{"id":"997ca27e-d9a0-48c2-8ba8-c8cdab66b6a9","arxiv_id":"2507.20337","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"PIVOTS, a point-cloud-based non-rigid volume-to-surface registration network with deformation-aware cross attention, outperforms existing methods on synthetic, phantom, and real liver data and ships with code and benchmarks.","lead":"A new neural network, PIVOTS, registers preoperative liver scans to partial intraoperative surfaces using point clouds, achieving lower registration error than existing methods. It was trained on synthetic biomechanical simulations and validated on real datasets, and its code and benchmarks are released for comparison.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-data evaluation never quantifies registration on actual intraoperative surfaces; target-domain superiority rests on clean MRI-derived surfaces plus a simulator whose fidelity the authors themselves flag as limited.","rationale":"The paper is a solid methods study: it releases code and data, retrains all learning-based baselines on the same synthetic training set, and its Phantom/HHLBM results are the only independent evidence for the central claim. I found no internal inconsistency or fatal flaw, and I agree with the reader's conditional assessment. My concern sharpens the reader's weakest assumption: the two quantitatively evaluated 'real' datasets are real in deformation source (probe compression, breathing) but not in surface acquisition, because the intraoperative surfaces are extracted from co-registered MR volumes and therefore lack stereo-reconstruction noise and occlusion. The paper itself acknowledges the sim-to-real gap in Sec. 5 and lists the missing physical effects. The laparoscopic dataset, which would be the true target-domain evidence, is used for qualitative comparison only. Consequently, the Synthetic/AMOS results partly measure fit to the authors' simulator, and the Phantom/HHLBM results measure performance on clean partial views rather than on actual laparoscopic reconstructions. This does not invalidate the within-protocol comparison—PIVOTS does outperform baselines under those conditions—but it means the clinically relevant claim 'for liver navigation' is not yet quantitatively demonstrated. An evaluation on a public intraoperative benchmark such as DePOLL would settle this. The reader's CONDITIONAL verdict remains appropriate; the condition should be made explicit: quantitative validation on real intraoperative surfaces with ground truth is required before the central claim is treated as established.","tokens_in":31628,"tokens_out":6658,"duration_ms":73301,"concrete_test":"Run PIVOTS and the strongest baselines (V2S-Net, BCF_FEM) on the DePOLL porcine dataset (Modrzejewski et al., 2019), which provides 12 real in vivo intraoperative deformed states of a liver with landmark ground truth, using an identical initial rigid alignment protocol; report mean TRE and the fraction of cases below 10 mm. If PIVOTS does not beat V2S-Net or its mean TRE exceeds 10 mm, the claim of superiority for real intraoperative surfaces is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PIVOTS provides accurate, robust volume-to-surface registration for laparoscopic liver navigation. The least secure condition is that the synthetic training distribution—random organ-like shapes, homogeneous Neo-Hookean material, simplified springs/fixed boundaries, and Perlin/Gaussian noise—matches real intraoperative deformation and reconstruction well enough for performance to transfer. The paper's own Discussion (Sec. 5) concedes a 'noticeable sim-to-real gap' and lists missing effects: tool-tissue interaction, self-collision, heterogeneous material properties, pneumoperitoneum volume changes, and 'authentic surgical noise.' More importantly, the quantitative real-data tests do not close this gap: the Phantom and HHLBM 'intraoperative' surfaces are clean partial views extracted from co-registered MRI volumes (Sec. 4.1), so they contain none of the stereo-reconstruction noise, illumination artifacts, or instrument occlusion of the target setting. The only real laparoscopic surfaces (n=6) are evaluated qualitatively (Sec. 4.8). Thus the headline numbers—3.11, 4.06, 5.96, and 6.38 mm—are either produced by the simulator itself (Synthetic/AMOS) or measured on clean MRI-derived surfaces (Phantom/HHLBM). No quantitative evidence demonstrates superiority on actual intraoperative point clouds. If real reconstruction noise has a different structure than Perlin/Gaussian noise, or if the simulator's deformation modes omit pneumoperitoneum and tool loading, the reported margins could shrink or vanish.