KR100952668B1 - Apparatus and method for tracking retinal vessel using Canny edge detection method - Google Patents

Abstract

The present invention is directed to an apparatus and method for measuring and tracking retinal vessels with accuracy within one pixel. The present invention is based on a modified Canny edge detection method, a Gaussian model of the vascular profile and a zero-crossing of the second derivative of the Gaussian model. In the present invention, the canny edge detection method is first used to detect the edge from the input blood vessel image. The edge direction is then set, which is used to obtain the cross-sectional profile of the vessel. The vessel sample profile is then retrieved from the vessel cross-sectional profile using the local maximum and local minimum. The Gaussian model is then used to fit the vascular sample profile. Next, the sign transform point of the second derivative of the vascular sample profile obtained in Gaussian is used to indicate the vessel boundary. Next, the peak value of the vascular sample profile obtained in Gaussian is used as the position of the vessel centerline and used to calculate the width of the vessel at this position. According to the present invention, the position of the vessel wall and the width of the vessel can be obtained with accuracy within one pixel.

Retinal Vessel Tracking, Retinal Vessel Centerline Tracking, Vessel Width, Canny Edge Detection

Description

Apparatus and method for tracking retinal vessel using Canny edge detection method}

The present invention relates to a retinal vessel tracking device and method, and more particularly, to a device and method for tracking the vessel of the retina using a modified Canny edge detection method.

Tracking technique is one of the important vessel segmentation algorithms. In retinal vessel detection, two important issues exist with respect to vessel measurement and tracking. One is to detect the vessel boundary location, and the other is to track the vessel's centerline at each point and measure the vessel's width. Edge detection and matched filters are used to track retinal vessels. To detect the left and right vessel edges, the parallel edge detection technique uses a bar model line model. Conventional line detection methods have been used to track parallel edges or detect ridges. Edge detectors such as canny detectors and Sobel detectors are widely used to detect vessel boundaries. Morphological detectors, gradient calculators, direct matched low pass difference templates and optimized canny detectors were also used to detect blood vessels. Vascular profiles are described by a modified Gaussian model that takes into account the central light reflection of the small arteries, and the vascular measurements are performed using a size-corrected secondary Gaussian filter.

Canny edge detection is based on the size and orientation of the local gradient, not directly on the image intensity. Canny edge detectors can robustly detect important anatomical features in retinal images without compensating for irregularities in the amount of light by any enhancement process or Gaussian smoothing, extracting boundaries, detecting edges of blood vessels, and detecting different configurations and backgrounds It has been used to separate.

The present invention is developed based on the properties observed for blood vessels in the retinal image as follows.

1) The density distribution of the cross-sectional shape of blood vessels can be evaluated using a function of Gaussian shape.

2) The vascular segment direction changes continuously, and the change in direction between the fragments is a smooth continuous function.

3) The width of the vessel changes continuously, and there is no instant change in the width of the vessel fragment. There is always a smooth change between adjacent fragments. Given the experimental assumption that the vessel width is a function of a matched filter when the size coefficient of the filter is properly given, the absolute value of the vessel diameter can be determined using a simply pre-set line.

In prior art vessel tracking, the detected vessels describe vessel profiles using a Gaussian model tracked and modified by combining Kalman and Gaussian filters. The width of the vessel is also obtained by data fitting.

Kalman filter is a recursive algorithm for obtaining an optimal estimate of a state vector in a linear dynamic system interrupted by noise. Kalman filters track position of vessels using position filters and velocity filters, and describe vessel profiles using Gaussian filters. The retinal vessel tracking by the combination of the Kalman filter and the Gaussian filter has a high possibility of error deviation of the vessel position and has a problem in accurate vessel position and vessel width calculation by obtaining the vessel width by data fitting.

SUMMARY OF THE INVENTION The present invention has been made in an effort to provide an apparatus and method for tracking blood vessels of the retina, which are guaranteed to be within one pixel using a modified Canny edge detection method.

Another technical problem to be solved by the present invention is to use a computer-readable recording medium that records a program for executing a method of tracking blood vessels of the retina, which is guaranteed to be within one pixel using a modified Canny edge detection method. To provide.

