REVIEW 4 major objections 6 minor 20 references
RibPull: Implicit Occupancy Fields and Medial Axis Extraction for CT Ribcage Scans
T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read 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
desk verdict A workmanlike application of SparseOcc to ribcage CT with a solid reconstruction comparison but an unvalidated skeletonization claim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (4)
- [§3.2, Table 1] 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.
- [§3.2, Fig. 3 and §5] 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.
- [Table 1] 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.
- [§2.2, §4] 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.
minor comments (6)
- [Abstract] Typo: 'compesate' should be 'compensate'.
- [Section 3.2 / Table 1] 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 2.2, Eq. (1)] 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.
- [Figure 2 caption] 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 2.3] 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.
- [General] 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.
Circularity Check
No significant circularity; RibPull is an application of self-authored SparseOcc with independent comparative evaluation.
full rationale
The paper's core reconstruction method, SparseOcc, is the authors' own prior work (ref [15], co-authored by the second author), which is a notable self-citation. However, the paper does not merely assert SparseOcc's superiority; it provides a head-to-head comparison against DiGS and NeuralPull on 20 ribcage scans, with quantitative Chamfer and Hausdorff metrics. These experiments are real measurements, not consequences of the citation, and the paper even reports that NeuralPull wins on Hausdorff distance. The reconstruction task is to encode a given point cloud; the evaluation measures fit to that same cloud, which is a standard (if limited) way to assess representation capacity, though it does not test generalization. That limitation is acknowledged in Section 5 ('our evaluation was performed on datasets that have undergone manual curation... which may not reflect real-world clinical scenarios'). The skeletonization via Laplacian contraction is an existing algorithm, applied without fitting, and is presented as a qualitative demonstration; the lack of quantitative validation is a correctness risk, not circularity. No load-bearing argument reduces to a self-citation or a renamed known result. The paper's contribution is an application and empirical study, not a fresh derivation from first principles.
Assumptions & free parameters
free parameters (2)
- entropy regularization weight λ =
not reported
- Laplacian contraction weights W_L and W_H =
not reported
assumptions (4)
- standard math The eikonal equation property |∇f|=1 is used to motivate SDF conversion (Section 2.2).
- domain assumption The Laplacian contraction system (3) preserves geometry and converges to a medial axis (Section 2.3).
- domain assumption The radiologist-annotated RibSeg point clouds are a complete and accurate ground truth for ribcage anatomy (Sections 2.1, 3.1).
- domain assumption Neural occupancy fields trained by SparseOcc generalize to thin, articulated bone structures like ribs.
Cite this review
Pith. "Pith review of RibPull: Implicit Occupancy Fields and Medial Axis Extraction for CT Ribcage Scans." pith.science (2026). https://pith.science/paper/FLIOBHVG
@misc{pith2026250901402,
author = {Pith},
title = {Pith review of: RibPull: Implicit Occupancy Fields and Medial Axis Extraction for CT Ribcage Scans},
year = {2026},
howpublished = {\url{https://pith.science/paper/FLIOBHVG}},
note = {Machine review of arXiv:2509.01402}
}
read the original abstract
We present RibPull, a methodology that utilizes implicit occupancy fields to bridge computational geometry and medical imaging. Implicit 3D representations use continuous functions that handle sparse and noisy data more effectively than discrete methods. While voxel grids are standard for medical imaging, they suffer from resolution limitations, topological information loss, and inefficient handling of sparsity. Coordinate functions preserve complex geometrical information and represent a better solution for sparse data representation, while allowing for further morphological operations. Implicit scene representations enable neural networks to encode entire 3D scenes within their weights. The result is a continuous function that can implicitly compesate for sparse signals and infer further information about the 3D scene by passing any combination of 3D coordinates as input to the model. In this work, we use neural occupancy fields that predict whether a 3D point lies inside or outside an object to represent CT-scanned ribcages. We also apply a Laplacian-based contraction to extract the medial axis of the ribcage, thus demonstrating a geometrical operation that benefits greatly from continuous coordinate-based 3D scene representations versus voxel-based representations. We evaluate our methodology on 20 medical scans from the RibSeg dataset, which is itself an extension of the RibFrac dataset. We will release our code upon publication.
