acceptodds
Under review as a conference paper at ICLR 2027

Tubularity-Aware Self-Supervised Pretraining for 3D Vessel Segmentation

Abstract

3D vessel segmentation requires delineating sparse, slender, and intricately branching tubular structures, yet acquiring annotations for learning these structures is costly. Self-supervised pretraining offers an opportunity to learn transferable representations from unlabeled images, but a key question remains: how can pretraining encourage vascular structure learning without manual annotations? Our premise is that, even when vessel locations are unknown, their elongated, curved morphology and fine-scale characteristics can guide the construction of self-supervised tasks. We propose Tubularity-Aware Pretraining (TAP), a self-supervised framework for 3D vessel segmentation that uses these priors to construct explicit restoration, localization, and detail prediction targets. Specifically, TAP combines cubic and curvilinear masking to learn contextual restoration of missing content; localizes intensity-matched donor content implanted into synthetic tubular regions to learn dense region discrimination; and predicts removed wavelet detail from coarse views to provide dedicated supervision for fine-scale information. Together, these tasks introduce explicit learning objectives for structural information relevant to vessel segmentation without requiring manual vessel annotations. Experiments demonstrate the effectiveness of TAP. Across four CT vessel datasets, TAP achieves state-of-the-art average DSC and clDice scores of 85.00% and 80.72%, respectively, outperforming the strongest competing method by 2.40 and 4.04 percentage points. Further experiments demonstrate its transferability to ASOCA, which is excluded from pretraining, and its effectiveness under limited-label fine-tuning. Code and models will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Self-supervised learning, 3D vessel segmentation, Tubular structure

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_DEjgOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Tubularity-Aware Pretraining (TAP) for 3D vessel segmentation. It combines mixed cubic and curvilinear reconstruction (HIDE), intensity-matched implant localization (IMPLANT), and wavelet detail prediction (SHARPEN) to learn representations from unlabeled CT images before supervised fine-tuning.

Critical strengths
  1. The motivation is clear and practically relevant. The sparse, elongated morphology of vessels and the difficulty of preserving small branches provide reasonable grounds for designing structure-aware pretraining tasks. The framework also requires no additional modules during downstream inference.
  2. The experimental evaluation is reasonably comprehensive. The component ablations support the complementary value of reconstruction, localization, and detail prediction. The controlled comparison in Table 8 is particularly useful because it aligns the pretraining corpus, backbone, and training protocol across competing self-supervised methods.
Critical weaknesses
  1. The methodological novelty needs clearer positioning. VasoMIM (Huang et al., 2026), already discussed in the paper, incorporates vascular anatomy into masked image modeling. Wavelet reconstruction targets have also been explored in WaMIM (Xiang et al., 2025). These approaches differ from TAP in their inputs, objectives, and applications, but the paper should explain more precisely what advances beyond adapting and combining these established ideas.
  2. The evidence for vessel-specific representation learning remains partly indirect. Curvilinear masks are generated independently of actual vessel locations, while wavelet detail targets include nonvascular boundaries and noise. The downstream improvements support the usefulness of these tasks, but do not fully establish that they specifically improve vascular connectivity rather than general local reconstruction and discrimination. The component effects also vary across datasets, and adding an individual component does not consistently improve every setting.
  3. The evidence for cross-dataset transfer is limited. All datasets in Tables 1 and 8 contribute training images to TAP pretraining, so these experiments primarily demonstrate the benefit of pretraining within the evaluated datasets. ASOCA is the only dataset excluded from pretraining and is evaluated after supervised fine-tuning. This supports transfer to one additional dataset, but is insufficient to establish broad transferability across vascular anatomies and acquisition settings. A leave-one-dataset-out evaluation using the existing datasets would substantially strengthen this claim.
Questions for authors
  1. Can the authors compare TAP with adaptations of VasoMIM’s anatomy-guided masking and WaMIM’s wavelet targets, using the same 3D backbone, pretraining data, and training budget, to isolate TAP’s contribution?
  2. Can the authors evaluate cubic masking, mixed masking, and full TAP using branch detection rate and centerline discontinuity counts, stratified by vessel diameter, to test whether the gains reflect improved vascular connectivity?
  3. Can the authors assess cross-dataset transfer through leave-one-dataset-out pretraining on the existing datasets, or evaluation on additional vessel datasets excluded from pretraining, using identical fine-tuning label budgets for TAP and the strongest SSL baseline?
5 comments
AuthorsOpenReview ↗

Response to Reviewer DEjg (1/2)

Thank you for the positive assessment and the concrete suggestions. We ran all three experiments you proposed and report them below.

