Retinal Vein Cannulation (RVC): Advances & Challenges

• Retinal Vein Cannulation is a microsurgical procedure that targets sub-millimeter retinal veins (approximately 60–150 μm) to deliver therapeutic agents and restore flow in occluded vessels.
• Research demonstrates that robotic and autonomous methods, employing precise visual servoing and state-machine guided workflows, achieve needle-tip accuracies within 20 μm and complete procedures in under 35 seconds.
• Integrating multiple imaging modalities such as monocular video and OCT, along with force-controlled techniques, enhances safety by preventing double puncture and minimizing scleral stress.

Retinal vein cannulation (RVC) is a vitreoretinal microsurgical procedure in which a microneedle is inserted into a retinal vein in order to deliver an anticoagulant or other agent intraluminally, typically upstream of an occlusion. In contemporary research, RVC is treated as a paradigmatic high-precision intraocular manipulation problem because it combines sub-millimeter vascular targets, stringent remote-center-of-motion constraints at the scleral port, limited depth perception under microscope imaging, and a narrow safety margin against double puncture, extravasation, retinal injury, and globe distortion. The recent literature spans manual, robot-assisted, and autonomous formulations, with ex vivo porcine eyes, live chicken embryo vasculature, and eye phantoms serving as principal validation models (Kim et al., 2023, Zhang et al., 2024, Wang et al., 29 Jul 2025).

1. Clinical task definition and geometric constraints

RVC is investigated primarily as a treatment pathway for retinal vein occlusion (RVO), in which a thrombus blocks venous outflow and targeted intravascular delivery may restore flow. The operative target is extremely small: one autonomous porcine study characterizes retinal veins for cannulation as approximately $60\text{–}120\,\mu m$ in diameter, while another study reports that the widest retinal veins average $151.32 \pm 15.22\,\mu m$. These dimensions are commensurate with, or smaller than, reported surgeon tremor amplitudes of approximately $182\,\mu m$ RMS or approximately $200\,\mu m$ mean amplitude, which makes unaided manual stabilization intrinsically difficult (Kim et al., 2023, Wang et al., 29 Jul 2025).

The task is constrained not only by scale but also by ocular access geometry. Instrument motion must pivot about a remote center of motion (RCM) at the scleral entry site in order to avoid scleral shear, corneal striae, eye rotation, and excessive extraocular loading. In the model-based autonomous formulation, this is written as

$(I-r_z r_z^T)(p_{rcm}-p)=0,$

which forces the tool longitudinal axis to align with the line connecting the tool tip and the RCM (Kim et al., 2023).

Depth management is the central perceptual difficulty. A top-down surgical microscope supplies strong lateral information but weak axial cues. In monocular pipelines this is handled indirectly through appearance change and event detection rather than explicit 3D reconstruction. In iOCT-integrated pipelines, B-scan imaging supplies axial information for contact and puncture confirmation. A recurring concern across the literature is double puncture: after the anterior vessel wall is breached, release of stored strain energy can drive the needle into the posterior wall or retina unless insertion is halted promptly (Kim et al., 2023, Zhang et al., 2024).

A related safety issue is scleral loading. In bimanual SHER telemanipulation work that is positioned as relevant to RVC, a suggested safe upper bound for scleral force is $120\,mN$. This force-centric view complements the image-centric literature by emphasizing that precise needle placement alone does not exhaust the safety problem (Esfandiari et al., 2024).

2. Procedural decomposition and workflow variants

Recent RVC research commonly decomposes the procedure into lateral alignment, axial approach, first contact, venous puncture, and infusion or withdrawal. A monocular autonomous system formalizes this as a supervisory state machine with the states Align, Lower, Contact, Insert, and Punctured. The operator specifies a target pixel $i_{goal}$ on the vein, the system aligns the needle tip laterally in the image, lowers incrementally along $Z$ while monitoring contact, inserts axially after contact, and stops immediately when puncture is detected, thereby reserving the post-puncture phase for injection while attempting to prevent posterior-wall breach (Kim et al., 2023).

A closely related autonomous formulation casts the same sequence as the canonical manipulation primitives “reach” and “push.” The “reach” phase consists of image-plane alignment and gentle landing on the vessel; the “push” phase consists of advancement along the needle axis until a learned puncture detector signals lumen entry. This framing is useful because it connects RVC to general robotic manipulation while preserving the task’s distinctive surgical constraints on speed, deformation, and failure tolerance (Kim et al., 2023).

