Fully automatic algorithm for the analysis of vessels in the angiographic image of the eye fundus
© Koprowski et al.; licensee BioMed Central Ltd. 2012
Received: 16 May 2012
Accepted: 5 June 2012
Published: 22 June 2012
The available scientific literature contains descriptions of manual, semi-automated and automated methods for analysing angiographic images. The presented algorithms segment vessels calculating their tortuosity or number in a given area. We describe a statistical analysis of the inclination of the vessels in the fundus as related to their distance from the center of the optic disc.
The paper presents an automated method for analysing vessels which are found in angiographic images of the eye using a Matlab implemented algorithm. It performs filtration and convolution operations with suggested masks. The result is an image containing information on the location of vessels and their inclination angle in relation to the center of the optic disc. This is a new approach to the analysis of vessels whose usefulness has been confirmed in the diagnosis of hypertension.
The proposed algorithm analyzed and processed the images of the eye fundus using a classifier in the form of decision trees. It enabled the proper classification of healthy patients and those with hypertension. The result is a very good separation of healthy subjects from the hypertensive ones: sensitivity - 83%, specificity - 100%, accuracy - 96%. This confirms a practical usefulness of the proposed method.
This paper presents an algorithm for the automatic analysis of morphological parameters of the fundus vessels. Such an analysis is performed during fluorescein angiography of the eye. The presented algorithm automatically calculates the global statistical features connected with both tortuosity of vessels and their total area or their number.
KeywordsImage processing Angiographic image Fully automatic algorithm
The available literature contains descriptions of manual, semi-automated and automated methods for analyzing angiographic images. The presented algorithms often refer to other visualization methods . They segment vessels, calculate their tortuosity or number in a given area. The accompanying changes in the analyzed area of the eye fundus are also colour-coded , or analyzed locally (analyzed in a declared area) . The local analysis of both the width of arterioles as well as the number of intersections between them is described in several papers [1, 4–7]. What is usually measured is the number of intersections (decussations) of vessels, their tortuosity and diameter. It is done by, for example, IVAN software . The measurement range is usually placed within the optic disc radius r - and ranges from 2r to 3r. The measurements of both tortuosity and meane diameter of arterioles in this range lead to reliable results. A good example is the Blue Mountain Eye Study which showed that the diameter of retinal arterioles and venules was decreasing with age regardless of other factors [9, 10]. On the other hand, in the Beaver Dam Study, it was observed that hypertension was the cause of retinal arteriolar narrowing. However, this phenomenon turned out to be less pronounced in older patients . This fact proves inadequate vascular response to hypertension in this population, which may be caused by atherosclerosis or increased vascular wall stiffness.
There exists a large number of papers connected with the analysis and processing of the eye fundus images. The methods of image analysis and processing described in them are profiled to enhance segmentation of vessels. In papers [12–16], the authors perform skeletonization of blood vessels and then present its successive approximation in accordance with the adopted algorithm. This method is extremely time-consuming because it requires the use of sophisticated techniques for the initial stage of image analysis and processing, and is also connected with a morphological operation (skeletonization) and approximation. For example, paper  presents a method for assessing tortuosity of vessels by their subsequent division into smaller and smaller segments. Individual vessels can also be compared with a pattern, as it was done in paper . Here the authors used a pattern in the form of a sine wave with variable amplitude in order to assess vascular tortuosity. Another method for obtaining correct results of segmentation is tracking the vessel outline which is described in papers [14, 17]. In paper , the authors describe the problems and methodology of measurements performed on images of the eye fundus and based on segmentation of vessels. In paper , the authors compare the results obtained from colour fundus photographs (FPs) and fluorescein angiographs (FAs). In both cases, however, the methods of image analysis refer to segmentation of vessels. Generally, the obtained results are compared with assessments of an expert whose task is to verify them - in paper , e.g., there are three experts. They are not always profiled algorithms for the analysis of just this type of images. Sometimes they are supported by additional software such as, for example, Paint-Shop Pro (Jasc Software) used in paper .
