{"id":"47c04ff7-2f08-4380-9172-2da969bb7c59","arxiv_id":"2505.19518","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A patient-specific VN-OccNet completion network reconstructs full liver surfaces from partial intra-operative point clouds, reducing rigid registration TRE from ~30-40 mm to ~3-5 mm on two in vitro phantoms.","lead":"This paper tests a patient-specific deep learning pipeline that fills in missing parts of a liver surface from a partial intra-operative point cloud, then uses the completed surface to improve rigid registration against pre-operative data. On two laboratory liver phantoms, registration error dropped from roughly 30-40 mm to 3-5 mm when the completed surface was used instead of the raw partial point cloud.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported TRE gains are measured on partial clouds generated by the same viewpoint-cropping pipeline used in training; real intra-operative occlusion/noise distributions are acknowledged as out-of-distribution, so the transfer of the 5.19/3.35 mm results to clinical inputs is the unsecured…","rationale":"I agree with the reader that the simulated-to-real deformation gap is the central risk, and the paper's own Discussion flags it. My stress-test refines that concern: the relevant distributional mismatch is not only the biomechanical deformation parameters, but also the entire partial-cloud generation process. Both training and the in vitro evaluation synthesize partial inputs by viewpoint-based nearest-point cropping from clean, segmented surfaces; no test uses independently acquired partial surfaces with realistic noise, holes, or arbitrary occlusion geometry. This makes the numerical improvement in Table 2 an in-distribution demonstration of the pipeline rather than evidence of clinical transferability. The paper is honest about this limitation, and the contribution is positioned as preliminary, so the concern does not invalidate the phantom-level result. It does, however, strengthen the case for a conditional verdict: the headline TRE numbers should not be read as evidence that the method will work on real intra-operative point clouds until the occlusion and noise distribution is varied or the method is tested on acquired data. The reader's conditional verdict therefore remains appropriate. A single concrete experiment, generating test targets with a different occlusion model, would directly settle whether the completion step is robust to the distribution shift or is an artifact of the synthetic cropping procedure.","tokens_in":9203,"tokens_out":6795,"duration_ms":72534,"concrete_test":"On phantoms No.1 and No.3, generate partial target point clouds from the segmented deformed surfaces using an occlusion model not present in training, e.g., randomly placed connected surface patches covering roughly 30% of the surface, plus Gaussian positional noise of 1-2 mm, while keeping the same VN-OccNet weights and Go-ICP registration protocol. Recompute the Table 2 TRE rows. If TRE with surface completion rises toward the 30%-visibility without-completion values, the central claim is tied to the training-time viewpoint-crop distribution; if TRE remains near 5.19 mm and 3.35 mm, the completion transfers beyond the synthetic pipeline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core claim is that surface completion from a partial intra-operative point cloud improves rigid registration. For that claim to hold in the intended setting, the completion network must generalize to partial observations that are not generated by the same synthetic pipeline used at training time. Section 3.2 generates training targets by cropping the posterior surface, downsampling to D points, and extracting the N nearest points to a randomly selected viewpoint. Section 3.4 applies the same five-viewpoint procedure to the in vitro phantoms, so Table 2's 'w/ surface completion' TRE values are obtained on inputs whose occlusion pattern exactly matches the training-time generative model. The Discussion explicitly concedes that 'real intra-operative data may exhibit more complex patterns, such as noise, holes, and occlusions' and that this 'remains an open challenge.' This is more specific than the FEM parameter range alone: even if the neo-Hookean deformation space (Young's modulus 2-5 kPa, forces up to 3 N) is realistic, the artificial viewpoint-crop representation of partial visibility is a second, equally load-bearing distributional assumption. If the network has learned to complete according to that synthetic crop distribution rather than according to the physical deformation field, the registration improvement (33.13 to 5.19 mm for phantom No.1) is an in-distribution result that may not transfer to the clinical partial-view problem the paper targets.