2.1.2. Vessel Segmentation

To segment the blood vessels, we use a simple but invaluable algorithm proposed by Saleh et al. [30] consisting of a modified threshold-based segmentation algorithm that extracts the blood vessels especially thick vessels of tree pattern.

2.2. Regions of Interest Definition and Extraction

As stated before, to compensate rotation and translation effects caused by eye-movement, the registration of retinal images is an essential part of many algorithms. This modification reduces the computational complexity of matching algorithm in the identification process. The best registration algorithms usually induced a slight error (a few pixels) in registering the retinal image that it causes the significant problems in the matching procedure. Therefore, the identification algorithms without the OD registration are preferred.

Since all blood vessels in the retina are originated from the OD and the vessels generally cover most of retina’s surface, as a part of solution to eye-movement issue and the problems of registration algorithm, we define surrounded regions by the blood vessels as ROI that is called Surrounded Regions (SRs). Moreover, in the tree pattern of blood vessels, there are dominant landmarks like bifurcations (wherever a blood vessel branch divides into thinner ones) and crossovers (wherever two different blood vessels cross each other) in the blood vessel tree [5, 31]. These two facts (blood vessels leave eye globe through OD and crossover points) guarantee the existence of our SR and also provide a good condition to extract strong and robust geometrical shape features. As far as SRs mostly appear between thicker blood vessels, not only the extraction of SR is simple but also the defined features based on them are more trustworthy for the identification process.

For this purpose, we perform morphological watershed algorithm based on dam construction strategy on the segmented image to extract the SRs. In the dam construction strategy, with dilation of the labeled regions sequentially, the dams construct between different labeled regions [32]. By using this dam construction strategy, we can eliminate unwanted extracted vessels among SRs also the boundary between two SRs locates in the middle of blood vessel. Fig. 3 shows the results of applying the morphological watershed algorithm on a sample of the retinal image. This process is efficient for our goal and there is no need to process each region separately. In addition, we can reduce the effect of thickness variety of the extracted blood vessels that may be caused by different illumination conditions, noise, rotation and translation. By

(a) (b) (c)

Fig. 3: Definition of the SRs in a retinal image. (a) A sample of the retinal images with a shown SR, (b) Extracted blood vessels from image (a) by the segmentation algorithm, (c) Result of the morphological watershed algorithm on the image (b) and the marked SR with *.

Table 1: Extracted region-based shape features from the marked SR in Fig. (3-a).

Features	Value	Normalized Value
Filled Area	3084	0.0366
Perimeter	300.6460	0.1741
Eccentricity	0.9122	0.8511
Major Axis Length	114.2352	0.2702
Minor Axis Length	46.7948	0.1431
Solidity	0.7035	0.5689
Equivalent Diameter	0.6632	0.1777

sorting the extracted SRs based on their area and omitting the greatest region which corresponds to the background, we can represent a retinal image by a set of the labeled SRs for further processes.

Let be a set of the extracted SRs of an image while , where is the number of extracted SRs and could be different in dissimilar images or even in the images belong to the same individual taken in different conditions. Therefore, we consider just the first ten of the largest SRs ( ) while the small SRs will not be too much important in the feature extraction procedure and in addition, it decreases the computational cost.