Ensuring a Connected Structure for Retinal Vessels Deep-Learning Segmentation

Abstract

Retinal vessels identification plays a critical role in computer-aided diagnosis and analysis of fundus images. While Deep-Learning-based segmentation methods have shown remarkable performances in handling detailed and pathological fundus, they produce disconnected components whereas retinal vessels are a connected structure. In this work, we developed a post-processing pipeline to ensure a connected structure for the retinal vessels networks. The proposed pipeline named VNR for Vessels Network Retrieval, generates segmentations with a single connected component (CC). This is performed by removing artifacts that are pixels-miss-classified as retinal vessels, and by reconnecting branches that are well-classified but disconnected. By retrieving the structural coherence in the retinal vessels networks, we enable measurements such as vessels length, tortuosity and depth of the vessels tree structure in a more reliable manner. We evaluate our results using pixel-wise and structural metrics, comparing against manually labelled groundtruth. Before applying VNR the predicted segmentations had an average Dice score of 0.839 with 174 CCs. As a result, 173 CCs need to be deleted or reconnected. After applying VNR, the segmentations have an average Dice score of 0.840 with only 1 CC. VNR is thus able to retrieve the connected structure of the retinal vessels networks while also keeping or increasing pixel information.