In order to achieve the above technical problem, a retinal blood vessel tracking device according to the present invention comprises: an edge detector for detecting an edge from an input image and outputting an edge image; A curve tracker for tracking a curve of a blood vessel wall expressed as an index of an edge and a distance between two neighboring edges from the edge image; A direction vector of an edge point is calculated based on the amount of change in each horizontal and vertical direction of each pixel constituting the tracked vessel wall curve, and the vessel wall profile is divided by dividing the curve of the vessel wall in a direction perpendicular to the direction vector. Blood vessel cross-sectional profile extraction unit for extracting; A blood vessel sample profile determiner determining a portion located between local minimum values from the blood vessel cross-sectional profile as a blood vessel sample profile; A Gaussian fit is generated based on the Gaussian model of each of the blood vessel sample profiles, and the vessels having an accuracy within one pixel are examined by examining the sign transform point of the second derivative of the generated Gaussian fit. An edge position obtaining unit for obtaining an edge position of the wall; And a data generator configured to generate at least one of a width of a blood vessel, a position of the blood vessel, and a centerline of the blood vessel based on the edge position and the peak value of the Gaussian pit.

In accordance with another aspect of the present invention, there is provided a retinal vessel tracking method including: an edge detection step of detecting an edge from an input image and outputting an edge image; A curve tracking step of tracking a curve of the vessel wall represented by the index of the edge and the distance between two neighboring edges from the edge image; A direction vector of an edge point is calculated based on the amount of change in the horizontal and vertical directions of each pixel constituting the tracked vessel wall curve, and the vessel cross section is divided by dividing the curve of the vessel wall in a direction perpendicular to the direction vector. Vascular cross-sectional profile extraction step of extracting the profile; A vessel sample profile determination step of determining a portion located between local minimums from the vessel cross-sectional profile as a vessel sample profile; A Gaussian fit is generated based on each vascular sample profile based on a Gaussian model, and the vascular wall has an accuracy within one pixel by investigating the sign transform point of the second derivative of the generated Gaussian fit. Obtaining an edge position of the edge position; And generating at least one of a blood vessel width, a blood vessel position, and a blood vessel centerline based on the edge position and the peak value of the Gaussian pit.

According to the retinal vessel tracking apparatus and method according to the present invention, the modified canny edge detection method can be used to efficiently track the retinal vessel centerline and measure the vessel width at each center point. In addition, more accurate vessel wall location can be found using curve smoothing or canny edge detection methods with accuracy within one pixel, and the vessel's centerline position and width of the vessel at the center point can be determined based on the vessel cross-sectional profile. . The present invention uses a Gaussian model to reorient the edges and fit the blood vessel cross-sectional profile to obtain an edge with an accuracy of one pixel or less from the Gaussian model, thereby ensuring accurate tracking of the retinal vessels.

Hereinafter, with reference to the accompanying drawings will be described in detail a preferred embodiment of the retinal vessel tracking device and method according to the present invention.

1 is a block diagram of a retinal blood vessel tracking apparatus according to the present invention, Figure 2 is a view showing a flow chart of the retinal vessel tracking method according to the present invention.

1 and 2, the retinal vessel tracking apparatus according to the present invention includes an edge detector 110, a curve tracer 120, a smoothing unit 130, a blood vessel cross-sectional profile extraction unit 140, and a blood vessel sample profile determination. The unit 150 includes an edge position obtaining unit 160 and a data generating unit 170.

The edge detector 110 detects an edge from the input image and outputs an edge image (S210). Edge detection uses Canny edge detection, which provides a priori knowledge of vessel boundaries and gradient direction. Canny edge detection is a method that finds a point where the change in intensity is greater than another point in the input image, tracks the point, and extracts points belonging to the formed line. Canny edge detectors minimize the likelihood of error variation with respect to edge location, detect edges close to the vessel wall, and finally provide an edge to the vessel wall point. In addition, the edge detector 110 may detect the left and right blood vessel edges based on a bar-shape line model. 3A shows an example of a vascular sample extracted from a small region of the retinal image, and FIG. 3B shows an example of a pixel position and a vascular wall direction that can be extracted from the vascular sample shown in FIG. 3A using a Canny edge detector. Is shown.

