{"id":"c5ce5807-b91a-4020-a90b-00ef5d6ed958","arxiv_id":"2411.18884","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A regression network trained on a new robotic ESD dataset predicts dissection trajectories and safety-margin confidence maps with a reported MAE of 3.18.","lead":"Researchers built a dataset of robotic endoscopic dissection videos and trained a model to suggest a safe cutting path and to show a confidence map of safe versus risky tissue areas. The goal is to reduce mistakes and injuries in a high-skill gastrointestinal procedure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ground-truth confidence map is not well-defined: Eq. (1) yields infinite values at the trajectory and 0 at the margin, contradicting the stated 1/0 labeling, so the reported MAE of 3.18 is not interpretable from the paper.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue: the ground-truth confidence map is not fully specified. My reading strengthens that concern: Eq. (1) is not merely missing a clamp or normalization; as written it contradicts the stated endpoint labels, giving an undefined value at the trajectory and 0 at the margin. This is a correctness risk in the central claim, not a stylistic or consensus disagreement. It does not necessarily invalidate the work if the released code implements a corrected formula, because the authors promise dataset and code availability. The other quantitative results, such as trajectory suggestion (BC ADE 9.10), are less central and rest on more standard metrics. Therefore the appropriate verdict remains conditional, pending code inspection of the label generation; no new verdict adjustment is needed beyond the reader's conditional acceptance.","tokens_in":10401,"tokens_out":4311,"duration_ms":39125,"concrete_test":"Check the released dataset-generation code (promised at github.com/FrankMOWJ/RCMNet) on a few annotated frames: verify (1) the empirical minimum and maximum of generated confidence maps, (2) label values at annotated trajectory and margin pixels, and (3) whether the implemented formula matches Eq. (1) or uses a normalized denominator such as d_margin + d_trajectory. If the maps are not in [0,1] with trajectory values equal to 1 and margin values equal to 0, recompute Table I after applying the intended normalization and see whether the MAE ranking and the 3.18 headline survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim -- RCMNet's MAE of 3.18 on confidence map-based safety margin prediction -- depends entirely on the definition of the ground-truth confidence maps. In Section II-A, Eq. (1) defines C = distance(area point, margin point) / distance(area point, trajectory point). For a point on the dissection trajectory, the denominator is zero, so C is undefined/infinite; at the margin, C = 0. Yet the preceding text states that the trajectory is assigned confidence 1 and the safety margin confidence 0. Thus Eq. (1) as written produces labels whose endpoint behavior is opposite to the stated design and whose range is unbounded near the trajectory. The 'distance threshold' mentioned for curved dissection areas only restricts which margin point is selected; it does not clamp or normalize C to [0,1]. The weighted MSE loss in Eq. (2) and all MAE/MSE numbers in Table I are computed against these labels. If the implementation uses a different formula, such as d_margin/(d_margin + d_trajectory), that formula must be stated and verified; as published, the paper does not establish that the confidence maps are a well-defined safety field, making MAE 3.18 uninterpretable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces ETSM, a dataset of 1,849 short clips from ex-vivo porcine robotic endoscopic submucosal dissection (ESD), with annotations of dissection trajectories and safety margins, along with an angular-difference algorithm to generate confidence-map safety margins. The authors propose RCMNet, a DINOv2-encoder plus All-MLP-decoder regression network, to predict these confidence maps, and evaluate trajectory suggestion using behavior cloning (BC), motion indeterminacy diffusion (MID), and implicit diffusion policy imitation learning (iDiff-IL). The central quantitative claims are that BC achieves ADE 9.10 for trajectory suggestion and that RCMNet achieves MAE 3.18 for confidence-map prediction at 532x532 resolution, outperforming MAN and SASNet in in-domain and robustness evaluations. The out-of-domain evaluation is qualitative, showing predictions on live porcine and patient