| Abstrakt: |
Retinal images are found as the major resources for the automatic diagnosis of diabetic retinopathy (DR). However, the advanced stages like proliferative diabetic retinopathy (PDR) cause the branch out of new and thin vessels in the retina. These vessels create a lot of confusion at the diagnosis of DR through retinal mages. Hence, this paper proposes a new blood vessels segmentation mechanism in three stages; they are major blood vessels segmentation, minor blood vessels segmentation and post-processing. In the first stage, the retinal image is enhanced for quality enhancement followed low pass filtering and top-hat transform to segment major blood vessels. In the second stage, the residue image left in the first stage is processed for pixel level classification. At this stage, each pixel is represented with a set of 19 features. Then this feature vector is fed to support vector machine (SVM) for classification. Finally, the resultant images obtained in the first two stages are combined to get the final retinal vessel structure. For simulation, we used two standard retinal image datasets namely STARE and CHASE_DB and the performance is measured through sensitivity, accuracy, specificity, Jaccard Similarity index and Dice similarity index. Simulation results show that the proposed method achieved better segmentation accuracy and it is of approximately 96.85% and 96.66% for STARE and CHASE_DB respectively. [ABSTRACT FROM AUTHOR] |
