Pith. sign in

REVIEW 1 major objections 4 minor 22 references

Inference-Time Orthogonal Seeding Enables Geometry-Aligned 3D Organ Segmentation for Slice-Propagation Methods

T0 review · 1 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that seed geometry—one axial, one coronal, one sagittal—rather than the registration model or training, drives slice-propagation segmentation quality, improving Dice by 21.9% and cutting surface distance by 53.5%.

desk verdict Real, cheap, reproducible gains from orthogonal seed propagation, but the 'orthogonality, not annotation count' claim rests on a 3-axial control that the fusion rule itself handicaps. read the letter →

arxiv 2608.12658 v1 pith:MOB7XN7A submitted 2026-08-12 cs.CV

Md Rakibul Haque , Tushar Kataria , Shireen Y. Elhabian This is my paper · ORCID
classification cs.CV
keywords slicepropagationorthogonalseedsdistance-weightedfusionlabel-freeregistrationself-supervisedsegmentationCTorganSli2Volmulti-planar
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Slice-propagation methods label a 3D volume by propagating one annotated slice through neighbors, but their accuracy degrades with distance from the seed. This paper tests whether the bottleneck is the registration model's training or the geometric placement of seed slices, and concludes it is the seed geometry. Propagating from one axial, one coronal, and one sagittal seed, then fusing the three label chains with a distance-weighted average, improves Dice by 21.9%, Normalized Surface Dice by 25.5%, and reduces Average Hausdorff Distance by 53.5% over the single-axis baseline on abdominal CT. A budget-matched control with three axial seeds provides no benefit and can degrade performance, so the paper attributes the gain to orthogonality rather than the number of annotated slices. If correct, the result is a purely inference-time upgrade: two extra annotations, no retraining, and no architectural change.

What carries the argument

The load-bearing object is the orthogonal seed set combined with the distance-weighted fusion rule of Eq. (5). Each propagation chain $c$ (axial, coronal, or sagittal) contributes at voxel $v$ with weight $\omega_c(v)=\exp(-|[v]_{a(c)}-s_c|/\sigma)$, where $a(c)$ is the chain's axis and $s_c$ its seed position, and the fused soft prediction is the weighted average over chains, thresholded at 0.5. The weights depend only on seed geometry, not on the network's predictions, so the fusion needs no labeled tuning data; the paper uses $\sigma=30$ slices. Orthogonal seeds work because every voxel is within a short effective distance of at least one seed along its own axis, shortening the propagation chains and reducing the error accumulation that the paper documents for single-axis propagation.

What would settle it

Run the same three-axial experiment with a fusion rule that gives each voxel to the chain whose own seed is closest (or otherwise prioritizes the chain carrying the exact mask at its seed), and check whether 3-axial then matches 3-orthogonal. If it does, the near-seed dip in Fig. 2 is a fusion artifact and the claim that orthogonality rather than count drives the gain fails; if it does not, the coplanarity explanation survives. A simpler observation: at the shared seed plane in 3-axial, the axial chain's prediction should equal the ground truth by construction, so any large error there directly indicts the fusion rule.

Watch

Extended reading notes

Core claim

The central claim is that, for slice-propagation segmentation, inference-time seed geometry—not the registration backbone or its self-supervised training—determines segmentation quality. The evidence is the consistent ordering across backbones: one axial seed and three coplanar axial seeds are both beaten by three orthogonal seeds, with the axial Sli2Vol backbone improving from Dice 0.595 to 0.725, NSD@2mm from 0.469 to 0.589, and AHD from 7.28 mm to 3.39 mm. The paper reads the failure of the budget-matched three-axial control as proof that the gain is orthogonality, not annotation count, and the per-slice curves in Fig. 2 as showing that orthogonal seeds keep both Dice and surface distance flat across the organ where single-axis and coplanar seeds decay or dip.

Load-bearing premise

The conclusion depends on the three-axial control being a fair test: if the control's poor performance near the shared seed plane comes from the fusion rule washing out the chain that holds the exact annotation, rather than from the seeds lying in one plane, then the claim that orthogonality rather than seed count drives the gain is not established.

