Abstract
Glaucoma is a chronic eye condition causing irreversible vision damage and presently stands as the second leading cause of blindness worldwide. Damaged optic disc and optic cup assessment in color fundus image has been shown to be a promising method for glaucoma screening. In this paper, we propose a new accurate and efficient optic disc and cup segmentation methodology using a residual deep learning approach. The proposed patch-based deep network (GlaucoNet) is trained with a large number of preprocessed input image patches, big enough to include the important discriminatory information around each pixel. The proposed architecture gives better segmentation results as extra skip connections are introduced in the framework which explicitly reformulate the layers such that the learning function is dependent on a residual function of the input layer. It addresses the degradation problem of the deep network by improving the information flow. The convex hull transformation is applied on the initial results to obtain the final segmentation output. The proposed method is tested on publicly available DRISHTI-GS, RIM-ONE and ORIGA-light datasets. The experimental results for optic disc segmentation show an overlapping score (OS) of 0.9106, 0.8972 and 0.8835 and the optic cup segmentation achieves an OS of 0.8229, 0.7401 and 0.8106 in DRISHTI-GS, RIM-ONE and ORIGA-light, respectively. Finally, the glaucoma risk index is obtained by computing the cup-to-disc height ratio obtained from the segmented regions. Experimental results showing better or comparable segmentation performance and decrease in cup-to-disc height ratio estimation errors demonstrate the efficacy of the proposed GlaucoNet-based segmentation methodology.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have