The curve tracker 120 tracks the curve of the blood vessel wall expressed as the distance between the index of the edge and two neighboring edges from the edge image (S220). C (u, r) = (x (u), y (u)) is the vessel wall curve, u is the index of the edge, r is the distance between two neighboring edges, and (x, y) is the Means the position of the pixel.

The smoothing unit 130 smoothes the curve of the blood vessel wall tracked by Equations 1 and 2 below (S230).

는 가우시안 스무딩 함수이다. Where X (u, r) is the convolution of the Gaussian smoothing function and function x (u), Y (u, r) is the convolution of the Gaussian smoothing function and function y (u), and r is two neighboring edges. Distance, u is the index of the edge, t is the integral variable, x (t) is the horizontal position function of the pixel, y (t) is the vertical position function of the pixel, and

Is a Gaussian smoothing function.

4 illustrates an edge direction vector field on a blood vessel image based on the blood vessel curve detected from the canny edge. Referring to FIG. 4, the direction of movement of the edge points located along the vessel wall is not smoothed, resulting in the cross-sectional profiles 410 and 420 of some neighboring vessels intersecting. In this case, both the edge position and the center point position of the blood vessel shape may be set incorrectly. In order to solve this problem, in the present invention, the edge direction is defined as the direction of the edge moving along the vessel wall in two-dimensional space. The filter is then used to smooth the direction of the neighboring edges.

5 is a diagram illustrating an edge direction vector field after alignment using a seven-step smoothing process. Referring to FIG. 5, the edge direction after alignment has a small change along the vessel wall. The smoothing process aligns the direction of movement of the vessel wall so that the cross-sectional profiles 510 and 520 of neighboring vessels do not cross each other. Unlike the gradient direction in the canny edge detection method, in the present invention, the edge direction is defined as a unit vector represented by the following equation.

Where u is the index of the edge and (x, y) is the position of the pixel in the image.

6 is a diagram illustrating a process of extracting a blood vessel cross-sectional profile while moving a search window along a blood vessel wall. Referring to FIG. 6, the blood vessel cross-sectional profile extractor 140 calculates a direction vector of an edge point based on a change amount of each pixel constituting the tracked blood vessel wall in the horizontal and vertical directions, and curves of the blood vessel wall. Is divided in the direction perpendicular to the direction vector to extract the blood vessel cross-sectional profile (S240). In this case, the blood vessel cross-sectional profile extractor 140 is defined as a blood vessel cross section of an image lying in a direction perpendicular to the direction 640 of the edge while moving the search window 630 having a predetermined size along the vessel walls 610 and 620. The blood vessel cross-sectional profile 650 is extracted. The window 630 extracts a number of vessel cross-sectional profiles while moving in the vertical direction of the edge. The window 630 is composed of a rectangle (the length of the side in the direction perpendicular to the edge direction: X, the length of the side in the edge direction: Y), and the length of the side in the direction perpendicular to the edge direction is the maximum width of the vessel (680). It is set to be more than twice.

The blood vessel sample profile determiner 150 determines a blood vessel sample profile based on a predetermined edge position point and a blood vessel cross-sectional profile corresponding to the Gaussian shape of the blood vessel cross-sectional profile (S250). At this time, it is necessary to find the peak of the blood vessel to search the sample position of the blood vessel sample profile from the blood vessel cross-sectional profile. In the present invention, the peak of the blood vessel sample is defined as the local maximum value closest to the edge point while having a size larger than that of the edge point. And the vascular sample profile is defined as the portion of the vascular cross-sectional profile between two neighboring minimum points of the peak. Accordingly, the blood vessel sample profile determiner 150 finds the peak point of the vessel defined by the local maximum value closest to the edge point while having an intensity greater than that of the peripheral edge point from the vessel cross-sectional profile, the peak point and the peak point. The vascular sample profile is determined based on the point with the minimum neighboring with. The vessel sample profile is a subset of the vessel cross-sectional profile that includes the points with two neighboring local minimums of the edge position.

