{"id":"599b132e-9fa4-47d8-b408-d6b287882e04","arxiv_id":"2412.02359","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"SurgiSim reconstructs a canonical 3D Gaussian scene from a monocular surgical video and runs soft-tissue MPM simulations using viscoelastic parameters estimated by matching the video.","lead":"SurgiSim turns ordinary one-camera surgery videos into a 3D simulation environment with soft tissue that deforms and springs back realistically. It automatically estimates tissue properties by comparing simulated motion to the original video, aiming to replace hand-built surgical simulators.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is that 2D point tracks lifted by per-frame monocular depth give scale-consistent 3D tissue trajectories (Eq.","rationale":"The reader's weakest_assumption is the one I would defend as most load-bearing. If the trajectories fed into Eq. 13 are not accurate and scale-consistent, the fitted stiffness/viscosity are not physically meaningful and the simulated response under new interactions is untrustworthy; this directly threatens the abstract's claim of 'realistic simulation' with parameters 'inferred by minimizing discrepancies' from video. The paper's own evaluation cannot expose such a bias because Table 1 replays the same manipulation used for fitting, and the user study also compares replayed/generated clips. A stereo-depth control is decisive here because EndoNeRF is binocular: it isolates the monocular-depth assumption while holding every other component fixed. I considered the algebraic inconsistency in Eq. 7 and the un-derived dissipation potential as secondary: those are localized and repairable, whereas a biased supervision signal invalidates the core quantitative claim if it lands. Since the reader already conditioned acceptance on this and related issues, my read does not change the verdict; I therefore set verdict_should_be to UNCHANGED, meaning the same conditional acceptance the reader recommended.","tokens_in":14422,"tokens_out":8475,"duration_ms":101560,"concrete_test":"On the EndoNeRF clips, recompute the 3D trajectories used in Eq. 13 from the stereo pairs (or from a metric depth source calibrated to the video), replacing only the Depth Anything v2 depth lifting in Sec. 3.3 while keeping all other modules, losses, and the replay protocol in Table 1 identical. Then compare the inferred (μ_E, η_v, γ_v) and the Table 1 metrics plus the free-rebound dynamics against the monocular-depth version. If the fitted parameters or simulated rebound damping change materially (e.g., >10% parameter shift or visibly different settling time), the monocular-depth assumption is load-bearing; if they are essentially unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 recovers 3D trajectories {p_t} by augmenting 2D dense optical tracks with per-frame monocular depth estimates from Depth Anything v2, then injects them into MPM as velocity boundary conditions: v_{B0}^t = (p_{n_t+1} - p_{n_t})/ΔT (Eq. 13). This is the primary physical supervision for the visco-elastic parameters (μ_E, η_v, γ_v) optimized by Eq. 14. The assumption that these trajectories are accurate and scale-consistent is load-bearing. Monocular depth is defined only up to an unknown per-image scale/shift, and the paper does not align or fuse depth scales across frames; EndoNeRF is a stereo dataset, but the paper deliberately uses only the left image and no binocular depth, so metric scale is never used. Temporal differences of per-frame depth estimates therefore mix scale drift with true tissue motion. Because the same video is used to fit the parameters and to produce the quantitative evaluation in Table 1 ('we reproduce the operation in the input videos'), the optimizer can absorb depth/track biases into the fitted constitutive parameters while still matching the training frames; PSNR/SSIM/LPIPS and user preference over replayed clips do not separate correct physics from appearance overfitting. Thus the paper's central claim of automatic, realistic simulation from monocular video rests on an unvalidated monocular trajectory assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"SurgiSim proposes an automatic pipeline that takes a monocular surgical video as input, reconstructs a canonical 3D Gaussian scene with a deformation field using multi-stage optimization with trajectory and anisotropic regularization plus surface thickening, and then runs MPM-based soft-tissue simulation with a proposed viscoelastic model. The physical parameters are inferred by minimizing the discrepancy between simulated renderings and the input video, with tissue motion obtained by lifting 2D dense point tracks using per-frame monocular depth estimates. Experiments