Mechanism for Diabetic Retinal Blood Vessel Profile Measurement and Analysis on Fundus Images

 

T. Jemima Jebaseeli^1*, C. Anand Deva Durai²

¹Assistant Professor, Department of Computer Sciences Technology, Karunya University, Coimbatore 641114, Tamilnadu, India.

²Assistant Professor, Department of Computer Science and Engineering, King Khalid University, Abha 61421, Saudi Arabia.

 

ABSTRACT:

Diabetic Retinopathy (DR) occurs due to Type II diabetes. At the early stage, if it is identified one can save their vision. Later stage, retinal detachment leads to 100% vision loss. An automatic computer based system is needed for diagnosis. There are diverse tools and automatic and semi-automatic systems are available. But the system is not identifying and measuring the narrow blood vessels accurately, because of the noise and imaging problems. Also, while tracking the retinal vessel, the narrow vessels are equally taken into consideration as wider vessels. Thus the proposed segmentation and classification techniques extract the blood vessels and measure the profile features of fundus images obtained from dissimilar modalities significantly.

 

 

 

INTRODUCTION:

Fundus images provide information about early signs of ophthalmic changes in the eye among diabetic patients. Analyzing the fundus image is the important task of ophthalmologists to predict the early signs of Diabetic Retinopathy and save the life of diabetic patient^1,2. At the later stage one cannot save from vision loss. The retinal vascular system of the human being is diagnosed by the characteristics of its arterioles and venules parents and its trunk vessels. Diabetic Retinopathy image analysis is the important field for the early diagnosis. A change in the micro vascular structure is the reason for many diseases^3,4. The lesions present in the retina causes alterations in the blood flow of cerebral vascular system which may cause cardio vascular problems. Hence there is a need of medical imaging system to compute various parameters needed to assist the ophthalmologists to diagnose the patients’ health^5,6.

 

 

 

Measurement of retinal vasculature depends on the accuracy of segmentation result. The Arteriolar-to-Venular (AVR) measurement is used to forecast the risk of hypertension among the patients⁷. There are various computation methods used for this calculations, but it reduces the real magnitude of clinical associations^8,9. The diameters of these vessels are measured from the fundus photographs. Also, the existing algorithm uses different constant variables in the measurement, intern dependent on the units¹⁰. The results diagnosed by two experts on the same image differently. Hence there is a need of a system to report the analysis in a unique way, incase if it is analyzed by multiple experts¹¹. There are several methods are undertaken to test the reliability of this fully automatic methodology. In this proposed system, retinal blood vessels are extracted by Kirsch’s template, segmented by Fuzzy C-Means, optic disc is removed by region based active contour method and then the vessel network is labeled. The retinal vessel tracking methods are intrinsically efficient and provide the meaningful description about the vessel network. The branches are detected at its bifurcation point and the vessel  networks are recorded. The vessel tracking is initiated from the optic disc region. The branching points which are identified during the tracking process are included in the vessel map. At the next level, this branching point acts as the starting point for tracking.

The iteration stop still it finds the end point. The following issues are addressed in this paper.

i.     There are some methods insufficient to provide the complete segmentation result in the case that the  blood vessels fade away from the middle to its extended direction.

ii.    Losing one branching point in tracking may lead to an incomplete blood vessel network and raises error.

iii.  There are algorithms which ignore the minor blood vessels at the branching point.

iv.  The premature termination of tracking process diminishes the vessel network.

v.    The aim of this research is to segment the retinal blood vessels without any discontinuities and quantitative measurement of all retinal blood vessels from its branching points.

 

LITERATURE REVIEW:

There are several automatic software tools produced by various researchers for measuring retinal vessel topology. The tools are explained as follows. ROPNET¹² is a semi-automatic tracking method and computes the tortuosity index narrow field images. ROPTOOL¹³ finds the retinal blood vessel path and measures its width and tortuosity. CAIAR¹⁴ is developed using Python and Pearl. This system is intended to measure the retinal blood vessel width and tortuosity of the school children. SIRUS¹⁵ is a web based system for retinal image analysis. It allows remote access and provides a four tier collaborative framework for experts. The system contains a web based client user interface and an application server for service delivery and the service module for the analysis of retinal microcirculation. A semi- automatic methodology based Arteriolar-to-Venular Ratio (AVR) system is implemented using J2EE and Open CV, XHTML, CSS, JavaScript and AJAX. IVAN¹⁶ is used to find the vessel path and analyze the AVR index. VAMPIR¹⁷ provides the retinal information of optic disc, vasculature, vessel width, branching coefficients and tortuosity measures. It is a semi- automatic method for quantifying the retinal vessel properties. SIVA¹⁸ extracts the retinal vascular structure and provides quantitative measure of retinal image. Live Vessel¹⁹ is a semi-automatic segmentation software application for locating vessels and vascular trees in 2D color retinal images. There are various limitations with the existing methodologies. The above systems analyze the information within the restricted area around optic disc. There are few parameters has been considered for differentiating retinal arterioles from venules. The vessel width and tortuosity was mainly considered. The processing time is more for a single image. IVAN takes 20min to process every single image. Clinically this solution is unfeasible. These systems are developed a small dataset. It cannot run automatically on large amount of fundus images.

 

SYSTEM OVERVIEW:

The proposed system is a tool for measuring the retinal vessel topology automatically on large number of image datasets. Here, the left and right eye of the patient’s images is considered for automatic multilevel measurement analysis. The proposed system helps the ophthalmologists in analyzing the diabetic retinal images. Fig 1 shows the overview of the proposed system.

 

INPUT FUNDUS IMAGE:

The fundus images are taken from publically available multimodal databases such as DRIVE, STARE, REVIEW, High Resolution Fundus (HRF) and DRIONS databases²⁰. There are 40 images from DRIVE, 20 images from STARE, 16 images from REVIEW, 15 images from HRF and 110 images from DRIONS databases used for the experimental research.

 

PREPROCESSING:

Fundus images obtained from different fundus camera have lower reflectance over the retinal vessels, while compared with its background surface. Hence retinal vessels appear darker than the background. Sometimes a light streak is included on the central length of the blood vessels. The green channel is extracted to remove the brighter strip. CLAHE (Contrast Limited Adaptive Histogram Equalization) removes the noise and to enhance the contrast of the image. It calculates the contrast transform function for each region individually. The contrast of each tile is enhanced in the histogram of the output region matches to the target parameter.

 

 

 

Fig 1. The overview of the proposed system for retinal blood vessel segmentation.

RETINAL BLOOD VESSEL SEGMENTATION:

The enhanced green channel of the input image obtained through CLAHE is used for the segmentation process. In the proposed retinal blood vessel segmentation process, Kirsch’s template contains the features of the retinal blood vessels. This template is used to find out the large and small blood vessels. Once the blood vessels are traced then it is segmented from the background using Fuzzy C-Means. The optic disc location is identified and removed using Region based Active contour model.

 

KIRSCH’S TEMPLATE:

The Kirsch’s template contains values of the retinal blood vessels found in 8 directions such as south, east, north, west, northeast, southeast, southwest and northwest is shown in Fig. 2. The Kirsch’s operator is used to find the edge magnitude at all direction.

 

Fig 2. Kirsch’s convolution kernels.

 

The Kirsch’s template produces an image containing grey level pixels of value 0 or 255. The pixel with value 0 indicates a black pixel and value 255 indicates a white pixel. Accordingly, the threshold is adjusted to find the edges of the retinal blood vessel image.

 

Fig 3. (a) Retinal image with vessel centerline and outline boundary edges,(b) Vessel centerline image, (c) 126 labelled vessel segment edges, (d) Location of segment-126.

FUZZY C-MEANS:

Fuzzy C-Means (FCM) method is used for clustering the data points. First the cluster centre is identified to connect other cluster groups. Iteratively the cluster centre will be updated to move towards the right locations within the data sets.

 

        (1)

           

        (2)

                (3)

 

·         denote an image with ‘’ pixels to be partitioned into  clusters.

·        represent multispectral features.

·         represent the membership of pixel in the cluster,

·         is the cluster centre

·         is a constant to control the fuzziness of the resulting partition.

 

The detected data points are marked on the vascular structure of the fundus image is shown in Fig. 3(c). The bifurcation point 126 is shown Fig. 3(d).