Regardless of the results, the presented methods of measuring the characteristics (diameter of the veins and arteries, the tortuosity and the number of intersections) are based on the repetitive course of the following proceedings:
· segmentation of veins and arteries as separate objects,
· calculation of average values of tortuosity or diameter of veins or arteries in a declared area.
Taking into consideration computational complexity of these two points, it can be noted that segmentation of veins and arterioles as separate objects is unnecessary as after this stage, features are combined and the average value is calculated. Bearing this in mind, the authors suggested a measurement method free of this defect [9, 10, 22]. When examining the number of intersections, vein and artery segmentation is not critical. However, it enables to include all vessels in the examined area in a more accurate and reliable way. In the angiographic images of the eye fundus, the vessel lumen diameter is evaluated. Whereas in the colour images, evaluation concerns the entire vessel together with the adventitia. The vessel lumen diameter better reflects the functional state of the vessel and directly influences the flow. The downside is the lack of a fully automatic distinction between arterioles and venules.
We describe a method of statistical analysis of the inclination of the vessels in the fundus related to their distance from the center of the optic disc. This method, however, differs from the classical methods published in [4, 6, 23]. Only the fractal analysis [24, 25] is similar to the approach presented in this article. However, this similarity concerns only the global approach to the analysis.
The algorithm created by the authors should have the following characteristics:
· fully automatic image analysis - without operator intervention - even in the batch mode,
· analysis of angiographic images of the eye fundus at any resolution, both spatial resolution (pixels per inch) and pixel depth (bits per pixel),
· automatic, statistical analysis of the diameter of vessels and their degree of tortuosity,
· automatic calculation of full statistics like:
· average gradient of all vessels at any distance from the optic disc,
· average volume occupied by all vessels at any distance from the optic disc,
· automatic analysis of a group of images (for consecutive patients or for the same patient, but between successive tests): an average measure, STD, median or an average value of changes in thickness of vessels,
· auto-save to disk of the received data and images.
The proposed algorithm was tested on a group of patients described below.
In order to verify the assumptions and correctness of the choice of the afore-mentioned image features, sample tests on patients have been performed. For this purpose, 12 healthy subjects and 40 patients with hypertension have been examined. They were aged 22 to 87 with a body weight between 53 to 92kg. Some of them were healthy and the others suffered from hypertension which was cardiologically stated. The Topcon funduscamera was used in the study. The camera is made by Zeiss and has angular width of 45 degrees and spatial resolution of 2136x3216 pixels. Fluorescein angiography was performed in a conventional way, and the images of the transit phase were analyzed as they best reflect the size of the vessel lumen. The patients agreed consciously to participate in the study which was conducted in accordance with the principles of the Declaration of Helsinki.
The algorithm suggested by the authors and the results of the analysis for the examined patients are described below.
The description of methodology for image analysis and processing is divided into a preliminary analysis of images (acquisition and filtration) and an appropriate algorithm.
LGRAY angiographic image in DICOM format with a resolution MxN = 2136x3216 pixels is filtered with a median filter whose mask size is h1, Mh1xNh1 = 3x3. The aim is to eliminate the noise. The mask size was chosen arbitrarily taking into consideration the minimum size of objects and optimization of the operation time. For the adopted size of the mask and the image resolution of 8 bits per pixel, the filtration time was 100ms. This part of the algorithm was implemented in the C language and Matlab for a PC with an Intel Xenon processor X5680@3.33GHz, 12GB RAM. LGRAY image after filtration with a median filter LMED undergoes successive stages of processing.
A method of this type (shown in Figure 1, Figure 2) allows for arbitrarily accurate approximation of the inclination angle (for the values k = 1,2,3 etc. up to the angle values α). By the same token, it enables to calculate the width of the vessel at a given point.
The method shown in Figure 2 shows the greatest accuracy but also the largest computational complexity. Thus, it will be still used. The set of values of the inclination angle α creates an inclination field, hereinafter referred to as Lα. It allows for an automated analysis of vessels width which is done as accurately as possible with the level of the resolution error ±1bit.