W1 / Q1. What TAP adds beyond VasoMIM and WaMIM

VasoMIM and WaMIM keep masked reconstruction as the task and change where or what the network reconstructs. VasoMIM masks more often the patches that a vesselness filter marks as vessels, and WaMIM reconstructs multi-level wavelet coefficients instead of pixels.

TAP differs in three designs. Hide gives part of its masks the shape of vessels instead of placing masks where a vesselness filter finds vessels, so it needs no vessel detector. Implant adds a localization task, which neither method has. Sharpen removes fine Haar detail from the input and asks the network to restore it, instead of reconstructing the wavelet coefficients of the image.

Controlled comparison. We replaced Hide's masks with VasoMIM-style masking and Sharpen with WaMIM-style targets, one at a time, keeping the 3D backbone, pretraining data, training budget and other components unchanged. Both methods are 2D, so we extended them to 3D. VasoMIM-style masking places cubic masks preferentially where a 3D Frangi vesselness map marks likely vessels, at TAP's masking ratio. WaMIM-style targets replace Sharpen with a 3D version of WaMIM's multi-level wavelet target, with the same total loss weight.

Component tested (rest as in TAP)DSC↑clDice↑ASD↓
Hide mask (Implant and Sharpen kept)
Random cubes (Table 4b, η = 0)82.0175.834.25
Vesselness-guided cubes (VasoMIM-style)82.3477.484.08
Cubes and tubular masks (TAP)85.0080.723.06
Detail target (Hide and Implant kept)
None (Table 10)84.0078.253.67
Wavelet coefficients of the image (WaMIM-style)84.3978.873.21
Removed detail from the coarse view (TAP)85.0080.723.06

TAP's tubular masks and Sharpen thus outperform the corresponding VasoMIM and WaMIM designs by 2.66/3.24 and 0.61/1.85 DSC/clDice, respectively.

W2 / Q2. Do the gains reflect vascular connectivity?

As you suggested, we evaluated cubic masking, mixed masking and TAP by branch detection and centerline gaps, stratified by vessel diameter. We split each ground-truth skeleton into branches at bifurcations. A branch counts as detected only if at least 50% of its centerline is covered by the predicted component that also covers the trunk of its tree. An internal gap is an uncovered centerline segment between two covered ends. Caliber is the mean diameter along a branch. Values are averaged over cases and then over the four test sets.

MethodRecall <2.5 mm↑2.5–4.5↑≥4.5↑Gaps <2.5 mm↓2.5–4.5↓≥4.5↓
Cubic masking40.1255.8782.1155.0019.683.69
Mixed masking42.1958.6383.9348.6617.373.43
SwinMM (Table 8)42.7859.0380.7551.6317.514.28
TAP47.4063.3286.0641.1813.852.21

In every diameter group, TAP has the highest connected-branch recall and the fewest gaps. Over all calibers, its connected-branch recall is 63.72, against 57.03 for cubic masking and 59.38 for SwinMM, and its Betti-0/1 errors are 83.40/12.04, against 170.41/19.32 and 122.66/14.47. The gains therefore reflect better vascular connectivity, not only better local reconstruction, including in small vessels.

The detail target also contains non-vascular boundaries and noise. Like the tubular masks, it carries the vessel prior through scale rather than location. Per voxel, vessels carry 2.81 times the background detail energy in CTA and 24.27 times in TOF-MRA, so the target emphasizes vessels over the background. In the table, adding Implant and Sharpen to mixed masking reduces gaps in every diameter group and raises connected-branch recall below 2.5 mm from 42.19 to 47.40.

Component effects do vary across datasets (Table 3), and we think this may relate to the vascular anatomy of each dataset. Tubular masking helps most on ImageCAS, possibly because its coronary arteries are a few long, thin vessels, close to the shape of TAP's masks. On Parse2022 and SEG.A, it changes DSC and clDice by less than 0.7, whereas adding Implant or Sharpen to it raises DSC by 3.5–6.8 and clDice by 6.0–8.2. In Parse2022, small pulmonary branches run among veins and airways of similar caliber, and in SEG.A small branches leave the much larger aorta, so telling a structure apart from its surroundings (Implant) and recovering fine detail (Sharpen) may matter more there than completing a path. TAP therefore uses the three tasks together, and the full model has the best DSC, clDice, Jaccard index and NSD on all four datasets (Tables 3 and 7).