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PIVOTS, a transformer-based point-cloud network for non-rigid volume-to-surface registration of the liver. The method takes a preoperative liver volume and a partial intraoperative surface as point clouds, uses a DGCNN+FPS encoder with multi-resolution feature extraction, and a decoder with novel deformation-aware cross-attention and upsampling cross-attention modules to predict a dense displacement field at multiple resolutions. Training is supervised on 193,563 synthetic pairs produced by a biomechanical simulator with randomized organ shapes, material parameters, boundary conditions, partial surface extraction, and Perlin/Gaussian noise. The method is evaluated on a synthetic test set, a semi-synthetic AMOS-based set, a six-deformation MRI phantom dataset, a 38-sample healthy human liver breathing motion (HHLBM) dataset, and qualitatively on six real laparoscopic videos. Reported mean errors are 3.11 mm (Synthetic), 4.06 mm (AMOS), 5.96 mm (Phantom), and 6.38 mm (HHLBM), outperforming all baselines on each dataset. The paper also provides robustness experiments for deformation magnitude, noise type and amplitude, and surface visibility, and releases code and datasets.","tokens_in":31925,"tokens_out":8256,"duration_ms":81985,"significance":"If the results hold, PIVOTS represents a strong step toward practical non-rigid liver registration in laparoscopic navigation. The paper's strengths are its extensive evaluation: twelve baselines are retrained or fine-tuned on the same synthetic data, the stress tests for deformation, noise, and visibility are systematic and include error distributions, and the code and datasets are publicly released, including a new real-data benchmark (HHLBM). The ablation study isolates the contributions of the proposed cross-attention modules and manual features. The authors are also transparent about remaining limitations, clearly stating the sim-to-real gap and the absence of a quantitative intraoperative evaluation. The main weakness is that the quantitative real-data experiments are conducted on clean MRI-derived surfaces rather than actual laparoscopic reconstructions, which limits the strength of the clinical generalization claim.","major_comments":[{"comment":"The quantitative real-data evaluation is performed only on MRI-derived surfaces: the Phantom and HHLBM 'intraoperative' partial surfaces are extracted by simulating a camera field of view from co-registered MRI volumes (Sec. 4.1), so they contain none of the reconstruction noise, illumination artifacts, or instrument occlusion of actual laparoscopic data. The only real laparoscopic surfaces (Laparoscopic Liver dataset, n=6) are evaluated qualitatively (Sec. 4.8). Consequently, the reported TREs of 5.96 mm (Phantom) and 6.38 mm (HHLBM) do not directly establish performance on actual intraoperative point clouds, and the abstract's claim that PIVOTS 'validates its performance on ... real datasets' is overstated for the target application. I recommend adding a quantitative metric on the laparoscopic dataset (e.g., surface-to-surface error between the deformed preoperative volume and the reconstructed intraoperative point cloud, with or without a small set of manual landmarks), or explicitly restricting the conclusion to the evaluated settings.","section":"§4.1, Table 2"},{"comment":"The synthetic training and test data are generated by the same simulation pipeline (Pfeiffer et al., 2025) with a homogeneous Neo-Hookean liver model, simplified spring/fixed boundary conditions, and Perlin/Gaussian noise; the Synthetic and AMOS test sets come from this same generator. The paper's own Discussion (Sec. 5) lists missing effects (tool-tissue interaction, self-collision, heterogeneous material properties, pneumoperitoneum volume changes, authentic surgical noise) and reports a sim-to-real gap of about 2 mm. Therefore, the robustness claims for large deformation (Sec. 4.4) and noise (Sec. 4.5) are demonstrated only under the simulator's deformation and noise model, not under the intraoperative conditions the method is intended for. I recommend adding an independent synthetic test set from a differently constructed simulator or a controlled physical phantom with larger deformations, and/or explicitly downgrading the claims to 'simulated intraoperative conditions' in the abstract and conclusion.","section":"§3.5, §4.3–4.6"}],"minor_comments":[{"comment":"The table header contains typos: 'Spearson' should be 'Spearman' and 'BCE_FEM' should be 'BCF_FEM'.","section":"§4.4, Table 3"},{"comment":"In the definition of the network inputs, the intraoperative point cloud is written as Sr = {Ps, Àv, Fs}, but the positional encoding for the intraoperative stream should be Às, not Àv. The subsequent dimensional equation also mixes the two: 'Às, Às ∈ Rn×(6É)' should read 'Àv, Às ∈ Rn×(6É)'.","section":"§3.1"},{"comment":"The claim of a 'critical visibility threshold' at 10% visibility is based on small samples: the Phantom dataset has only six deformation states and the HHLBM dataset 38 samples, and the 0–10% bin contains a fraction of these. The standard deviations in that bin are large (e.g., ±9.92 mm for PIVOTS on HHLBM). Please report