on EndoNeRF clips report simulation-reproduction metrics, a user study with surgeons and laypersons, and ablations of the reconstruction and parameter-estimation components.","tokens_in":14812,"tokens_out":7489,"duration_ms":90517,"significance":"The end-to-end goal is valuable: if the reconstruction and physics are sound, SurgiSim would reduce manual mesh building and parameter tuning for surgical simulation, and the attempt to infer constitutive parameters from video through differentiable MPM and rasterization is a useful research direction. The user study, the automatic pipeline, and the qualitative demonstration of pulling and cutting interactions are clear strengths. However, several load-bearing technical points need correction or additional validation before the central claims of geometric consistency and realistic viscoelastic tissue dynamics can be accepted.","major_comments":[{"comment":"The trajectory regularizer is algebraically L_traj = Σ_{Gj,Gk∈Ni} Δµ_j · Δµ_k, because the norm factors in the numerator and denominator cancel. As written, minimizing this objective encourages displacement pairs to be anti-parallel rather than parallel, which is the opposite of the stated goal that the deformation directions should be parallel. The second stated goal, that movement length should be small, is also incompatible with a raw dot-product term, whose magnitude grows with displacement length. Please correct the loss (for example, use negative cosine similarity with a separate length penalty) and rerun the affected experiments; the current Table 3 does not contain a geometry metric and therefore cannot establish the claimed geometric-consistency benefit.","section":"§3.2, Eq. (7)"},{"comment":"The viscous model is not consistently derived from the stated dissipation potential. Differentiating Ψ(∂F_v/∂t) = 1/2 η_v tr((∂F_v/∂t)^T ∂F_v/∂t) gives a stress proportional to η_v ∂F_v/∂t, not σ_v = det(F_v)·2η_vD, and the relation ∂F_v/∂t = γD introduces parameters that do not appear in a single constitutive equation: γ_v appears in the update of F_v while η_v appears in the stress. The paper calls the model a Maxwell model, but the additive stress σ = σ_E + σ_v corresponds to a Kelvin–Voigt (parallel) arrangement rather than a Maxwell (series) arrangement. Please provide a consistent finite-strain viscoelastic formulation and clarify which rheological element is actually intended.","section":"§3.3, Eqs. (11)–(12)"},{"comment":"The physical parameters are optimized by minimizing Eq. (14) against the input video, and Table 1 evaluates the simulation by reproducing the operation in the same input videos. This is an in-sample fitting test, not a test of predictive realism: errors in depth scale, tracking drift, and model misspecification can be absorbed into the fitted µE, ηv, and γv while still matching the training frames. To support the claim of realistic simulation under new interactions, the paper needs held-out validation, such as simulating an interaction not used for fitting, or comparing the predicted free rebound against frames not included in Eq. (14), together with a report of the fitted parameter ranges.","section":"§3.3 and Table 1"},{"comment":"The 3D trajectories {p_t} are obtained by lifting 2D dense tracks with per-frame Depth Anything v2 estimates, and the paper does not describe any alignment of depth scale or shift across frames. Monocular depth is defined only up to per-image scale and shift, so the time differences in Eq. (13) mix apparent depth-scale drift with true tissue motion, and any error directly biases the injected MPM velocities and the inferred constitutive parameters. Because EndoNeRF is a stereo dataset, the authors could validate the monocular trajectories against stereo-derived depths (as they do in the supplementary reconstruction comparison of Table 4), or add a scale-calibration step; without this, the primary physical supervision is unvalidated.","section":"§3.3, Eq. (13)"},{"comment":"The ablation reports that removing trajectory regularization changes PSNR, SSIM, and LPIPS only slightly (e.g., PSNR 37.114 vs. 37.036), and the paper argues that this regularizer affects geometry rather than rendering quality. However, no geometric metric is provided, such as point-to-surface distance, normal consistency, or a direct measure of trajectory parallelism. Without such a metric, the central claim that trajectory regularization improves the geometric consistency of the simulation environment is not quantitatively supported, especially given the algebraic issue in Eq. (7).","section":"§4.3, Table 3"}],"minor_comments":[{"comment":"The text says the surface thickening method pushes Gaussian kernels along the z-axis, but