 

OPTIC DISC SEGMENTATION:

The region based active contour model is used to find out the region of interest and to detect the optic disc. First, select the initial mask for the input segmented image and apply region based active contour to obtain the mask of the retinal image. Subtract the mask from the input image produces a new mask. At last, subtract the new mask from the old mask and the resulting image is non-masked segmented image.

 

Retinal blood vessels – width measurement:

The retinal vessel boundary is as shown in Fig. 4, P₁ and P₂ are the segment points. P₁ is the starting point of the vessel edge pixel and P₂ is the ending point of that rotated vessel edge. B₁ represents the upper boundary and B₂ represents the lower boundary of the vessel. C is the centreline vessel point in between the edges. W is the width of the vessel.

 

Fig 4. Retinal blood vessel – centreline width Image.

 

Retinal blood vessel – centreline identification:

Detection and quantitative measurement of variations in retinal blood vessel image helps to diagnose DR. The retinal blood vessel measurement process has to discover the vessel centreline. Each vessel contains a boundary upside and downside. The region in between these boundary edges are called as vessel centreline. The vessels have two parallel centre lines made by the central light reflex. Thus it appears as two lines. The identified vessel centreline of the retinal image is shown in Fig. 3 (a) and (b).

 

Retinal blood vessel–profile analysis:

The vessel is an arc to measure its characteristics the graph based approach is considered. The length of the vessel is the number of edges it contains. The distance between two vertices is the length of the shortest path between them. The diameter of the vessel is the longest minimal path.

 

 

Fig 5. Retinal blood vessel branches.

 

In case, if the diameter is not minimal, then the two vertices are connected at its endpoints by a shorter path. The centreline vessel pixel radius is the distance from the centre to the outer edge. The eccentricity of the vertex is the length of from the vertex to some vertex in the graph. The eccentricity of longest distance vertex represents the measurement of the larger vessel segment in the graph. The centre of the graph contains collection of vertices, whose eccentricity is least. That is, the collection of vertices whose longest distance to all other vertices is the smallest. The radius of a graph is determined by the length of the shortest eccentricity vertex in the graph. The vessel profile segment 126 of Fig. 3 (d) is shown in Fig. 6. The corresponding data points of this vessel are shown in Fig. 3 (c). It contains 12 connecting points. The diameters of every two connecting points are shown in Table 1.

 

SEGMENT LENGTH:

_{In order to extract useful information from the blood vessel image, a portion of the skeleton to be selected. This portion is called vessel segment. Then to isolate a vessel segment which is defined by two end points. The vessel segment is available from the interconnected points. The extracted segment of the fundus image edges are shown in Fig. 3. The retinal blood vessel branches shown in Fig. 5 are taken into account to describe various quantitative measurements. In Fig. 5, P0 is the} right most pixel of the contour. P₂ is the left most pixel of the contour. P₁, P₃ are the leaves are pointed by the skeleton end nodes. N₁, N₃ are extrapolated to the intersection with the contour line. L₁,……, L₅ are the length of the segments. Assume that the leaf lengths correspond to the lengths of end branches of the extrapolated skeleton. The starting branch pixel arrays are stored as Bi : i = 1, …, n_v, where n_v is the no.of skeleton in the starting nodes. The ending branch pixel arrays are stored as Ei : i = 1, …, n_v, where n_v is the no.of skeleton in the ending nodes. j is the start node and N is the end node. The length of the branch is calculated by the following equation.

 

 

Fig 6. Retinal vessel profileplot

 

                                    (4)

Ÿ  Li = Length of the skeleton branch Bi

Ÿ  Pj, Pj-1 = jth section between two consecutive pixels along the horizontal and vertical distances xj and yj respectively.

Ÿ  RS = no. of pixels in the skeleton branch Bi

Ÿ  The Euclidean distance (in pixels) of two consecutive branch dense pixels

 

 

 

 

Table 1 Retinal blood vessel diameter.

Offsets (px)	Diameters (px)
0	2.9867
1.37685	3.0201
2.50833	3.0806
3.50298	3.3089
4.44757	3.6992
5.39024	4.2799
6.3434	4.4965
7.29849	4.6329
8.24208	4.7264
9.17164	4.7689
10.1122	4.7559
11.129	4.7247

 

Mean Diameter:

An area is the number of white pixels on black and white image. The white pixels are retinal blood vessel pixels on the segment.