confidence intervals or a statistical test (e.g., paired comparison per sample) before asserting a threshold effect.","section":"§4.6, Table 4"},{"comment":"The training data exports the timestep with the highest deformation rather than the equilibrium state, so some training pairs correspond to transient dynamic states that may not represent stable intraoperative configurations. Please justify this choice physically, or provide an ablation comparing training on equilibrium states only.","section":"§3.5.2"},{"comment":"The baseline list includes surface-to-surface and rigid methods (e.g., Lepard, C2P-Net, LiverMatch+NDP) that are adapted to the volume-to-surface task, but the adaptation is not described. Please add a short description of the input sampling, loss formulation, and any modifications used for each learning-based baseline.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well within the scope of a medical image computing journal, and the dataset release and benchmark contribution are valuable. The central concern is that the paper's clinical claims for laparoscopic navigation outrun the quantitative evidence, which is limited to MRI-derived surfaces and simulator-generated noise. The requested additional intraoperative evaluation or a careful tightening of the claims should be feasible within the scope of a major revision. I would not reject the paper: the method, the HHLBM dataset, and the benchmark are solid contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine architectural step forward for volume-to-surface non-rigid registration, and the code and benchmark release is a real service to the community. But the headline millimeter numbers are mostly measured on synthetic data or clean MRI-derived surfaces, not on actual laparoscopic reconstructions, so the clinical superiority claim is still open.\n\nThe new part is a point-cloud architecture with deformation-aware cross-attention; unlike V2S-Net's voxel grid it consumes partial noisy surfaces directly. The ablation makes the case that the attention design matters: removing it raises MED from 3.90 to 6.60 mm, and swapping in KPConv hurts. The evaluation is far more thorough than the usual medical registration paper: four datasets, several baselines retrained on the same synthetic set, dedicated deformation/noise/visibility experiments, and qualitative real laparoscopic data. Shipping code and training/test data is exactly the kind of reproducible contribution the field needs.\n\nThe soft spots are real but proportionate. The synthetic training and part of the test data come from the same in-house biomechanical pipeline, so Synthetic and AMOS numbers partly validate against the simulator's assumptions. Phantom and HHLBM surfaces are clean partial views cut from co-registered MRI, not stereo-reconstructed point clouds with illumination artifacts and instrument occlusion; the only real laparoscopic surfaces (n=6) are qualitative. The authors explicitly concede a 'noticeable sim-to-real gap' and list missing effects: tool-tissue interaction, self-collision, heterogeneous material, pneumoperitoneum, authentic surgical noise. There are also no significance tests, and the phantom set has six deformations. The citation pattern is fine, though one key dependency (the simulation pipeline) is an unpublished companion paper; releasing the code mitigates that.\n\nWho is this for? Anyone working on surgical AR or non-rigid registration benchmarks. The architecture and dataset are worth engaging with. I would send it to peer review and let reviewers push the authors to either obtain quantitative intraoperative validation or carefully scope the claims as feasibility, not established clinical advantage.","headline":"Solid architectural contribution with open code and benchmarks; the headline accuracy numbers are mostly synthetic or clean-MRI, so the clinical superiority claim is still open.","tokens_in":32543,"tokens_out":2162,"would_cite":true,"duration_ms":24265,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Point-cloud network registers livers in under 7 mm from partial views","keywords":["intraoperative","liver navigation","point clouds","non-rigid registration","volume-to-surface registration","deformation prediction","augmented reality surgery","synthetic training data"],"falsifier":"A direct clinical test would settle it: collect intraoperative liver surfaces during laparoscopic surgery with tracked landmarks or embedded fiducials, run PIVOTS, and measure landmark target registration error. If mean TRE exceeds 10 mm on such unseen surgical data, or if errors grow steeply when pre-registration displacement exceeds 80 mm or visible surface falls below 10%, the claimed synthetic-to-real robustness does not transfer.","tokens_in":31438,"feed_emoji":"🩺","tokens_out":10874,"duration_ms":96994,"temperature":0.7,"pith_summary":"PIVOTS is a neural network that registers a full preoperative liver volume to a