line 6 of Algorithm 1 scales all position coordinates by (rand(3)+l)/1000; please clarify the intended transformation and correct the inconsistency.","section":"Algorithm 1"},{"comment":"The sentence \"These corrupted definitions are only used in this section, and their meaning will be re-defined\" appears to be a typo and should read \"redefined\"; the current wording is confusing.","section":"§3.3"},{"comment":"The table header and row labels for the two-alternative forced-choice comparison (\"Ours w.o. Guide\") are ambiguous; please present the four-way comparison and the paired comparison as separate blocks with clear column headings.","section":"Table 2"},{"comment":"Some citations appear mismatched: reference [42] is NR-SLAM rather than a tissue-mechanics or viscoelasticity reference, and references [14] and [15] in the MPM related-work sentence appear to be unrelated medical and chemistry papers; these should be corrected.","section":"References"},{"comment":"The supplementary reconstruction comparison uses stereo depth and the original dataset masks, whereas the main simulation pipeline uses monocular depth and SAM-refined masks; please state explicitly that Table 4 does not validate the monocular simulation setup.","section":"Table 4 and §4.2"},{"comment":"The line \"Published at Top Journal Arxiv\" and the superscript \"B\" after the last author appear to be formatting artifacts and should be removed.","section":"Header and formatting"}],"recommendation":"major_revision","confidential_remarks":"The algebraic error in Eq. (7) and the constitutive inconsistency in Eqs. (11)–(12) are the most serious technical issues, and I would want to see them corrected or explicitly justified before considering the paper further. The in-sample nature of the quantitative evaluation and the unvalidated monocular-depth supervision are also important; the strong qualitative results and the surgeon-inclusive user study are encouraging, but the current evidence does not establish generalization to new interactions. The citation errors and formatting artifacts are minor but should be cleaned up."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. Here's the short version: SurgiSim is a genuine end-to-end integration that works in practice, but the realism claim is undermined by a circular evaluation and a wrong equation in a stated contribution. The pipeline is not trivial: monocular video in, simulation-ready 3D Gaussian scene out, with estimated viscoelastic parameters and plausible novel interactions. The multi-stage optimization and surface thickening are sensible engineering, and the user study—especially with surgeons—gives some evidence that the results look more realistic than the baselines. That deserves credit.\n\nThe soft spots are real. Eq. 7 is algebraically wrong: the denominators cancel, leaving a sum of unnormalized dot products. Minimizing that sum encourages anti-parallel motion, not the parallel motion the text claims. The ablation shows it barely affects metrics, but as a stated contribution it needs fixing. Second, the quantitative evaluation in Table 1 fits parameters to the same videos it evaluates on. That makes the table a fit report, not a test of generalization. The user study partially compensates, but it also uses the same fitted simulations. The paper needs held-out validation—fit on one interaction, simulate a different one. Third, the 3D trajectories from monocular depth are load-bearing. Depth Anything v2 is affine-invariant per image; without cross-frame scale alignment, differences in depth can be dominated by scale drift. The paper doesn't address this, and it's fixable since the dataset is stereo. The viscoelastic model is ad hoc, but that's not unusual for this line of work; I wouldn't block on it. Missing code and excluding the cutting clip from the main evaluation are secondary.\n\nBottom line: this is a credible engineering paper that a serious referee could push into shape, but the central realism claim is not yet supported. I'd send it to review, with major revision: correct the loss, add a held-out experiment, and align or justify the depth scale.","headline":"SurgiSim is a genuine end-to-end engineering integration that works in practice, but the realism claim is undermined by a circular evaluation and a wrong equation in a stated contribution.","tokens_in":15293,"tokens_out":2716,"would_cite":false,"duration_ms":30946,"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":"SurgiSim claims a single monocular surgical video can become a physically simulated soft-tissue scene, with stiffness and viscosity inferred from the footage itself.","keywords":["3D Gaussian Splatting","surgical simulation","Material Point