AuthorsOpenReview ↗

Response to Reviewer DEjg (2/2)

W3 / Q3. Transfer beyond the pretraining corpus

Leave-one-dataset-out. We pretrained TAP without Parse2022, the only pulmonary dataset in the corpus, and fine-tuned it on Parse2022 under the same protocol. The baselines are the Table 8 checkpoints, pretrained with Parse2022 included. Each entry is DSC/clDice.

Pretraining (fine-tuned on Parse2022)100% labels10% labels
DAE, with Parse2022 (best Table 8 baseline on Parse2022)84.17/73.32–
SwinMM, with Parse202283.76/71.49–
TAP, with Parse202287.39/81.6384.46/74.89
TAP, without Parse202285.17/75.4683.90/73.64

Unlabeled images of the target dataset do help, and the usual self-supervised setting has them. Even without them, TAP exceeds every Table 8 baseline, and with 10% of the labels it stays within 0.56/1.25 of TAP pretrained with Parse2022.

Datasets excluded from pretraining. We also evaluated two more datasets excluded from pretraining, with identical label budgets for TAP and SwinMM, the strongest self-supervised baseline in Table 8. Both start from their Table 8 checkpoints, which share corpus, backbone and schedule. COSTA-IXI-HH, the Hammersmith Hospital subset of COSTA, has 60 brain TOF-MRA images of healthy volunteers (48 for training, 12 for testing). AirRC-GE, the GE-scanner subset of AirRC, has 142 chest CT images with pulmonary artery annotations (114 for training, 28 for testing). Each entry is DSC/clDice.

Unseen datasetLabelsScratchSwinMMTAP
ASOCA (coronary CTA)100%80.71/79.0984.75/83.0486.13/84.54
10%69.34/64.1971.56/70.3974.40/72.60
COSTA-IXI-HH (brain TOF-MRA)100%93.09/92.5894.02/93.1494.53/94.23
10%88.63/85.6089.27/87.5390.11/88.63
AirRC-GE (pulmonary arteries, CT)100%87.72/83.0390.00/87.4091.58/90.08
10%84.54/79.1786.30/83.0888.41/85.49

TAP leads SwinMM in all six comparisons, by 1.16/1.76 DSC/clDice on average with full labels and by 1.93/1.91 with 10% labels, and on every dataset the DSC margin grows as labels decrease. On AirRC-GE, TAP fine-tuned with 10% of the labels (88.41/85.49) exceeds training from scratch with all labels (87.72/83.03). Both experiments are now in App. D.2 of the revision.

Thank you again for these suggestions, which have strengthened the paper.

Thank you for conducting all three experiments. These are substantial experiments to complete within a week, and the results are impressive and provide strong support for your claims. I think TAP is well designed for this task and makes an important contribution to vessel segmentation. I would therefore raise my recommendation from Weak Accept to Accept at the final stage.

One point I hope you will address in the revision concerns the connectivity evaluation raised by Reviewer jPY9 in W2 / Q3. The additional evaluation and the improvement in connected-branch recall are also impressive. However, this criterion may count some incomplete or incorrect connections as successful. For example, i) a branch could count as detected even if its actual junction is broken but a wrong path connects it to the trunk, or ii) if only half of its centerline is recovered. Other types of errors may also go undetected by this criterion. I hope the revision will explain whether these cases were checked and show examples where the model succeeds and fails to recover the connections. Sharing these cases openly would help readers understand what the model does well, where it still fails, and how it could be improved.

Reviewer_jPY9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TAP pretrains Swin UNETR on 1,066 unlabeled CT volumes using mixed cubic and curvilinear masking, tubular implant localization, and Haar-detail prediction. It is then fine-tuned for vessel segmentation. The four-source-dataset means are 85.00 DSC and 80.72 clDice, with additional transfer to ASOCA, which is excluded from pretraining. Appendix C includes a common-corpus and common-backbone comparison, implant shortcut controls, and alternative detail targets. Appendix B examines vessel caliber and false positives.

