Proliferative retinopathies, such as diabetic retinopathy and retinopathy of prematurity, are leading causes of vision impairment. A common feature is a loss of retinal capillary vessels resulting in hypoxia and neuronal damage. The oxygen-induced retinopathy model is widely used to study revascularization of an ischemic area in the mouse retina. The presence of endothelial tip cells indicates vascular recovery; however, their quantification relies on manual counting in microscopy images of retinal flat mount preparations. Recent advances in deep neural networks (DNNs) allow the automation of such tasks. We demonstrate a workflow for detection of tip cells in retinal images using the DNN-based Single Shot Detector (SSD). The SSD was designed for detection of objects in natural images. We adapt the SSD architecture and training procedure to the tip cell detection task and retrain the DNN using labeled tip cells in images of fluorescently stained retina flat mounts. Transferring knowledge from the pretrained DNN and extensive data augmentation reduced the amount of required labeled data. Our system shows a performance comparable to the human level, while providing highly consistent results. Therefore, such a system can automate counting of tip cells, a readout frequently used in retinopathy research, thereby reducing routine work for biomedical experts.
Keywords: convolutional neural networks; deep learning; flat mount images; oxygen-induced retinopathy; proliferative retinopathy; tip cell detection; transfer learning.