Method","viscoelasticity","monocular video reconstruction","physical parameter estimation","soft tissue deformation","deformable scene reconstruction"],"falsifier":"Record a surgical scene with a calibrated stereo rig and a tracked tool applying a known force, run SurgiSim's estimation, and compare the inferred Young's modulus and viscosity with independently measured values or with the observed rebound time; if simulated and real displacement profiles diverge under the same manipulation, the trajectory-lifting step is the prime suspect.","tokens_in":14210,"feed_emoji":"🩺","tokens_out":7771,"duration_ms":73926,"temperature":0.7,"pith_summary":"The paper claims that a single monocular surgical video is enough to build a simulation-ready, physically behaving soft-tissue scene. SurgiSim reconstructs a canonical 3D Gaussian model with a deformation field, using trajectory and anisotropy regularizations so the geometry stays consistent over time, and thickens the surface to give the simulator a volume to deform. It then fits tissue stiffness and viscosity by comparing rendered simulations against the observed video, guided by 3D trajectories obtained from point tracking and monocular depth. If the claim holds, realistic surgical simulators could be produced automatically from routine recordings, without hand-built meshes or manually tuned physics parameters.","feed_headline":"Surgical videos become soft-tissue simulators with learned physics","feed_subtitle":"Tissue stiffness and viscosity come from footage, so surgeons can rehearse new tool actions on the reconstructed scene.","key_machinery":"The load-bearing pieces are (1) a canonical 3D Gaussian Splatting scene coupled to a multi-resolution feature-plane and MLP deformation field, trained in stages with trajectory regularization (deformations in a small neighborhood stay parallel and small) and anisotropic regularization (Gaussians stay compact) to keep the canonical geometry physically consistent; (2) a surface-thickening step that turns the reconstructed shell into a volumetric particle set; (3) a Material Point Method simulator with a visco-elastic stress combining fixed corotated elasticity and a Maxwell-type viscous dissipation potential; and (4) automatic physics-parameter estimation in which 2D dense point tracks near the tool contact, lifted by monocular depth into 3D trajectories, are injected as velocity boundary conditions in the MPM update, and stiffness, viscosity, and damping are optimized by rasterizing the simulated Gaussians and minimizing the difference from the input video frames.","core_discovery":"On the paper's own terms, the central discovery is that a monocular surgical video contains enough information to build both a simulation-ready scene and the tissue's material response. The system reconstructs a canonical 3D Gaussian scene whose deformations over time are explained by a deformation field, regularized so that neighboring Gaussian trajectories stay parallel and the kernels stay compact; it thickens this shell into a volume; and it then fits a visco-elastic material model (fixed corotated elasticity plus a Maxwell-style viscous dissipation) by lifting dense 2D point tracks near the tool into 3D trajectories with monocular depth, injecting those as velocity boundary conditions in the material-point simulation, and minimizing the difference between rendered simulation frames and the input video. The result is a scene that reproduces the recorded manipulation and can respond to new tool interactions with damped, tissue-like motion rather than the unbounded elastic oscillations of the baselines.","pith_inferences":["The paper leaves implicit that, if inferred parameters transfer across clips of the same tissue type, the method could accumulate a per-procedure library of tissue material models; that is a natural next step the authors do not pursue.","A testable extension would be calibrating the monocular depth scale against a known instrument dimension in the frame; the paper's estimated stiffness and viscosity inherit whatever scale error the depth model has, so a calibration check would show whether absolute material values are recoverable.","The paper's own future-work note concedes that invisible side textures and cut surfaces still need manual correction and that topological inversions are unsupported, so the demonstrated interactions are limited to pulling and compressing rather than tearing tissue.","The parameter-estimation loop currently drives motion from prescribed trajectories; coupling the simulator with a rigid tool model so forces emerge from contact would let the system generalize to tool actions not present in the input video."],"forward_implications":["Surgical training simulators could be generated automatically from archived operation videos, removing the manual mesh-modeling step.","A robotic surgery system could rehearse a planned manipulation on the reconstructed tissue and observe the predicted deformation before acting.","Because stiffness, viscosity, and damping are fit per video, the simulation inherits the specific tissue response of the recorded procedure instead of generic preset values.","Under the paper's visco-elastic model, released tissue quickly returns to rest with brief damping, matching observed surgical behavior where elastic-only baselines keep oscillating."