Real-Time Iris Tracking Using Deep Regression Networks for Robotic Ophthalmic Surgery

Abstract

Robotic-assisted platforms are expected to guarantee the accuracy of surgical operation and accelerate its learning curve. Iris tracking can guide the robotic manipulator during the operation. However, few researches focused on it during surgery. It is a big challenge due to the deformation of the iris and occlusion caused by instruments. A novel real-time iris tracking method based on a regression network are proposed to meet the speed and accuracy requirements of the ophthalmic robotic system. It utilizes the low-level visual features and high-level semantic meanings from different layers to capture the discriminative representation of the iris target. Then the bottleneck layers are added to improve computation efficiency. Furthermore, a multi-loss function is designed by jointly learning Absolute loss and Euclidean loss. Finally, the experimental results under the typical surgical scene demonstrate that iris tracker achieves an accuracy of 89.16% and a real-time speed of 134fps with GPU, which is suitable for the ophthalmic robotic system to perform real-time robotic manipulation.