ϵma, ϵmi-- the maximum and minimum values of the mask h2,
θ -- the inclination angle of the mask h2,
mh2,nh2 -- the coordinates of individual values in the mask h2,
σ -- standard deviations of an average.
pr –decimal-to-binary conversion threshold.
The image Lk represents the final form of the image of the algorithm for image analysis and processing. One of the possible applications of the presented algorithm, that is the measurements made in the image Lk, is presented in next section.
The algorithm presented in the previous chapter is profiled for automatic analysis of the width of vessels. The analysis is made for each pixel of the image, with an accuracy described by the relation (1) (for the angular values with a resolution of one degree). The use of this automated method of analysis of data obtained in the image Lk will be suggested below.
The presented analysis applies to:
fully automatic measurement without any operator intervention,
measurement of the number of vessels in a declared area further on denoted as z,
the average angular value for individual objects in the image Lk – further on denoted as φsr,
standard deviation of mean angular values in the image Lk - further on denoted as φSTD,
calculation of the maximum value of the angle in the image Lk - further on denoted as φmax,
calculation of the percentage of the ratio of vessels surface area in relation to the measured area (calculated as the ratio of the total number of pixels that make up the vessels in the measured area to the total number of pixels of the area) – denoted as ps,
designation of histogram of pixel inclination of objects.
Fragment of results of automated measurement of morphometric parameters of vessels
Analyzing these results, we can, for example, read them for measurement no. 4. This is a patient without hypertension for whom 11 separated objects have been detected automatically in the area 2r to 3r. Their average angle of inclination with respect to the axis was 16.1°. In addition, most pixels of objects were found for a 13° angle. All the detected pixels of objects constituted 6.6% of the total measured area (2r to 3r).
A group of 52 patients (including 40 healthy and 12 hypertensive ones) was divided in equal proportions into learning, validation and test group). Then, the cross-validation method was used, with particular reference to the ratio between the decision classes (stratified cross-validation). A classifier in the form of a decision tree was implemented, assuming five attributes: the angle for the maximum number of pixels - ϕmax, standard deviation of the angle average ϕSTD, the average value of the angle - ϕsr, the number of vessels in a declared area – z, and the percentage of the ratio of vessels surface area to the measured area - ps. It was recognized that these attributes are equally privileged. On this basis, six decision trees were constructed; five trees for each of the attributes occurring independently and one for all of them together.
Comparison with other methods
The presented algorithm is designed to collect statistical information about the number and tortuosity of vessels in the analyzed area of the eye. The value of the angle ϕ obtained from the image is, by definition, the measure of tortuosity and number of vessels. According to the information given in the introduction, the aim of the algorithm is not the most accurate segmentation of vessels. In general, objects (vessels) visible in the picture Lw or Lk (Figure 9, Figure 10) do not have to be continuous. Furthermore, the algorithm does not have to segment all the vessels. Therefore, high accuracy is not required. The only common feature is related to the fact that the ratio of thickness of vessels must be preserved. The reason is that each pixel of an object (vessel) affects the shape of the histogram (Figure 12) and thus the value ϕmax.
The results of the presented algorithm were compared with the results obtained with DRIVE database . The assessment of quality of segmentation of vessels should be distinguished from the evaluation of the adopted statistical methods. Minor differences were obtained in the detected level of details of vessels in favor of the method described in paper . Underestimation and lack of vessels continuity were the main reasons for these differences. For 20 verified cases, the underestimation was less than 15% of the total area of all objects. It should be noted that despite receiving seemingly worse results of segmentation of vessels, the method described above has not been profiled for this purpose. This method enables to obtain directly the measurement results of tortuosity and percentage of vessels in the analyzed area. Therefore, in contrast to the methods described in the introduction [12–21], there is no need to perform additional analyzes like zooming vessels with a curve [12, 16] or using patterns of tortuosity  etc. This is the biggest advantage of the presented algorithm over other methods described in all the publications [12–21]. An additional advantage of this algorithm is negligible sensitivity to change of image acquisition parameters - for different operators, different camera settings and different patient settings. This is due to the characteristics of the algorithm: automatically corrected unevenness of lighting and acceptance of lack of vessels continuity.