Figures
Reference graph
Works this paper leans on
-
[15]
Unsupervised occupancy learning from sparse point cloud,
Ouasfi, A. and Boukhayma, A., “Unsupervised occupancy learning from sparse point cloud,” in [ Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ], 21729–21739 (2024)
work page 2024
-
[9]
Corona-Figueroa, A., Frawley, J., Bond-Taylor, S., Bethapudi, S., Shum, H. P., and Willcocks, C. G., “Mednerf: Medical neural radiance fields for reconstructing 3d-aware ct-projections from a single x-ray,” in [ 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) ], 3843–3848, IEEE (2022)
work page 2022
-
[1]
Teasar: tree-structure extraction algorithm for accurate and robust skeletons,
Sato, M., Bitter, I., Bender, M. A., Kaufman, A. E., and Nakajima, M., “Teasar: tree-structure extraction algorithm for accurate and robust skeletons,” in [ Proceedings the Eighth Pacific Conference on Computer Graphics and Applications ], 281–449, IEEE (2000)
work page 2000
-
[2]
A new fuzzy skeletonization algorithm and its applications to medical imaging,
Jin, D. and Saha, P. K., “A new fuzzy skeletonization algorithm and its applications to medical imaging,” in [International Conference on Image Analysis and Processing ], 662–671, Springer (2013)
work page 2013
-
[3]
Qiu, A., Bitouk, D., and Miller, M. I., “Smooth functional and structural maps on the neocortex via orthonormal bases of the laplace-beltrami operator,” IEEE Transactions on Medical Imaging 25(10), 1296– 1306 (2006)
work page 2006
-
[4]
Anisotropic laplace-beltrami eigen- maps: bridging reeb graphs and skeletons,
Shi, Y., Lai, R., Krishna, S., Sicotte, N., Dinov, I., and Toga, A. W., “Anisotropic laplace-beltrami eigen- maps: bridging reeb graphs and skeletons,” in [ 2008 IEEE computer society conference on computer vision and pattern recognition workshops ], 1–7, IEEE (2008)
work page 2008
-
[5]
Lesa: Longitudinal elastic shape analysis of brain subcortical structures,
Zhang, Z., Wu, Y., Xiong, D., Ibrahim, J. G., Srivastava, A., and Zhu, H., “Lesa: Longitudinal elastic shape analysis of brain subcortical structures,” Journal of the American Statistical Association 118(541), 3–17 (2023)
work page 2023
-
[6]
Nerf: Rep- resenting scenes as neural radiance fields for view synthesis,
Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., and Ng, R., “Nerf: Rep- resenting scenes as neural radiance fields for view synthesis,” Communications of the ACM 65(1), 99–106 (2021)
work page 2021
Show all 20 references
-
[7]
3d gaussian splatting for real-time radiance field rendering,
Kerbl, B., Kopanas, G., Leimk¨ uhler, T., and Drettakis, G., “3d gaussian splatting for real-time radiance field rendering,” ACM Transactions on Graphics 42(4) (2023)
2023
-
[8]
Gaspct: Gaussian splatting for novel brain cbct projection view synthesis,
Nikolakakis, E., Gupta, U., Vengosh, J., Bui, J., and Marinescu, R., “Gaspct: Gaussian splatting for novel brain cbct projection view synthesis,” in [ Medical Imaging 2025: Image Processing ], 13406, 318–325, SPIE (2025)
2025
-
[10]
Deep- learning-assisted detection and segmentation of rib fractures from ct scans: Development and validation of fracnet,
Jin, L., Yang, J., Kuang, K., Ni, B., Gao, Y., Sun, Y., Gao, P., Ma, W., Tan, M., Kang, H., et al., “Deep- learning-assisted detection and segmentation of rib fractures from ct scans: Development and validation of fracnet,” EBioMedicine 62 (2020)
2020
-
[11]
Ribseg dataset and strong point cloud baselines for rib segmentation from ct scans,
Yang, J., Gu, S., Wei, D., Pfister, H., and Ni, B., “Ribseg dataset and strong point cloud baselines for rib segmentation from ct scans,” in [ Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 2...
2021
-
[12]
Pointnet++: Deep hierarchical feature learning on point sets in a metric space,
Qi, C. R., Yi, L., Su, H., and Guibas, L. J., “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” Advances in neural information processing systems 30 (2017)
2017
-
[13]
Implicit geometric regularization for learning shapes,
Gropp, A., Yariv, L., Haim, N., Atzmon, M., and Lipman, Y., “Implicit geometric regularization for learning shapes,” (2020)
2020
-
[14]
Digs : Divergence guided shape implicit neural representation for unoriented point clouds,
Ben-Shabat, Y., Koneputugodage, C. H., and Gould, S., “Digs : Divergence guided shape implicit neural representation for unoriented point clouds,” (2023)
2023
-
[16]
L1-medial skeleton of point cloud.,
Huang, H., Wu, S., Cohen-Or, D., Gong, M., Zhang, H., Li, G., and Chen, B., “L1-medial skeleton of point cloud.,” ACM Trans. Graph. 32(4), 65–1 (2013)
2013
-
[17]
Neural skeleton: Implicit neural representation away from the surface,
Cl´ emot, M. and Digne, J., “Neural skeleton: Implicit neural representation away from the surface,” Com- puters & Graphics 114, 368–378 (2023)
2023
-
[18]
Point cloud skeletons via laplacian based contraction,
Cao, J., Tagliasacchi, A., Olson, M., Zhang, H., and Su, Z., “Point cloud skeletons via laplacian based contraction,” in [2010 Shape Modeling International Conference ], 187–197, IEEE (2010)
2010
-
[19]
Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces,
Ma, B., Han, Z., Liu, Y.-S., and Zwicker, M., “Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces,” arXiv preprint arXiv:2011.13495 (2020)
2011 arXiv
-
[20]
Ribseg v2: A large-scale benchmark for rib labeling and anatomical centerline extraction,
Jin, L., Gu, S., Wei, D., Adhinarta, J. K., Kuang, K., Zhang, Y. J., Pfister, H., Ni, B., Yang, J., and Li, M., “Ribseg v2: A large-scale benchmark for rib labeling and anatomical centerline extraction,” IEEE Transactions on Medical Imaging 43(1), 570–581 (2023)
2023
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.