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a patient-specific pipeline for rigid registration of pre-operative liver surfaces to partial intra-operative point clouds. A VN-OccNet occupancy network is trained on simulated deformations of a patient-specific liver model to complete partial point clouds into full watertight surfaces, and the completed surface is registered to the pre-operative source with Go-ICP. The authors evaluate the rotation equivariance of VN-OccNet against OccNet under four train/test rotation settings, and report Target Registration Error (TRE) on two in vitro phantoms, showing large improvements when using completed surfaces (5.19 ± 1.34 mm and 3.35 ± 0.61 mm for phantoms No. 1 and No. 3) over the original partial clouds at 30% visibility (33.13 ± 13.89 mm and 39.87 ± 15.07 mm).","tokens_in":9506,"tokens_out":5829,"duration_ms":48792,"significance":"If the reported TRE gains hold under realistic intra-operative conditions, the approach would offer a useful step toward addressing partial visibility in image-guided liver interventions without manual initialization or rigid pre-registration. The systematic comparison of VN-OccNet and OccNet under different rotation settings is a genuine empirical contribution, and the patient-specific simulation strategy is well motivated. However, the empirical support is narrow: only two in vitro phantoms are used, the training and test partial-cloud generation pipelines are matched, and no comparison is made with prior completion-based registration methods. These limitations do not invalidate the proof-of-concept, but they substantially constrain the strength of the clinical claim as currently stated.","major_comments":[{"comment":"The TRE values in Table 2 and the Wilcoxon Rank Sum test in Section 4.2 are computed over 500 rotated target clouds per phantom, but these 500 clouds are generated by applying 100 random rotations to only five viewpoint-specific partial clouds. The repeated rotations are not independent samples; treating them as such inflates the effective sample size and the reported significance. Please report statistics per unique viewpoint or per phantom (n = 5 per phantom), or use an analysis that accounts for the repeated-measures structure, and state the number of independent cases explicitly.","section":"§3.4, Table 2"},{"comment":"The partial point clouds used to obtain the strong TRE reductions in Table 2 are generated by the same pipeline at training and test time: posterior cropping, downsampling to D points, and selecting the N nearest points to a viewpoint. The in vitro test uses the same five-viewpoint procedure, so the 30% visibility condition is exactly the distribution the network was trained on. The Discussion concedes that real intra-operative data may contain noise, holes, and occlusions not represented in this simulation. To support the claim that completion improves registration in the targeted clinical setting, the authors should evaluate on partial inputs produced by a different occlusion/noise model or on real intra-operative data, or explicitly delimit the claim to the matched synthetic occlusion setting.","section":"§3.2 and §3.4"},{"comment":"The paper positions its contribution relative to Jia et al. [5] and Foti et al. [2], stating that those methods require rigid initialization or manual correspondences, but it does not compare against them empirically. Since the central claim is that the proposed completion-based registration is an improvement over existing approaches, the absence of any such baseline makes the comparison claim unsubstantiated. At minimum, run the earlier methods on the same phantom data or include a learning-based registration baseline; otherwise soften the claim.","section":"§5 and Related Work"},{"comment":"The registration evaluation is limited to two in vitro phantoms (No. 1 and No. 3), and no real patient data are used. The two phantoms may not capture the anatomical and deformation variability needed to support the general statement that completion 'significantly enhance[s] registration performance.' Please add per-phantom variability, justify the choice of these two phantoms, and discuss how the results would transfer to other patients.","section":"§3.1 and §4.2"},{"comment":"The network is trained on deformations from a neo-Hookean FEM simulation with stiffness 2-5 kPa and forces up to 3 N, while the in vitro phantoms are deformed by physical wedges. The manuscript does not quantify the similarity between the simulated deformation space and the actual phantom deformations, so the statement in Section 5 that the model 'generalize[s] effectively to realistic deformations observed in in vitro phantoms' is supported only by indirect TRE measurements on completed surfaces. Provide a direct comparison (e.g., deformation magnitudes, surface displacement fields) or acknowledge this as an assumption.","section":"§3.1 and §5"}],"minor_comments":[{"comment":"In the second term of the Chamfer Distance definition, 