Vision for Vision — Deep Learning in Retinal Image Analysis

摘要

Automated, fast and large-scale computer-aided diagnosis of medical images has become reality. The greatest breakthrough is Deep Learning. It already has huge impact in self-driving cars, industrial product inspection, surveillance, robotics and translation services, and in the medical arena it is outperforming human experts already in many domains. However, it is still largely a black box. What can we learn from recent insights in the functionality, nanometer-scale connectivity and self-organization of the human visual brain? We will discuss several recent breakthroughs in our understanding of visual perception and visual deep learning. We apply these techniques in the RetinaCheck project, a large screening / early warning project for eye damage due to diabetes. In China now an alarming 11.6% of the population has developed diabetes, due to genetic factors and fast lifestyle changes. In this project large amounts of retinal fundus images are acquired, and the e-cloud deep learning system successfully learns to identify early biomarkers of retinal disease. The circle is round: we can prevent blindness by learning from the visual system: vision for vision.