{"id":"57d58a3a-372d-4671-80ae-9dcbe1d48757","arxiv_id":"2507.21965","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-robot system with deep learning models performs autonomous retinal vein cannulation in chicken embryos, classifying puncture outcomes with 85% accuracy on 27 attempts.","lead":"This paper reports a robotic system that autonomously navigates to a vein, detects contact, and punctures the vein in chicken embryos using microscope and OCT imaging with deep learning models. It is a step toward automated retinal vein surgery, though validation on human-sized vessels is still needed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 85% accuracy is a per-attempt classifier score, not a procedural cannulation success rate; with 3 false-positive puncture predictions in 27 attempts, the full-pipeline autonomy claim is not yet supported.","rationale":"The reader's conditional verdict is appropriate. The reader listed the per-attempt-classification-vs-procedural-success gap in the rationale, but their stated weakest assumption was B-scan plane alignment and depth control accuracy. My stress-test elevates the classifier-vs-procedure issue: even with a perfectly aligned B-scan, the reported evidence does not establish that the full pipeline succeeds autonomously, because the stopping criterion is the model's own success prediction and false positives terminate the procedure. The proposed check would settle this by reconstructing per-procedure outcomes from the attempt log. This does not move the verdict: the paper still contains a coherent feasibility demonstration, but the headline claim needs to be reworded or supplemented with a procedural success rate.","tokens_in":10362,"tokens_out":5075,"duration_ms":62727,"concrete_test":"Reconstruct each procedure from the raw per-trial logs as a sequence of attempts until the system stops on predicted success or the protocol ends. Count the number of procedures ending in verified cannulation (air inflation and/or blood exit) divided by total procedures, and report false-positive terminations separately. If the per-procedure success rate is below 85%, or if the three estimated false positives correspond to terminated failures, revise the abstract to state 'per-attempt puncture classification accuracy' and report the verified procedural success rate. Also report retries per procedure and puncture-stage timing separately from navigation and contact time.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the integrated system reliably completes the retinal vein cannulation pipeline autonomously. The only end-to-end outcome metric reported is 85% classification accuracy for puncture detection (abstract, Section V). Table II's precision/recall/support values imply a confusion matrix of roughly TN=9, FP=3, FN=1, TP=14 out of 27 attempts. Because the control policy stops only when the puncture model confirms success, false positives are the dangerous errors: when the model declares success on an unpunctured vein, the procedure is terminated without cannulation. The 82% precision on Class 1 means about 3 of 17 predicted successes were false alarms, and Section IV-C states such cases were marked as failures. Thus 85% accuracy and 93% recall on successes do not translate into a per-procedure success rate; with retries, the actual fraction of completed cannulations is not reported. In addition, the timing comparison (26.97 s vs 469 s) appears to measure only the puncture stage after contact, and Section VI concedes that depth control accuracy was not quantified. The feasibility insight remains, but the headline autonomous-cannulation claim is currently supported by a classifier metric, not by a demonstrated procedural success rate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an automated robotic system for retinal vein cannulation (RVC) that combines a top-down microscope with B-scan optical coherence tomography (OCT) and deep-learning modules for needle tip navigation, contact detection, and puncture recognition. Validation is carried out on a chicken embryo model, where the system reports 85% classification accuracy for puncture detection, reduced navigation and puncture times compared to manual (keyboard) control, and an independent outcome check by air injection into the vein. The authors acknowledge limitations, including sensitivity to lighting and occlusion, lack of depth-control accuracy quantification, and absence of a systematic failure-mode analysis. The central feasibility claim is that deep learning plus B-scan OCT can support real-time navigation, contact detection, and puncture verification for a microsurgical task in a living animal model.","tokens_in":10570,"tokens_out":6443,"duration_ms":75613,"significance":"If the result holds, the system is a meaningful step toward automating a delicate microsurgical procedure; the use of an independent physical outcome check (air-induced vein inflation or blood exit) is a strength, as is the integration of two robotic platforms with real-time OCT imaging. However, the evidence presented is currently at the level of a per-attempt binary classifier evaluated on 27 attempts, without a per-procedure success rate, confidence interval, or quantitative depth-accuracy metric. The feasibility insight