 

               (5)

 

Standard Deviation Diameter:

The standard deviation measure is the amount of data spread on the mean value. The standard deviation is nearly equivalent to the average deviation of the mean value.

 

               (6)

 

Algorithm to calculate Standard deviation diameter

1    Load the blood vessels segmented image I.

2    Find the size of the image, m as row and n as column.

3    Calculate the standard deviation of the vessel pixel p1.

 

Tortuosity:

Tortuosity is used to measure the ratio of the average length of all blood vessel path lines passing through the width of the given cross section point. The distance traversed by the vessel centreline is calculated by summing up the distance between the consecutive points in the segment. Let the vessel segment has consecutive points. The distance of the curved vessel is given as,

 

       (7)

 

where

are the coordinates of the pixel in the vessel segment. The straight vessel distance is measured by finding out the distance between the starting point and ending points of the vessel.

 (8)

The vessel tortuosity is calculated as follows,

 

                              (9)

 

Error Rate:

The error rate of total length measurement of the proposed method in relation to the manual method is described as follows,

 

                   (10)

 

E_i (m) = Lenth of i^th branch of skeleton measured by the method m

E_j (l) = Length of i^th branch of skeleton measured by the manual method

 

Table 2 Quantitative measures of retinal blood vessel.

 

EXPERIMENTAL RESULTS:

The DRIVE database is used to assess the quantitative measurement results. The segmented retinal vessel for the fundus image is shown in Fig.3. The segmentation process is implemented using MATLAB. A successful detection is defined as tracking the vessel without leaving any obvious segment. A partly successful detection must acquire more than half of a vessel path; otherwise, it is categorized as failed detection. An erroneous detection is a part of the detection result of a vessel is actually not blood vessel. An accurate measurement of vessel width is the most essential and challenging problem in the quantitative analysis of retinal blood vessel. The image quality will affect the blood vessel detection result significantly. If severe diseases or errors appear on the image, it will impact the result. The proposed method is capable of handle depigmented images with severe diseases. Starting from one end point, all the retinal blood vessels are traced until the other end point is reached and each vessel is labelled. Table 2 give the overall measures of segment 126 and its connecting points. The proposed method achieved the average vessel classification accuracy of 99.50% with their corresponding ground truth images. The error rate is 0.01% for all the image sets used in this paper. Hence our proposed method successfully estimates the vessel profile for the diagnosis of Diabetic Retinopathy.

 

CONCLUSIONS:

The proposed system is used to track all the blood vessels without losing it continuity. The system is perfectly identified the next source point to initiate tracking at the branch points. Also produces the continuous vessel structure throughout the process in the fundus image. The absolute value of the vessel diameter is also calculated by using various parameter measurements. Significantly it increased the rate of successive detection and measurements of retinal blood vessels in the DR image.

 

REFERENCES:

1.     R. Keerthikanth, A. Rajan, Detection of Retinal Blood Vessels using Curve let Transform, ICSEC, May 2014, Vol. 2, Issue 3.

2.     Argyrios Christodoulidis, Thomas Hurtut, Houssem Ben Tahar, Farida Cheriet, A multi-scale tensor voting approach for small retinal vessel segmentation in high resolution fundus images, Computerized Medical Imaging and Graphics, 2016, 28–43.

3.     Priyanka Yadao, Sayali Naval, Animesh Maheshwari, Blood Vessel Segmentation in Retinal Fundus Images using Matched Filter, Journal of Harmonized Research in Engineering, 2014.

4.     Abdolhossein Fathi,  Ahmad  Reza  Naghsh-Nilchi1, Automatic wavelet-based retinal blood vessels segmentation and vessel diameter estimation, Biomedical Signal Processing and Control, 2013, 71–80.

5.     Luo Gang, Opas Chutatape, Shankar M. Krishnan, Detection and Measurement of Retinal Vessels in Fundus Images Using Amplitude Modified Second- Order Gaussian Filter, IEEE Transactions on Biomedical Engineering, February 2002, Vol. 49.

