{"id":"680d94e0-545e-4c4a-9070-34650e68b78c","arxiv_id":"2412.19328","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A patches-to-partial matching module improves complete-to-partial rigid point cloud registration for liver surgery at low visibility ratios, with benchmark datasets released.","lead":"This paper shows that standard learning-based point-cloud registration methods fail when the intraoperative liver surface is only partially visible, and introduces a plug-and-play patches-to-partial module that improves registration errors by about 29% (in silico) and 40% (in vitro) at low visibility.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The P2P gain hinges on the visibility-score localization (S_vis) in Sec. III-C1; under realistic anterior-only cropping, this localization could fail and remove the entire benefit. A viewpoint-based cropping test would settle it.","rationale":"The reader's weakest assumption identified exactly this dependency: the visibility score S_vis = sum_j xS xT^T must be informative enough to localize the visible region. The paper's own Sec. VI-D admits failure when features become highly inconsistent. My read agrees that this is the most load-bearing point because every downstream component (patch node generation, candidate patch construction, and final selection) presupposes that the correct anatomical region survives the top-M cutoff. The current experiments do not stress this assumption: both training and test cropping are uniform random, so the base networks have never seen the viewpoint-dependent, anterior-only visibility pattern that characterizes laparoscopic surgery. A concrete test with anterior-only cropping or in vivo porcine data would directly probe whether the reported 29-40% error reductions persist under the intended use case. Because the reader already issued a CONDITIONAL verdict and the concern is a gap to be closed by additional experiments rather than a demonstrated internal inconsistency, no change to that verdict is needed. The agreement is complete: the reader's weakest assumption is the same load-bearing concern I identify.","tokens_in":19827,"tokens_out":4452,"duration_ms":45916,"concrete_test":"Using the released code and datasets (or a faithful reimplementation), generate an additional in silico test set in which target point clouds are cropped from the anterior liver surface only, e.g., by simulating a laparoscope frustum or using the visibility masks from Heiselman et al. [17], rather than uniform random cropping. For each case in the [0.2, 0.3) visibility band, compute (a) the recall of the true visible source region among the top-M points selected by S_vis in Sec. III-C1, and (b) RMS-TRE for Lepard+P2P versus Lepard alone. If recall is substantially below 1 (e.g., <0.6) and the P2P improvement vanishes or reverses, the central claim fails under the intended clinical scenario. If recall remains high and the improvement persists, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire P2P pipeline depends on Visible Source Point Estimation (Sec. III-C1): the top-M source points by S_vis = sum_j xS xT^T are the only candidates from which patch nodes are drawn (III-C2). If the true visible region is not among these top-M points, no later step can recover it; Eq. 4 only selects among wrong hypotheses. The paper itself concedes in Sec. VI-D that when source and target features become highly inconsistent, this localization fails and the correct anatomical region may be absent from all candidate patches. This is load-bearing because the reported gains (Lepard 9.47 to 6.73 mm in silico; 20.71 to 12.45 mm in vitro) are attributed to P2P, and the experimental protocol crops targets uniformly at random (IV-A1, IV-B6), which is the same distribution used to train the frozen base networks. In laparoscopic liver surgery, the visible surface is predominantly the anterior surface from a limited viewpoint [17], a systematic domain shift. Under that shift, base-network features may rank wrong anatomical regions first, so the correct patch is never proposed and P2P cannot improve over baseline. This is not merely a deployment gap; it targets the mechanism responsible for the observed improvement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses rigid registration of a complete preoperative liver surface to a partial intraoperative point cloud, which the authors term complete-to-partial ambiguity. They construct in silico (1,320 training and 121 testing liver models with deformations) and in vitro (four phantoms, 800 samples) benchmarks, evaluate GO-ICP, RegTr, RoITr, LiverMatch, and Lepard, and propose a plug-and-play patches-to-partial (P2P) module. P2P estimates visible source points via feature similarity (S_vis), generates K candidate patches, matches each to the target using dual-softmax and weighted SVD, and selects the best transformation