7 shows the position of the canny edge point, a local maximum and a local minimum. Referring to FIG. 7, the point P with the local maximum of the peak of the vascular sample profile is found first, and the positions of the points U and V with two neighboring minimums with respect to the point P are obtained. Thus, the curve between the points U and P is extracted from one vessel sample profile from the vessel cross-sectional profile shown in FIG. 7. The curve is at both ends with points U and V having the minimum value, and includes the local maximum value P.

The edge acquisition unit 160 generates a Gaussian pit based on each vascular sample profile based on a Gaussian model, examines the sign transform point of the second derivative of the generated Gaussian pit, and then estimates an edge position of a vascular wall having an accuracy within one pixel. Hereinafter, referred to as a "lower pixel edge position" (S260). FIG. 8A is a diagram illustrating a vascular sample profile and Gaussian fit results with the local minimum value points U and V at both ends shown in FIG. 7. 8B shows the lower pixel edge position compared to the canny edge position for the corresponding vessel cross-sectional profile. Referring to FIG. 8B, the Gaussian model is a smoothing function that is slightly magnified than the vascular sample. The lower pixel edge position is calculated by examining the sign transform point of the second derivative of the Gaussian model, and is indicated by a point (·) on the Gaussian fit curve. Compared to the canny edge position marked with an asterisk (*) on the vessel sample profile curve, the lower pixel edge position is shifted left by about one pixel from the canny edge position.

9 is a diagram illustrating lower pixel edge positions and canny edge positions on a local blood vessel image. Referring to FIG. 9, the square mark indicates the position 910 of the canny edge, and the triangle mark indicates the lower pixel edge position 920. FIG. 10 is a diagram illustrating a canny edge detector and different vessel boundaries detected in accordance with the present invention. Referring to FIG. 10, it can be seen that the lower pixel edge positions obtained according to the present invention form a vascular wall having a gentler shape compared to the Kenny edge position.

The data generator 170 generates at least one of the width of the blood vessel, the position of the blood vessel, and the centerline of the blood vessel based on the lower pixel edge position and the peak value of the Gaussian pit (S270). The peak of the Gaussian model is used as the centerline of the vessel, and the width of the vessel at this point is calculated based on the centerline and the new edge point. The curve smoothing or the Canny edge detection method with accuracy within one pixel can be used to find more accurate vessel wall location, and to determine the position of the vessel's centerline and the width of the vessel at the center point based on the vessel cross-sectional profile.

After the method according to the present invention was implemented by MATLAB, experiments were performed with real images. For the experiment, a set of 53 test images with a size of 1032 × 1032 pixels were randomly selected, and the selected test images are shown in FIG. 11. In the experiment, a total of 16,775 edge points were detected by a Canny edge detector having a threshold value set to 0.1 after smoothing an image using a Gaussian filter with a kernel of 1 for the test image shown in FIG. 11. We also used a function to track the edge position in the set of curves to record the edge points along the other curve, and performed a seven-step smoothing process to smooth each curve. The direction vector was calculated using Equation 3 based on the edge positions smoothed along the curve, and each edge point was assigned. And the direction vector corresponding to each edge point was used to select the vessel cross-sectional profile. In this experiment, the cross-sectional profile was selected using a 50 pixel window. The sample portion of the profile was retrieved using the local minimum for each profile, followed by a Gaussian fit. The sign change point of the second derivative of the Gaussian pit was investigated to calculate a new edge with an accuracy within one pixel.

The results according to the invention include the lower pixel edge position estimated from the Kenny edge position and the Gaussian pit, and represent the edge with accuracy within one pixel using the peak position of the Gaussian pit, where the width of the vessel is It was calculated based on the distance between the lower pixel edge position and the peak value of the Gaussian fit model.

12 is a diagram illustrating an extracted center line position of an input blood vessel image. The generated data sample set is described in Table 1 below with respect to the position of the center line illustrated in FIG. 12.