Editorial extensions

If this is right

  • Switching from one axial seed to three orthogonal seeds improves Dice by 21.9%, NSD@2mm by 25.5%, and reduces AHD by 53.5% on the axial Sli2Vol backbone over the single-axis baseline.
  • Explicitly training the correspondence network on three axes adds little once orthogonal seeds are used; the axially trained network applied to off-axis seeds captures nearly all the gain.
  • A budget-matched three-axial seed control does not beat the single axial seed and can degrade performance, so annotation count alone is not the driver.
  • The pattern is consistent across correspondence and deformable backbones for Dice and AHD; the deformable backbone's lack of a soft prediction weakens its NSD gain under hard voting, which the paper attributes to fusion rather than to orthogonal seeding itself.
  • The improvement requires only two additional annotated slices and no architectural changes or retraining, making it a plug-and-play inference strategy for slice-propagation segmentation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the orthogonality result holds beyond abdominal CT, the same inference-only recipe—adding one coronal and one sagittal seed to any single-axis slice-propagation pipeline—could transfer to MRI or other modalities, provided the backbone's slice correspondence remains reliable off-axis.
  • The paper's own assumption that targets are compact with comparable extent across all three axes predicts a testable boundary: for thin, elongated, or branching structures such as vessels or bronchi, the orthogonal-seeding gain should shrink, and seed count might then matter more than orthogonality.
  • The near-seed dip of the three-axial control in Fig. 2 suggests an untested explanation: the distance-weighted fusion rule may be the culprit. A nearest-seed-priority fusion could close much of the gap, which would reframe the contribution as fusion design rather than seed geometry.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

1 major / 4 minor

Summary. The paper proposes an inference-time change to slice-propagation segmentation: instead of propagating from a single axial seed, propagate from three orthogonal seeds (axial, coronal, sagittal) and fuse the resulting soft maps with a label-free distance-weighted rule (Eq. 5). The authors evaluate this on three backbones (axial Sli2Vol, triplanar Sli2Vol, and TransMorph) across multi-organ CT cohorts, reporting large gains in Dice, NSD, and AHD over the single-axis baseline. They argue that the gains stem from seed orthogonality rather than annotation count, citing a budget-matched 3-axial control that provides no benefit and can degrade performance. Statistical significance is assessed with paired Wilcoxon tests on per-case Dice.

Significance. If the causal claim holds, the contribution is significant: a training-free, architecture-free inference modification yields large accuracy gains on surface-sensitive metrics, with public code and a controlled comparison across training regimes and inference seed configurations. The use of multiple backbones, an unseen organ (spleen), and paired nonparametric tests are strengths. However, the central interpretation—that orthogonality, not annotation count, drives the improvement—rests on a 3-axial control that is potentially confounded by the distance-weighted fusion rule. The paper also promises an equal-weight averaging ablation that is not reported in the results. These issues require additional experiments or a more cautious claim.

major comments (1)
  1. [Section 3, last paragraph of Method (Sec. 2)] The text states that "for the correspondence backbone we also report equal-weight averaging (ωc≡1) as an ablation isolating the distance weighting," but no such result appears in Section 3. This missing experiment is exactly what would isolate the contribution of distance weighting in Eq. (5) and is needed to interpret the 3-axial control. The authors should either add the equal-weight ablation results for all configurations or explicitly remove the promise from the Method section.
minor comments (4)
  1. [Section 3, Implementation Details] The phrase "analytical ∼!3× increase in forward passes" appears to contain a typographical artifact ("!3"); please correct to "3×".
  2. [Section 3, Quantitative Results] The paper mentions the largest Dice gain on pancreas (+28%) but does not provide per-organ results in a table or figure; please include per-organ breakdowns for all metrics.
  3. [Section 3, Statistical tests] Paired Wilcoxon results are reported only for Dice; since the headline claims include NSD and AHD, report significance tests for those metrics as well.
  4. [Fig. 2 and surrounding text] Figure 2's caption identifies the backbone as triplanar Sli2Vol, but the text around it sometimes refers to the axial Sli2Vol backbone; please ensure consistent labeling of which backbone is shown.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the orthogonal-seeding gains are empirical measurements; self-citations are contextual, and the only concern is a missing fusion ablation, which is a reporting gap rather than circularity.