'min_{hat t in P}' should be 'min_{hat t in hat T}'; the set P is undefined.","section":"§3.5, Eq. (1)"},{"comment":"In the MISE description, the notation '313 voxels' and '323' should be formatted as '31^3 voxels' and '32^3' for readability.","section":"§2.2"},{"comment":"Several numerical entries have an errant space after the decimal point (e.g., '0 .47', '3 .51', '0 .86'); these formatting issues should be corrected.","section":"Table 1"},{"comment":"The caption refers to 'five target point clouds generated for each intra-operative liver model during testing,' but it is unclear whether the figure shows in silico or in vitro models; clarify how the five viewpoints are selected and reused in Section 3.4.","section":"Figure 3 caption"},{"comment":"The phrase '100 random SO(3) rotations within the range [-pi/2, pi/2]' is imprecise; specify whether the range applies to Euler angles or axis-angle parameters.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid proof-of-concept with a clear pipeline, but the evaluation is too narrow for a journal-level claim of clinical utility. The main risk is the matched-synthetic-occlusion setup; the additional experiments suggested (out-of-distribution occlusion, proper statistical replication, baseline comparisons) should be within reach using the existing phantom data. I would not reject, but I would require those revisions before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid proof-of-concept for patient-specific surface completion as a pre-registration step in liver surgery. The new bit is pairing VN-OccNet's rotation-equivariant occupancy prediction with Go-ICP and testing on physical phantoms. The reported TRE gains are real in the narrow setting tested: 5.19 ± 1.34 mm and 3.35 ± 0.61 mm versus 33-40 mm without completion at 30% visibility. That is a big improvement, and the authors are honest that it is an in-distribution result.\n\nWhat it does well: the rotation-equivariance analysis is thorough—OccNet collapses under Z/SO(3) while VN-OccNet stays flat, and the comparison is clean. The mesh generation at 0.25 s per surface is a useful practical detail. The phantom evaluation with fiducials is the right way to measure TRE, and the paper discusses limitations openly, including the per-patient training burden and the gap to real intra-operative noise and occlusion.\n\nThe soft spots, in order. First, the test partial clouds are generated by the same viewpoint-crop pipeline used to create training targets: posterior crop, downsample to D, select N nearest points to a viewpoint. So the 30% visibility test is a same-distribution test. The authors admit this in the Discussion, but it means the headline TRE numbers should not be read as predicting performance on real laparoscopic capture. Second, there is no comparison against prior completion-based registration methods—Jia et al. and Foti et al. are discussed but not benchmarked. Third, the Wilcoxon test treats repeated random rotations of the same five viewpoint clouds as independent samples, which inflates significance. Fourth, only two of four available phantoms were used; No. 2 and No. 4 are missing without explanation. No code or data release either.\n\nStill, the central engineering claim—that completion can make rigid registration work at 30% visibility—is supported within the stated distribution. The missing baselines and distribution generalization are addressable. This deserves a serious referee, but the revision bar should include a baseline comparison and an out-of-distribution test, even a synthetic one with simulated noise and holes.\n\nWho for: groups working on image-guided liver registration or point cloud completion for medical data. I would not cite it in my own immediate work, but if I worked in this niche, it would be a useful reference. Recommended for peer review, conditional on revisions.","headline":"Patient-specific VN-OccNet completion plus Go-ICP cuts TRE from ~30-40 mm to ~3-5 mm on two phantoms, but the test partial clouds come from the same synthetic pipeline used in training, so clinical transfer is unproven.","tokens_in":10015,"tokens_out":2051,"would_cite":false,"duration_ms":17414,"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":"Completing a partially visible liver surface with a rotation-equivariant occupancy network reduces rigid-registration error from roughly 33–40 mm to 3–5 mm in phantom tests.","keywords":["image-guided liver surgery","point cloud completion","occupancy networks","SO(3)-equivariant networks","pre- to intra-operative registration","target registration error","patient-specific training","phantom validation"],"falsifier":"Take a real intra-operative liver surface where the true deformation is known from postoperative imaging or tracked fiducials, feed the partial cloud to the patient-specific network, and compare the completed surface to the true deformed