K _x K _y ψ _x ψ _y φ _x φ _y λ 669 26 669.1626 27.0188 666 27.0900 6.3252 508 27 509..222 27.9960 506 28.5370 6.0443 669 27 669.1078 28.0254 666 28.1140 6.2156 669 28 669.1409 29.0292 666 29.1357 6.2819 669 29 669.1584 30.0332 666 30.1580 6.3167 508 30 509.6108 30.9230 507 31.2521 6.2216 669 30 669.1737 31.0381 666 31.1842 6.3473 669 31 669.1921 32.0440 666 32.2177 6.3841 669 32 669.2188 33.0508 666 32.2601 6.4375 509 33 511.2986 33.7165 508 34.4367 6.5972 669 33 669.3123 34.0535 666 34.3113 6.6246 669 34 669.3459 35.0604 666 35.3696 6.6918 510 35 512.1432 35.5679 509 36.7559 6.2865 669 35 669.3555 36.0696 666 36.4323 6.7110 669 36 669.4276 37.0710 666 37.4961 6.8552 868 36 869.0395 37.0200 866 35.4838 6.0790 669 37 669.4639 38.0748 666 38.5582 6.9279 867 37 868.3686 38.1705 865 36.6121 6.7373 866 39 867.4067 40.1525 864 38.8748 6.8134

In Table 1, (K _x , K _y ) is the vessel wall position detected by the Canny edge detection method, (ψ _x , ψ _y ) is the vessel position detected by the method according to the present invention, (φ _x , φ _y ) is the centerline position obtained according to the present invention and λ is the vessel width obtained according to the present invention.

The invention can also be embodied as computer readable code on a computer readable recording medium. The computer-readable recording medium includes all kinds of recording devices in which data that can be read by a computer system is stored. Examples of the computer-readable recording medium include a ROM, a RAM, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like, and may be implemented in the form of a carrier wave (for example, transmission via the Internet) . Computer-readable code can also be stored and executed.

Although the preferred embodiments of the present invention have been shown and described above, the present invention is not limited to the specific preferred embodiments described above, and the present invention belongs to the present invention without departing from the gist of the present invention as claimed in the claims. Various modifications can be made by those skilled in the art, and such changes are within the scope of the claims.

1 is a block diagram of a retinal vessel tracking device according to the present invention;

2 is a flowchart illustrating a retinal vessel tracking method according to the present invention;

3A is a diagram illustrating an example of a blood vessel sample extracted from a small region of a retinal image, and FIG. 3B is a diagram illustrating an example of a pixel position and a vessel wall direction that can be extracted using a canny edge detector.

4 illustrates an edge direction vector field on a blood vessel image based on a blood vessel curve extracted from a canny edge;

5 shows that there is a small change in the direction of the edge located along the vessel wall after edge direction alignment,

6 illustrates extracting a blood vessel cross-sectional profile while moving the search window along the vessel wall;

7 shows the position, local maximum and local minimum of the canny edge point;

FIG. 8A shows the vascular sample profile and Gaussian fit results with the local minimum points U and V at both ends shown in FIG. 7, and FIG. 8B compares the canny edge position for the corresponding vessel cross-sectional profile. A drawing showing the subpixel edge locations,

9 illustrates lower pixel edge position and canny edge position on a local vessel image;

10 shows a canny edge detector and different vessel boundaries detected in accordance with the present invention;

FIG. 11 is a diagram illustrating an image tested by randomly selecting 53 test image sets having a size of 1032 × 1032 pixels.

12 is a diagram illustrating an extracted center line position of an input blood vessel image.

Claims (14)

An edge detector for detecting an edge from an input image and outputting an edge image; A curve tracker for tracking a curve of a blood vessel wall expressed as an index of an edge and a distance between two neighboring edges from the edge image; A direction vector of an edge point is calculated based on the amount of change in the horizontal and vertical directions of each pixel constituting the tracked vessel wall curve, and the vessel cross section is divided by dividing the curve of the vessel wall in a direction perpendicular to the direction vector. A blood vessel cross-sectional profile extraction unit for extracting a profile; A blood vessel sample profile determiner determining a portion located between local minimum values from the blood vessel cross-sectional profile as a blood vessel sample profile; A Gaussian fit is generated based on each vascular sample profile based on a Gaussian model, and the vascular wall has an accuracy within one pixel by investigating the sign transform point of the second derivative of the generated Gaussian fit. An edge position obtaining unit for obtaining an edge position of the edge unit; And And a data generator configured to generate at least one of a width of a blood vessel, a position of the blood vessel, and a centerline of the blood vessel, based on the edge position and the peak value of the Gaussian pit. The method of claim 1, The edge detection unit detects retinal vessel tracking device, characterized in that for detecting the left and right vessel edges based on a bar-shape line model. The method of claim 1, Retinal vessel tracking device further comprises a smoothing unit for smoothing the curve of the vessel wall by the following equations (1) and (2): [Equation 1]

