{"id":"0328cf34-d8ca-4aa7-b3e7-ed8a12382914","arxiv_id":"2607.19174","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A Bayesian sequential mapping method using learned classifiers and optimized retractions localizes tissue attachment points for robotic dissection, outperforming random retractions in simulation.","lead":"This paper introduces a Bayesian active-sensing method that maps where soft tissue is attached to underlying surfaces during robotic dissection, choosing where to retract next to reduce uncertainty. It reports improved attachment localization over random retraction in simulation and shows four real ex vivo chicken-breast dissections on a da Vinci robot.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BRO's advantage over random retraction may be an artifact of unverified goal-shape heuristic: selected q* may not be exposed to camera.","rationale":"The reader's weakest_assumption identified the q* visibility issue as the first load-bearing concern; my analysis agrees and sharpens it. The central quantitative claim is that BRO outperforms random retraction. This claim rests on the assumption that the acquisition function selects a point whose observation will reduce uncertainty. If the vertical-plane/DeformerNet pipeline does not reliably expose q*, the entire loop degenerates into random retraction with a possibly confounded selection signal. The paper provides no verification—either in simulation or real experiments—that q* is visible after retraction. The concrete test would settle whether the mechanism is real. I do not see a need to change the reader's CONDITIONAL verdict; the conditions the reader lists should include this visibility verification. The threshold tuning issue is real but secondary, because it affects only the zero-shot claim, not the core simulation comparison. Overall, the paper is promising but the BRO advantage should be considered unverified until the visibility check is performed.","tokens_in":10615,"tokens_out":2165,"duration_ms":25822,"concrete_test":"Re-run the BOX/CYLINDER/ELLIPSOID dissection experiments (30 trials each) while logging, after each DeformerNet retraction, whether q* is visible in the rendered depth image (e.g., raycast against the deformed mesh). Compute the AUPRC gain of BRO over random restricted to trials where q* is visible vs. not. If the gain disappears or reverses on visible trials, BRO's advantage is not caused by exposing q*. As a stronger test, replace the DeformerNet planning with an oracle that retracts to make q* visible by construction (e.g., lift the exact sub-region around q* along the camera axis); if BRO no longer beats random, the acquisition function's information gain is not the driver.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central simulation claim is that BRO beats random q* selection. But the mechanism that makes BRO informative is that the selected q* becomes visible after retraction, so the classifier observes the true state at/around q*. Algorithm 1 calls 'Robotic retraction to reveal q*' and Section IV-E solves Eq. (7) by random sampling of vertical plane parameters, yet no check anywhere confirms that the executed retraction exposes q* to the camera. In fact, the objective in Eq. (7) penalizes lifting attached tissue; it does not encode visibility of q* except through the heuristic half-plane rotation. If DeformerNet often fails to achieve the goal shape, or if the vertical-plane heuristic does not produce a fold that reveals the query point, then q* is just a random index into the map, and BRO's acquisition function (EI/nEI/UCB) has no causal link to the data actually collected. The reported AUPRC advantage over 'no acquisition' could then be an artifact of unintended biases (e.g., q* often near the lifted edge) rather than information-driven selection. Since the baseline 'no acquisition' selects uniformly at random from Q, any correlation between the q* heuristic and visibility could also produce spurious gains. A secondary issue is the incision threshold adjustment (0.9→0.83, 50%→41%) tuned on real-world point clouds, which undermines the 'zero-shot' claim, but the simulation claim is more load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an active-sensing framework for identifying tissue attachment points during robotic dissection. A Sequential Bayesian Hilbert Map (SBHM) represents a distribution over attachment labels at query points on the resection surface. At each iteration, an ensemble of PointNet classifiers produces noisy, partially occluded labels from a partial-view point cloud after tissue retraction; these labels update the SBHM via variational Bayesian logistic regression. The next retraction is selected by Bayesian Retraction Optimization (BRO) using EI/nEI/UCB acquisition functions, with the retraction planned by DeformerNet toward a vertical-plane goal shape chosen to expose the selected query point. Incisions are made where the SBHM predictive probability exceeds thresholds. Experiments in Isaac Gym on