6.     Rebekka Heitmar, Angelos A. Kalitzeos, Reliability of retinal vessel caliber measurements using a retinal oximeter, Heitmar and Kalitzeos, BMC Ophthalmology, 2015,15:184.

7.     Behdad Dashtbozorg, Ana Maria Mendonça, Aurelio Campilho, An Automatic Graph-Based Approach For Artery/Vein Classification In Retinal Images, IEEE Transactions On Image Processing, March 2014, Vol. 23, No. 3.

8.     Retinal Vessel Centerline Extraction Using Multiscale Matched Filters, Confidence and Edge Measures Michal Sofka, Charles V. Stewart, IEEE Transactions on Medical Imaging, December 2006, Vol. 25, No. 12.

9.     Bashir Al-Diri, Andrew Hunter, David Steel, Maged Habib, Manual Measurement of Retinal Bifurcation Features, 32nd Annual International Conference of the IEEE EMBS Buenos Aires, Argentina, August 31 - September 4, 2010.

10.   Bashir Al-Diri, Andrew Hunter, David Steel, An Active Contour Model for Segmenting and Measuring Retinal Vessels, IEEE Transactions on Medical Imaging, September 2009, Vol.28, No. 9.

11.   Andrew Hunter, David Steel, Ansu Basu, Robert Ryder, and R. Lee Kennedy, Measurement of Retinal Vessel Widths From Fundus Images Based on 2-D Modelling James Lowell, IEEE Transactions On Medical Imaging, October 2004, Vol. 23, No. 10.

12.   Rothaus, K, Jiang, X, Rhiem P, Separation of the retinal vascular graph in arteries and veins based upon structural knowledge, Image and Vision Computing, 2009, 864–875.

13.   Dashtbozorg B, Mendonca A. M, Campilho A, An automatic graph-based approach for artery/vein classification in retinal images, IEEE Transactions on Image Processing, 2013.

14.   Owen C. G, Rudnicka A. R, Mullen R, Barman S.A, Monekosso D, Whincup P. H, Measuring retinal vessel tortuosity in 10-year-old children: validation of the computer-assisted image analysis of the retina (CAIAR) program, Investigative Ophthalmology & Visual Science, 2004–2010.

15.   Ortega M, Barreira N, Novo J, Penedo M. G, Pose-Reino A, Gomez-Ulla F, Sirius: A web-based system for retinal image analysis. International Journal of Medical Informatics, 2010, 722–732.

16.   Grisan E, Ruggeri A, A divide et impera strategy for automatic classification of retinal vessels into arteries and veins. In Engineering in medicine and biology society, 2003. Proceedings of the 25th annual international conference of the IEEE, 2003, Vol. 1, pp. 890–893.

17.   Perez-Rovira A, MacGillivray T, Trucco E, Chin K. S, Zutis K, Lupascu C, Tegolo D, Giachetti A, Wilson P. J, Doney A, Dhillon B, VAMPIRE: Vessel assessment and measurement platform for images of the Retina. Engineering in medicine and biology society, EMBC, 2011 annual international conference of the IEEE, 2011, pp. 3391–3394.

18.   Vazquez S. G, Cancela B, Barreira N, Penedo M. G, Rodriguez-Blanco M, Pena Seijo M, Improving retinal artery and vein classification by means of a minimal path approach. Machine Vision and Applications, 2013, 24, 919–930.

19.   Kelvin P, Ghassan H, Rafeef A, Live-vessel: extending livewire for simultaneous extraction of optimal medial and boundary paths in vascular images. In Proceedings of the 10th international conference on Medical image computing and computer-assisted intervention. Brisbane, Australia: Springer- Verlag, 2007.

20.   DRIVE: http://www.isi.uu.nl/Research/Databases/DRIVE, STARE: http://cecas.clemson.edu/~ahoover/stare/, REVIEW: http://reviewdb.lincoln.ac.uk/Image%20Datasets/Review.aspx, HRF: https://www5.cs.fau.de/research/data/ fundus-images/, DRIONS: http://www.ia.uned.es/~ejcarmona/DRIONS-DB.html.

 

 

 

 

Accepted on 12.07.2018      © RJPT All right reserved

Research J. Pharm. and Tech 2019; 12(1): 21-26.