Robot-assisted workflows retain surgeon supervision but externalize tremor suppression, RCM enforcement, and verification. In an ex vivo porcine protocol using keyboard-controlled SHER systems and iOCT, the procedure comprises preoperative setup, needle navigation, a programmed “quick push” puncture, slight retraction for verification, iOCT- and flush-based confirmation of intraluminal placement, infusion, and final withdrawal. The workflow explicitly allows Steps 3–4 to be repeated if iOCT does not confirm entry into the lumen, underscoring that puncture and successful cannulation are not treated as synonymous events (Zhang et al., 2024).

An OCT-guided autonomous workflow validated in a chicken embryo model preserves the same stage-wise structure but changes the sensing assignments. The top-down microscope image is used for 2D navigation, whereas B-scan OCT is used for contact detection and puncture recognition. If puncture is not confirmed with high confidence, the robot retracts by $2/5$ of the puncture depth and retries. This suggests a policy architecture in which progression through the workflow is conditioned not only on geometric state but on confidence in event recognition (Wang et al., 29 Jul 2025).

3. Perception, imaging, and control formulations

The most technically distinctive line of work pursues full autonomy using only monocular microscope video. In that system, needle-tip localization is performed by a ResNet-18 encoder-decoder that predicts the tip image coordinates as per-column and per-row classification on $640\times480\times3$ RGB input. Training used approximately 1200 labeled images and approximately 300 test images, with augmentations including random crops and resizing near the tip, random rotations, pixel dropout, and hue jitter. Visual servoing then uses the image error $151.32 \pm 15.22\,\mu m$0 and an online-estimated $151.32 \pm 15.22\,\mu m$1 calibration Jacobian $151.32 \pm 15.22\,\mu m$2. The update law is a Broyden update with $151.32 \pm 15.22\,\mu m$3, the XY goal tolerance is $151.32 \pm 15.22\,\mu m$4 pixel, and each lowering step uses $151.32 \pm 15.22\,\mu m$5 along $151.32 \pm 15.22\,\mu m$6. Motion planning is generated by Differential Dynamic Programming with trajectory length $151.32 \pm 15.22\,\mu m$7, while the cost penalizes terminal tracking error, control effort, and RCM deviation (Kim et al., 2023).

Event detection is likewise vision-only in the monocular system. Needle–vein contact is inferred from needle deflection at the elbow by normalized cross-correlation (NCC) template matching. A contact event is declared when the percent drop in maximum NCC exceeds $151.32 \pm 15.22\,\mu m$8, tuned on 14 contact exemplars. Venipuncture is detected by a recurrent CNN implemented as a ResNet-18 encoder with an auxiliary image-reconstruction decoder; inputs are $151.32 \pm 15.22\,\mu m$9 sequences over a 15-frame temporal horizon, corresponding to approximately 3 s at 7 Hz after down-sampling from 30 Hz. Training used approximately 200 puncture events with approximately 50 held out for testing, and labels were delayed by 1–2 frames during training to avoid premature stopping (Kim et al., 2023).

OCT-guided autonomy adopts a different sensing partition. In the chicken embryo study, microscope-based tip localization is handled by Detectron2 keypoint Mask R-CNN with a ResNet-101 backbone, FPN, and Keypoint R-CNN head, trained on 17,000 microscope images from a total dataset of 47,342 images. Navigation stops when $182\,\mu m$0 px at a microscope ROI scale of $182\,\mu m$1 mm/pixel. Contact detection on B-scan OCT uses a ResNet-18 binary classifier trained on 4,353 labeled B-scans and augmented to 11,350 images, reaching 98% accuracy on a 200-image test set. Puncture detection uses YOLOv8x on 4,730 labeled B-scans and reports $182\,\mu m$2 and $182\,\mu m$3. The OCT stream provides axial resolution $182\,\mu m$4, which is used specifically for contact and puncture discrimination rather than general 3D scene reconstruction (Wang et al., 29 Jul 2025).

