{"id":"ac4d60a4-bc33-459b-893b-38639392ec2a","arxiv_id":"2509.01402","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Applying SparseOcc occupancy fields to 20 segmented CT ribcages gives better Chamfer reconstruction than DiGS and NeuralPull, and the resulting field can be skeletonized via Laplacian contraction.","lead":"This paper applies an existing neural occupancy field method, SparseOcc, to reconstruct ribcages as continuous 3D fields and then extracts ribcage skeletons with Laplacian contraction. It reports better Chamfer distances than two neural field baselines on 20 CT scans, but the contribution is largely a new application of known techniques.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Medial axis extraction is presented with no quantitative validation; RibSeg v2 centerlines exist but are unused, so the skeletonization half of the central claim is unverified.","rationale":"The reader's weakest assumption was the same-cloud evaluation protocol. That is a real limitation, but for the narrow surface-reconstruction task it is a standard benchmark and the paper explicitly acknowledges the curated-data limitation. The more decisive deficiency is the complete absence of any quantitative evaluation for the medial axis output, which is half of the stated contribution. The paper even thanks RibSeg v2 for 'anatomical centerlines,' demonstrating that a suitable ground truth exists and was not used. This gap directly undermines the title, abstract, and Section 4's claim of being a methodology for medial axis extraction. The concrete test I propose—comparing the extracted skeleton to RibSeg v2 centerlines—would settle the concern; if the skeleton aligns with true centerlines, the central claim holds, and if it does not, the paper's contribution is substantially weakened. Since the reader already issued CONDITIONAL, my focused concern does not change the verdict but sharpens the condition: require quantitative skeleton validation.","tokens_in":5311,"tokens_out":4825,"duration_ms":59352,"concrete_test":"Obtain the RibSeg v2 centerline annotations for the 20 evaluated ribcages. For each scan, compute the medial axis from the SparseOcc reconstruction using the paper's Laplacian contraction, then measure the symmetric point-to-curve distance between each extracted rib branch and the corresponding anatomical centerline. Compare this error to a baseline skeleton (e.g., L1-medial skeleton computed directly on the input point cloud) and report mean/std. If the RibPull skeleton error is not substantially lower than the baseline or exceeds a clinically relevant threshold (e.g., 2× voxel spacing), the medial-axis claim in Section 3.2/Figure 3 is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim comprises two pillars: accurate reconstruction and medial axis extraction via Laplacian contraction. The reconstruction pillar is supported by Chamfer metrics in Table 1, though these are computed against the same curated point cloud used as training input and SparseOcc actually loses on Hausdorff distance (12.6 vs 8.25 for NeuralPull), suggesting missing geometry. The skeletonization pillar rests entirely on Figure 3, with no quantitative metric. This is a load-bearing gap because the paper's title, abstract, and Section 4 ('first ISR methodology ... to reconstruct complex anatomical structures') explicitly foreground medial axis extraction. Crucially, the RibSeg v2 dataset—which the authors thank in the Acknowledgements—provides anatomical centerline annotations for each rib, an ideal ground truth for exactly this evaluation. Without comparing the extracted skeleton to these centerlines, the visual result could be a plausible-looking but anatomically incorrect curve, especially given the higher Hausdorff error. The claim in Section 5 that 'Laplacian-based contraction performs well for medial axis extraction' is therefore unsupported by evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RibPull, a pipeline that applies the existing SparseOcc neural occupancy field method to radiologist-annotated ribcage point clouds from the RibSeg/RibFrac dataset, extracts an isosurface via Marching Cubes, and then applies Laplacian-based contraction to obtain a medial-axis skeleton. The method is compared with DiGS and NeuralPull on 20 CT scans using Chamfer-L1, Chamfer-L2, and Hausdorff distances; the reported Chamfer scores favor SparseOcc while Hausdorff favors NeuralPull. The skeletonization result is shown only qualitatively in Figure 3. The paper claims to be the first implicit surface reconstruction methodology for medical imaging that leverages occupancy fields and emphasizes the benefits of continuous coordinate-based representations for geometrical operations.","tokens_in":5578,"tokens_out":4580,"duration_ms":55053,"significance":"If the reconstruction and skeletonization claims were properly validated, this would be a useful demonstration that occupancy-field representations designed for sparse point clouds can handle thin, tubular ribcage geometry and enable downstream morphological analysis. The use of a public benchmark (RibSeg/RibFrac), the reporting of multiple distance metrics, and the authors' explicit acknowledgment of limitations are strengths. However, the current evidence is not sufficient to support the central claims: the reconstruction evaluation is self-referential, and the medial-axis extraction is not quantitatively assessed. The reported Hausdorff result and the large variances of DiGS also complicate the comparison. The contribution is