Bachelor Thesis: Automated method to segment retinal blood vessels from color fundus photographs.
Supervisor: Professor Dr. G M Atiqur Rahaman, Co-supervisor: Dr. Sajib Saha (CSIRO, Australia)
Segmentation of retinal blood vessels is very important diagnostic procedure in ophthalmology. Segmenting blood vessels in presence of pathological lesion is one of the major challenges while segmentation. In this thesis, we propose a method to segment retinal blood vessel to overcome this challenge. The proposed method is also able to overcome the most of the other challenges such as segmentation in presence of central vessel reflex, crossover and bifurcation regions etc. We evaluate the proposed method on three publicly available and popular dataset: DRIVE, STARE and CHASE DB1. We get promising performance for each of the dataset. For DRIVE dataset, the Accuracy, Area Under Curve (AUC), Sensitivity, Specificity are 0.961, 0.847, 0.711 and 0.983 respectively. For STARE dataset, they are 0.960, 0.878, 0.790 and 0.973 respectively and for CHASE DB1 they are 0.951, 0.854, 0.742 and 0.967. To extract trainable feature, we propose a descriptor and compare the descriptor with a novel descriptor (SURF). We get nearly identical performance for the proposed descriptor respect to SURF descriptor.