{"id":"7ab5211a-38e8-4d34-9d79-a62ef8407111","arxiv_id":"2504.12552","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On an internal dataset of 38 simulated surgical trials, operating room event detection from digital twin inputs (segmentation masks plus depth maps) matches or slightly exceeds detection from raw RGB video, while removing identifiable visual content.","lead":"This paper replaces raw operating room videos with geometric 'digital twin' representations, made from segmentation masks and depth maps, and shows that an event-detection model trained on these representations performs about as well as one trained on the original videos. The work could let hospitals share surgical workflow data without sharing video that shows patients and staff.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline parity result lacks statistical support: one split, 38 trials, no variance or significance test; the 2.18 mAP-point gap may be noise.","rationale":"The reader's weakest_assumption concerns foundation-model accuracy, but that is indirectly tested by the downstream mAP: if masks and depth were badly wrong, DT mAP would likely be poor. The more fragile assumption is that the reported comparison is statistically meaningful. With 38 trials and no variance estimate, the 2.18-point gap is within plausible noise. This is a concrete, checkable flaw and is central to the paper's main empirical claim. I therefore partially agree with the reader: their rationale mentions statistical rigor, but their stated weakest_assumption does not. I also note the privacy/de-identification claim is entirely unverified, but the immediate load-bearing issue for the strongest claim is the missing statistical evidence. The verdict remains CONDITIONAL as the reader said; the contribution may be salvageable with a rigorous multi-split evaluation.","tokens_in":3603,"tokens_out":5887,"duration_ms":64165,"concrete_test":"Perform a per-trial evaluation: compute mAP per video (or per trial) for Mask-Depth DT and RGB under identical splits, then run a paired bootstrap (10,000 resamples) or Wilcoxon signed-rank test across the 38 trials to estimate the 95% confidence interval for the mean mAP difference. Also run 5-fold cross-validation. If the CI includes zero (p > 0.05), the paper should state 'no statistically significant difference' and remove 'sometimes better.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 and Table 1 report a single average-mAP comparison (Mask-Depth DT 72.93 vs RGB 70.75) from an internal set of 38 simulated trials, with no confidence intervals, no cross-validation, and no significance test. Table 2's temporal prediction errors show large per-class standard deviations (e.g., 'Loading patient to gurney': 32.62 ± 66.69 s start, 30.57 ± 75.03 s end), indicating high variability. A 2.18-point mAP gap on a small dataset can easily arise from sampling noise. Since the paper's central claim is that DT inputs are on par with, and sometimes better than, RGB, this lack of statistical grounding is load-bearing: if the gap is not significant, the 'sometimes better' assertion is unsupported, and even 'on par' is not rigorously demonstrated. The absence of released code or data prevents independent verification. This is a missing-support flag on the strongest claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-stage pipeline for privacy-preserving operating room (OR) workflow analysis. In stage one, RGB videos are converted into a Digital Twin (DT) representation consisting of semantic segmentation masks and monocular depth maps, generated with vision foundation models (DETR-prompted SAM and Depth Anything v2). In stage two, a two-stream Transformer model with cross-attention processes the mask and depth streams to detect five perioperative events. Evaluation on an internal dataset of 38 simulated surgical trials compares RGB input against Mask DT, Depth DT, and the fused Mask-Depth DT, reporting mean Average Precision (mAP) at temporal IoU thresholds and per-class temporal errors. The authors claim that the DT-based approach performs on par with, and sometimes better than, raw RGB, and that the DT representation is fully de-identified.","tokens_in":3855,"tokens_out":5096,"duration_ms":50980,"significance":"If the empirical claim were convincingly established, this work would be a useful contribution to privacy-preserving OR workflow analysis: it could enable event detection from conventional RGB cameras without optical privacy hardware, facilitate cross-institutional data sharing through de-identified scene abstractions, and potentially reduce domain shift between institutions. Strengths of the paper include the use of a standard temporal action localization metric (mAP at tIoU thresholds), per-class boundary error reporting, and qualitative visualizations. The combination of modern foundation models for scene abstraction is timely and well motivated. However, the evidence presented is currently too weak to support the central parity claim, and the privacy-preservation assertion is not validated.","major_comments":[{"comment":"The central claim that DT inputs are on par with, and sometimes better than, RGB is supported only by a single average-mAP comparison: Mask-Depth DT achieves 72.93% versus 70.75% for RGB, a gap of 2.18 points on a dataset of 38 simulated trials. No confidence intervals, cross-validation, or significance tests are reported. The high per-class variability in Table 2 (e.g., 'Loading patient to gurney' start error of 32.62 ± 66.69 s) suggests that the observed difference could easily arise from sampling noise. The authors should provide error bars, perform