data.","tokens_in":10756,"tokens_out":3830,"duration_ms":35562,"significance":"If the confidence-map formulation is made well-defined and reproducible, the paper would contribute a useful dataset and a novel regression-based approach for safety-margin visualization in robot-assisted ESD. Strengths include the release of dataset and code, the use of a video-level train/test split, the inclusion of robustness evaluation under several corruption types, and the first application of a regression decoder with a pre-trained DINOv2 encoder to this task. The trajectory-suggestion comparison with three baselines is a reasonable contribution. However, the central quantitative claim—RCMNet's MAE of 3.18—depends entirely on the ground-truth confidence map definition, which as written is internally inconsistent and not a proper [0,1] confidence field. The concern raised in the stress-test note directly lands on this manuscript: Eq. (1) does not produce bounded confidence values and the text itself acknowledges that values exceed 1. This must be resolved before the reported numbers can be interpreted.","major_comments":[{"comment":"The definition of the ground-truth confidence map is not well-defined. The text assigns confidence 1 to the dissection trajectory and 0 to the safety margin and its exterior, but Eq. (1) computes C = d_margin / d_trajectory. At a point on the trajectory, the denominator is zero, making C undefined or infinite; at the margin, C = 0. The sentence 'the confidence exceeds 1' in the following paragraph confirms that values are not clamped to [0,1]. Since the weighted MSE loss in Eq. (2) and all MAE/MSE values in Table I are computed against these labels, the reported MAE of 3.18 is not interpretable as an error on a confidence map. Please state the exact label-generation formula actually used (e.g., d_margin / (d_margin + d_trajectory)), or describe and justify a clamping/normalization step, and regenerate all reported results with the corrected formulation.","section":"Section II-A, Eq. (1)"},{"comment":"The 'distance threshold between the margin point and the area point' is introduced to handle curved dissection areas, but its value is never specified. Because this threshold directly affects which margin point is selected and hence the resulting confidence values, every number in Table I depends on this unspecified hyperparameter. This makes the algorithm non-reproducible. Please provide the threshold value, the units (pixels or millimeters), and the exact selection rule (e.g., the margin point with the smallest angular difference among those within T pixels).","section":"Section II-A, distance threshold"},{"comment":"The evaluation reports single MAE/MSE values per model with no standard deviations, confidence intervals, or repeated-seed experiments. The test set consists of 369 clips from four videos, so the observed difference between RCMNet (MAE 3.18) and SASNet (MAE 4.19) could be within run-to-run variability. Please report results over multiple seeds or provide a statistical test, and specify the numerical range of the confidence maps used in the metrics (e.g., 0–1 or 0–255) so that the MAE values are interpretable.","section":"Section III-C-2, Table I"}],"minor_comments":[{"comment":"The word 'Intermediant' should be 'Intermediate'.","section":"Section II-B-2-a"},{"comment":"The decoder name appears inconsistently as 'All-MLP' and 'ALL-MLP'; please unify the terminology.","section":"Section II-B-2-b"},{"comment":"The phrase 'weight Mean Square Error' should be 'weighted Mean Square Error'.","section":"Section II-B-2-c"},{"comment":"The number '1, 849' contains an odd space; write '1,849'.","section":"Section III-A"},{"comment":"The table header 'Constrast' should be 'Contrast'.","section":"Table I"},{"comment":"The baselines MID and iDiff-IL are compared with BC, but the manuscript does not describe how these models were adapted to the ESD task or what hyperparameters were used; a brief implementation note would improve reproducibility.","section":"Section III-C-1"},{"comment":"The predicted and ground-truth confidence maps are shown with color gradients but no colorbar or value labels, making it difficult for the reader to judge quantitative agreement; adding colorbars would improve interpretability.","section":"Figures 5(b) and 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially interesting for the surgical robotics and medical image analysis community, and the public release of dataset and code is a positive step. However, the core label-generation formula must