Vessel Segmentation in Coloured Retinal Fundus Images Based on Multi-scale Analysis

Retinal diseases are already the most common cause of childhood blindness worldwide. Accordingly, it would be extensively beneficial to humans and health-related communities if we could automate the procedure of diagnosis thoroughly or at least partially by exploiting capabilities of computer-aided diagnosis (CAD). This paper proposes two segmentation methods, a supervised method and an unsupervised one, which shall expertly tackle the problem of vessel segmentation in retinal fundus images. Our unsupervised method exploits the power of multi-scale spatial filters to locate and detect different types of vessels in terms of vessel diameter. Furthermore, we proposed a novel denoising filter to overcome a challenge called “FOV’s tangential ring” effectively. In our supervised algorithm, we combined the unsupervised method with a support-vector machine (SVM) classifier, in which samples’ features are produced using a feature-fusion technique. Dataset used in this research is the public DRIVE (Digital Retinal Images for Vessel Extraction) dataset. We have also addressed another challenging problem with a solution that is dataset independent, the challenge of generating mask for retinal coloured images. Our supervised method has achieved a higher accuracy of 94.48%, and our unsupervised method has achieved an accuracy of 94.28% with a response time of 1.65 second providing human operators or automatic systems with fast and reliable results.