Edge profile C_ILM and C_RPE, and make fitting of a polynomial, i.e. obtain original macula retinae image H0W0 In C_ILM_R and C_RPE_R edge.

Retina optics coherent image dividing method the most according to claim 1, it is characterised in that described original Image resolution ratio is 2000*2048.

Retina optics coherent image dividing method the most according to claim 1, it is characterised in that described average Wave filter is the mean filter of 9 × 9.

Retina optics coherent image dividing method the most according to claim 1, it is characterised in that use average Described image H0W0_GAU is filtered by wave filter, then by filtered for mean filter image H0W0_MEAN changes on vertical and horizontal yardstick limit of gradually doing, and i.e. carries out image down sampling, is first vertically On direction, H0W0_MEAN is compressed half, obtain image H1W0, then repeat yardstick several times and become Change, obtain follow-up image H2W0, image H3W1 and last image H4W2.

Retina optics coherent image dividing method the most according to claim 1, it is characterised in that obtain edge After profile C_ILM and C_RPE, use the method for mobile polynomial regression that edge contour is carried out fairing processing, Obtain C_ILM_R and C_RPE_R.

Retina optics coherent image dividing method the most according to claim 5, it is characterised in that fairing processing Method particularly includes: it is current point with the certain point on edge contour, before and after current point, respectively takes 16 points, always Totally 33 data carry out polynomial regression, and polynomial exponent number is 5 rank.