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation that serves as both the scene model and the simulation particles.","marker":"[23]"},{"why":"Provides the MPM-in-3DGS simulation framework that SurgiSim adapts, and serves as the main elastic-only baseline.","marker":"[50]"},{"why":"Supplies the dense optical point tracking that produces the 2D tissue trajectories near the tool contact point.","marker":"[25]"},{"why":"Provides the monocular depth estimation that lifts 2D tracks to the 3D trajectories used as velocity boundary conditions.","marker":"[55]"},{"why":"Supplies the Maxwell viscoelastic model on which the paper's viscous dissipation potential is based.","marker":"[21]"},{"why":"Provides the MPM particle-to-grid and grid-to-particle B-spline transfer used in the simulation.","marker":"[44]"},{"why":"Shows dynamic surgical scene reconstruction with a canonical Gaussian scene and deformation field that SurgiSim's environment setup builds on.","marker":"[34]"},{"why":"Supplies the EndoNeRF dataset of surgical videos used for evaluation and acts as a dynamic reconstruction baseline.","marker":"[47]"}],"fun_headline_variants":["Monocular video yields soft-tissue physics for surgical sim","Surgical videos teach a simulator tissue stiffness and viscosity","Learn tissue physics from video to rehearse surgery","From one camera: surgical simulation with real tissue physics","SurgiSim: turn any surgical clip into a touchable soft-tissue model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The inferred tissue properties are only as trustworthy as the 3D trajectories, which are built from single-camera depth estimates and 2D point tracks; any depth-scale error or tracking drift directly biases the estimated stiffness and viscosity, and therefore the simulated tissue response.","fun_headline_variants_meta":{"raw":{"variants":["Monocular video yields soft-tissue physics for surgical sim","Surgical videos teach a simulator tissue stiffness and viscosity","Learn tissue physics from video to rehearse surgery","From one camera: surgical simulation with real tissue physics","SurgiSim: turn any surgical clip into a touchable soft-tissue model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1381,"prompt_tokens":980,"completion_tokens":401,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":318}},"tokens_in":596,"tokens_out":401,"duration_ms":4234,"temperature":1.0,"reasoning_tokens":318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:35:10.976570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a surgical scene with a calibrated stereo rig and a tracked tool applying a known force, run SurgiSim's estimation, and compare the inferred Young's modulus and viscosity with independently measured values or with the observed rebound time; if simulated and real displacement profiles diverge under the same manipulation, the trajectory-lifting step is the prime suspect.","supporting_citations":[{"cited_title":"Physgaussian: Physics-integrated 3d gaussians for generative dynamics","cited_arxiv_id":null,"evidence_quote":"Provides the MPM-in-3DGS simulation framework that SurgiSim adapts, and serves as the main elastic-only baseline."},{"cited_title":"Dense optical tracking: connecting the dots","cited_arxiv_id":null,"evidence_quote":"Supplies the dense optical point tracking that produces the 2D tissue trajectories near the tool contact point."},{"cited_title":"A viscohyperelastic maxwell model for rubber viscoelasticity","cited_arxiv_id":null,"evidence_quote":"Supplies the Maxwell viscoelastic model on which the paper's viscous dissipation potential is based."},{"cited_title":"A material point method for snow simulation","cited_arxiv_id":null,"evidence_quote":"Provides the MPM particle-to-grid and grid-to-particle B-spline transfer used in the simulation."},{"cited_title":"Neural rendering for stereo 3d reconstruction of deformable tissues in robotic surgery","cited_arxiv_id":null,"evidence_quote":"Supplies the EndoNeRF dataset of surgical videos used for evaluation and acts as a dynamic reconstruction baseline."}],"review_version":1}