, [Equation 2]

, Where X (u, r) is the convolution of the Gaussian smoothing function g (u, r) and function x (u), and Y (u, r) is the Gaussian smoothing function g (u, r) and function y (u) Convolution of, r is the distance between two neighboring edges, u is the index of the edge, t is the integral variable, x (t) is the horizontal position function of the pixel, y (t) is the vertical position function of the pixel, And the Gaussian smoothing function.

to be. The method of claim 1, The vessel section profile extracting unit extracts the vessel section profile defined as a vessel section of an image lying in a direction perpendicular to an edge direction while moving a search window having a predetermined size along the vessel wall. Device. The method of claim 4, wherein The length of the sides in the direction perpendicular to the edge direction of the window is greater than twice the maximum width of the vessels retinal vessel tracking device. The method according to any one of claims 1 to 5, The blood vessel sample profile determiner finds the peak point of the vessel defined by the local maximum value closest to the edge point having an intensity greater than the strength of the peripheral edge point from the vessel cross-sectional profile, and the peak point and the peak point and the neighbor Retinal vessel tracking device, characterized in that for determining the blood vessel sample profile based on the point having a minimum value. The method according to any one of claims 1 to 5, And the data generator calculates a width of the blood vessel based on a distance between the edge position and the peak value of the gaussian pit. An edge detection step of detecting an edge from an input image and outputting an edge image; A curve tracking step of tracking a curve of the vessel wall represented by the index of the edge and the distance between two neighboring edges from the edge image; A direction vector of an edge point is calculated based on the amount of change in the horizontal and vertical directions of each pixel constituting the tracked vessel wall curve, and the vessel cross section is divided by dividing the curve of the vessel wall in a direction perpendicular to the direction vector. Vascular cross-sectional profile extraction step of extracting the profile; A vessel sample profile determination step of determining a portion located between local minimums from the vessel cross-sectional profile as a vessel sample profile; Gaussian pits are generated based on the respective vascular sample profiles, and the sign positions of the second derivatives of the generated Gaussian pits are examined to determine the edge positions of the vascular walls having an accuracy within one pixel. Obtaining an edge position; And And generating at least one of the width of the blood vessel, the position of the blood vessel, and the centerline of the blood vessel, based on the edge position and the peak value of the Gaussian pit. The method of claim 8, Retinal vessel tracking method further comprises a smoothing step of smoothing the curve of the vessel wall by the following equations (1) and (2): [Equation 1]

, [Equation 2]

, Where X (u, r) is the convolution of the Gaussian smoothing function g (u, r) and function x (u), and Y (u, r) is the Gaussian smoothing function g (u, r) and function y (u) Convolution of, r is the distance between two neighboring edges, u is the index of the edge, t is the integral variable, x (t) is the horizontal position function of the pixel, y (t) is the vertical position function of the pixel, And the Gaussian smoothing function.

to be. The method of claim 8, In the step of extracting the blood vessel cross-sectional profile, the blood vessel cross-sectional profile defined as the blood vessel cross-section of the image lying in the direction perpendicular to the direction of the edge while moving a search window having a predetermined size along the vessel wall is extracted. Vessel tracking method. The method of claim 10, The length of the side in the direction perpendicular to the edge direction of the window is greater than twice the maximum width of the vessels retinal vessel tracking method. The method according to any one of claims 8 to 11, The blood vessel sample profile determination step, Finding a peak point of the vessel defined by the local maximum value closest to the edge point having an intensity greater than that of the peripheral edge point from the vessel cross-sectional profile; And Determining the vascular sample profile based on the peak point and a point having a minimum value adjacent to the peak point. The method according to any one of claims 8 to 11, And the data generating step calculates the width of the blood vessel based on the distance between the edge position and the peak value of the Gaussian pit. A computer-readable recording medium having recorded thereon a program for executing the retinal vessel tracking method according to any one of claims 8 to 11.

Images