Dilated convolutions in retinal blood vessels segmentation

Abstract

Segmentation of retinal blood vessels allows a quantitative analysis of vessels, hence it helps to diagnose several cardiovascular and ophthalmologic diseases. Manual segmentation is time-consuming, therefore, an automatic method is needed. Dilated convolutions have been recently adopted for semantic segmentation task and higher performances have been achieved. In this paper, we investigate the use of dilated convolutions for retinal vessel segmentation. The proposed architectures are evaluated in the DRIVE dataset. The context information provided by dilated convolutions demonstrated to be valuable for the presented task, leading to more accurate segmentations. Our best model achieves accuracy, specificity and sensitivity of 0.9567, 0.9813 and 0.7903, respectively.