paired significance tests (e.g., bootstrap or permutation tests), and describe the train/validation/test split. Without this, the 'sometimes better' assertion is unsupported, and even 'on par' is not rigorously demonstrated.","section":"Section 3, Table 1"},{"comment":"The train/validation/test split is not described. The reader does not know how the 38 trials were partitioned, whether the split is by trial or by OR, or whether the same subjects or ORs appear in multiple splits. This information is essential for interpreting the reported metrics and for assessing the variability of the comparison. Please specify the exact split, consider leave-one-OR-out evaluation or repeated cross-validation, and state whether the same hyperparameters were used for all input representations.","section":"Section 3, Dataset"},{"comment":"The statement that 'since these DT only contain depth and semantic information, they are fully de-identified' is an assertion without supporting analysis. Depth maps and semantic masks can still encode identifying information (e.g., body geometry, gait, clothing shape, scene layout), and the paper provides no re-identification risk assessment, no privacy metric, and no comparison to existing anonymization methods such as face blurring or structured-light depth imaging. Because privacy-preservation is the paper's stated motivation and is central to its title, this claim is load-bearing and needs at least a careful discussion of residual risks or a quantitative privacy analysis.","section":"Section 2, Stage 1"},{"comment":"The pipeline relies on the accuracy of Depth Anything v2 and DETR-prompted SAM in OR scenes, but no sensitivity or error-propagation analysis is provided. If the foundation models produce inaccurate masks or depth maps, the DT representation will be degraded, which could directly affect event detection. The authors should report the quality of the generated DTs (e.g., segmentation/depth accuracy against manually annotated examples) and analyze how such errors affect downstream event detection, or compare against oracle masks/depth to bound the effect.","section":"Section 2, Stage 1"}],"minor_comments":[{"comment":"The caption is truncated: 'Our two-stage pipeline use Digi-' should be completed, for example, as 'Our two-stage pipeline uses Digital Twins for privacy-preserving OR workflow analysis.'","section":"Figure 1 caption"},{"comment":"The notation 'Emf' and 'Edf' is not defined; please clarify what the subscripts denote (e.g., mask frame embedding and depth frame embedding) and define the temporal window size T.","section":"Section 2, Stage 2"},{"comment":"The event detection model is called 'SafeOR', but the architecture is attributed to MUST (Pérez et al., 2024). Please clarify whether SafeOR is a new name, a variant, or a direct adoption of the MUST model, and cite the appropriate source.","section":"Section 2, Stage 2"},{"comment":"No hyperparameters are reported (e.g., T, number of transformer layers L, embedding dimension, learning rate, training epochs). Providing these details is important for reproducibility, even in a short paper.","section":"Section 2 and 3"},{"comment":"Only Mask-Depth DT and RGB are compared in the temporal-error table; including single-modality DT results would help interpret the benefit of fusion.","section":"Table 2"},{"comment":"The claim that DTs 'facilitate cross-institutional model training and improves model generalizability' is not experimentally tested. The paper should either add a supporting experiment (e.g., cross-institution or domain-shift evaluation) or soften the claim to a hypothesis.","section":"Section 4"},{"comment":"Minor language issues: 'model relations between both modalities' should be 'model the relations between both modalities', and 'ground truth (GT)' is used before the abbreviation is introduced.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short conference paper, but the evaluation limitations are substantial relative to the central claim. The dataset is internal and not released, and no code is provided, so independent verification is impossible. The authors' prior work is heavily cited for both the DT framework and the event detection architecture; this is not circularity, but the novelty of this pipeline relative to those prior works should be made explicit. The main risk is that the empirical parity claim will not withstand a more rigorous evaluation. A major revision with additional statistical analysis, a clear split description, and a privacy discussion could make the claim credible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the MIDL short paper on DT-based OR event detection. The core idea is sensible: use monocular depth and segmentation foundation models to strip RGB video down to a de-identified geometric-semantic representation, then run event detection on that. Doing this for event detection rather than just phase recognition is a natural extension of the authors' prior work, and the evaluation with mAP at multiple tIoU thresholds is a reasonable choice. I also appreciate that they compare against an RGB baseline rather than only showing the DT works in isolation.\n\nThat said, the central claim—that the DT approach is on par with, and sometimes better than, RGB—is not statistically grounded. The numbers in Table 1 come from a single split of 38 simulated trials, with no confidence intervals, cross-validation, or significance test. The 2.18 mAP-point gap between Mask-Depth DT (72.93) and RGB (70.75) is well within the noise one would expect at this sample size. The per-class temporal errors in Table 2 are huge (e.g., 30.57 ± 75.03 s for 'Loading patient to gurney' end time), which reinforces that the dataset is small and variable. So the paper's own summary 'sometimes better' is an overstatement; 'not clearly worse' is what the data actually show. Even the 'on par' part needs a proper multi-split or bootstrapped evaluation to be convincing.