Comparing the obtained results with those of other authors, a similar global approach in the fractal analysis can be found. The fractal dimension shown in papers [31, 32] allows for a group division into healthy subjects and those with hypertension. In paper , the value, i.e. fractal dimension, is fixed at 1.437 with a standard deviation of 0.025. However, this method is semi-automatic. Paper , on the other hand, presents an interesting method based on nonlinear orthogonal projection approach. This method uses the afore-mentioned DRIVE database. The authors have obtained 96.1% accuracy. The results are similar to those obtained in this paper (96% accuracy). However, they were obtained with a slightly different method. The differences consist in the fact that the algorithm presented in paper  is not fully automatic (the differences are thus related to the segmentation method). The algorithm presented in paper  does not apply directly to the detection of patients with hypertension. The level of accuracy at 96.1% indicated by the authors concerns the quality of segmentation of vessels. It is not a measure of the quality of separation of patients with hypertension from healthy subjects. Yet in paper  a global fractal analysis applies only to chronic kidney disease. The results in the diagnosis of this condition are at 95%. Therefore, a comparison with the use of the fractal dimension was carried out for the images obtained in this study. For this purpose, Fracllac software (Local Connected Fractal Dimension Analysis function) was used, which is, for example, described in paper . However, using only Fracllac software, no correlation between hypertension and the fractal dimension in the angiographic image was obtained. The reason was a major influence of lighting unevenness and artifacts visible in the image, which were not filtered. Whereas using the image pre-processing suggested in this paper, the accuracy was 81%. However, this result was obtained for the hybrid method which combines a filtration method suggested in this paper with the fractal analysis made with Fracllac software.
The algorithm described here can also be divided into functions related to each analysis phase. Then it will be possible to make a comparison with other GUI profiled to Matlab. Such an example is Fraclab  which is a set of functions extending the functionality of Matlab. It has common features with the presented algorithm only in terms of filtration. The main way of calculating the matrices Lma and Lθ, and on their basis Lw, which is presented above (4), is not available there. Of course the angle of the mask h2 can be changed manually, and then this manual method is similar to the one presented above.
This paper presents a tool (algorithm) designed for automatic analysis of morphological parameters of vessels in the fundus watched during fluorescein angiography. The presented algorithm automatically calculates the global statistical features connected with both tortuosity of vessels as well as their total area or their number. On the basis of preliminary studies, we have shown correlations between the total value of the inclination angle of vessels and hypertension. This result confirms the usefulness of the described algorithm for image analysis and processing in medical practice. However, further research in a larger population is needed. The algorithm can also operate in a batch mode where the operator only selects a folder with images for analysis. Currently, with no time optimization, the analysis of one image takes a few seconds on the Intel Core 2 Quad Q9300 2.5 GHz CPU with 8 GB RAM.
No outside funding was received for this study.