be corrected and fully specified before the quantitative claims can be trusted. I would recommend the editor require the authors to either state and validate the actual formula used or reproduce the experiments with a properly normalized confidence map (e.g., d_margin/(d_margin+d_trajectory)). The lack of error bars is a secondary concern but should also be addressed. The out-of-domain evaluation being qualitative is acceptable for a first study, but the in-domain numbers are the main claim and they currently rest on an undefined label definition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: the dataset is the real contribution; the confidence-map label definition is not. The stress-test note lands. Eq. (1) defines C as distance-to-margin divided by distance-to-trajectory, which is infinite on the trajectory and zero on the margin — the opposite of the stated 1/0 design. No clamping or normalization is described, and the distance threshold for curved areas does not fix the endpoint behavior. That makes the headline MAE of 3.18 uninterpretable as published. This is not a cosmetic issue: every confidence-map experiment is graded against those labels.\n\nWhat deserves credit: 1849 clips from robotic ESD with trajectory and safety-margin annotations are genuinely new and useful. The angular-difference algorithm for picking the margin point is a sensible fix to the opposite-side nearest-neighbor error. RCMNet is straightforward engineering — DINOv2 plus LoRA plus an All-MLP decoder — but the comparison against MAN and SASNet, including corruption robustness, is a reasonable evaluation for a task paper. The trajectory suggestion part is honest: behavior cloning beats the diffusion baselines, and the paper does not oversell it.\n\nSoft spots beyond Eq. (1): Table I has no error bars or repeated runs, so the 3.18 versus 3.46 gaps may be noise. The out-of-domain evidence is qualitative only. The ETSM labels come from expert annotation converted through the authors' own algorithm, so the benchmark is self-contained; that is acceptable for a dataset paper, but it does not clinically validate the safety field. The 0.4 mm offset and the distance threshold are described loosely and act as free parameters.\n\nWho this is for: people building intraoperative visual guidance and autonomous ESD tools; the dataset alone could seed follow-up work. I would send this to peer review, but the reviewers should demand a corrected Eq. (1) or released code that reproduces the labels, error bars, and at least one quantitative out-of-domain number. If the label definition is fixed, this is a solid dataset-plus-baseline contribution.","headline":"The ETSM dataset is the real contribution, but the confidence-map label definition is broken as written and the headline MAE is not interpretable until it is fixed.","tokens_in":11220,"tokens_out":4421,"would_cite":false,"duration_ms":41770,"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":"A regression network predicts safety margins for robotic endoscopic dissection with a mean absolute error of 3.18, enabling intraoperative visual guidance.","keywords":["endoscopic submucosal dissection","robot-assisted surgery","trajectory prediction","safety margin","confidence map","regression network","DINOv2","behavior cloning"],"falsifier":"Recompute the ETSM ground-truth labels by explicitly clamping or normalizing the Eq. (1) ratio to $[0,1]$, or by specifying the missing distance threshold, and re-run the RCMNet versus MAN/SASNet comparison. If the reported MAE of 3.18 at $532\\times532$ shifts materially, or if the relative ordering of the models changes, the paper's central quantitative claim is not stable under the correct label definition. An even more direct check is to calculate the fraction of labeled pixels whose raw confidence exceeds 1 in the released dataset; a non-negligible fraction would confirm that the ground truth is not a proper confidence field.","tokens_in":10268,"feed_emoji":"🩺","tokens_out":11059,"duration_ms":82864,"temperature":0.7,"pith_summary":"This paper aims to make robot-assisted endoscopic submucosal dissection (ESD) safer by giving the surgeon, or an autonomous system, two kinds of intraoperative guidance: a suggested dissection trajectory and a confidence map that grades the dissection area by safety margin. To do this, the authors create the ETSM dataset of 1,849 short clips from dual-arm robotic ESD on ex-vivo porcine models, with annotations for optimal trajectories and safety margins. They define a confidence