\n\nThe privacy claim also needs more care. The paper calls the DT 'fully de-identified,' but it doesn't analyze what information could be recovered from depth maps or segmentation masks (e.g., person-specific geometry or re-identification from body shape). That's not necessarily fatal—the representation is clearly much less identifiable than RGB—but 'fully' is stronger than what is demonstrated.\n\nThe other soft spots are the missing description of the train/validation split and the lack of code/data release. With only 38 trials, the exact split matters; a different split could easily flip the ranking.\n\nThe citation pattern is fine. The authors lean on their own prior work (Ding et al. 2024a,b; Pérez et al. 2024), but that's legitimate here because the DT framework and the MUST architecture are genuinely theirs. The comparison against RGB is an independent evaluation, so there's no circularity.\n\nBottom line: this is a plausible, well-scoped idea that deserves a serious referee, but the empirical case is currently thin. A revised version with multi-split evaluation, confidence intervals, and a more careful privacy discussion would make the contribution solid. I'd send it to review, with the expectation of major revision.","headline":"A useful privacy-preserving pipeline idea, but the headline parity claim rests on a single small split with no significance testing; treat the 'sometimes better' as anecdotal.","tokens_in":4347,"tokens_out":2113,"would_cite":false,"duration_ms":21336,"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":"Operating-room workflow events can be detected from de-identified digital twins—semantic masks plus depth maps—just as accurately as from raw RGB video, making surgical footage safe to share across institutions.","keywords":["Operating Room","Workflow Analysis","Digital Twins","Workflow Optimization","Privacy-Preserving Video Analysis","Event Detection","Semantic Segmentation","Depth Estimation"],"falsifier":"Run the same SafeOR training on a real, multi-site OR video dataset (not the 38-trial role-play simulation) comparing raw RGB against Mask-Depth DT; if the DT average mAP falls materially below the RGB baseline, the parity claim is falsified. Separately, a successful attempt to re-identify a patient or reconstruct an identifiable scene from the mask-and-depth representation would falsify the privacy claim.","tokens_in":3462,"feed_emoji":"🏥","tokens_out":9080,"duration_ms":77806,"temperature":0.7,"pith_summary":"This paper argues that privacy doesn't have to cost accuracy in operating-room (OR) workflow analysis. The authors propose converting ordinary RGB video into a de-identified 'digital twin' representation—per-frame semantic segmentation masks plus monocular depth maps—and then detecting events from that representation instead of from the video itself. On an internal dataset of 38 simulated surgical trials covering five event classes, the mask-plus-depth twin matches or beats raw RGB video for temporal event detection (average mAP 72.93 vs 70.75). The reason to care is practical: if masks and depth maps carry enough information for workflow analysis, then sensitive footage can be discarded and only the de-identified twin shared, enabling multi-institution training without exposing patients or staff.","feed_headline":"De-identified OR twins detect events as well as raw video","feed_subtitle":"Mask and depth data keep OR event detection accurate while removing identifiable video.","key_machinery":"The central object is the Digital Twin of the OR scene, a de-identified 2D representation in which each frame is replaced by a semantic segmentation mask and a depth map. The generation pipeline is the first mechanism: a fine-tuned object detector (DETR with collaborative hybrid assignments) proposes bounding boxes for 14 OR object classes, those boxes prompt the Segment Anything Model (SAM) to produce masks, and Depth Anything v2 produces monocular depth. The second mechanism is the SafeOR event detector, a two-stream architecture that encodes a temporal window of mask frames and depth frames with separate video backbones, then fuses the two streams' embeddings with cross-attention over a long-term sequence (extending the MUST multi-scale transformer) before a classification head labels events per frame. Together, the two mechanisms convert a privacy-sensitive RGB stream into a geometric-semantic abstraction and show that event detection can be done entirely on that abstraction.","core_discovery":"On the paper's own terms, the central discovery is that a Digital Twin composed only of semantic segmentation masks and depth estimates—obtained by prompting SAM with bounding boxes from a fine-tuned DETR and running Depth Anything v2 on conventional RGB frames—preserves enough spatio-temporal information for OR event detection to match, and in aggregate slightly exceed, detection from raw RGB video. The authors report that their fused mask-and-depth twin (Mask-Depth DT) reaches an average mAP of 72.93 across temporal IoU thresholds, versus 70.75 for the RGB baseline, with the largest gains at stricter thresholds (46.89 vs 43.09 mAP@0.75). They also report smaller temporal prediction errors on several event classes, such as gurney entering, suggesting the geometric abstraction gives cleaner event boundary estimates. Under this view, the DT is not