Robot-assisted but non-autonomous systems use imaging differently. The porcine keyboard workflow couples Leica Proveo 8 microscopy with integrated iOCT so that the top-down view supports gross targeting while B-scan verification establishes whether the tip has crossed the vessel wall into the lumen. The paper reports no full camera–robot calibration matrices or iOCT resolution details, and B-scan positioning is adjusted manually, but it does formalize RCM orientation control through the SO(3) logarithm $182\,\mu m$5, with rotation halting when the error norm is below $182\,\mu m$6 degrees (Zhang et al., 2024).

4. Robotic platforms and empirical performance

Autonomous monocular RVC in cadaveric porcine tissue has been reported with 24 back-to-back cannulations on 4 pig eyes, all successful by the study’s criterion of accurate placement, successful puncture, and timely stop to prevent over-insertion. Air injection confirmed lumen access in every trial. The mean final needle-tip error to the clicked goal was $182\,\mu m$7, the maximum error was $182\,\mu m$8, the complete procedure took less than 35 s on average when model-predictive control was used in trials 13–24, and all procedures were completed in under 1 minute. Maximum RCM deviation was at most $182\,\mu m$9, mean extra tip travel after human-labeled contact was $200\,\mu m$0 with a maximum of $200\,\mu m$1, and mean extra $200\,\mu m$2-travel after puncture detection was $200\,\mu m$3 with a maximum of $200\,\mu m$4 (Kim et al., 2023).

An earlier autonomous report from the same research trajectory presents the system as the first autonomous RVC strategy using only a robotic arm, a needle, and monocular images. In that account, 24 autonomous insertions were performed on 3 ex vivo porcine eyes, with mean XY error $200\,\mu m$5, average duration approximately 35 s, mean contact-detection overtravel $200\,\mu m$6, and mean post-puncture overtravel $200\,\mu m$7. No visible tissue damage or double puncture was observed (Kim et al., 2023).

Robot-assisted ex vivo porcine RVC with keyboard teleoperation and iOCT verification has been validated on 12 eyes, achieving 10 successful cannulations and infusions, or 83.33%. For the 10 successful cases, the mean time for navigation was 57.45 s, the mean time for puncture plus slight retraction was 43.55 s, the mean infusion time was 67.58 s, and the mean final retraction time was 19.99 s, yielding a total active time of approximately 188.57 s for Steps 2–6. Infusion was performed at 12 mmHg using Viscous Fluid Control mode (Zhang et al., 2024).

Autonomous RVC on a live chicken embryo model reports 20 validation trials in which manual keyboard teleoperation and autonomous navigation were both performed before puncture. Across trials, autonomous navigation averaged 36.74 s and autonomous puncture averaged 26.97 s. Relative to manual keyboard control, autonomous navigation reduced time by 68.6% from a manual mean of 117 s, and autonomous puncture was markedly faster than manual procedures, which averaged 469 s. In real experiments, puncture detection accuracy was 85%; class-wise metrics were reported as Precision 0.90, Recall 0.75, and F1 0.82 for failure, and Precision 0.82, Recall 0.93, and F1 0.87 for success (Wang et al., 29 Jul 2025).

These figures are not directly commensurate across studies because the experimental models, sensing stacks, success definitions, and endpoints differ. A plausible implication is that RVC performance should be interpreted as workflow-specific rather than as a single scalar benchmark.

5. Safety logic, verification, and common points of confusion

The literature treats safety as a multilayer problem. At the motion-planning level, RCM constraints reduce scleral shear and eye rotation. At the event-detection level, immediate stopping on contact or puncture aims to avoid vessel compression, posterior-wall breach, and retinal injury. At the verification level, iOCT or injection tests are used to distinguish true intraluminal placement from wall engagement or near-miss puncture (Kim et al., 2023, Zhang et al., 2024).

One common misunderstanding is to equate apparent vessel penetration with successful cannulation. The robot-assisted porcine workflow explicitly requires iOCT B-scan visualization of the tip inside the lumen and a small “water flush” before full infusion proceeds. Post-infusion iOCT B-scans showing “clear, empty” vessel lumens provide an additional confirmation layer. In the autonomous porcine monocular studies, lumen access was confirmed by air injection rather than cross-sectional OCT (Zhang et al., 2024, Kim et al., 2023).

