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

Tianle Wu, Mojtaba Esfandiari, Peiyao Zhang, Russell H. Taylor, Peter Gehlbach, Iulian Iordachita · 2025

Read Deep Learning-Enhanced Robotic Subretinal Injection with Real-Time Retinal Motion Compensation by Tianle Wu, Mojtaba Esfandiari, Peiyao Zhang, Russell H. Taylor, Peter Gehlbach, Iulian Iordachita (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-03939.txt.

Introduction

Tianle Wu and co-authors build a robot that injects drugs under the retina while tracking retinal motion live, in work from twenty twenty-five. Subretinal shots treat diseases such as macular degeneration at their source. Breathing and heartbeat shift the retina and endanger precise needle placement. A slip can tear the pigment layer that vision depends on. Intraoperative scans image the retina during the procedure.

A memory network predicts the motion of the inner retinal surface. It beats a frequency-based baseline at forecasting that motion. A live registration step aligns needle tip and robot coordinates. Speed control adapts smoothly as the needle advances toward target. Tests in simulation and pig eyes achieve successful injections. Tracking error stays in the tens of micrometres before insertion. Safer, sharper retinal surgery is the promise shown.

Catalog record

CREATORTianle Wu, Mojtaba Esfandiari, Peiyao Zhang, Russell H. Taylor, Peter Gehlbach, Iulian Iordachita
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-03939

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deep Learning-Enhanced Robotic Subretinal Injection with Real-Time Retinal Motion Compensation Tianle Wu, Mojtaba Esfandiari, Peiyao Zhang, Russell H. Taylor, Peter Gehlbach, Iulian Iordachita 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 μ \,\mu m in pre-insertion phases. These results show the potential of AI-driven robotic assistance to improve the safety and accuracy of retinal microsurgery. I Introduction Age-related macular degeneration (AMD) is the third main cause of visual impairment and blindness worldwide [ 1 ] . According to a meta-analysis study, the projected number of people with AMD would increase to 288 million in 2040 [ 2 ] . The main current standard for treating AMD is based on symptomatic mitigation by intravitreal injections of anti-vascular endothelial growth factor (VEGF) antibodies, but it has issues such as ”resistance to anti-angiogenic therapy” [ 3 ] and systemic toxicity [ 4 ] . Recent advances, such as stem cell transplantation [ 5 ] and gene therapy [ 6 ] , offer more effective long-term solutions to treat the cause of AMD, through direct delivery of therapeutic agents into the subretinal space. Subretinal injection requires cannulation of a microsurgical needle into the retina’s internal limiting membrane (ILM), positioning the needle tip precisely above the retinal pigment epithelial (RPE) cells without damaging the non-regenerative RPE, as it may cause irreversible damage to the eye vision. Performing a free-hand subretinal injection is challenging due to the delicacy of RPE, physiological hand tremor (about 182 \,\mu m [ 7 ] ), small thickness of retina (within the range of about 100 to 570 \,\mu m [ 8 ] ), improper visualization of relative depth of the needle to the target site, to name a few. Recent advances in robotic technology enable removing hand tremor, providing required positioning accuracy for retinal microsurgeries. Some examples include the Preceyes Surgical System [ 9 ] , the Intraocular Robotic Interventional and Surgical System (IRISS) [ 10 ] , the Steady-Hand Eye Robot (SHER) [ 11 ] , and others [ 12 , 13 , 14 ] . Two versions of this robot, namely SHER 2.0 and SHER 2.1, are 5 Degrees-Of-Freedom (DOF) serial manipulators that can be used in either cooperative or teleoperation modes, both unimanual [ 15 ] and bimanual [ 16 ] frameworks. Human-in-the-loop control modes are limited due to constrained visualization, making accurate positioning of the needle tip inside subretinal space extremely challenging. Several image-guided robot navigation and control strategies are proposed to autonomously guide the robot toward the target area in ex vivo porcine eyes using intraoperative optical coherence tomography (iOCT) imaging and deep learning methods [ 17 , 18 , 19 , 20 ] . Despite promising outcomes in autonomous subretinal injection research, these methods, however, are based on the simplistic assumption that there was no movement in the eye. Neglecting the reciprocating movement of the retina, in actual human eye, caused by respiration, heartbeat, snoring, etc., can deteriorate the needle positioning control accuracy, damaging delicate RPE layer. In a study, the axial movement of retina (along the perpendicular axis to the retina tissue) resulted by heartbeat and respiration is measured as 81 \,\pm\, 3.5 \,\mu m in amplitude and about 1 Hz in frequency [ 21 ] . Therefore, it is necessary to consider such a reciprocating retina movement in the robot control algorithm to enhance the robotic system’s adaptability with real situations and improve patient safety during autonomous robotic subretinal injection procedures. This paper presents a fully automatic robotic system for subretinal injection, integrating a motion-adaptive control framework with deep learning-based motion prediction. Our approach consists of five structured phases to ensure precise needle positioning despite retinal fluctuations. A U-Net segmentation model extracts the ILM position from iOCT scans, while a Long Short-Term Memory (LSTM) motion prediction module forecasts ILM displacement, outperforming a baseline Fast Fourier Transform (FFT) sine wave model in capturing non-periodic motion patterns. To enhance needle positioning accuracy, we employ a registration step that maps the estimated needle tip position from the image frame to the robot’s local frame, applying noise reduction and a 1D registration transformation for depth alignment. Finally, a proportional speed control algorithm dynamically adjusts needle velocity to synchronize with ILM motion, ensuring smooth and precise insertion. By integrating predictive modeling, adaptive registration, and synchronized control, our method enhances the safety and accuracy of autonomous robotic subretinal injections in dynamic retinal environments. The key contributions are: • Fully automated robotic subretinal injection with real-time motion compensation, adapting to varying retinal motion amplitudes and frequencies in a setting close to clinical conditions. • A dynamic proportional control strategy for motion synchronization, replacing constant velocity insertion with an adaptive approach that enhances precision. • A robust and efficient LSTM-based motion prediction model, achieving both high accuracy and fast inference with minimal input data. The model significantly outperforms the baseline FFT sinewave approach, enabling real-time prediction within 0.25 s. • Comprehensive validation in both simulation and ex vivo porcine eye experiments, demonstrating the system’s effectiveness under ideal conditions and real biological environments. II Methods Fig. 1: Flowchart of the proposed deep learning-based autonomous subretinal injection method with retinal motion synchronization and compensation algorithm. Our control algorithm follows a structured workflow with five distinct phases (Fig. 1 ) to ensure precise and synchronized robotic needle insertion. Phase 1: Motion Estimation - A segmentation network [ 20 ] is used to generate the ILM z -position. Phase 2: Needle Registration - The needle moves to a preparation position, approximately 500 \,\mu m above the ILM layer. An error report is generated based on the difference between the predicted ILM z -position (from Phase 1) and the ground truth ILM z -position. The absolute maximum difference is denoted as e . A 1D registration process maps the needle tip position from the image frame to the robot’s local frame, generating the registration matrix M . Phase 3: Sanity Check - The prediction error e is compared against a predefined threshold. Sanity checks on the registration are performed. If it detects the needle in the lower half of the OCT B-scan, which indicates a potential misalignment, the result is a failure in the sanity check. If both checks pass, the process continues; otherwise, manually reposition the needle and restart the process. Phase 4: Motion Synchronization - The robot performs synchronized motion while maintaining a fixed

Read the complete text