surface; if the Chamfer distance or per-point error on real data is much larger than the in silico test values, or if Go-ICP on the completed surface gives TRE comparable to the raw partial cloud, the central claim would be contradicted.","tokens_in":9026,"feed_emoji":"🩺","tokens_out":8105,"duration_ms":53300,"temperature":0.7,"pith_summary":"During image-guided liver surgery, the intra-operative surface is only partially visible, which makes aligning it with the pre-operative model unreliable. This paper argues that the missing surface can be synthesized by a patient-specific occupancy network, VN-OccNet, trained on simulated deformations of the patient's own pre-operative liver model, and that registration performed on the completed surface is far more accurate than registration on the raw partial cloud. In phantom experiments, the completed surfaces produced target registration errors of 5.19 ± 1.34 mm and 3.35 ± 0.61 mm, compared to 33.13 ± 13.89 mm and 39.87 ± 15.07 mm for the original 30%-visibility point clouds. The paper also shows that VN-OccNet's rotation equivariance keeps completion quality consistent across arbitrary test rotations, where a standard OccNet degrades sharply.","feed_headline":"Completing partial liver surfaces cuts registration error sixfold","feed_subtitle":"Patient-specific completion turns 30%-visible liver clouds into full meshes, cutting alignment error to ~3–5 mm.","key_machinery":"VN-OccNet is a vector-neuron occupancy network: the encoder represents point-cloud features as vectors in SO(3)-equivariant layers, and the decoder predicts occupancy probabilities for query points, with the architecture guaranteeing that a rotation of the input point cloud merely rotates the latent features rather than changing the reconstruction. The decoder outputs occupancy for a query grid; the Multiresolution Isosurface Extraction (MISE) algorithm adaptively subdivides surface-intersecting voxels down to a 128³ resolution, and Marching Cubes turns those voxels into a watertight mesh. A threshold c = 0.4 converts occupancy probabilities into inside/outside labels. The completed mesh's vertices are sampled as a dense target point cloud, and Go-ICP performs globally optimal rigid registration against the pre-operative source cloud. The full chain—completion in about 0.25 s plus registration in about 16 s—is what the experiments evaluate.","core_discovery":"The central claim is that surface completion, rather than a better registration objective, is the effective remedy for partial intra-operative visibility. Using VN-OccNet—a rotation-equivariant occupancy network—the authors generate a complete watertight liver mesh from a partial point cloud, extract its vertices, and feed them to the Go-ICP rigid registration algorithm. On two in vitro phantoms, this reduces target registration error from roughly 33–40 mm at 30% visibility to 3–5 mm, a statistically significant improvement that persists even when the raw partial cloud is given more visibility (40% and 50%). The same network maintains its reconstruction accuracy across I/I, Z/Z, Z/SO(3), and SO(3)/SO(3) train/test configurations, whereas the non-equivariant OccNet baseline fails in the cross-rotation settings. The paper presents the result as a patient-specific pipeline: each patient's pre-operative model is deformed in simulation to create training data, so the network learns that patient's geometry and deformation patterns rather than a generic liver shape.","pith_inferences":["If the completion step transfers to real intra-operative acquisition, initial rigid registration could become fully automatic: no manual correspondences, no fiducial initialization, and no pre-alignment, because the reconstruction is rotation-agnostic.","The same recipe—simulate deformations of the patient's own pre-operative anatomy, train an occupancy network, complete, then register—could be applied to other partially visible organs or to laparoscopic video surfaces, though the FEM simulation fidelity would be the main risk.","The paper itself acknowledges (Discussion) that real intra-operative clouds contain noise, holes, and occlusions absent from the training simulation, and that training a fresh network per patient may limit practicality; these caveats define the main risk to the central claim."],"forward_implications":["At 30% intra-operative visibility, registration with a completed surface achieves TRE of 5.19 ± 1.34 mm (phantom No. 1) and 3.35 ± 0.61 mm (phantom No. 3), several times smaller than the 33–40 mm TRE obtained with the raw partial cloud.","Increasing the raw cloud's visibility from 30% to 50% only reduces TRE to about 26 mm, still far worse than the completed-surface result, so completion buys more accuracy than acquiring more surface.","Because VN-OccNet's accuracy is stable across Z-axis and SO(3) rotations without rotation data augmentation, the method