is real but the full-pipeline autonomy claim is not yet supported by the reported metrics. The paper also offers a useful comparison with prior robotic RVC work and is transparent about several limitations.","major_comments":[{"comment":"The headline 85% accuracy is a per-attempt classification accuracy on 27 puncture attempts, not a procedural cannulation success rate. The precision and recall values imply a confusion matrix of roughly TN=9, FP=3, FN=1, TP=14. Because the control policy in Section III stops and fully retracts the needle once the model predicts puncture, a false-positive prediction would terminate the procedure without actual cannulation. The paper does not report how many of the 20 embryos were ultimately confirmed by the air-injection check to have been successfully cannulated by the autonomous pipeline. Please report the per-procedure success rate, the number of attempts terminated by false positives, and a confidence interval for the per-attempt accuracy.","section":"Section V, Table II"},{"comment":"The timing comparison (automated puncture time 26.97 s vs manual 469 s) is not clearly defined: it is unclear whether the manual baseline includes the same stages (contact detection, puncture, verification) and from which event each timing measurement starts and ends. The navigation time reduction (36.74 s vs 117 s) is also reported without the corresponding navigation accuracy, such as the distribution of final pixel-distance errors relative to the 3-pixel threshold. Time savings alone are not evidence of improved overall performance unless the success rates of the automated and manual conditions are also reported.","section":"Section V, Table I"},{"comment":"The conclusion states that 'we did not quantify the needle's depth control accuracy nor perform a systematic analysis of failure modes.' This is load-bearing because the abstract claims 'precise depth sensing' via B-scan OCT, and both contact detection and puncture recognition rely on B-scan images that must capture the needle-tissue interaction in the imaging plane. Without reporting a metric for B-scan plane alignment, needle tip localization error in depth, or z-axis positioning error, the 'precise depth sensing' claim is unsupported. The paper should provide at least one quantitative measure of depth positioning accuracy or an explicit characterization of the B-scan plane registration error.","section":"Section VI and Section III"},{"comment":"The validation description says the procedure was 'conducted across 20 embryos,' but Table II reports 27 puncture attempts (support of 12 for Class 0 and 15 for Class 1). The relationship between the 20 embryos and the 27 attempts is not explained (e.g., how many retries occurred and whether retries were counted as independent samples). This ambiguity affects the interpretation of the reported accuracy, the validity of any statistical inference, and the reader's ability to assess whether the three false positives correspond to three distinct embryos or repeated attempts on the same embryo.","section":"Section IV-C and Table II"}],"minor_comments":[{"comment":"There is a typo in the sentence 'improved puncture time to an average of 26.97 s,significantly faster' — the comma should be followed by a space before 'significantly.'","section":"Section V"},{"comment":"The dataset notation D = {(Ii, bi)}N i=1 is malformed; it should be typeset as D = {(I_i, b_i)}_{i=1}^{N} for clarity.","section":"Section II-D"},{"comment":"The navigation model is described as 'a ResNet-101 backbone with a key point RCNN head' but Figure 3(a) labels the model as 'Mask RCNN'; these names should be reconciled.","section":"Section III-A and Figure 3"},{"comment":"The comparison to Willekens et al.'s 73% procedural success rate is apples-to-oranges because that metric is a per-procedure cannulation success, whereas the 85% here is a per-attempt classification accuracy; the comparison should be clearly framed as non-equivalent or replaced with a procedural success rate.","section":"Section V"},{"comment":"No confidence intervals or statistical tests are provided for the time comparisons or the classification metrics; reporting 95% confidence intervals for the accuracy and the time differences would help assess precision given the small sample sizes.","section":"Table I and Table II"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth reading. This is a genuine attempt at a full autonomous pipeline for retinal vein cannulation, with real experiments and an external outcome check. The new thing is the integration: prior work from this group covered teleoperated RVC and autonomous subretinal injection, but here they combine needle navigation, contact detection, and puncture detection into one workflow, using one robot to stabilize the vein and another to drive the needle. Validation on 20 live chicken embryos with outcome confirmed by visible vein inflation or blood exit after air injection is a solid, honest experimental design. The models are standard (keypoint RCNN, ResNet-18, YOLOv8x), but the pipeline contribution stands.