Abstract
The retina features the only blood vessel network in humans that is visible in a non-invasive imaging method. This, along with uniqueness and stability throughout life in healthy subjects, makes it an ideal target for personal identification methods in biometric systems and also for the screening and diagnosis of diseases. However, retinal images usually present low contrast of the vessels in relation to the retinal background and high level of noise stemming mainly from the acquisition process. This work aims to reduce noise and improve contrast to increase the accuracy of retinal vessel segmentation. 2D Gabor wavelet (GW) is usually employed to reduce noise and improve vessel contrast in relation to the background. In this work, it is proposed that, before the thresholding, the GW output images are partitioned into 20 sub-images in such a way that each can be treated independently. The images used were obtained from two public databases, DRIVE and STARE, and the algorithm was developed in MatLab® environment. The proposed approach reached an accuracy of 96.15%, sensitivity of 73.42%, and specificity of 98.30% in DRIVE. In STARE, the accuracy was 94.87%, sensitivity 71.74%, and specificity 96.93%. The methods proposed by the authors indicate gains in accuracy and specificity in the automatic detection of retinal vessels, in both databases used, when compared with those in the main published works. The accuracy is also higher than the 94.73% in interobserver accuracy previously determined as the gold standard.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have