Retinal Vessel Segmentation In Fundus Images Using Convolutional Neural Network

The structure of retinal vessel can reflect health status of patients, and a clear representation of retinal vessel map helps ophthalmologist to make diagnosis of some disease, such as diabetic retinopathy (DR) and hypertension. However most automatic methods for the task cannot produce a good performance, they always misclassify pixels in vessel boundaries and thin vessels. In this paper, we propose a deep learning-based model for automatic accurate retinal vessel segmentation. We cascaded two U-net to construct a multi-model network and then obtain a coarse-to-fine segmentation. We introduced residual learning and added sufficient skip connections to reuse feature maps. We adopted dilated convolution and arranged dilation rates carefully to enable the model to capture more context information. Finally, we conducted intensive experiments on DRIVE, STARE, and CHASE_DB1 databases. Our proposed model can produce an accuracy of 0.9552/0.9699/0.9642, an AUC of 0.9787/0.9852/0.9846, a sensitivity of 0.8211/0.8466/0.8395 on DRIVE, STARE and CHASE_DB1 databases, respectively.