Eye fundus retinal vessels are the only blood vessels of the body that can be visible with non-invasive imaging (Miri et al., 2017; Fraz et al., 2014). Retinal images are commonly used in manual and in automatic vessel structure analysis. This analysis leads to disease detection which is especially important in the early stages of the disease (Li and Wee, 2014). First clinical findings in early diabetic retinopathy are often made in abnormalities of retinal blood vessels (Miri et al., 2017). One of these abnormalities is the ratio between the artery and vein diameters (AVR). AVR is approximately equal to 0.7 for a healthy person (Sun et al., 2009). Ratio increases in the evaluation of sick persons, arteries become larger. Blood vessel width is measured for main vessels in an AVR measurement region – the circular region from 1.5 to 3 optic nerve disc (OD) radius from its centre (Knudtson et al., 2003). The ratio is calculated for vessels above and below the OD centre separately. There are also some other methods for measuring this ratio, including more vessels for calculations, selecting other places for measurements, etc. (see e.g. Sun et al., 2009; Niemeijer et al., 2011). The timely evaluation of disorders can lead not only to the better evaluation of the disease but also to protection of disorder progression.

The measurement making and other eye fundus analysis steps are time consuming for an expert, and sometimes ophthalmologists cannot spend much time for analysis if they have many patients. Here, computer aided systems help to do most of the analysis and act as advisory or tracking systems (see, e.g. Li and Wee, 2014; Bankhhead et al., 2012). When making such systems, there are many specific situations to be considered. For example, when an expert does measurements, the measurement place is selected very differently for the specific image – the most important factor is the ability to make precise measurements. When the computer makes automatic measurements the place where they are done is always concrete, and in many cases, these places are not the best for measurements to make. The automatic system does not know the quality of the image and even whether the person is sick or not. In many cases (Muramatsu et al., 2011; Bankhhead et al., 2012) computer aided systems are adapted for specific technical equipment or even to the specific image database and when a new type of images is gathered, the system must update to meet the new requirements. Sometimes this leads even to an information loss because of reduced image resolution (Ravishankar et al., 2009).

Many eye fundus analysis algorithms are tested on high quality image databases only. However, in some specific circumstances, it is hard to get high quality images, or it is hard to obtain another image if the image is not of high enough quality. Image quality is affected by the illness also. These situations lead to the requirements for the automatic system to be able to deal with images of lower quality.

Classification is required for AVR calculations. In order to do this, main vessels must be detected and classified into arteries and veins. Then, with the view to evaluate the artery and vein ratio from eye fundus images at the specific place, the following steps can be applied. For example, OD should be detected in order to make measurements at a specific place. Then, places for measurements must be selected.

The OD detection is a very important stage in eye fundus analysis (Buteikienė et al., 2012). As it was mentioned before, OD detection is used for measurement place selection for AVR calculations. It is also used in vessel tracing algorithms in disease evaluations based on OD analysis like OD excavation and in other cases. Main problems arising in precise OD detection are uneven lightening, disease based changes in eye fundus and noise produced by vessels inside of OD. In this paper, a two stage OD detection algorithm is used. This OD detection algorithm is based on the algorithm described in Stabingis et al. (2018) with several improvements described in the next section.

In the next stage, the vessel width measurement algorithm (Stabingis et al., 2018) is applied. Profile analysis is a technique used in vessel measurement (see e.g. Li et al., 2003; Bankhhead et al., 2012). The vessel width measurement algorithm, used in the proposed algorithm, is based on vessel profile analysis described in the paper (Stabingis et al., 2018). Profile information is often extracted according to the Bresenham algorithm which extracts information fast but makes distances between points distributed unevenly. In the proposed algorithm, profile information extraction uses spatial distance function which enables to use the same analysis algorithm with images of different resolution. Measurement algorithm is capable of detecting vessels without prior information about vessel existence at a specific point. The extracted vessel tree is used only for the detection of the initial measurement direction and only when this tree information is available. AVR measurement zone vessels are measured, and then vessels classification is applied.

