Efficient CNN-CRF network for retinal image segmentation

Abstract

The research of retinal vessel segmentation prevails since retinal vessels well indicate the diseases, such as diabetic retinopathy, glaucoma and hypertension. This paper proposes an efficient CNN-CRF framework to segment the vessels from digital retinal images. Our approach combines the prediction ability of CNN and the segmentation ability of CRF, and trains an end-to-end deep learning segmentation model for retinal images. Unlike pixel-wise segmentation, our network is able to segment one image during once network forward computation. When applying our CNN-CRF to the DRIVE database, the average accuracy achieves 0.9536 with the average recall rate of 0.7508, outperforming the state-of-art approaches. And our approach requires only 0.53 s per image, the fastest among deep learning approaches.