A second misunderstanding is that tremor suppression alone solves RVC. Force-control studies indicate otherwise. In bimanual SHER telemanipulation, adaptive sclera force control uses FBG-based sensing of $200\,\mu m$8 and $200\,\mu m$9 and switches from kinematic teleoperation to force regulation when thresholds are exceeded. In sitting posture, BMAT reduced dominant-hand mean sclera force from $(I-r_z r_z^T)(p_{rcm}-p)=0,$0 in BMAC to $(I-r_z r_z^T)(p_{rcm}-p)=0,$1, and reduced the percentage of time above the $(I-r_z r_z^T)(p_{rcm}-p)=0,$2 limit from $(I-r_z r_z^T)(p_{rcm}-p)=0,$3 to $(I-r_z r_z^T)(p_{rcm}-p)=0,$4. Although this study used an eye phantom and a vessel-following task rather than live venous cannulation, it isolates scleral-load regulation as a separate safety axis that image-based targeting does not directly address (Esfandiari et al., 2024).

The studies also make clear that prompt stopping is not instantaneous in practice. In monocular autonomous porcine trials, contact detection lagged human assessment by tens of milliseconds, and actual extra tip travel after contact averaged $(I-r_z r_z^T)(p_{rcm}-p)=0,$5. Puncture detection was faster, with mean extra $(I-r_z r_z^T)(p_{rcm}-p)=0,$6-travel $(I-r_z r_z^T)(p_{rcm}-p)=0,$7, and the authors judged this to be within safe bounds for $(I-r_z r_z^T)(p_{rcm}-p)=0,$8 veins. In robot-assisted mode, occasional vein bumps attributable to depth misperception were observed, though no visible damage was reported (Kim et al., 2023).

6. Limitations, disputed assumptions, and research directions

No current RVC research program resolves all of the task’s sensing, biomechanics, and translational issues simultaneously. Monocular-only autonomy demonstrates that explicit stereo, OCT, or force sensing is not a strict prerequisite for ex vivo success, but the same studies acknowledge that depth uncertainty still contributes to small contact and puncture detection lags. Their stated future directions include integrating stereo, OCT, or tool-tip force sensing, incorporating motion compensation for breathing-induced retinal motion, enlarging puncture datasets beyond approximately 200 events, and exploring end-to-end temporal policies with safety validation (Kim et al., 2023).

Ex vivo porcine workflows with iOCT address depth verification but remain limited by cadaveric physiology. The studies note absence of physiologic blood pressure or thrombus, vessel collapse, corneal opacity, and the need to simulate venous patency by local compression or air pressurization. They also emphasize that a 100 $(I-r_z r_z^T)(p_{rcm}-p)=0,$9 metal tip is unlikely to achieve full lumen catheterization consistently; partial intraluminal placement may still permit infusion but risks some fluid entering the vitreous. Proposed next steps include smaller needles, tangential approaches, improved iOCT guidance such as volume scans with intelligent real-time slicing, image-based eye-motion compensation such as optical flow, progression to in vivo experimentation, and eventual clinical translation (Zhang et al., 2024).

The chicken embryo line improves imaging richness and data volume but raises a different generalization issue: all models were trained on a single imaging setup and domain, and vessel selection in that study included substantially larger structures than human retinal veins. The authors therefore propose validation on ex vivo porcine eyes with smaller and more realistic anatomy, unification of contact and puncture detection into one model, adaptive image preprocessing, confidence-based decision logic, external-dataset validation, and exploration of more generalizable architectures such as self-supervised vision transformers or models with geometry priors (Wang et al., 29 Jul 2025).

Force-control work identifies yet another unresolved assumption: even when scleral forces are regulated, the evidence base is still largely phantom-based and expert-user-specific. Planned extensions include multi-user studies, biological validation, haptic feedback, and tighter integration with imaging, visual servoing, and automated puncture assistance (Esfandiari et al., 2024).

Taken together, the field does not converge on a single canonical RVC stack. One branch prioritizes minimal hardware and monocular inference; another prioritizes iOCT-based confirmation and staged teleoperation; a third emphasizes force-aware bimanual manipulation. This suggests that future clinically deployable RVC systems may be hybrid rather than doctrinally monocular, OCT-only, or force-only, combining event-driven autonomy, geometric RCM enforcement, depth verification, and force-aware supervision in a single safety architecture.