via a closest-point-distance rule. The module improves RMS-TRE at visibility 0.2-0.3 from 9.47 to 6.73 mm (Lepard, in silico) and from 20.71 to 12.45 mm (in vitro), with reported statistical significance.","tokens_in":20084,"tokens_out":7716,"duration_ms":66924,"significance":"If the results hold, this is a useful contribution: it identifies a practical failure mode of learning-based registration methods for liver surgery, provides a benchmark with fiducial-based RMS-TRE evaluation, and proposes a simple, differentiable, and parallelizable module with modest computational overhead. The release of code and datasets is a concrete strength. The evaluation includes ablations, sensitivity analysis, a Procrustes reference, and tests on two datasets. However, the central claim that the module resolves the ambiguity in liver surgery is only tested under random-cropped target clouds, which may not reflect the clinically relevant anterior-only visibility, and one statistical test appears misapplied.","major_comments":[{"comment":"The target point clouds in both datasets are generated by random cropping of the full liver surface. The paper itself notes in VI-C that in laparoscopic surgery only the anterior liver surface is visible, but no experiment uses viewpoint-based or anterior-only cropping. This is load-bearing because P2P's visible-source-point estimation (S_vis) ranks source points by the sum of feature similarities; under a systematic shift from random crops to anterior-only views, the top-M points may not cover the true visible region, so the candidate patches (III-C2) would not contain the correct anatomy and Eq. 4 would select among incorrect hypotheses. I recommend adding an anterior-only or camera-viewpoint-based cropping test on the existing meshes, or explicitly limiting the claims to the random-crop scenario.","section":"IV-A1 and VI-C (also III-C1)"},{"comment":"The paper uses the Wilcoxon Rank Sum test to compare baseline versus baseline+P2P. Because the same target point clouds are used in both conditions, the observations are paired; the appropriate test is the Wilcoxon signed-rank test (or a paired permutation test). Please re-run the statistical analysis with a paired test and report the resulting p-values or effect sizes.","section":"IV-B5"},{"comment":"The paper does not directly evaluate whether the visibility score S_vis correctly localizes the true visible region. Since every downstream step (patch node generation, patch matching, and selection in Eq. 4) depends on this localization, I suggest reporting a diagnostic such as the recall of the ground-truth visible source points among the top-M candidates, or the fraction of cases in which the correct patch is among the K candidates. This would directly address the failure mode acknowledged in VI-D and would strengthen the claim that the observed improvement is attributable to the proposed mechanism.","section":"III-C1 and V-C"}],"minor_comments":[{"comment":"The random cropping procedure is referenced to previous work but not described; please specify how the crop is generated (e.g., random sphere center and radius) so that the benchmark is reproducible.","section":"IV-A1"},{"comment":"Please clarify whether the 200 target point clouds per in vitro phantom are generated with the same random-cropping protocol and visibility range as the in silico data.","section":"IV-A2"},{"comment":"The statement 'confirmed to be statistically significant (p < 0.05)' is not sufficient; report the test statistic or a confidence interval for the key comparisons.","section":"Tables I and VII"},{"comment":"The justification for K=5 based on 20% visibility is heuristic; consider softening the wording or providing quantitative support.","section":"VI-A"},{"comment":"The Distmin operator should be defined explicitly as the nearest-neighbor distance from each target point to the transformed source point cloud, and tie-breaking or handling of empty sets should be stated.","section":"Eq. (4)"},{"comment":"Minor typographical issues: 'learnable-parameter-free' is awkward; 'In Vitro' appears inconsistently capitalized; Section VI-B refers to 'P2P' without redefinition.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound within its controlled setting, but the central claim is broader than the evidence. The main concern is the domain gap between random cropping and clinically realistic anterior-only visibility; the authors already acknowledge this in VI-C and VI-D, so the missing experiment is a gap rather than an internal inconsistency. The statistical test issue is also fixable. Given the available code and datasets, major revision seems appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. This is a genuinely useful plug-and-play fix for a real failure mode: on complete-to-partial liver registration at 20-30% visibility, adding the patches-to-partial (P2P) module to Lepard cuts RMS-TRE from 9.47 to 6.73 mm in silico and from 20.71 to 12.45 mm in vitro, with statistical significance, and it helps LiverMatch too. Second, the evaluation is honest but narrower than the clinical framing: the synthetic target clouds are random crops, not anterior-only camera views, and there is no in vivo validation.