BOX/CYLINDER/ELLIPSOID geometries compare BRO with random acquisition and show AUPRC gains; a noise experiment indicates SBHM updates improve over single classifier outputs. Four dVRK chicken-breast trials are presented as zero-shot sim-to-real transfer.","tokens_in":10949,"tokens_out":6306,"duration_ms":69804,"significance":"The strength of the work is a clean probabilistic formulation: uncertainty over attachment labels is propagated online without an explicit deformable tissue model, using a standard variational approximation (Eqs. 5-6) and a Monte-Carlo predictive covariance. If the central active-sensing advantage is established, the framework would be a useful step toward automating dissection, and the use of an ensemble as a noisy sensor for a Bayesian map is a sensible way to handle out-of-distribution classifier outputs. The manuscript also makes falsifiable predictions (BRO beats random acquisition; SBHM noise robustness) that are testable. However, several load-bearing validations are missing; the most serious is that the planned retraction is never checked to actually expose the selected query point, so the causal mechanism of BRO is unverified.","major_comments":[{"comment":"The load-bearing claim is that BRO's selected q* 'reveals' informative evidence. Algorithm 1 calls 'Robotic retraction to reveal q*', but Eq. (7) minimizes the expected attachment probability over the lifted half-plane and contains no visibility or occlusion term; it only enforces q* lies in the half-plane that is rotated upward. The camera is fixed, and DeformerNet may fail to achieve the goal shape. The paper never reports the fraction of simulated retractions in which q* is actually visible (unoccluded) in the post-retraction point cloud, nor compares BRO vs random under visibility filtering. Without this, the AUPRC advantage in Fig. 5 may be an artifact of correlation between the heuristic and occlusion rather than information-driven acquisition. This should be measured and, if necessary, the planning objective revised.","section":"§IV-E, Algorithm 1, Eq. (7)"},{"comment":"The abstract and introduction claim the method 'outperforming two competitive baselines in simulation.' The only policy baseline in §V-C is 'no acquisition' (uniform random q*). Fig. 6 compares SBHM versus a single noisy classifier, which is an ablation of the Bayesian update, not a retraction baseline. There is no comparison to a hand-crafted retraction policy or simulation-based active inference baseline. The claim should be corrected to one baseline, or a second baseline should be added.","section":"Abstract; §V-C, Fig. 5"},{"comment":"'Zero-shot sim-to-real transfer' is not supported as stated. §IV-D states that for real-world experiments the incision confidence threshold is lowered from 0.9 to 0.83 and the neighborhood rule relaxed from 50% to 41%, and that 'these hyperparameters are chosen manually based on real-world point clouds.' This is adaptation using target-domain data, not zero-shot. Additionally, §VI reports only four trials with qualitative figures and no quantitative localization error or uncertainty estimates. Please either present a held-out protocol, report quantitative metrics, or weaken the claim to 'threshold-adapted transfer.'","section":"§IV-D, §VI"},{"comment":"The text states BRO 'significantly outperforms' the random baseline, but Fig. 5 shows only mean AUPRC curves over 30 trials with no error bars, confidence intervals, or paired significance tests across the EI/nEI/UCB variants. Given the small trial count and high variance in attachments, the significance claim needs explicit statistical support.","section":"§V-C, Fig. 5"}],"minor_comments":[{"comment":"The denominator is written as (1 + π(φ^T S φ)^2 / 8)^{1/2}, which mismatches the standard probit approximation σ((m^Tφ)/(1+(π/8) φ^T S φ)^{1/2}). Please check the formula and use consistent notation.","section":"Eq. (2)"},{"comment":"The loop condition 'tissue attached(Q)' is never defined; it should be specified (e.g., based on ground truth in simulation, or on SBHM confidence in deployment).","section":"Algorithm 1"},{"comment":"Each non-occluded query point contributes one label per ensemble member; these correlated labels are treated as independent observations in the VB update. This may overcount ensemble agreement; consider weighting or a single fused label.","section":"§IV-C, Eq. (4)"},{"comment":"Attached-class precision/recall of 0.343/0.584 is quite low; it would be useful to report ensemble disagreement and calibration, since they drive the noise model.","section":"§V-A"}],"recommendation":"major_revision","confidential_remarks":"The core idea is publishable after revision, but the visibility verification is essential. The claimed 'two competitive baselines' and 'zero-shot transfer' should be corrected or supported. I would also encourage the editor to ask for code/data to check reproducibility, as the simulation pipeline is central to the paper's claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: useful simulation result and a reasonable probabilistic framing, but the abstract overclaims and the key mechanism — whether the chosen retraction actually reveals the point the acquisition function wants — is never checked.