Classification problem is still very significant part in medical image analysis (see e.g. Miri et al., 2017; Renukalatha and Suresh, 2018; Morkunas et al., 2018). Many different classification algorithms are used for vessel classification in eye fundus images (see e.g. Miri et al., 2017; Dashtbozorg et al., 2014; Sun et al., 2009; Niemeijer et al., 2011; Treigys et al., 2008). If the classification is intended to be done automatically and with images of different type, unsupervised classification methods are more suitable. Supervised classification methods are more accurate than unsupervised methods, but they must be adapted for different type images. Different medical image analysis methods have specific image quality issues (see e.g. Prasath, 2017). Eye fundus vessel classification struggles with different or uneven lightening issues. The proposed algorithm uses k-means clustering method with spatially recalculated features decreasing the influence of uneven lightening.

There are many different features which are commonly used in vessel classification. These features are often extracted from RBG, HSI, grey level image information using statistical functions, texture features and others (see e.g. Renukalatha and Suresh, 2018; Muramatsu et al., 2011; Stabingis et al., 2016). Such features are commonly used for many other image analysis problems. New type features, based on the profile information gathered from the vessel width measurement algorithm, are proposed in this paper. These features are extracted from the vessel measurement stage and use information from vessel inner part and average it along the vessel adaptively to the vessel width.

Classification accuracy is compared with other common classification algorithms of an eye fundus blood vessels according to publicly available eye fundus databases. IVSIPIRE-AVR (Niemeijer et al., 2011) and DRIVE (Staal et al., 2004) databases are used for comparison. DRIVE database is used only for a demonstration that algorithm works with images of any resolution without the need for adaptation. The accuracy results from this database are not important, because such small resolution images are not used any more. The proposed algorithm is compared with algorithms proposed in papers (Dashtbozorg et al., 2014; Muramatsu et al., 2011; Mirsharif et al., 2013; Niemeijer et al., 2011). Algorithms used for vessel patch classification are selected only. Algorithms used for the whole vessel tree classification are not used for comparison, because our algorithm is built only for AVR evaluation at this stage.

The method described in Muramatsu et al. (2011) is used especially for AVR evaluation. This method consists of the following stages. First, vessel segmentation method is applied to the retinal image. The segmentation method involves top-hat transformation, double-ring filtering, thresholding techniques and it is performed on green colour channel. Vessel centreline pixels are extracted. According to vessel crossing and bifurcation points, centreline pixels are divided into separate vessel parts. Small vessels (with diameter smaller than 3 pixels) are removed from calculations. Thresholding and ellipse fitting is applied for OD detection. The longer ellipse axis is used as OD radius for AVR measurement zone. Segmentation and OD detection parameters were selected for specific image resolution. Centreline pixels from AVR measurement zone were used in classification. For each centreline pixel, five features are extracted for classification. These features are RGB values corresponding each pixel and contrast values gathered from red and green channels. The contrast features are calculated from red and green channels as the difference in average pixel values in a 5 × 5 pixel region around the pixel of interest inside the vessel, and in a 10 × 10 pixel region around the pixel of interest outside the vessel. LDA classifier was used for each centreline pixel classification and each vessel segment is classified into an artery or a vein by selecting a majority class from classified centreline pixels. Then, major vessels are selected for AVR evaluation. Vessel measurements are done according to segmented vessels, averaging along the vessel patches. This method was tested on DRIVE image database and main results used for comparison are presented in Table 2.

The method used in Mirsharif et al. (2013) consists of the following steps: image enhancement, vessel segmentation, thinning, feature extraction, artery/vein classification and post processing. Histogram matching, several histogram equalization and multi scale retinex techniques were used for image enhancement. The method proposed in Soares et al. (2006) is used for vessel segmentation. This method uses Gabor wavelets. Vessels thinner than three pixels were removed from the vessel network. Vessels are divided into smaller segments and for each segment many features were extracted. Features included information from RGB, HSL and LAB colour spaces and their statistics. From all these features, 8 best discriminating features were selected. LDA classifier was used for this method. Then post processing, using structural knowledge, is applied for reclassification of the vessels unconnected to the tree. The method was used for classification of all vessels and for classification of vessels in AVR measurement zone. OD was marked manually here. AVR measurement zone data from DRIVE image database is used for comparison.