\n\nWhat is actually new: the P2P module itself, which turns a complete-to-partial problem into a set of patch-to-partial matching hypotheses. It uses the base network's frozen point-wise features to estimate a visibility score, proposes K patches from the top-scoring source points, matches each patch to the target, and selects via closest-point distance. No new learnable parameters, negligible runtime, and it slots into existing correspondence pipelines without retraining. That is a clean idea and the paper demonstrates it on two datasets with fiducial-based RMS-TRE, ablations, sensitivity analysis, and statistical tests. The benchmark itself, 13,200 simulated deformations plus phantom data, is a useful community resource if it is actually released.\n\nThe soft spots are real but mostly acknowledged by the authors. The stress-test worry about Sec. III-C1 is on target: the whole pipeline depends on Visible Source Point Estimation localizing the right region, and the paper's own Sec. VI-D concedes that when source and target features are inconsistent, the correct anatomical region may not appear in any candidate patch. Because the synthetic training and testing both use random cropping, the reported gains come from the distribution the base networks are already tuned to. Anterior-only cropping, which is the actual laparoscopic scenario, is a systematic shift that could break the localization and erase most of the benefit. The manually cropped phantom examples in Fig. 10 suggest the module does not instantly collapse, but they are qualitative and few. I would want a viewpoint-based cropping experiment before believing the clinical claims. No code or data was available at review time despite the release promise, which is a reproducibility gap.\n\nWho this is for: people working on point cloud registration for image-guided surgery, especially liver, will get real value from the benchmark and the module design. It deserves a serious referee. My recommendation: send it to peer review, and ask for the anterior-crop experiment, code release, and ideally a small in vivo feasibility study. The central idea is sound and the limitations are disclosed, not hidden.","headline":"Genuinely useful plug-and-play fix for low-visibility liver registration; the open question is whether it survives realistic anterior-only crops.","tokens_in":20613,"tokens_out":2740,"would_cite":true,"duration_ms":25473,"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":"The paper claims that a patches-to-partial matching module, plugged into existing correspondence-based registration networks, cuts rigid alignment error at 20–30% liver surface visibility from 9.47 mm to 6.73 mm in silico and from 20.71…","keywords":["point cloud registration","complete-to-partial ambiguity","liver surgery","image-guided surgery","patches-to-partial matching","rigid registration","laparoscopic surgery","deep learning"],"falsifier":"Construct a low-visibility test pair (visibility ratio in [0.2, 0.3)) where the target is a genuinely unseen region of the liver with large deformation, and inspect whether the top-$M$ source points chosen by $S_{\\mathrm{vis}}$ contain the true target region; if the chosen set misses it and the mean closest-point selection rule cannot recover the correct transformation, the module's core premise is refuted. A direct ablation replacing the visibility-score localization with randomly sampled patches would also show whether localization, rather than the multi-hypothesis matching, is the reason for the gains.","tokens_in":19636,"feed_emoji":"🧩","tokens_out":6852,"duration_ms":54850,"temperature":0.7,"pith_summary":"The paper targets a specific failure mode in image-guided liver surgery: before any deformation correction, the preoperative full liver surface (source) must be rigidly aligned to a partial intraoperative view (target), often covering only 20–30% of the organ. It argues that state-of-the-art learning-based correspondence registration methods, built and trained for partial-to-partial natural scenes, break down under this complete-to-partial ambiguity because the smooth liver surface offers many regions that can look like the visible patch. To fix this, it proposes a patches-to-partial (P2P) module that cuts candidate regions out of the source cloud, matches the target to each candidate patch separately, and keeps the rigid transformation that minimizes the mean closest-point distance to the target. Plugging this module into Lepard and LiverMatch reduces RMS-TRE at visibility