Two-Stage U-Net for Optic Disc/Cup Segmentation

Optic disc/cup segmentation is the key step to calculate cup-to-disc ratio automatically which plays a substantial character in the clinical marker investigation of glaucoma early screening and intelligent diagnosis. However, conventional segmentation methods have disadvantages in determining the boundary of optic disc, identifying optic cup and vessel. This investigation proposes a novel two-stage U-Net to segment optic disc/cup gradually. Domain knowledge of fundus image and widely used morphological processing techniques are integrated in input and output block, which enable the model to improve the optic disc/cup's contrast and correct the unbalanced illumination among all images. Besides, he neural network introduced deep supervision block to facilitate the low-level layers to acquire further semantic knowledge. The proposed algorithm attains pixel-wise AUC of 99.70% and 96.83% for optic disc and cup segmentation separately, which achieves the state-of-the-art representation for these two assignments on the Drishti-GS1 dataset.