Tezislər / Theses

THE 3
rd
 INTERNATIONAL SCIENTIFIC CONFERENCES OF STUDENTS AND YOUNG RESEARCHERS 
dedicated to the 99
th
anniversary of the National Leader of Azerbaijan Heydar Aliyev
178
Various algorithms are present for retinal image processing. In our work 
the used technique is a type of Supervised methods. In supervised methods, 
the rule for vessel extraction is learned by the algorithm on the basis of a 
training set of manually processed and segmented reference images often 
termed as the gold standard. This vascular structure in these ground truth or 
gold standard images is precisely marked by an ophthalmologist. However, 
there is significant disagreement in the identification of vessels even amongst 
expert observers. In a supervised method, the classification criteria are 
determined by the ground truth data based on given features. Therefore, the 
prerequisite is the availability of the already classified ground truth data, 
which may not be available in real life applications. As supervised methods 
are designed based on pre-classified data, their performance is usually better 
than that of unsupervised ones and can produce very good results for healthy 
retinal images. The use of Principal Component Analysis (PCA) followed by 
neural networks reported a success rate of 99.56% for the training data and 
96.88% for the validation data, respectively, with an overall sensitivity and 
specificity of 83.3% (standard deviation 16.8%) and 91% (standard deviation 
5.2%), respectively. The result of the approach was compared with an 
experienced ophthalmologist manually mapping out the location of the blood 
vessels in a random sample of seventy three 20 × 20 pixel windows and 
requiring an exact match between pixels in both images.
To evaluate the performance of proposed method the first step was the 
collection of relevant data. As taking image of retinal blood vessels is not 
easy to achieve some ready databases are offered publicly for those who are 
going to conduct research on relevant topic. We made use of The DRIVE 
(Digital Retinal Images for Vessel Extraction) database for assessment.
After getting the data to work with, methodology has been determined 
for handling segmentation task. Methodology is composed of multiple stages 
since different concept of image processing will be applied to process the 
data. Each stage has found its application in coding part. Image processing 
toolbox of Matlab software has been utilized to deal with various subtasks 
(conversion between image formats and colorspaces, reshaping, adaptive 
histogram equalization, separation of color channels, etc) to achieve the final 
result. Apart from the image processing techniques principal component 
analysis algorithm has also been applied to obtain the required data portion 
from the image for utilization in further stages.
In the testing stage written code has been evaluated on 8 images and 
after the analysis of obtained results certain conclusions were derived to 
summarize the procedure and performance.
The flow diagram for the procedure is demonstrated below and we have 
gone through it step by step (Figure 1):

---