value at each point in the dissection area as the ratio of its distance to the safety margin over its distance to the optimal trajectory, generated by an angular-difference-based algorithm that avoids picking the wrong side of a curved margin. On this task, their proposed regression network RCMNet, which combines a DINOv2 encoder fine-tuned with LoRA and an All-MLP decoder, reaches a mean absolute error of 3.18 at $532\\times532$ resolution, lower than the crowd-counting-derived baselines MAN and SASNet. For trajectory suggestion, behavior cloning outperforms two diffusion-based methods, and the paper frames the regression formulation as a better fit for safety-margin delineation than segmentation.","feed_headline":"Safety-margin predictor for robotic endoscopy hits MAE 3.18","feed_subtitle":"Confidence-field regression maps safe dissection zones, beating segmentation baselines.","key_machinery":"The load-bearing machinery is the confidence-map generation rule of Eq. (1): for each point in the dissection area, the confidence is the ratio of its distance to the nearest safety-margin point over its distance to the nearest trajectory point, $\\frac{\\sqrt{(x_e-x_a)^2+(y_e-y_a)^2}}{\\sqrt{(x_t-x_a)^2+(y_t-y_a)^2}}$. Because a naive nearest-margin search can pick the wrong side of a curved margin, the paper adds an angular-difference search: find the nearest trajectory point, cast a ray from it through the area point, then select the margin point whose direction has the smallest angular difference from that ray; a distance threshold is introduced for curved cases. The second pillar is RCMNet, which uses a pre-trained DINOv2 ViT-B/14 encoder with LoRA fine-tuning, extracts multi-scale features from transformer blocks 3, 6, 9, and 12, and decodes them with an All-MLP network (channel unification, upsampling, concatenation, fusion MLP, and a 1-channel prediction head). Training uses a weighted MSE loss with weight 10 on pixels outside the safety margin, so the model learns to output zero confidence there. The trajectory-suggestion comparisons are carried out with behavior cloning, MID, and iDiff-IL.","core_discovery":"The paper's central claim is that the safety margin of an endoscopic submucosal dissection area can be represented as a continuous confidence field — value 1 on the optimal dissection trajectory, 0 on and outside the safety margin, smoothly decreasing in between — and that this field can be predicted directly by a regression network. On the ETSM test set, their RCMNet reaches a mean absolute error of 3.18 and a mean squared error of 344.40 at $532\\times532$ resolution, beating the crowd-counting baselines MAN (MAE 3.68) and SASNet (MAE 4.19), and it also attains the lowest mean MAE under the four corruption categories. Qualitatively, the predicted maps track the gradient from trajectory to margin even in difficult cases, and the method is also applied to out-of-domain live porcine and patient frames. In the trajectory-suggestion half of the framework, behavior cloning with a CNN-MLP gives average displacement error 9.10, final displacement error 11.20, and Fréchet distance 20.14, outperforming MID and iDiff-IL. The authors frame these results as the first regression-based approach for delineating varying safety levels of dissection areas.","pith_inferences":["Because the confidence value is a dimensionless pixel-distance ratio, converting it to physical units by calibrating pixel size from the endoscope's optics would make the reported MAE interpretable as a true margin error in millimeters; the paper does not perform that calibration.","The current confidence module uses a single input frame while the trajectory module uses three frames; fusing temporal information into the confidence branch is a natural extension the paper lists as future work and is likely to reduce errors during tool-induced occlusions.","The dataset comprises 21 ex-vivo porcine procedures with annotations by expert endoscopists; a multi-expert study measuring inter-annotator agreement on trajectory and margin placement would clarify how much of the 3.18 MAE is annotation noise rather than model error."],"forward_implications":["RCMNet's predicted confidence maps can be overlaid on the endoscopic view in real time, giving the surgeon a graded 'safe–caution–danger' display during dissection.","The angular-difference-based confidence generation handles curved dissection areas by avoiding opposite-margin misassignment, a property the authors demonstrate in Fig. 2.","The regression