largely an application of the authors' prior SparseOcc method; as such, the paper's novelty and evaluation need substantial strengthening before the claims can be accepted.","major_comments":[{"comment":"The evaluation is circular: the same manually annotated point clouds are used both as training input and as ground truth for the Chamfer/Hausdorff distances. The reported Chamfer-L1/L2 therefore measure how well the network fits its own training samples, not reconstruction accuracy on unseen or noisy clinical segmentations. This is load-bearing for the central claim that SparseOcc 'handle[s] sparse and noisy data' (Section 2.2). Please evaluate on held-out point subsets, on noisified inputs, or against an independent dense ground truth (e.g., high-resolution CT meshes), and report per-case distributions.","section":"§3.2, Table 1"},{"comment":"Medial axis extraction is validated only by a single visual example. The statement in Section 5 that 'Laplacian-based contraction performs well for medial axis extraction' is unsupported. Since the authors acknowledge RibSeg v2, which provides anatomical centerline annotations for each rib, a quantitative comparison between the extracted skeleton and these centerlines is both feasible and necessary. Without it, the skeletonization half of the paper's title and abstract remains unverified.","section":"§3.2, Fig. 3 and §5"},{"comment":"SparseOcc has a worse Hausdorff distance than NeuralPull (12.6 ± 4.39 vs. 8.25 ± 3.22), indicating that some surface regions are missed in the worst case. This conflicts with the claim in Section 5 that SparseOcc 'accurately preserv[es] the topological and geometrical structure.' The authors should analyze this discrepancy and discuss whether the lost regions correspond to clinically important anatomy. DiGS also exhibits extreme variance (Chamfer-L1 8.04 ± 10.1); reporting only means and standard deviations hides what is likely instability. Per-case results or a more robust aggregation are needed.","section":"Table 1"},{"comment":"The method described in Section 2.2 is SparseOcc [15]; Equations (1) and (2) are presented without modification or a new derivation. Section 4 nevertheless introduces 'RibPull' as 'the first ISR methodology for medical imaging that leverages occupancy fields.' The novelty should be clarified: either provide a concrete adaptation, ablation, or improvement over SparseOcc for the ribcage domain, or reframe the paper as an application study. A comparison with existing medical-imaging implicit representations (beyond acquisition-oriented works such as MedNeRF [9]) is needed to support the 'first' claim.","section":"§2.2, §4"}],"minor_comments":[{"comment":"Typo: 'compesate' should be 'compensate'.","section":"Abstract"},{"comment":"Inconsistent naming: 'Neurall-Pull' in the text and 'N-Pull' in the table caption; please unify. Also clarify what 'normalization applied' means (e.g., bounding sphere scale) and whether the 100,000 sampled points are from the reconstructed surface or the ground truth.","section":"Section 3.2 / Table 1"},{"comment":"In Equation (1), U_theta(q) appears to be a scalar probability in the text but is used in a vector expression. Please clarify the notation, or state that the gradient term is the vector direction and U_theta is scalar.","section":"Section 2.2, Eq. (1)"},{"comment":"The caption says 'The progression shows surface reconstruction to skeleton extraction,' but the figure shows only three reconstruction results, not skeletonization. The caption should match the figure content.","section":"Figure 2 caption"},{"comment":"The Laplacian contraction weights W_L and W_H are mentioned in Equations (3) and (4) but no values or selection procedure are given. Please report the hyperparameters used for the ribcage experiments to enable reproducibility.","section":"Section 2.3"},{"comment":"No network architecture, training iterations, learning rate, or runtime are reported. Since the paper promises code release, these details are important for reproducibility even before code is available.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core method is the authors' own prior work: co-author Ouasfi is also an author of SparseOcc [15], and the manuscript does not explicitly disclose this relationship beyond the citation. The paper is essentially an application study of SparseOcc to ribcage CT data, with a qualitative skeletonization demonstration. The technical evaluation is currently circular for reconstruction and absent for skeletonization; both are fixable with additional experiments, but the novelty framing must also be recalibrated. If the authors can add quantitative centerline evaluation and a non-circular reconstruction protocol, and clearly position the contribution with respect to SparseOcc, the paper may become acceptable; in its current form it does not support its central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an application paper. The core reconstruction piece is SparseOcc (Ouasfi and Boukhayma, 2024), and the medial axis extraction is standard Laplacian contraction (Cao et al., 2010). What is new is an empirical comparison of SparseOcc against DiGS and NeuralPull on 20 ribcage point clouds from RibSeg, plus a visual demonstration of skeletonization on the resulting field. The reconstruction numbers are real for that narrow setting, and the skeletonization idea is a reasonable use of a continuous representation. But the evaluation protocol is self-referential, the skeletonization is not validated at all, and the novelty claim in Section 4 is overstated.