noisy, partially visible point cloud of the liver surface captured during laparoscopic surgery. The paper's central claim is that a network trained exclusively on synthetic biomechanical simulations of liver deformation can predict a dense displacement field for every point of the volume, and that this prediction is more accurate and more robust than existing rigid, non-rigid, and biomechanics-based baselines. Across the four evaluation datasets the reported mean errors are 3.11 mm on Synthetic, 4.06 mm on AMOS, 5.96 mm on Phantom, and 6.38 mm on the real breathing-motion dataset HHLBM, the best in the comparison table on each. If the claim holds, augmented-reality overlays of tumors and vessels could be refreshed in about 230 ms while tolerating the large deformations, noise, and occluded views that occur in real surgery.","feed_headline":"Point-cloud network registers livers in under 7 mm from partial views","feed_subtitle":"Trained only on simulated deformations, the network maps full liver volumes onto noisy partial surfaces in 230 ms.","key_machinery":"The load-bearing machinery is the PIVOTS architecture paired with a synthetic training-data generator. The encoder uses farthest-point sampling and Dynamic Graph CNN layers to build multi-resolution geometric features from the preoperative and intraoperative clouds. The decoder centers on deformation-aware cross-attention: the intraoperative surface first attends to the volume, and the volume then attends to the updated surface features, letting information flow in both directions at every resolution; an upsampling cross-attention module then propagates coarse features back to fine levels, and a displacement field is regressed at each level under a weighted mean-squared-error loss. The generator creates 193,563 training samples from random organ-like shapes, a homogeneous Neo-Hookean material with randomized Young's modulus (3–30 kPa) and Poisson ratio (0.45–0.48), gravity, fixed, and spring boundary conditions, and partial surfaces corrupted with Perlin and Gaussian noise; this simulator is the source of the network's claimed generalization to real patients.","core_discovery":"The central claim is that volume-to-surface non-rigid registration can be solved end-to-end directly on point clouds: given a preoperative volume cloud $V$ and an intraoperative partial surface cloud $S$, PIVOTS outputs a dense displacement field $\\Phi$ for every point of $V$, including internal points, in a single forward pass. The paper reports the lowest mean registration error on every tested dataset—$3.11\\pm3.38$ mm on Synthetic, $4.06\\pm3.35$ mm on AMOS, $5.96\\pm1.47$ mm on Phantom landmarks, and $6.38\\pm2.56$ mm on HHLBM landmarks—with the next-best competing method typically 2–6 mm worse. PIVOTS also keeps mean errors below 10 mm even for the largest pre-registration displacement group (above 80 mm), across Perlin noise amplitudes up to 15 mm and Gaussian noise up to $\\sigma=5$ mm, and for visible surface fractions above roughly 10%; the paper reports that all methods struggle below 10% visibility. The authors attribute this stability to the deformation-aware cross-attention decoder and to synthetic training data that randomizes organ shape, material parameters, boundary conditions, partial views, and noise.","pith_inferences":["A testable extension is to apply the same architecture and synthetic-data recipe to other soft organs such as prostate or lung; the paper suggests this is plausible, but its evidence is limited to liver-shaped geometry and liver material ranges.","The ablation, in which removing the deformation-aware cross-attention raises error by about 70%, suggests that bidirectional cross-attention, rather than the point-cloud backbone, is the main source of accuracy; other registration networks might close the gap by adding this mechanism.","A natural next validation would be quantitative evaluation on true intraoperative laparoscopic surfaces with measured landmarks, since the paper's laparoscopic dataset currently supports only qualitative comparison; a clinical test with tracked tools or preoperative fiducials would directly expose the synthetic-to-real gap."],"forward_implications":["If the reported errors hold, a laparoscopic augmented-reality system could update its overlay of vessels and tumors roughly every 230 ms from a single forward pass, without patient-specific retraining.","Because the network generalizes from random synthetic shapes to real livers, phantom deformations, and in-vivo breathing motion, the same trained weights could serve as the non-rigid core of a navigation system across different patients.","Mean errors below 10 mm even at the largest pre-registration displacement group and with Perlin noise up to 15 mm imply the registration can tolerate imperfect surface reconstruction and substantial pneumoperitoneum-induced deformation.","The visibility experiments imply that reliable registration requires