removes the need to pre-align intra-operative point clouds before completion.","The watertight mesh output enables uniform surface sampling, which the authors state is important for registration methods that assume consistent point density.","The patient-specific training strategy, using simulated neo-Hookean deformations of the pre-operative model, is what allows the network to generalize to the in vitro phantom deformations without generic liver training data."],"supporting_citations":[{"why":"Defines VN-OccNet, the rotation-equivariant occupancy network used to complete partial surfaces.","marker":"[1]"},{"why":"Introduces occupancy networks and the MISE isosurface-extraction scheme used to build the mesh from predicted occupancies.","marker":"[10]"},{"why":"Supplies Go-ICP, the globally optimal rigid registration algorithm used to align the completed surface with the pre-operative source.","marker":"[16]"},{"why":"Provides the in silico and in vitro phantom dataset used for training and for the registration evaluation.","marker":"[19]"},{"why":"Describes the neo-Hookean finite-element deformation simulation that generates the patient-specific training deformations.","marker":"[11]"},{"why":"Defines the point-cloud-plus-occupancy data representation used to train the occupancy network.","marker":"[15]"},{"why":"Reports prior evaluation showing conventional point completion methods are unsuitable, motivating the rotation-equivariant choice.","marker":"[12]"}],"fun_headline_variants":["Patient-specific completion cuts liver registration error 6x","Partial liver clouds get full surfaces via VN-OccNet","Rotation-equivariant completion reduces liver registration error","Completing livers from 30% visibility cuts error to 3–5 mm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated deformations used for training—a soft-elastic finite-element model of the patient's liver, with stiffness 2–5 kPa and applied forces up to 3 N—match the way a real liver deforms during surgery; if real deformations fall outside this simulated range, the completed surfaces will be wrong and the registration gains will not transfer to patients.","fun_headline_variants_meta":{"raw":{"variants":["Patient-specific completion cuts liver registration error 6x","Partial liver clouds get full surfaces via VN-OccNet","Rotation-equivariant completion reduces liver registration error","Completing livers from 30% visibility cuts error to 3–5 mm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3387,"prompt_tokens":990,"completion_tokens":2397,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2326}},"tokens_in":606,"tokens_out":2397,"duration_ms":17053,"temperature":1.0,"reasoning_tokens":2326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:12:05.701465+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real intra-operative liver surface where the true deformation is known from postoperative imaging or tracked fiducials, feed the partial cloud to the patient-specific network, and compare the completed surface to the true deformed surface; if the Chamfer distance or per-point error on real data is much larger than the in silico test values, or if Go-ICP on the completed surface gives TRE comparable to the raw partial cloud, the central claim would be contradicted.","supporting_citations":[{"cited_title":"In: Proceedings 14 N","cited_arxiv_id":null,"evidence_quote":"Defines VN-OccNet, the rotation-equivariant occupancy network used to complete partial surfaces."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Introduces occupancy networks and the MISE isosurface-extraction scheme used to build the mesh from predicted occupancies."},{"cited_title":"IEEE transactions on pattern analysis and machine intelligence 38(11), 2241–2254 (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies Go-ICP, the globally optimal rigid registration algorithm used to align the completed surface with the pre-operative source."},{"cited_title":"IEEE Transactions on Medical Imaging pp","cited_arxiv_id":null,"evidence_quote":"Provides the in silico and in vitro phantom dataset used for training and for the registration evaluation."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part IV 23","cited_arxiv_id":null,"evidence_quote":"Describes the neo-Hookean finite-element deformation simulation that generates the patient-specific training deformations."},{"cited_title":"In- ternational Journal of Computer Vision128, 1162–1181 (2020)","cited_arxiv_id":null,"evidence_quote":"Defines the point-cloud-plus-occupancy data representation used to train the occupancy network."},{"cited_title":"In: Medical Imaging 2025: Image-Guided Procedures, Robotic Interventions, and Modeling","cited_arxiv_id":null,"evidence_quote":"Reports prior evaluation showing conventional point completion methods are unsuitable, motivating the rotation-equivariant choice."}],"review_version":1}