full rationale

The paper's load-bearing claim is that 3-orthogonal seeding outperforms 1-axial and 3-axial seeding, and this is established by direct evaluation on 342 test cases rather than by derivation. No equation in the paper defines the reported Dice, NSD, or AHD gains; the fusion rule in Eq. 5 weights chains by exp(-distance/sigma) with sigma=30 fixed and stated to be stable over [20,50], and sigma is not fit to the reported metrics. Seed slices are chosen by maximum ground-truth area only to standardize the selection, not to optimize the measured outcome, so there is no fitted-input-called-prediction step. The 3-axial budget-matched control is a legitimate comparison design; its interpretation could be confounded by the distance-weighted fusion because near one axial seed the exact ground-truth chain is diluted by the other two axial chains, but that is an experimental-validity concern, not circularity: the conclusion is not equivalent to the construction of the control. The self-citations ([3], [14]) are contextual references to distance-aware multi-atlas weighting and slice-propagation error accumulation, and they are not the source of the central claim. The paper promises an equal-weight averaging ablation (omega_c = 1) for the correspondence backbone in Sec. 2 but does not present it in Sec. 3; this missing ablation weakens the attribution of the gain to distance weighting, but it is an omitted experiment rather than a circular reduction. Under the stated hard rules, no step reduces by construction to its input, so the circularity score is 1.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its central claim rests on two hand-chosen inference choices (sigma, 3-axial seed positions) and several domain assumptions about propagation error behavior and organ geometry. The main free parameter is sigma, which the paper argues is stable.

free parameters (2)
  • sigma (distance-weight decay) = 30 slices (stable in [20,50])
    Chosen by hand; not tuned on labels, but it controls how quickly each propagation chain's influence decays, and the 3-axial versus 3-orth comparison may depend on it.
  • 3-axial seed positions = 25%, 50%, 75% of organ extent
    Chosen as a budget-matched control; different positions could change the comparison, and the paper does not sweep them.
assumptions (4)
  • domain assumption A label-free slice-reconstruction objective (Eq. 2) yields a correspondence map that transfers masks accurately between adjacent slices.
    The whole method depends on adjacency learned from intensity reconstruction generalizing to mask transport; this is the established Sli2Vol premise.
  • domain assumption Propagation error accumulates monotonically with distance from the seed, so distance-weighted fusion is a valid way to combine chains.
    The fusion rule in Eq. 5 assumes reliability decays with distance; if errors accumulate non-monotonically, this weighting could be suboptimal.
  • domain assumption The target organ is compact with comparable extent in all three axes, so off-axis seeds reduce effective propagation distance.
    Stated in Assumptions and Limitations; false for thin, elongated, or branching structures.
  • domain assumption The max-area slice is a reasonable proxy for an informative user-selected seed.
    Stated in Section 2; user-driven selection and its effort are not evaluated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inference-Time Orthogonal Seeding Enables Geometry-Aligned 3D Organ Segmentation for Slice-Propagation Methods." pith.science (2026). https://pith.science/paper/MOB7XN7A

@misc{pith2026260812658,
  author       = {Pith},
  title        = {Pith review of: Inference-Time Orthogonal Seeding Enables Geometry-Aligned 3D Organ Segmentation for Slice-Propagation Methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MOB7XN7A}},
  note         = {Machine review of arXiv:2608.12658}
}
read the original abstract

Dense voxel-level annotation remains a major bottleneck in 3D medical image segmentation. Single-slice propagation methods such as Sli2Vol reduce this burden by propagating one annotated seed slice through a volume using label-free registration. However, axial-only propagation accumulates errors with distance from the seed, especially in surface-distance metrics, because it ignores coronal and sagittal evidence and therefore underuses the 3D information available in CT/MRI volumes. To better leverage volumetric geometry, we study how key training and inference choices affect slice-propagation models, including single-axis versus multi-axis label-free registration, single-seed versus multi-seed propagation, and orthogonal seed configurations. Instead of propagating from a single axial seed, we use three orthogonal seeds---one axial, one coronal, and one sagittal---and fuse their propagated labels with a simple label-free rule. Our results show that the training paradigm has limited impact: an axially trained network applied to off-axis seeds captures nearly all the improvement, while explicit three-axis training adds little. Instead, performance is driven by inference-time seed geometry, especially orthogonality rather than the number of annotated slices, as a budget-matched three-axial control provides no benefit and can even degrade performance. On a multi-organ CT cohort, orthogonal seeding with the axial Sli2Vol backbone improves Dice by 21.9%, Normalized Surface Dice by 25.5%, and reduces Average Hausdorff Distance by 53.5% over the single-axis baseline.