Method used in Niemeijer et al. (2011) consists of the following steps: pre-processing, vessel segmentation and centreline extraction, OD detection, vessel width measurement, vessel classification, AVR evaluation. Pre-processing involves FOV extraction and Gaussian filtering. Vessel segmentation is performed with classification method. This stage was constructed for DRIVE database, and the use with larger images involves reducing the image size. Pixel classification produces likelihood map and threshold produces vessel structure. The method includes “tubogganing” technique from Fairfield (1990). Then, the structure was thinned. A supervised position regression method was used to detect the centre of the OD (Niemeijer et al., 2009). The OD radius was considered as constant for all images. Vessel measurements were done perpendicular to detected local centreline angle. Artery/vein classification was done with LDA method, using 27 different features. Features were extracted from different image channels using different statistics. Statistics were applied to vessel centreline pixels and to vessel patches. During this classification stage, vessel likelihood was calculated. Final class label for vessels was obtained only during the AVR calculation stage. During this stage, vessels were classified into arteries and veins including prior artery/vein structural knowledge and classified vessel pairs were used for AVR measurements. This method was tested with INSPIRE-AVR database, which was first introduced in this paper (Niemeijer et al., 2011).

Finally, the method (Dashtbozorg et al., 2014), which includes graph analysis, is selected for the comparison. This method uses the characteristics of retinal vessel tree. The method described in Mendonca et al. (2013) is used for vessel segmentation. Vessel centreline structure is extracted and then used for graph construction. The obtained graph is then modified, to solve common error issues. Graph analysis leads to initial node classification. Then, LDA classifier is applied with features extracted from different image channels and with their statistics. 19 features were used for this classification part. Whole vessel tree classification was performed there, also, calculation only for AVR measurement region vessels are presented. This method was tested with DRIVE and INSPIRE-AVR databases.

The proposed algorithm is also tested on the high resolution image database OPTO-AVR described in Section 3. Here, images are acquired with Optomed OY digital mobile eye fundus camera Smartscope M5 PRO from soldiers before and after sports load. The database consists of 86 different images.

The structure of this article is as follows. Method implementation chapter describes the proposed method which consists of scale parameter evaluation, OD detection, vessel measuring, feature extraction, and classification. Method comparison with other similar methods is presented in the result section. Conclusions and further development opportunities are described in the conclusion section.

Main adaptive features of the proposed method, compared with less adaptive methods are presented in Fig. 1. The proposed method is adaptive to different size images, no image downsampling is applied. Any parameters don’t need to be tuned manually for different image data sets, no different models, with different parameters are created. Image noise and uneven lightening of the image is considered while creating the algorithm. All main steps: OD detection, blood vessel measurements, artery/vein classification and AVR evaluation, are done automatically.

Main steps of an automatic algorithm for evaluation of artery and vein ratio are presented in Fig. 2. Green colour channel I G from eye fundus image is used for eye fundus mask I M extraction. For mask extraction, common image analysis techniques are applied. After thresholding, the higher intensity part from I G , mathematical morphology closing and flood fill operations are used to form a circular mask of eye fundus part in the image. I M is then used to calculate the scale parameter p s according to Eq. (1) combining logistic growth model with a linear model. (1) p s = β 1 1 + β 2 exp ( − W M β 3 ) + W M β 4 , where β 1 , β 2 , β 3 are parameters of the logistic growth model part, β 4 is the parameter of the linear model part ( β 1 = 4, β 2 = 20, β 3 = 500 and β 4 = 3000 in our case), W M is the width of fundus mask I M . Parameter p s , evaluated by Eq. (1), is used in almost all steps of the algorithm. This parameter allows to use images of different sizes. Eq. (1) is a modified parameter version from Stabingis et al. (2018) and includes adaptation for different FOV values of images. Newer fundus cameras acquire images with larger resolution and also use optics with larger FOV value. The obtained mask I M is further used in other steps of the algorithm like norming operation and others.