ratio [0.2, 0.3) from 9.47 mm to 6.73 mm (in silico) and from 20.71 mm to 12.45 mm (in vitro), with significance at p < 0.05. If this holds, fully automatic initial rigid registration becomes substantially more reliable for laparoscopic liver procedures.","feed_headline":"Patches-to-partial matching cuts low-visibility registration error 40%","feed_subtitle":"Adding the module to Lepard drops RMS error from 9.5 mm to 6.7 mm in silico and 20.7 mm to 12.5 mm in vitro.","key_machinery":"The central mechanism is the patches-to-partial (P2P) module, which converts the complete-to-partial registration into several smaller complete-to-complete registrations. It computes a visibility score per source point by summing target-wise feature similarities, selects the most promising source points, samples k patch centers by farthest-point sampling, forms k patches with the same number of points as the target, and runs dual-softmax matching plus weighted SVD per patch; a distance-based rule then picks the best rigid transform. The module is learnable-parameter-free, fully differentiable, and parallelizable, so it can wrap any correspondence-based registration network that outputs point-wise features and a candidate transformation.","core_discovery":"The central claim is that complete-to-partial ambiguity is not an information problem that needs a new network architecture; it is a matching-scope problem that can be fixed by converting the complete-to-partial registration into patches-to-partial registration. The P2P module computes a visibility score per source point as the summed similarity between that source feature and all target features, keeps the top source points, applies farthest-point sampling to choose k patch centers, builds k patches with as many points as the target cloud, matches each patch to the target with dual-softmax and weighted SVD, and selects the transformation with the smallest mean closest-point distance between the transformed full source and the target. The paper reports that this module, added without changing losses, architectures, or outlier rejection, consistently improves Lepard and LiverMatch across visibility ratios, noise levels, and deformation ranges, and that it outperforms a RANSAC baseline in low visibility. The paper also constructs and releases an in silico benchmark of over 1,000 liver models with more than 10,000 simulations and an in vitro phantom test set, and it uses fiducial-based RMS-TRE rather than correspondences to measure subsurface registration quality.","pith_inferences":["The same patches-to-partial conversion may transfer to other smooth, low-texture organs such as kidney and prostate, since the ambiguity is geometric rather than liver-specific; this would be a straightforward cross-organ re-run of the benchmark.","Because the module only needs point-wise features and a candidate transformation, a testable extension is to apply it to correspondence-based methods not fine-tuned on liver data, to see whether the visibility score still localizes enough to help.","The K=5 default rests on the 20% minimal visibility ratio; in even lower-visibility settings, the same logic would push K higher, and the paper's sensitivity curve suggests doing so costs only compute.","If the visibility-score localization is the true load-bearing part, then improving feature consistency (for example through stronger cross-attention in the host network) should yield further gains without changing the P2P selection logic."],"forward_implications":["At visibility ratio [0.2, 0.3), adding P2P to Lepard lowers RMS-TRE from 9.47 mm to 6.73 mm on the in silico test set and from 20.71 mm to 12.45 mm on the in vitro phantoms.","The same plug-in cuts LiverMatch's errors from 12.85 mm to 8.74 mm in silico and from 20.54 mm to 14.97 mm in vitro, with p < 0.05.","Because the module reuses the network's extracted features and only resamples and rematches, it adds about 0.02–0.03 seconds per case and is fully differentiable and parallelizable, so it can wrap future correspondence-based networks unchanged.","The success rate at a 20 mm threshold rises in low-visibility, low-deformation cases (LiverMatch from 84.38% to 95.54%; Lepard from 94.49% to 98.66%), supporting use as an automatic initialization for downstream non-rigid registration.","The constructed in silico benchmark of over 1,000 liver models with more than 10,000 simulations, together with the in vitro phantom test set, gives other researchers a common platform for measuring complete-to-partial liver registration."],"supporting_citations":[{"why":"Lepard provides one of the two baseline feature extractors the P2P module is plugged into; its dual-softmax matching and point features carry the module's matching step.","marker":"[13]"},{"why":"LiverMatch, the liver-specific baseline, is the other host network; its feature descriptors are