Figures

Figures reproduced from arXiv: 2608.12658 by the authors.

Figure 1
Figure 1. Overview. (A) A self-supervised registration backbone is trained on adjacent slice pairs with no labels, giving axial Sli2Vol, triplanar Sli2Vol, and deformable Trans￾Morph. (B) At inference we compare one axial seed, three coplanar axial seeds, and three orthogonal seeds. (C) The per-axis propagated maps are combined by a label-free distance-weighted fusion into a 3D segmentation. provided along axial, coronal, and… view at source ↗
Figure 2
Figure 2. Segmentation quality versus distance to the nearest seed, on the triplanar Sli2Vol backbone. Left: per-slice Dice. Right: per-slice surface distance (mm). The single-axis baseline (1-axial) decays away from its lone seed, the coplanar control (3- axial) fills the mid-range but starts poorly near the shared seed plane, and the orthog￾onal configuration (3-orthogonal) stays comparatively high and flat throughout. All … view at source ↗
Figure 3
Figure 3. Representative pancreas case on the triplanar Sli2Vol backbone. Green: correct; red: missed; blue: leakage. 1-axial and 3-axial miss the head and tail far from the axial seed; 3-orthogonal recovers the full organ. All rows share network weights and differ only in seed geometry. Assumptions and Limitations. Orthogonal seeding assumes a compact target with comparable extent across all three axes, so that off-axis seed… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

  1. [1]

    IEEE Transactions on Medical Imaging28(8), 1266–1277 (2009)

    Artaechevarria, X., Muñoz-Barrutia, A., Ortiz-de Solórzano, C.: Combination strategies in multi-atlas image segmentation: application to brain mr data. IEEE Transactions on Medical Imaging28(8), 1266–1277 (2009)

    work page 2009

  2. [2]

    IEEE Transac- tions on Medical Imaging38(8), 1788–1800 (2019)

    Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Voxelmorph: A learning framework for deformable medical image registration. IEEE Transac- tions on Medical Imaging38(8), 1788–1800 (2019)

    2019

  3. [3]

    In: Proceedings of the Winter Conference on Applications of Computer Vision

    Bin Aziz, A.Z., Karanam, M.S.T., Kataria, T., Elhabian, S.Y.: Efficientmorph: Parameter-efficient transformer-based architecture for 3d image registration. In: Proceedings of the Winter Conference on Applications of Computer Vision. pp. 1330–1341 (2025)

    work page 2025

  4. [4]

    In: Medical Image Computing and Computer Assisted Intervention (MICCAI)

    Bitarafan, A., Azampour, M.F., Bakhtari, K., Soleymani Baghshah, M., Keicher, M., Navab, N.: Vol2flow: Segment 3d volumes using a sequence of registration flows. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). Springer (2022). https://doi.org/10.1007/978-3-031-16440-8_58

    work page doi:10.1007/978-3-031-16440-8_58 2022

  5. [5]

    Medical Image Analysis101, 103478 (2025)

    Bitarafan, A., Mozafari, M., Azampour, M.F., Soleymani Baghshah, M., Navab, N., Farshad, A.: Self-supervised 3d medical image segmentation by flow-guided mask propagation learning. Medical Image Analysis101, 103478 (2025)

    work page 2025

  6. [6]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Cai, H., Li, S., Qi, L., Yu, Q., Shi, Y., Gao, Y.: Orthogonal annotation benefits barely-supervised medical image segmentation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3302–3311. IEEE (2023)

    work page 2023

  7. [7]

    Medical image analysis82, 102615 (2022)

    Chen, J., Frey, E.C., He, Y., Segars, W.P., Li, Y., Du, Y.: Transmorph: Transformer for unsupervised medical image registration. Medical image analysis82, 102615 (2022)

    2022

  8. [8]

