Automatic Optic Cup Segmentation Using Affine Snakes in Gradient Vector Field

We propose an automated technique for optic cup segmentation using affine snakes in gradient vector field. Localization of the optic disc is done using the superpixel approach. Subsequently, the crop out portion of the optic disc region with a diameter just above the normal optic disc is utilized during segmentation to avoid any false positives. The segmentation of the optic cup is done using affine snakes, which evolves using an affine transformation and requires a precedent knowledge of the desired shape. Firstly, a force field is computed around the image and deform the snake until the net force is minimized. Thereafter, the six parameters of affine transformation is updated to obtain the optimal fit, using the gradient descent algorithm. Moreover, our validation results report on four publicly available datasets, amounting to a total of 2514 images for automatic optic disc localization, and 101 fundus images for optic cup segmentation. The proposed method results in optic disc detection accuracy of 94.9%, 94.05%, 96.75%, and 98.5% on IDRID, DRISHTI-GS, MESSIDOR, and REFUGE respectively, and optic cup segmentation Dice index of 0.8055 on DRISHTI-GS dataset.