re-sampled into candidate patches for P2P.","marker":"[22]"},{"why":"Medical Segmentation Decathlon CT volumes are the anatomical source for generating the in silico liver models and test set.","marker":"[43]"},{"why":"The non-rigid biomechanical deformation simulation pipeline generates the deformed liver surfaces used for training and testing.","marker":"[21]"},{"why":"This clinical characterization supplies the premise that only about 20–30% of the liver surface is visible in laparoscopic surgery.","marker":"[17]"},{"why":"The phantom dataset provides the four undeformed/deformed phantoms with embedded fiducial markers used for in vitro testing.","marker":"[48]"},{"why":"RoITr is the node-to-group baseline that fails at low visibility, showing that coarse-scale patches alone do not solve the complete-to-partial ambiguity.","marker":"[28]"}],"fun_headline_variants":["Patches-to-partial matching resolves low-visibility ambiguity","P2P module converts complete-to-partial to patches-to-partial","Visibility-aware patches improve liver registration under occlusion","New matching scope cuts error in low-visibility liver registration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the source and target point features produced by the frozen registration network are similar enough that the visibility score $S_{\\mathrm{vis}}=\\sum_j \\mathbf{x}^S(\\mathbf{x}^T)^\\top$ localizes the true visible region of the source liver; the paper itself notes in Section VI-D that when features become highly inconsistent this localization can fail and exclude the correct anatomical region from all candidate patches.","fun_headline_variants_meta":{"raw":{"variants":["Patches-to-partial matching resolves low-visibility ambiguity","P2P module converts complete-to-partial to patches-to-partial","Visibility-aware patches improve liver registration under occlusion","New matching scope cuts error in low-visibility liver registration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1505,"prompt_tokens":1031,"completion_tokens":474,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":647,"tokens_out":474,"duration_ms":5052,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:40:33.217533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a low-visibility test pair (visibility ratio in [0.2, 0.3)) where the target is a genuinely unseen region of the liver with large deformation, and inspect whether the top-$M$ source points chosen by $S_{\\mathrm{vis}}$ contain the true target region; if the chosen set misses it and the mean closest-point selection rule cannot recover the correct transformation, the module's core premise is refuted. A direct ablation replacing the visibility-score localization with randomly sampled patches would also show whether localization, rather than the multi-hypothesis matching, is the reason for the gains.","supporting_citations":[{"cited_title":"Lepard: Learning partial point cloud matching in rigid and deformable scenes,","cited_arxiv_id":null,"evidence_quote":"Lepard provides one of the two baseline feature extractors the P2P module is plugged into; its dual-softmax matching and point features carry the module's matching step."},{"cited_title":"Learning feature descriptors for pre-and intra-operative point cloud matching for laparoscopic liver registration,","cited_arxiv_id":null,"evidence_quote":"LiverMatch, the liver-specific baseline, is the other host network; its feature descriptors are re-sampled into candidate patches for P2P."},{"cited_title":"The medical segmentation decathlon,","cited_arxiv_id":null,"evidence_quote":"Medical Segmentation Decathlon CT volumes are the anatomical source for generating the in silico liver models and test set."},{"cited_title":"Non-rigid volume to surface registration using a data-driven biomechanical model,","cited_arxiv_id":null,"evidence_quote":"The non-rigid biomechanical deformation simulation pipeline generates the deformed liver surfaces used for training and testing."},{"cited_title":"Characterization and correction of intraoperative soft tissue deformation in image-guided laparoscopic liver surgery,","cited_arxiv_id":null,"evidence_quote":"This clinical characterization supplies the premise that only about 20–30% of the liver surface is visible in laparoscopic surgery."},{"cited_title":"Boundary constraint- free biomechanical model-based surface matching for intraoperative liver deformation correction,","cited_arxiv_id":null,"evidence_quote":"The phantom dataset provides the four undeformed/deformed phantoms with embedded fiducial markers used for in vitro testing."},{"cited_title":"Rotation- invariant transformer for point cloud matching,","cited_arxiv_id":null,"evidence_quote":"RoITr is the node-to-group baseline that fails at low visibility, showing that coarse-scale patches alone do not solve the complete-to-partial ambiguity."}],"review_version":1}