    In: MICCAI Workshop on 3D Segmentation in the Clinic: A Grand Challenge

    van Ginneken, B., Heimann, T., Styner, M.: 3d segmentation in the clinic: A grand challenge. In: MICCAI Workshop on 3D Segmentation in the Clinic: A Grand Challenge. pp. 7–15 (2007) 10 Md Rakibul Haque, Tushar Kataria, and Shireen Y. Elhabian

    work page 2007

Show all 22 references
  1. [9]

    arXiv preprint arXiv:1904.00445 (2019)

    Heller, N., Sathianathen, N., Kalapara, A., Walczak, E., Moore, K., Kaluzniak, H., et al.: The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes. arXiv preprint arXiv:1904.00445 (2019)

    2019 arXiv

  2. [10]

    Nature Methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods18(2), 203–211 (2021)

    2021

  3. [11]

    Medical Im- age Analysis69, 101950 (2021)

    Kavur, A.E., Gezer, N.S., Barış, M., Aslan, S., Conze, P.H., Groza, V., et al.: Chaos challenge – combined (ct-mr) healthy abdominal organ segmentation. Medical Im- age Analysis69, 101950 (2021)

    2021

  4. [12]

    In: IEEE/CVF International Conference on Computer Vision (ICCV)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015– 4026 (2023)

    2023

  5. [13]

    Nature Communications15(1), 654 (2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications15(1), 654 (2024)

    2024

  6. [14]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Nihalaani, R., Kataria, T., Adams, J., Elhabian, S.Y.: Estimation and analysis of slice propagation uncertainty in 3d anatomy segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 273–285. Springer (2024)

    2024

  7. [15]

    In: Medical Image Computing and Computer-Assisted Intervention (MICCAI)

    Prasoon, A., Petersen, K., Igel, C., Lauze, F., Dam, E., Nielsen, M.: Deep fea- ture learning for knee cartilage segmentation using a triplanar convolutional neu- ral network. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 246–253. Springer (2013)

    2013

  8. [16]

    In: Medical Image Computing and Computer Assisted Intervention (MICCAI)

    Roth, H.R., Lu, L., Farag, A., Shin, H.C., Liu, J., Turkbey, E.B., Summers, R.M.: Deeporgan: Multi-level deep convolutional networks for automated pancreas seg- mentation. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 556–564. Springer (2015)

    2015

  9. [17]

    In: Medical Image Computing and Computer Assisted Intervention (MICCAI)

    Roth, H.R., Lu, L., Seff, A., Cherry, K.M., Hoffman, J., Wang, S., Liu, J., Turkbey, E., Summers, R.M.: A new 2.5d representation for lymph node detection using random sets of deep convolutional neural network observations. In: Medical Image Computing and Computer Assisted Int...

    2014

  10. [18]

    IEEE Transactions on Medical Imaging35(5), 1160–1169 (2016)

    Setio, A.A.A., Ciompi, F., Litjens, G., Gerke, P., Jacobs, C., van Riel, S.J., Wille, M.M.W., Naqibullah, M., Sánchez, C.I., van Ginneken, B.: Pulmonary nodule de- tection in ct images: false positive reduction using multi-view convolutional net- works. IEEE Transactions on Me...

    2016

  11. [19]

    arXiv preprint arXiv:1902.09063 (2019)

    Simpson,A.L.,Antonelli,M.,Bakas,S.,Bilello,M.,Farahani,K.,vanGinneken,B., et al.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv preprint arXiv:1902.09063 (2019)

    2019 arXiv

  12. [20]

    Medical Image Analysis73, 102184 (2021)

    Sundaresan, V., Zamboni, G., Rothwell, P.M., Jenkinson, M., Griffanti, L.: Tripla- nar ensemble u-net model for white matter hyperintensities segmentation on mr images. Medical Image Analysis73, 102184 (2021)

    2021

  13. [21]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wang, X., Jabri, A., Efros, A.A.: Learning correspondence from the cycle- consistency of time. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2566–2576 (2019)

    2019

  14. [22]

    In: Medical Image Computing and Com- puter Assisted Intervention (MICCAI)

    Yeung, P.H., Namburete, A.I.L., Xie, W.: Sli2vol: Annotate a 3d volume from a single slice with self-supervised learning. In: Medical Image Computing and Com- puter Assisted Intervention (MICCAI). pp. 69–79. Springer (2021)

    2021

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.