\n\nThe paper builds a Sequential Bayesian Hilbert Map over candidate attachment points, feeds it noisy labels from an ensemble of PointNet classifiers, and uses Bayesian optimization to select which retraction to execute next. That integration is genuinely new. Prior work either used hand-crafted incision policies or differentiable simulators; nobody had treated attachment mapping as a probabilistic active-sensing problem with a classifier-driven map.\n\nThe simulation evidence is the strongest part. BRO beats random retraction across three tissue geometries and three acquisition functions, and the SBHM significantly improves over a single noisy classifier. Both results look internally consistent, and the variational Bayesian logistic regression is standard and correctly derived. The paper is also honest in places: it flags the planar-surface limitation, the noise in real point clouds, and the fact that the real-world thresholds were set by hand.\n\nThe soft spots are real but not fatal. First, the abstract claims the method outperforms \"two competitive baselines\" when only random selection is compared. That is a straight overstatement. Second, the real-world \"demonstration\" is four qualitative trials, with thresholds tuned on the real clouds, so \"zero-shot\" is a stretch. Third, and my main concern: the paper never checks that the selected acquisition point q* is actually visible after the retraction. The vertical-plane heuristic can fail, and if it does, the acquisition function has no causal link to the observation. I don't think this sinks the simulation result — the same retraction procedure is used for both BRO and random, so the difference in behavior likely still comes from the retraction geometry — but the paper should verify visibility or at least measure the correlation between acquisition score and information gained.\n\nWho gets value from this: someone in surgical robotics or active perception who wants a concrete example of closing the loop between probabilistic mapping and action selection. It deserves a serious referee, but I would send it back for revisions: fix the overclaiming, add error bars and statistical tests to the simulation, and close the visibility loop. Then it would be a solid paper.","headline":"A useful integration of Bayesian mapping and active sensing for surgical dissection, but the abstract overclaims and the paper never verifies that the chosen retraction actually reveals the selected query point.","tokens_in":11452,"tokens_out":3504,"would_cite":false,"duration_ms":42426,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that treating tissue dissection as an active-sensing problem—maintaining a Bayesian map of attachment likelihoods and choosing retractions by expected information gain—can localize attachment points without deformable tissu","keywords":["tissue attachment mapping","surgical dissection","Sequential Bayesian Hilbert Map","Bayesian retraction optimization","active sensing","classifier ensemble","zero-shot transfer","information gain"],"falsifier":"Record the camera-view point cloud immediately after each retraction and check whether the selected acquisition point is visible and whether the tissue was lifted as intended. If retractions routinely fail to expose the chosen point, BRO's information-gain objective is disconnected from the observation and the reported advantage over random retraction should vanish.","tokens_in":10458,"feed_emoji":"🔪","tokens_out":5651,"duration_ms":54210,"temperature":0.7,"pith_summary":"The paper is trying to establish that a robot can find where soft tissue is attached to underlying anatomy by actively pulling on it and watching, without needing a physical model of tissue deformation. It represents attachment state as a probability map, updated by an ensemble of learned classifiers that label each candidate point as attached, not attached, or occluded from each retraction view. A Bayesian optimization step chooses the next retraction specifically to reduce uncertainty about attachment locations, and once the map is confident, the robot cuts the likely attachment points. If correct, this would replace hand-crafted incision policies and simulation-heavy approaches with a probabilistic loop that improves with each manipulation. The authors report that the approach outperforms random retraction in simulation and identifies attachment disks correctly in real ex vivo chicken breast tissue after simulation-only training.","feed_headline":"Bayesian active sensing maps tissue attachments without a tissue model","feed_subtitle":"Retractions chosen to maximize information gain localize attachment points and outperform random retraction, in simulation and on real tissu","key_machinery":"The central object is the Sequential Bayesian Hilbert Map (SBHM), a