a lossy compromise but a privacy-preserving representation that can be shared across institutions while retaining the signal needed for workflow analysis.","pith_inferences":["If masks and depth truly suffice, then training data for OR event detection could be generated synthetically by rendering simulated operating rooms into semantic-depth twins, removing the need to collect or de-identify any real video.","The same de-identification-by-abstraction recipe could transfer to other privacy-sensitive video domains—home care, retail, elder monitoring—where the question is whether a task's signal survives the loss of appearance.","The privacy claim is here asserted rather than adversarially tested; a natural stress test would try to re-identify individuals or reconstruct recognizable scenes from masks and depth alone, since room geometry and equipment layout can themselves be identifying.","Because Depth Anything v2 and SAM were trained on natural images, their masks and depths may degrade on domain-shifted OR scenes; a testable extension is to measure how much DT quality (and downstream mAP) drops on unseen real ORs."],"forward_implications":["Hospitals can share OR workflow data as masks and depth maps instead of video, avoiding the identifiability that blocks multi-institution training.","Event detection no longer requires specialized depth cameras; monocular depth from ordinary RGB cameras suffices to build the privacy-preserving representation.","Because the DT strips away appearance, models trained on it may generalize better across hospitals with different equipment, lighting, and room layouts, a benefit the paper argues for.","The SafeOR model on Mask-Depth DT beats single-modality masks or depth alone, showing the two modalities carry complementary event signals.","The DT representation yields comparable or better temporal boundary estimates for several event classes, so workflow bottleneck measurement does not degrade under anonymization."],"supporting_citations":[{"why":"Supplies Depth Anything v2, the monocular depth estimator that produces the depth map stream of the digital twin.","marker":"Yang et al., 2024"},{"why":"Supplies SAM, the segmentation model that, prompted by detector boxes, generates the semantic mask stream.","marker":"Kirillov et al., 2023"},{"why":"The fine-tuned DETR variant whose bounding-box predictions prompt SAM to segment the 14 OR objects.","marker":"Zong et al., 2023"},{"why":"The MUST multi-scale transformer whose cross-attention fusion design SafeOR extends to fuse mask and depth streams.","marker":"Pérez et al., 2024"},{"why":"Establishes depth images as a privacy-preserving input modality for OR workflow analysis, but requires structured-light hardware.","marker":"Jamal and Mohareri, 2022"},{"why":"Prior face-anonymization approach for OR video that only hides facial features, motivating the fuller de-identification of the digital twin.","marker":"Bastian et al., 2023"},{"why":"Defines the digital twin concept in surgical data science and frames it as a geometric scene abstraction that supports de-identification.","marker":"Ding et al., 2024a"}],"fun_headline_variants":["OR digital twins match raw video in event detection","Privacy-preserving twins detect OR events like raw video","Mask-depth OR twins rival raw video for event detection","De-identified OR twins perform on par with raw video","Twin depth maps match video for OR workflow analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the foundation models' predicted segmentation masks and depth maps are accurate and stable enough to faithfully represent the OR scene, since any error in those predictions propagates directly into the digital twin and degrades the event detector.","fun_headline_variants_meta":{"raw":{"variants":["OR digital twins match raw video in event detection","Privacy-preserving twins detect OR events like raw video","Mask-depth OR twins rival raw video for event detection","De-identified OR twins perform on par with raw video","Twin depth maps match video for OR workflow analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000804,"raw_usage":{"total_tokens":3513,"prompt_tokens":910,"completion_tokens":2603,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":2527}},"tokens_in":526,"tokens_out":2603,"duration_ms":19345,"temperature":1.0,"reasoning_tokens":2527,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:28:22.645305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same SafeOR training on a real, multi-site OR video dataset (not the 38-trial role-play simulation) comparing raw RGB against Mask-Depth DT; if the DT average mAP falls materially below the RGB baseline, the parity claim is falsified. Separately, a successful attempt to re-identify a patient or reconstruct an identifiable scene from the mask-and-depth representation would falsify the privacy claim.","supporting_citations":[{"cited_title":"Detrs with collaborative hybrid assignments training","cited_arxiv_id":null,"evidence_quote":"The fine-tuned DETR variant whose bounding-box predictions prompt SAM to segment the 14 OR objects."},{"cited_title":"Multi-modal unsupervised pre-training for surgical operating room workflow analysis","cited_arxiv_id":null,"evidence_quote":"Establishes depth images as a privacy-preserving input modality for OR workflow analysis, but requires structured-light hardware."},{"cited_title":"Disguisor: holistic face anonymization for the operating room","cited_arxiv_id":null,"evidence_quote":"Prior face-anonymization approach for OR video that only hides facial features, motivating the fuller de-identification of the digital twin."}],"review_version":1}