Deep Learning-Enhanced Robotic Subretinal Injection with Real-Time Retinal Motion Compensation

TL;DR

AMD treatment involves subretinal injections performed under retinal motion; this study introduces a fully automated robotic system that uses iOCT imaging, an LSTM-based ILM motion predictor, 1D needle registration, and dynamic proportional control to synchronize needle motion with retinal displacement. The LSTM predictor significantly outperforms a FFT baseline, enabling real-time motion compensation with predictions on the order of $0.25$ s and achieving mean pre-insertion tracking errors of $16.4\,
40m$. Validation in simulation and ex vivo porcine eyes demonstrates precise motion synchronization and successful subretinal injections, highlighting potential safety and accuracy gains for retinal microsurgery. Overall, the approach advances AI-enhanced robotic assistance for dynamic ocular environments and supports translation toward clinical practice.

Abstract

Subretinal injection is a critical procedure for delivering therapeutic agents to treat retinal diseases such as age-related macular degeneration (AMD). However, retinal motion caused by physiological factors such as respiration and heartbeat significantly impacts precise needle positioning, increasing the risk of retinal pigment epithelium (RPE) damage. This paper presents a fully autonomous robotic subretinal injection system that integrates intraoperative optical coherence tomography (iOCT) imaging and deep learning-based motion prediction to synchronize needle motion with retinal displacement. A Long Short-Term Memory (LSTM) neural network is used to predict internal limiting membrane (ILM) motion, outperforming a Fast Fourier Transform (FFT)-based baseline model. Additionally, a real-time registration framework aligns the needle tip position with the robot's coordinate frame. Then, a dynamic proportional speed control strategy ensures smooth and adaptive needle insertion. Experimental validation in both simulation and ex vivo open-sky porcine eyes demonstrates precise motion synchronization and successful subretinal injections. The experiment achieves a mean tracking error below 16.4 μm in pre-insertion phases. These results show the potential of AI-driven robotic assistance to improve the safety and accuracy of retinal microsurgery.

Figures (8)

• Figure 1: Flowchart of the proposed deep learning-based autonomous subretinal injection method with retinal motion synchronization and compensation algorithm.
• Figure 2: The surgical needle and the ILM and RPE layers of the retina are segmented using deep learning algorithms and the B$^5$-scans technique: 5 evenly spaced B-scans arikan2024real. Our proposed method synchronizes the robot velocity with retina's alternating motion along the $z$ axis, providing adaptive synchronized positioning of the needle tip in between ILM and RPE layers without damaging sensitive RPE cells.
• Figure 3: Neural network process to predict ILM position: (a) sequence of segmented ILM $z$-position, given by segmentation results, (b) LSTM network, and (c) one-step-ahead prediction of ILM $z$-position.
• Figure 4: The experimental setup includes the SHER 2.0, the Galil controller, the Leica iOCT microscope (left), the piezo-actuated linear stage, the syringe pump (middle), a 42-gauge needle, and an open-sky porcine eye (right).
• Figure 5: Regions of retina motion collection.