more than about 10% visible liver surface, so intraoperative mapping must prioritize wide coverage, not just a narrow initial view.","The released synthetic training and test sets, plus the HHLBM real-breathing dataset, give future volume-to-surface liver registration methods a common benchmark for fair comparison."],"supporting_citations":[{"why":"It establishes the non-rigid volume-to-surface liver registration task and V2S-Net, the primary voxel-based baseline PIVOTS is compared against and outperforms.","marker":"Pfeiffer et al., 2020"},{"why":"It supplies the simulation pipeline from which the synthetic training data, the Synthetic test set, and the AMOS deformations are generated.","marker":"Pfeiffer et al., 2025"},{"why":"It provides Dynamic Graph CNN, the encoder backbone that extracts multi-resolution local geometric features from the input point clouds.","marker":"Wang et al., 2019"},{"why":"It provides the real CT liver segmentations used to construct the semi-synthetic AMOS evaluation set and the AMOS noise benchmark.","marker":"Ji et al., 2022"},{"why":"It describes the abdominal phantom and deformation setup that produces the Phantom dataset with known internal landmark positions.","marker":"Zhylka et al., 2024"},{"why":"It supplies PBSM, the physics-based shape-matching baseline whose real-data performance PIVOTS is compared against.","marker":"Suwelack et al., 2014"},{"why":"It supplies BCF_FEM, the boundary-constraint-free biomechanical baseline that PIVOTS beats on all four datasets.","marker":"Yang et al., 2024"},{"why":"It supplies Lepard, a learning-based non-rigid point-cloud registration baseline used to show that surface-to-surface methods do not transfer directly to volume-to-surface alignment.","marker":"Li and Harada, 2022a"}],"fun_headline_variants":["Tackles liver registration from sparse, noisy point clouds","Robust liver alignment from partial views via cross-attention","Fast volume-to-surface registration for liver navigation","Synthetic-trained net beats baselines on liver registration","PIVOTS: point-cloud liver alignment under deformation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic biomechanical simulator—random organ-like shapes, a homogeneous Neo-Hookean liver with randomized material parameters, simplified gravity/fixed/spring boundary conditions, and Perlin/Gaussian surface noise—produces deformation patterns close enough to real intraoperative liver motion that a network trained only on those samples keeps its accuracy on patients.","fun_headline_variants_meta":{"raw":{"variants":["Tackles liver registration from sparse, noisy point clouds","Robust liver alignment from partial views via cross-attention","Fast volume-to-surface registration for liver navigation","Synthetic-trained net beats baselines on liver registration","PIVOTS: point-cloud liver alignment under deformation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000396,"raw_usage":{"total_tokens":2134,"prompt_tokens":1060,"completion_tokens":1074,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":998}},"tokens_in":676,"tokens_out":1074,"duration_ms":10480,"temperature":1.0,"reasoning_tokens":998,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:44:31.487668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct clinical test would settle it: collect intraoperative liver surfaces during laparoscopic surgery with tracked landmarks or embedded fiducials, run PIVOTS, and measure landmark target registration error. If mean TRE exceeds 10 mm on such unseen surgical data, or if errors grow steeply when pre-registration displacement exceeds 80 mm or visible surface falls below 10%, the claimed synthetic-to-real robustness does not transfer.","supporting_citations":[{"cited_title":", author Riediger, C","cited_arxiv_id":null,"evidence_quote":"It establishes the non-rigid volume-to-surface liver registration task and V2S-Net, the primary voxel-based baseline PIVOTS is compared against and outperforms."},{"cited_title":", author Güttner, B","cited_arxiv_id":null,"evidence_quote":"It supplies the simulation pipeline from which the synthetic training data, the Synthetic test set, and the AMOS deformations are generated."},{"cited_title":", author Smit, J","cited_arxiv_id":null,"evidence_quote":"It describes the abdominal phantom and deformation setup that produces the Phantom dataset with known internal landmark positions."},{"cited_title":"o hl, S. , author Bodenstedt, S. , author Reichard, D. , author Dillmann, R. , author dos Santos, T. , author Maier-Hein, L. , author Wagner, M. , author W \\","cited_arxiv_id":null,"evidence_quote":"It supplies PBSM, the physics-based shape-matching baseline whose real-data performance PIVOTS is compared against."},{"cited_title":", author Simon, R.A","cited_arxiv_id":null,"evidence_quote":"It supplies BCF_FEM, the boundary-constraint-free biomechanical baseline that PIVOTS beats on all four datasets."}],"review_version":1}