\n\nThe load-bearing soft spot is that the headline 85% accuracy is a per-attempt classification metric, not a procedural success rate. From Table II, the support values imply roughly 27 attempts, and about 3 false positives where the model predicted puncture but the physical check failed. Because the control loop stops on predicted success, those are failed procedures, and the paper never reports how many of the 20 trials ended in successful cannulation. The timing comparison (26.97 s vs 469 s) also covers only the puncture phase after contact, and the manual baseline is keyboard teleoperation, which the paper does acknowledge. Depth control accuracy is unquantified, as Section VI admits. There is also an internal inconsistency: Figure 1 gives vein diameter as 300–400 µm while the training section says the average was 1.27 mm. These are fixable but real.\n\nWho is this for? People working on robotic microsurgery, especially OCT-guided automation. The paper deserves a serious referee; the experiment is real, the limitations are acknowledged, and the central feasibility claim—that deep learning plus B-scan OCT can drive a full cannulation pipeline in a living model—probably holds up, though the abstract oversells it. Send it to review, but ask for a procedural success rate, a confidence interval, and a correction of the diameter inconsistency.","headline":"A real integrated pipeline for autonomous retinal vein cannulation on chicken embryos, but the headline 85% accuracy is a classifier metric rather than a demonstrated procedural success rate.","tokens_in":11138,"tokens_out":2051,"would_cite":true,"duration_ms":24155,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep-learning-driven robot performs the full retinal vein cannulation pipeline in a live chicken embryo model, detecting punctures with 85% accuracy.","keywords":["retinal vein cannulation","autonomous microsurgery","optical coherence tomography","deep learning","chicken embryo model","surgical robotics","puncture detection","needle navigation"],"falsifier":"Deliberately offset the OCT B-scan plane from the needle tip by about one needle diameter during autonomous punctures and measure puncture-detection accuracy; if accuracy does not collapse, the claim that B-scan depth information drives the decision is not supported. A second check is to compare the robot's commanded depth against the OCT-measured needle-tip depth across trials, a quantity the paper does not report.","tokens_in":10150,"feed_emoji":"🤖","tokens_out":8543,"duration_ms":95612,"temperature":0.7,"pith_summary":"Retinal vein cannulation is a microsurgical procedure for retinal vein occlusion that is hard to perform by hand because veins are narrow, tissue moves, and surgeons' tremor exceeds the vein diameter. This paper tries to establish that a robot can automate the whole procedure end to end: navigate a needle to a target vein, detect contact, puncture the vein, and verify the puncture, all using microscope images for lateral position and B-scan OCT cross-sections for depth. The authors report 85% puncture-detection accuracy on 20 live chicken embryos and average autonomous navigation and puncture times of 36.74 s and 26.97 s, much faster than keyboard-teleoperation baselines of 117 s and 469 s. If the result holds, it suggests that deep-learning-driven visual servoing plus OCT depth sensing could turn retinal vein cannulation from a tremor-limited manual skill into a repeatable automated procedure.","feed_headline":"Robot autonomously cannulates retinal veins in embryos","feed_subtitle":"OCT-guided deep learning cuts navigation and puncture time while keeping 85 percent detection accuracy.","key_machinery":"The load-bearing mechanism is the pairing of B-scan OCT depth information with three trained vision models. A keypoint-detection network on microscope images locates the needle tip and computes a velocity command toward the target. A binary contact classifier on B-scan cross-sections decides whether the needle has reached the vein wall, and if not, the robot increments downward. A puncture detector on B-scans labels each attempt as success or failure and triggers a retract-and-retry loop when the puncture is not confirmed; the retry retracts by two-fifths of the puncture depth before reattempting. This closed loop of optical sensing, learned classification, and robot motion is what makes the procedure autonomous once the surgeon selects the target.","core_discovery":"The central discovery is that the full retinal vein cannulation sequence can be executed autonomously by coupling two image sources with three learned decisions. A microscope image gives the lateral needle-tip position for navigation; a B-scan OCT cross-section gives the depth information needed to judge when the needle touches the vein wall and when the wall has been punctured. The system first drives the needle to a user-selected 2D target, then lowers it until a binary contact classifier confirms contact, then advances at higher speed and uses a puncture detector to confirm success, retracting and retrying when the puncture is not confirmed. On 20 chicken embryos it classified puncture outcomes with 85% accuracy, with an F1 score of 0.87 for successful punctures, and completed navigation and puncture in 36.74 s and 26.97 s on average.","pith_inferences":["If the 85% accuracy transfers to smaller vessels, adding a confidence threshold before puncturing could reduce the