- Farsiu S, Chiu SJ, Izatt JA, Toth CA: Fast detection and segmentation of drusen in retinal optical coherence tomography images. Proceedings of Photonics West. Proc SPIE 2008, 68440D: D1-D12.Google Scholar
- Khan AU, Rabya BK, Bilal S, Asad J, Shah M: Enhancement of Angiogram Images Using Pseudo Color Processing. Information Technology Journal 2008, 7: 210-214. 10.3923/itj.2008.210.214View ArticleGoogle Scholar
- Hughes S, Gardiner T, Hu P, Baxter L, Rosinova E, Chan-Ling T: Altered pericyte-endothelial relations in the rat retina during aging: Implications for vessel stability. Neurobiol Aging 2006, 27: 1838-1847. 10.1016/j.neurobiolaging.2005.10.021View ArticleGoogle Scholar
- Chu C, Delp E, Buda A: Detecting left ventricular endocardial and epicardial boundaries by digital two-dimensional echocardiography. IEEE Trans Med Imaging 1988, 7: 81-90. 10.1109/42.3932View ArticleGoogle Scholar
- Grattoni P, Bonamini R: Contour detection of the left ventricular cavity from angiographic images. IEEE Trans Med Imaging 1985, 4(2):72-78.View ArticleGoogle Scholar
- Herna´ndez-Hoyos M, Orkisz M, Puech P, Mansard-Desbleds C, Douek P, Magnin IE: Computer-assisted Analysis of Threedimensional MR Angiograms. Radiographics 2002, 22: 421-436.View ArticleGoogle Scholar
- Xinchun L, Shidong C, Mouyan Z, Zhenming C: Edge-detection based on the local variance in angiographic images. Journal of Electronics 2000, 17(4):338-344.Google Scholar
- Wong TY, Knudtson MD, Klein R, et al.: Computer-assisted measurement of retinal vessel diameters in the Beaver Dam Eye Study: methodology, correlation between eyes, and effect of refractive errors. Ophthalmology 2004, 111(6):1183-1190. 10.1016/j.ophtha.2003.09.039View ArticleGoogle Scholar
- Wong TY, Klein R, Klein BE, Meuer SM, Hubbard LD: Retinal vessel diameters and their associations with age and blood pressure. Invest Ophthalmol Vis Sci 2003, 44(11):4644-4650. 10.1167/iovs.03-0079View ArticleGoogle Scholar
- Wong TY, Shankar A, Klein R, Klein BE, Hubbard LD: Prospective cohort study of retinal vessel diameters and risk of hypertension. BMJ 2004, 10(7457):79. 329View ArticleGoogle Scholar
- Wong TY, Shankar A, Klein R, Klein BE, Hubbard LD: Retinal arteriolar narrowing, hypertension, and subsequent risk of diabetes mellitus. Arch Intern Med 2005, 165: 1060-1065. 10.1001/archinte.165.9.1060View ArticleGoogle Scholar
- Wilson CM, Cocker KD, Moseley MJ, et al.: Computerized analysis of retinal vessel width and tortuosity in premature infants. Invest Ophthalmol Vis Sci 2008, 49(8):3577-3785. 10.1167/iovs.07-1353View ArticleGoogle Scholar
- Gelman R, Martinez-Perez ME, Vanderveen DK, Moskowitz A, Fulton AB: Diagnosis of Plus Disease in Retinopathy of Prematurity Using Retinal Image multiScale Analysis. Invest Ophthalmol Vis Sci 2005, 46(12):4734-4738. 10.1167/iovs.05-0646View ArticleGoogle Scholar
- Kochner B, Schulmann D, Michaelis M, Mann G, Englemeier KH: Course tracking and contour extraction of retinal vessels from colour fundus photographs: most efficient use of steerable filters for model based image analysis. Proceedings of the SPIE Conference on Medical Imaging 1998, 3338: 755-761.Google Scholar
- Heneghan C, Flynn J, O’Keefe M, Cahill M: Characterization of changes in blood vessel width and tortuosity in retinopathy of prematurity using image analysis. Medical Image Analysis 2002, 6: 407-429. 10.1016/S1361-8415(02)00058-0View ArticleGoogle Scholar
- Koreen S, Gelman R, Martinez-Perez ME, et al.: Evaluation of a computer-based system for plus disease diagnosis in retinopathy of prematurity. Ophthalmology 2007, 114(12):59-67. 10.1016/j.ophtha.2007.10.006View ArticleGoogle Scholar