Image pre-processing allows to unify different images and to reveal some image features required for specific operations. Image pre-processing is treated as a vital stage for many eye fundus image analysis methods. The proposed method consists of several different large stages (see Fig. 2.), and for each stage, a different image pre-processing algorithm is applied. Pre-processing allows controlling different stages separately because some methods can enhance the extraction of the blood vessel tree but reduce OD extraction reliability. Vessel tree I t r extraction stage is not very important in the proposed algorithm, because the I t r information is used in other stages only as the auxiliary. In some other eye fundus analysis algorithms, I t r extraction is very important, and the extracted tree is used as the reference for almost all other stages. Such dependence on extracted I t r can lead to some errors, especially for sick patients, when some artefacts in the image can be treated as blood vessels or when vessel and background contrast is very low, producing breakage in a vessel tree. There are several different technics for I t r extraction like thresholding, mathematical morphology, Gabor filtering, wavelets, etc. More precise methods are also developed. These methods produce higher extraction accuracy but require a larger amount of computations. Mathematical morphology based method is applied here and is described in more detail in Stabingis et al. (2018). All other stages work even without the extracted I t r and in places where I t r is incorrectly extracted.

The proposed method is tested on three eye fundus image databases. DRIVE (Staal et al., 2004) image database is one of the most popular databases for different method comparison, but this database is old, and images in this database are of small resolution – 565 × 584 pixels. There are 40 eye fundus images in this database. This database was selected for the comparison in order to show the method’s ability to work automatically on smaller images. Another database used for comparison is INSPIRE-AVR (Niemeijer et al., 2011). There are also 40 eye fundus images in this database, and the resolution of images is 2394 × 2048 pixels. The method is also tested with OPTO-AVR database where images are acquired in a complicated situation involving physical exercises. Images are taken before the physical load, right after the physical load and after 45 minutes. Such an experiment is performed in order to test whether changes in AVR can show tiredness or stress. Photographs of both eyes are taken with Optomed OY digital mobile eye fundus camera Smartscope M5 PRO. Image resolution is 1536 × 1152 pixels, and there are 86 images in this database. Vessel extraction accuracy is the proportion of detected vessels. Vessels with widths larger than 3 × p s and in AVR measurement zone are selected only. Extracted vessels were classified in two classes. Classification accuracy, artery class specificity ( S p), sensitivity ( S e) and area under ROC curve (AUC) statistics were calculated. Vessel detection and classification results for different databases are presented in Table 1.

Methods used for vessel patch classification in AVR measurement zone were selected for the comparison. The selected methods are tested on INSPIRE-AVR (Niemeijer et al., 2011) and on DRIVE (Staal et al., 2004) image databases. All data used for comparison is gathered from publications that describe the methods. Methods used for comparison are described in more detail in introduction section and performance information of these methods is presented in Table 2.

Vessel classification results compared with other methods show similar classification results. After analysing misclassification cases, main classification errors occur due to vessel crossings and overlapping and due to lightening issues. In all images analysed, 15 non-vessel places were recognized like vessels due to a vessel-like reflections. AVR evaluation was compared with publicly available INSPIRE-AVR (Niemeijer et al., 2011) data and mean absolute error MAE = 0.093 was obtained.  AVR evaluations were compared with experts measurements performed with OPTO-AVR fundus database, and MAE = 0.156 was obtained.

The method proposed in this paper is used for fully automatic eye fundus image based blood vessel measurements, vessel classification and evaluation of artery-vein ratio according to this information. This parameter can be used for early detection of some diseases. Here, novel features, based on vessel spatial information, are used for vessel classification. Features are normalized taking into account the spatial distance between vessels. It reduces the influence of uneven image lightening. Extracted features are used with the k-means method, but the same features can also be used with other common classification methods. The method is tested with DRIVE, INSPIRE-AVR and with OPTO-AVR eye fundus image databases. The accuracy of vessel classification is obtained equal to 0.854, 0.859 and 0.835, respectively. The proposed unsupervised method is competitive to other state-of-the-art supervised classification methods from this field. As it is common for unsupervised methods, the proposed method produces slightly less accurate classification results, but has the advantage due to its universality, because the proposed method does not require adjustment for new image sets, is adaptive to different image sizes and deals with noisy or unevenly lightened images.