misclassified failure cases the model currently produces.","The speed comparison is against keyboard teleoperation, not against a surgeon's hands; a fair clinical comparison would include setup and decision time and a tremor model.","Because chicken embryo veins are roughly an order of magnitude wider than human retinal veins, the contact and puncture classifiers would likely need retraining on smaller-caliber vessels before human use, a transfer the paper lists as future work."],"forward_implications":["An operator's only input is a 2D target on the microscope image; the robot then handles navigation, contact, puncture, and verification without manual intervention.","Autonomous navigation averaged 36.74 s compared with 117 s for keyboard teleoperation, a 68.6% reduction.","Autonomous puncture averaged 26.97 s compared with 469 s for keyboard teleoperation, roughly a 17-fold reduction.","Puncture classification reached 85% accuracy, with precision 0.90/0.82 and recall 0.75/0.93 for failure/success classes.","Failures came from lighting variation, tissue motion, and needle occlusion degrading B-scan contrast, so the paper identifies robustness to these conditions as necessary next work."],"supporting_citations":[{"why":"Describes the needle-manipulation robot used to perform autonomous navigation and puncture.","marker":"[7]"},{"why":"Describes the second robot that stabilizes the target vein with a silicone holder.","marker":"[25]"},{"why":"Establishes the chicken embryo chorioallantoic membrane as a model for comparing manual and robot-assisted retinal vein cannulation.","marker":"[22]"},{"why":"Shows the prior teleoperated robotic retinal vein cannulation workflow on porcine eyes that this work extends to full autonomy.","marker":"[21]"},{"why":"Demonstrates autonomous needle navigation in subretinal injections using intraoperative OCT, the depth-sensing approach the system relies on.","marker":"[17]"},{"why":"Reports the 73% success rate for robot-assisted retinal vein cannulation in a porcine model that this paper uses as a comparison baseline.","marker":"[6]"}],"fun_headline_variants":["Autonomous robot cannulates retinal veins in embryos","Deep learning robot punctures retinal veins autonomously","OCT-guided AI robot automates retinal vein cannulation","Robot uses AI to cannulate retinal veins in embryos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system assumes the B-scan OCT cross-section shows the actual needle-to-vein interaction; if the scan plane drifts or lighting and tissue motion hide the needle, the contact and puncture classifiers are deciding on images that do not show what they are supposed to classify.","fun_headline_variants_meta":{"raw":{"variants":["Autonomous robot cannulates retinal veins in embryos","Deep learning robot punctures retinal veins autonomously","OCT-guided AI robot automates retinal vein cannulation","Robot uses AI to cannulate retinal veins in embryos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1266,"prompt_tokens":889,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":505,"tokens_out":377,"duration_ms":4442,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:10:11.749759+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deliberately offset the OCT B-scan plane from the needle tip by about one needle diameter during autonomous punctures and measure puncture-detection accuracy; if accuracy does not collapse, the claim that B-scan depth information drives the decision is not supported. A second check is to compare the robot's commanded depth against the OCT-measured needle-tip depth across trials, a quantity the paper does not report.","supporting_citations":[{"cited_title":"New steady-hand eye robot with micro-force sensing for vitreoretinal surgery,","cited_arxiv_id":null,"evidence_quote":"Describes the needle-manipulation robot used to perform autonomous navigation and puncture."},{"cited_title":"Toward clinically applicable steady-hand eye robot for vitreoretinal surgery,","cited_arxiv_id":null,"evidence_quote":"Describes the second robot that stabilizes the target vein with a silicone holder."},{"cited_title":"A comparison of manual and robot-assisted retinal vein cannulation in chicken chorioallantoic membrane,","cited_arxiv_id":null,"evidence_quote":"Establishes the chicken embryo chorioallantoic membrane as a model for comparing manual and robot-assisted retinal vein cannulation."},{"cited_title":"A feasible workflow for retinal vein cannulation in ex vivo porcine eyes with robotic assistance,","cited_arxiv_id":null,"evidence_quote":"Shows the prior teleoperated robotic retinal vein cannulation workflow on porcine eyes that this work extends to full autonomy."},{"cited_title":"Autonomous needle navigation in subretinal injections via ioct,","cited_arxiv_id":null,"evidence_quote":"Demonstrates autonomous needle navigation in subretinal injections using intraoperative OCT, the depth-sensing approach the system relies on."},{"cited_title":"Robot-assisted retinal vein cannulation in an in vivo porcine retinal vein occlusion model,","cited_arxiv_id":null,"evidence_quote":"Reports the 73% success rate for robot-assisted retinal vein cannulation in a porcine model that this paper uses as a comparison baseline."}],"review_version":1}