- Wallace DK, Jomier J, Aylward SR, Landers MB: Computer-automated quantification of plus disease in retinopathy of prematurity. J AAPOS 2003, 7: 126-130. 10.1016/S1091-8531(02)00015-0View ArticleGoogle Scholar
- Patton N, Aslam TM, MacGillivray T, et al.: Retinal image analysis: concepts, applications and potential. Prog Retin Eye Res 2006, 25: 99-127. 10.1016/j.preteyeres.2005.07.001View ArticleGoogle Scholar
- Pakter HM, Fuchs SC, Maestri MK, Moreira LB, Dei Ricardi LM, Pamplona VF, Oliveira MM, Fuchs FD: Computer-assisted methods to evaluate retinal vascular caliber: what are they measuring? Invest Ophthalmol 2011, 52(2):810-815. Vis Sci 10.1167/iovs.10-5876View ArticleGoogle Scholar
- Kiely AE, Wallace DK, Freedman SF, Zhao Z: Computer-assisted measurement of retinal vascular width and tortuosity in retinopathy of prematurity. Arch Ophthalmol 2010, 128(7):847-852. 10.1001/archophthalmol.2010.133View ArticleGoogle Scholar
- Swanson CR, Cocker KD, Parker KH, Moseley MJ, Wren SME, Fielder AR: Semi-automatedcomputer analysis of vessel growth in preterminfants without and with ROP. Br J Ophthalmol 2003, 87: 1474-1477. 10.1136/bjo.87.12.1474View ArticleGoogle Scholar
- Sekiguchi H, Sugimoto N, Kawahito M, Lee JD, Nakano A, Fujita M, Eiho S: Image processing on regular coronary angiograms for myocardial perfusion measurements. Computers Cardiol 2006, 33: 821-824.Google Scholar
- Klinder T, Ostermann J, Ehm M, Franz A, Kneser R, Lorenz C: Automated model-based vertebra detection, identification, and segmentation. Medical Image Analysis 2009, 13: 471-482. 10.1016/j.media.2009.02.004View ArticleGoogle Scholar
- Mancardi D, Varetto G, Bucci E, Maniero F, Guiot C: Fractal parameters and vascular networks: facts & artifacts. Theor Biol Med Model 2008, 5: 12. 10.1186/1742-4682-5-12View ArticleGoogle Scholar
- Legrand P, Lévy-Vehel J: Signal and image processing with Fraclab.: FRACTAL04, Complexity and Fractals in Nature, 8th International Multidisciplinary Conference Thinking in Patterns: fractals and related phenomena in nature. , ; 2004:321-322.Google Scholar
- Gonzalez R, Woods R: Digital Image Processing. Addison-Wesley Publishing Company, New York; 1992:471-482. Chap. 4Google Scholar
- Sonka M, Michael Fitzpatrick J: Volume 2, Medical Image Processing and Analysis. In Handbook of Medical Imaging. SPIE, Belligham; 2000.Google Scholar
- Figueriredo M, Leitao J: Bayesian estimation of ventricular contours in angiographic images. IEEE Trans Med Imaging 1992, 11: 416-429. 10.1109/42.158946View ArticleGoogle Scholar
- Otsu N: A threshold selection method from gray-level histograms. IEEE Trans. Sys., Man., Cyber 1979, 9(1):62-66.MathSciNetView ArticleGoogle Scholar
- Staal JJ, Abramoff MD, Niemeijer M, Viergever MA, Ginneken B: Ridge based vessel segmentation in color images of the retina. IEEE Trans Med Imaging 2004, 23: 501-509. 10.1109/TMI.2004.825627View ArticleGoogle Scholar
- Liew G, Wang JJ, Cheung N, Zhang YP, Hsu W, Lee ML, Mitchell P, Tikellis G, Taylor B, Wong TY: The retinal vasculature as a fractal: methodology, reliability, and relationship to blood pressure. Ophthalmology 2008, 115(11):1951-1956. 10.1016/j.ophtha.2008.05.029View ArticleGoogle Scholar
- Sng CCA, Sabanayagam C, Lamoureux EL, Liu E, Lim SC, Hamzah H, Lee J, Tai ES, Wong TY: Fractal analysis of the retinal vasculature and chronic kidney disease. Nephrol Dial Transplant 2010, 25(7):2252-2258. 10.1093/ndt/gfq007View ArticleGoogle Scholar
- Zhang Y, Hsu W, Lee ML: Detection of retinal blood vessels based on nonlinear projections. J Sign Process Syst 2009, 55: 103-112. 10.1007/s11265-008-0179-5View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.