formulation for safety-margin prediction outperforms the segmentation formulation, since segmentation methods achieve only roughly 0.45 IoU and are deemed unreliable for this task.","Behavior cloning's trajectory suggestions (ADE 9.10) are accurate enough to serve as a starting point for automated dissection subtasks, though full autonomy is not claimed."],"supporting_citations":[{"why":"Supplies the pre-trained DINOv2 encoder that RCMNet fine-tunes with LoRA.","marker":"[12]"},{"why":"Inspires the All-MLP decoder design used in RCMNet.","marker":"[14]"},{"why":"Crowd-counting baseline (MAN) that RCMNet must beat on confidence-map regression.","marker":"[23]"},{"why":"Second crowd-counting baseline (SASNet) used for in-domain and robustness comparisons.","marker":"[24]"},{"why":"State-of-the-art dissection trajectory prediction method (iDiff-IL) used as a baseline in the trajectory-suggestion task.","marker":"[7]"},{"why":"Diffusion-based trajectory prediction baseline (MID) used in the trajectory-suggestion comparison.","marker":"[8]"},{"why":"Provides the behavior-cloning method that achieves the best trajectory-suggestion results.","marker":"[11]"},{"why":"Describes the dual-arm robotic system and procedure setting from which the ETSM videos are recorded.","marker":"[9]"},{"why":"Defines the corruption taxonomy and severity levels used for the robustness evaluation.","marker":"[25]"}],"fun_headline_variants":["Regression predicts robotic ESD safety margins with 3.18 MAE","First regression model for ESD safety zones hits MAE 3.18","RCMNet maps dissection safety margins via regression, MAE 3.18","Robotic ESD safety-map regression achieves MAE 3.18"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ground-truth confidence maps are a well-defined field whose values lie in $[0,1]$; the paper defines confidence as a distance ratio that can exceed 1 and introduces a distance threshold for curved cases without stating how the ratio is clamped or normalized.","fun_headline_variants_meta":{"raw":{"variants":["Regression predicts robotic ESD safety margins with 3.18 MAE","First regression model for ESD safety zones hits MAE 3.18","RCMNet maps dissection safety margins via regression, MAE 3.18","Robotic ESD safety-map regression achieves MAE 3.18"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000658,"raw_usage":{"total_tokens":3085,"prompt_tokens":1093,"completion_tokens":1992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":709,"completion_tokens_details":{"reasoning_tokens":1910}},"tokens_in":709,"tokens_out":1992,"duration_ms":12785,"temperature":1.0,"reasoning_tokens":1910,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:47:15.889630+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the ETSM ground-truth labels by explicitly clamping or normalizing the Eq. (1) ratio to $[0,1]$, or by specifying the missing distance threshold, and re-run the RCMNet versus MAN/SASNet comparison. If the reported MAE of 3.18 at $532\\times532$ shifts materially, or if the relative ordering of the models changes, the paper's central quantitative claim is not stable under the correct label definition. An even more direct check is to calculate the fraction of labeled pixels whose raw confidence exceeds 1 in the released dataset; a non-negligible fraction would confirm that the ground truth is not a proper confidence field.","supporting_citations":[{"cited_title":"Boosting crowd counting via multifaceted attention,","cited_arxiv_id":null,"evidence_quote":"Crowd-counting baseline (MAN) that RCMNet must beat on confidence-map regression."},{"cited_title":"To choose or to fuse? scale selection for crowd counting,","cited_arxiv_id":null,"evidence_quote":"Second crowd-counting baseline (SASNet) used for in-domain and robustness comparisons."},{"cited_title":"Imitation learning from expert video data for dissection trajectory prediction in endoscopic surgical procedure,","cited_arxiv_id":null,"evidence_quote":"State-of-the-art dissection trajectory prediction method (iDiff-IL) used as a baseline in the trajectory-suggestion task."},{"cited_title":"Stochastic trajectory prediction via motion indeterminacy diffusion,","cited_arxiv_id":null,"evidence_quote":"Diffusion-based trajectory prediction baseline (MID) used in the trajectory-suggestion comparison."},{"cited_title":"Exploring the limitations of behavior cloning for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Provides the behavior-cloning method that achieves the best trajectory-suggestion results."}],"review_version":1}