\n\nThe good part: Table 1 gives us a clean head-to-head on a real medical dataset. SparseOcc wins Chamfer-L1 and L2 by a large margin (1.55 vs 5.81 vs 8.04 for L1). That is not nothing. The paper also shows that a continuous field can support Laplacian contraction to produce a plausible ribcage skeleton, which is the sort of morphological operation voxel grids do poorly. If the authors release the code, others can build on this.\n\nThe soft spots are significant. First, the same annotated point cloud is used both to train the field and as ground truth for the metrics. The networks overfit to the input cloud, so the Chamfer numbers measure fitting capacity, not generalization to unseen or noisy CT data. The authors acknowledge the curation issue in the conclusion but do not adjust the claims. Second, SparseOcc loses on Hausdorff distance (12.6 vs 8.25 for NeuralPull), which suggests the reconstruction misses parts of the geometry. That is worth discussing. Third, and most important: the skeletonization pillar of the central claim has no quantitative evaluation. Figure 3 is a picture. The acknowledgements thank RibSeg v2, which provides anatomical centerlines for each rib. Those centerlines are the obvious ground truth for exactly this comparison, and they are unused. Without that, the claim that 'Laplacian-based contraction performs well for medial axis extraction' is unsupported. A plausible-looking curve could be anatomically wrong.\n\nThe novelty overclaim is avoidable. SparseOcc is cited, but calling the combination 'the first ISR methodology for medical imaging that leverages occupancy fields' is not defensible given the paper's own references. A revised version should reposition the contribution as an application study.\n\nBottom line: the paper is a useful application note for people working on medical shape representation, but it needs major revision before it supports its own title. The reconstruction experiment is worth reporting, and the skeletonization can be made rigorous quickly using existing centerline annotations. I would send it to peer review with clear instructions to fix the evaluation and the skeletonization validation.","headline":"A workmanlike application of SparseOcc to ribcage CT with a solid reconstruction comparison but an unvalidated skeletonization claim.","tokens_in":6053,"tokens_out":3090,"would_cite":false,"duration_ms":34785,"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":"This paper introduces RibPull, which represents CT-scanned ribcages as continuous neural occupancy fields and claims that these implicit fields reconstruct sparse rib point clouds more accurately than SDF-learning baselines, then support me","keywords":["Implicit neural representation","Occupancy fields","CT ribcage reconstruction","Medial axis extraction","Laplacian contraction","Signed distance function","Point cloud surface reconstruction","RibSeg dataset"],"falsifier":"Run RibPull on raw, uncurated CT segmentations with added noise or artifacts, or evaluate the extracted skeleton against RibSeg v2's anatomical centerlines with a quantitative distance metric; if the Chamfer advantage over NeuralPull disappears or the skeleton deviates substantially from the centerlines, the central claim is weakened.","tokens_in":5238,"feed_emoji":"🦴","tokens_out":4350,"duration_ms":46543,"temperature":0.7,"pith_summary":"RibPull argues that implicit occupancy fields—not voxel grids—are the right representation for CT ribcage geometry. The method takes radiologist-annotated rib point clouds, trains a neural occupancy network using SparseOcc's margin-uncertainty loss, converts the learned field into a signed distance function, and extracts the medial axis with Laplacian-based contraction. On 20 scans from the RibFrac-derived RibSeg dataset, the paper reports Chamfer-L1 of 1.55 ± 0.24, substantially lower than NeuralPull's 5.81 ± 0.38, and a roughly 57% memory reduction from input point cloud to network weights. If the claim holds, continuous coordinate-based representations make morphological operations like skeletonization practical on medical anatomy, with downstream uses in fracture detection, scoliosis assessment, and surgical planning.","feed_headline":"Implicit fields rebuild CT ribcages four times closer to ground truth","feed_subtitle":"A neural occupancy field turns sparse rib point clouds into smooth, skeletonizable surfaces.","key_machinery":"The load-bearing object is the margin uncertainty function Uθ(x) of a neural occupancy field, defined as the difference between the probability of a query point being outside versus inside the shape; sampling its zero set approximates the surface, and a Newton-Raphson update pulls those samples toward the input point cloud. This converts sparse, noisy point clouds into a continuous field, after which Laplacian-based contraction—solving the linear system with cotangent Laplacian and balance weights—collapses the shape toward its medial axis.","core_discovery":"The central claim is that a binary neural occupancy field trained with SparseOcc's pull-and-regularize loss can encode a CT ribcage as a continuous function accurate enough that Marching Cubes reconstruction beats SDF-learning baselines on Chamfer distances, and smooth enough that Laplacian-based contraction yields the ribcage skeleton. The field's margin uncertainty—the difference between predicted outside and inside probabilities—is used to sample surface points via Newton-Raphson iteration, and those samples are