Retinal Blood Vessels Segmentation: Improving State-of-the-Art Deep Methods

Retinal blood vessels segmentation is an important step for computer-aided early diagnosis of several retinal vascular diseases, in particular diabetic retinopathy. This segmentation is necessary to evaluate the state of the vascular network and to detect abnormalities (aneurysms, hemorrhages, etc). Many image processing and machine learning methods have been developed in recent years in order to achieve this segmentation. These methods are difficult to compare with one another since the evaluation conditions vary greatly. Moreover, public databases often provide multiple ground truths. In this paper, we implement a competitive state-of-the art method and evaluate it on the DRIVE (Digital Retinal Images for Vessel Extraction) public database. Based on this method, we test and present several improvements which are evaluated using a dedicated performance evaluation protocol. This protocol uses five criteria and three different evaluations in order to assess the robustness of the methods’ performances.