Critical strengths

S1 The central empirical result survives meaningful controls

Table 8 is the strongest evidence. With the same 1,066 pretraining volumes, Swin UNETR backbone, crop size, and optimization schedule, TAP reaches 85.00/80.72 DSC/clDice versus 82.55/75.21 for SwinMM, the strongest controlled comparator. Table 3 also separates cubic masking, mixed masking, and the auxiliary tasks. These comparisons substantially reduce the concern that the gain comes entirely from a different corpus or architecture. [Sec. 4.3; Appendix C.1, Table 8.]

S2 The analyses address the intended vascular failure modes

The intensity-matching, boundary-feathering, and volume-matched blob controls in Table 9 probe plausible implant shortcuts. In Figure 4, improved small-vessel centerline recall accompanies lower false-positive fraction (17.70% to 11.21%) and volume (25.27 to 15.60 mL) relative to cubic masking. ASOCA transfer adds evidence beyond source-dataset images, with full-label DSC improving from the strongest competing result of 84.99 to 86.13. Together, these are useful checks beyond average overlap. [Table 2; Appendix B; Table 9.]

Critical weaknesses

W1 The controls establish a useful recipe more clearly than a tubularity specific mechanism

Table 9 is a worthwhile control, but matching implant volume does not isolate tubular structure. The random blobs and fixed-radius tubes can still differ in surface-to-volume ratio, boundary orientation, spatial frequency, and connectedness. Any of these changes the localization problem. The result therefore supports the tested tube generator over the tested blob generator; it does not identify which vascular property transfers. This distinction matters because the geometric construction is the central motivation. The hard-blending tube/blob difference is 0.45 DSC points, while the blob variant already reaches 84.55. Similarly, Table 10 shows that generic Sobel supervision reaches 84.72 DSC, compared with 85.00 for Haar. These are positive results for localized perturbation and high-frequency supervision, but a narrower case for the particular anatomical interpretation. [Appendix C.2-C.3, Tables 9-10.]

Table 8 convincingly rules out a different corpus or backbone as the sole explanation. It does not resolve the interpretation above: its comparison changes the complete objective, and TAP uses two input views. I would therefore read the main result as an effective combination of pretext tasks, with evidence suggestive of a useful geometric bias. The manuscript should distinguish that conclusion from attributing the full gain to learning vessel-specific structure.

W2 The vascular analysis measures coverage more directly than connectivity

The caliber analysis in Appendix B addresses an important concern: higher centerline recall is accompanied by lower false-positive fraction and volume, so the benefit is not simply indiscriminate foreground expansion. However, these measurements cannot tell whether the recovered centerline belongs to the correct connected branch. A short missing junction or an incorrect bridge can alter the vascular graph substantially while changing relatively few voxels; aggregate clDice and false-positive volume do not distinguish those errors from less consequential local errors. The appendix correctly acknowledges this limitation. The stronger language about continuity in the main discussion should be read with the same qualification.

There is also a gap between the strongest comparison and the most interpretable analysis. Figure 4 examines Parse2022 against cubic masking, whereas Table 8 identifies SwinMM as the strongest controlled comparator. It therefore locates a benefit over a relatively simple reconstruction baseline, but does not show that the advantage over competing self-supervision specifically comes from small-vessel recovery. Finally, Appendix A computes ASD and NSD with unit voxel spacing, whereas Appendix B uses physical spacing for caliber and false-positive volume. These quantities answer different geometric questions; the surface results cannot be treated as a common physical-distance measure across datasets. None of this negates the DSC result, but it limits the anatomical conclusions drawn from it.

Questions for authors

Q1 What does the geometry control identify

Apart from foreground volume, which properties of the tubular and random-blob masks were matched? In particular, how do their boundary area, connected components, and effective spatial scales differ? Please explain which part of the existing evidence supports a tubular-structure explanation over a more general localized-corruption explanation.

Q2 What is being transferred to ASOCA

Table 2 evaluates ASOCA after supervised fine-tuning, whereas the common-corpus objective comparison in Table 8 covers the four source datasets. Were the ASOCA comparators initialized from the controlled Table 8 checkpoints or the released checkpoints used in Table 1? Were the synthetic tube scale and other task settings fixed before examining ASOCA results? The answers determine whether Table 2 isolates the pretraining objective on an unseen dataset or compares complete initialization pipelines.