probabilistic map that represents each candidate attachment point's binary state through a sigmoid function of a Gaussian-distributed weight vector; the Gaussian mean and covariance encode the current belief and enable principled fusion of noisy classifier evidence across retraction views. The action-selection mechanism is Bayesian Retraction Optimization (BRO), which uses acquisition functions—Expected Improvement, noisy Expected Improvement, or Upper Confidence Bound—computed from the map's predictive mean and covariance to choose which query point to expose to the camera. The retraction itself is generated by a learned sh","core_discovery":"The paper's central claim is that tissue attachment identification during dissection can be framed as a probabilistic active-sensing problem, not a tissue-modeling problem. A Sequential Bayesian Hilbert Map (SBHM) maintains a continuous belief over candidate attachment points, and after each retraction an ensemble of learned classifiers provides noisy labels—attached, not attached, or occluded—that update the map through variational Bayesian logistic regression. A Bayesian Retraction Optimization (BRO) step then picks the next retraction by maximizing an acquisition function over the map's predictive distribution, subject to a safety heuristic that avoids lifting strongly attached tissue. In","pith_inferences":["A natural extension the authors do not pursue: the same BRO information-gain objective could be applied to other surgical subtasks where manipulation exposes hidden structures, such as tumor boundary localization or freeing adhesions.","The paper's vertical-plane goal-shape heuristic assumes a planar attachment surface; if goal-shape generation were generalized to curved anatomy, the identical SBHM plus BRO loop should in principle apply without changing the inference machinery.","Because the acquisition function is defined over the map's uncertainty, BRO could be made adaptive to real-world noise by treating classifier quality itself as an estimated quantity, rather than depending on hand-tuned incision thresholds.","A cheap testable extension would be to evaluate whether the chosen acquisition point is actually visible after each retraction and to feed that visibility back into the acquisition function as a constraint."],"forward_implications":["If the central claim holds, dissection robots can localize attachment points without building a deformable tissue model, removing a major sim-to-real bottleneck.","Choosing retractions by expected information gain yields faster, more accurate attachment maps than random retraction in simulation, suggesting active sensing is worth the planning cost.","Sequential Bayesian fusion over classifier outputs keeps attachment estimates reliable even when individual predictions are noisy or out-of-distribution, which is the condition encountered in real tissue.","The confidence-driven incision rule closes the loop by removing attachment regions once the map is sufficiently certain, without requiring a separate motion planner for cutting.","Zero-shot transfer on real ex vivo tissue suggests the simulation-trained classifier ensemble plus Bayesian filtering can generalize beyond the training distribution, at least for planar attachment surfaces."],"fun_headline_variants":["Bayesian active sensing locates tissue attachments without a model","Probabilistic dissection: Bayesian map guides retraction to find attachment points","Bayesian retraction optimization targets tissue attachments with noisy labels","No tissue model? Bayesian active sensing pinpoints attachment points"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The approach assumes that after a planned retraction, the chosen acquisition point is actually visible to the camera; if the retraction does not reveal it, the information-gain calculation has no connection to what the robot observes.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian active sensing locates tissue attachments without a model","Probabilistic dissection: Bayesian map guides retraction to find attachment points","Bayesian retraction optimization targets tissue attachments with noisy labels","No tissue model? Bayesian active sensing pinpoints attachment points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3291,"prompt_tokens":691,"completion_tokens":2600,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":2539}},"tokens_in":435,"tokens_out":2600,"duration_ms":16632,"temperature":1.0,"reasoning_tokens":2539,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T13:12:49.716664+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the camera-view point cloud immediately after each retraction and check whether the selected acquisition point is visible and whether the tissue was lifted as intended. If retractions routinely fail to expose the chosen point, BRO's information-gain objective is disconnected from the observation and the reported advantage over random retraction should vanish.","supporting_citations":[],"review_version":1}