Diseases such as diabetes and high blood pressure are growing in incidence, generating new needs for microvascular diagnosis. The retina, which offers an observation window of the human microvascular network, can be imaged in high resolution thanks to Adaptive optics ophthalmoscopy (AOO), giving a unique opportunity for deciphering the anatomy and physiology of human microvessels.
In this article, we propose a complete framework, AOV, to segment vessels in AOO and compute biomarkers characterizing arteries and arterial bifurcations. Our segmentation method combines deep-learning and active contour models: deep-learning allows to deal with high variability of images and vascular topologies while parametric active contours allow for modeling structural information and reach better accuracy. To this end, we trained an convolutional neural network, a U-Net optimized to be robust to various vessel sizes and orientations. This initial result is refined in a second step, with new parametric active contour models designed to precisely delineate artery walls and handle bifurcations. Our models integrate structural a priori knowledge such as approximate parallelism or symmetry properties. Estimates of vessel diameter and arterial wall thickness are extracted from the segmentation results and biomarkers are derived to characterize arterial branches and bifurcations.
Our experiments demonstrate that our algorithm provides reliable and accurate segmentations (mse=1.75±1.24 pixel), with high reproducibility. Vessel diameters are well measured with an overall error of 0.19 +/- 2.92 pixel, leading to good biomarker estimates.
AOV also proposes a supervised usage mode to deal with the most complex images. To our knowledge, it is the only framework that allows large-scale analysis of AOO images, with good accuracy and acceptable user involvement. Biomarkers are provided with the corresponding interval of error, which is another important contribution.