pulled toward the input point cloud while entropy regularization keeps the field stable. Converting this occupancy field to a signed distance representation then permits arbitrary","pith_inferences":["Because the evaluation reconstructs from and measures against the same manually curated point cloud, the reported Chamfer numbers likely reflect fitting capacity on clean annotations; the paper itself flags that real-world noisy CT segmentations may not behave the same.","The same occupancy-field pipeline could extend to other thin tubular anatomies—coronary arteries, bronchi, colon—where voxel representations suffer similar discretization artifacts.","The skeletonization result is shown qualitatively; comparing extracted skeletons against RibSeg v2's anatomical centerlines would quantify clinical usefulness.","The continuous field could be queried for normals and curvature, enabling biomechanical analyses the paper does not evaluate."],"forward_implications":["Continuous occupancy-to-SDF representation allows resolution-independent queries, avoiding the staircase artifacts and topological inconsistencies of voxel grids.","Laplacian contraction applied to the implicit field produces ribcage medial axes, a morphological operation that is impractical on discrete voxel representations.","The network weights store the ribcage scene in about 1.8 MB versus 4.2 MB for the input point cloud, a roughly 57% memory reduction.","If the reconstruction accuracy holds, the same pipeline could support fracture detection, scoliosis assessment, and surgical planning from CT scans.","The method is positioned as a first step toward compressed-sensing CT reconstruction and broader anatomical structure reconstruction."],"supporting_citations":[{"why":"Supplies the SparseOcc occupancy-learning method, including the margin-uncertainty loss and entropy regularization that RibPull adapts.","marker":"[15]"},{"why":"Provides the RibFrac CT chest scans that form the underlying dataset source.","marker":"[10]"},{"why":"Provides the RibSeg ribcage segmentation model and the radiologist-annotated point clouds used as training input and ground truth.","marker":"[11]"},{"why":"Serves as the NeuralPull baseline whose Chamfer and Hausdorff distances RibPull compares against.","marker":"[19]"},{"why":"Serves as the DiGS baseline, another implicit neural representation for unoriented point clouds.","marker":"[14]"},{"why":"Provides the Laplacian-based contraction algorithm used for medial axis extraction from the continuous field.","marker":"[18]"}],"fun_headline_variants":["RibPull: neural occupancy fields rebuild ribcages 4x sharper","Implicit fields turn sparse CT rib points into smooth skeletons","Occupancy fields beat SDF baselines on CT ribcage accuracy","Neural fields extract ribcage medials from sparse CT data","RibPull: 4x tighter ribcage meshes via implicit occupancies"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported accuracy is measured by reconstructing from and comparing against the same manually curated, clean point cloud, so the numbers show how well the network fits clean annotations rather than how it handles noisy, real-world CT segmentations.","fun_headline_variants_meta":{"raw":{"variants":["RibPull: neural occupancy fields rebuild ribcages 4x sharper","Implicit fields turn sparse CT rib points into smooth skeletons","Occupancy fields beat SDF baselines on CT ribcage accuracy","Neural fields extract ribcage medials from sparse CT data","RibPull: 4x tighter ribcage meshes via implicit occupancies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1196,"prompt_tokens":747,"completion_tokens":449,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":353}},"tokens_in":491,"tokens_out":449,"duration_ms":5180,"temperature":1.0,"reasoning_tokens":353,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:33:34.563613+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RibPull on raw, uncurated CT segmentations with added noise or artifacts, or evaluate the extracted skeleton against RibSeg v2's anatomical centerlines with a quantitative distance metric; if the Chamfer advantage over NeuralPull disappears or the skeleton deviates substantially from the centerlines, the central claim is weakened.","supporting_citations":[{"cited_title":"Unsupervised occupancy learning from sparse point cloud,","cited_arxiv_id":null,"evidence_quote":"Supplies the SparseOcc occupancy-learning method, including the margin-uncertainty loss and entropy regularization that RibPull adapts."},{"cited_title":"Deep- learning-assisted detection and segmentation of rib fractures from ct scans: Development and validation of fracnet,","cited_arxiv_id":null,"evidence_quote":"Provides the RibFrac CT chest scans that form the underlying dataset source."},{"cited_title":"Ribseg dataset and strong point cloud baselines for rib segmentation from ct scans,","cited_arxiv_id":null,"evidence_quote":"Provides the RibSeg ribcage segmentation model and the radiologist-annotated point clouds used as training input and ground truth."},{"cited_title":"Digs : Divergence guided shape implicit neural representation for unoriented point clouds,","cited_arxiv_id":null,"evidence_quote":"Serves as the DiGS baseline, another implicit neural representation for unoriented point clouds."},{"cited_title":"Point cloud skeletons via laplacian based contraction,","cited_arxiv_id":null,"evidence_quote":"Provides the Laplacian-based contraction algorithm used for medial axis extraction from the continuous field."}],"review_version":1}