Segmentation of Optic Nerve Head Using Two-stage Snakes in Generalized Gradient Vector Field

We propose a generalized gradient vector field based two-stage affine snakes method for the segmentation of optic nerve head in retinal fundus images. We obtain the course segmentation using affine transformation of an elliptical shape prior that tends to converge iteratively based on the force computations in generalized gradient vector field of the fundus image. We optimize six affine parameters out of which two for scaling, two for shearing, and two for translation of the shape prior. We obtain the fine segmentation using point snakes algorithm. The two-stage idea resulted in fast convergence of the snake on the optic disc with great segmentation accuracy. The detection of the optic nerve head is done using normalized template matching technique. The proposed method is validated on various publicly available fundus image datasets such as Drions-DB, IDRiD, RIM-ONE, Messidor, and Drishti-GS, resulting in Dice similarity index of 0.91, 0.94, 0.94, 0.91 and 0.95, respectively.