Deep Learning-Based Segmentation of Retinal Vessels in Adaptive Optics Ophthalmoscopy Images

Adaptive Optics Ophthalmoscopy (AOO) is a non-invasive imaging technique that produces high-resolution images of the retinal vessels from which morphometric biomarkers can be extracted. However, the related segmentation techniques must cope with the variability in image quality and complexity, including regions with thin branches or arteriovenous crossings. Considering these challenges, we proposed in [1] a two-step method where dedicated active contour models were initialized from a binary mask of the vessels, extracted by deep learning. In this article, we investigate several deep neural network architectures to improve the quality of the binary mask, as the biomarker accuracy strongly depends on this step. Experiments show that the DeepLabV3+ model combined with the Xception backbone enables us to accurately recover the vessel binary mask (mean accuracy of 98.59%).