Q3 How should the geometric improvements be interpreted

Which reported result directly measures preservation of branch connections, beyond centerline coverage? If there is none, can the continuity claim be limited accordingly? Please also state the resampling grid on which voxel-valued ASD/NSD are computed and explain how that grid relates to the physical-spacing calculations in Appendix B. This would make the anatomical meaning of the existing results clear.

4 comments
AuthorsOpenReview ↗

Response to Reviewer jPY9 (1/2)

Thank you for the careful and constructive review. We address each point below.

W1 / Q1. What the geometry control identifies

In the Table 9 control, tube and blob implants have the same foreground volume in every crop, and they share the same intensity matching (Eq. 3), boundary treatment, losses and training. Their shapes differ by design. Thin, elongated, connected regions have more boundary per voxel, fewer and longer components and finer spatial scales than compact blobs, and these are the properties the tubular masks are meant to introduce. The table quantifies the differences, as means over 1,000 tube masks and 1,000 volume-matched blob masks on a 96³ grid. Local thickness is the diameter of the largest sphere that contains a voxel and fits inside the mask, and the spectral centroid is the mean spatial frequency of the mask's Fourier spectrum.

PropertyTube (default)Volume-matched blob
Foreground voxels70,77970,779
Surface / volume (1/voxel)0.8200.701
Connected components (26-conn.)11.39169.74
Mean local thickness (voxels)4.435.16
Skeleton length / volume0.04830.0120
Spectral centroid (cycles/voxel)0.1460.080

Per mask voxel, tubes have 17% more surface and 4.0 times as much skeleton length. They are thinner, form about 11 connected components per crop rather than about 170, and their spectrum is centered at finer spatial scales. These properties vary together by construction, so, as you note, the control compares two classes of shape, a few thin components with long skeletons against many compact fragments, rather than isolating a single property.

The existing evidence for tubular structure over generic localized corruption comes from the comparisons that keep the corruption localized and change only its shape. In Hide, at the same masking ratio (ρ = 0.5), replacing 30% of the cubic mask volume with tubes raises DSC/clDice from 82.01/75.83 to 85.00/80.72 (Table 4b). In Implant, tubes add 0.71/2.61 over no implant and volume-matched blobs 0.26/0.99 (Table 9). Tubes have more boundary per voxel, so abrupt paste edges might make them easier to find, but with feathered edges the difference between tubes and blobs stays at 0.45/1.50, close to 0.45/1.62 with hard pasting (Table 9).

The blob variant (84.55 DSC) and the Sobel variant of Table 10 (84.72) each change a single component and keep the tubular masks in Hide. Their gaps to TAP therefore measure the implant shape and the choice of detail target, respectively, whereas removing the tubes from Hide lowers DSC by 2.99. On the second input view: all these controls keep both views, so their differences do not come from it. Without Sharpen, TAP uses a single view per crop, and this variant (Table 3, row 4: 84.00/78.25) still exceeds every Table 8 baseline (best, SwinMM: 82.55/75.21).

These results support the paper's premise that vessel morphology can guide self-supervised tasks even when vessel locations are unknown, and show that the tubular design improves on generic localized corruption, but they do not identify which single property of the tubes is responsible. The revision no longer attributes the full gain to learning vessel-specific structure. For example, it changes "learns transferable vascular representations" in Sec. 4.2 to "learns representations that transfer to vessel segmentation".

AuthorsOpenReview ↗

Response to Reviewer jPY9 (2/2)

W2 / Q3. Connectivity beyond centerline coverage

None of the reported results measures branch connections directly, so we computed three such measures on the existing test predictions.

  • Connected-branch recall. We split the ground-truth skeleton into branches at bifurcations. A branch counts as detected only if at least 50% of its centerline is covered by the predicted component that also covers the trunk of its tree, so fragments disconnected from their tree do not count.
  • Internal gaps. These are ground-truth centerline segments left uncovered between two covered ends, such as missing junctions and breaks.
  • Betti-0/1/2 errors. These count errors in the number of components, loops and cavities. An incorrect bridge, for example, creates a spurious loop.

Each measure is averaged over cases and then over the four test sets, and recall in each diameter group excludes cases without branches in that group. SwinMM is the Table 8 checkpoint, with the same corpus, backbone and schedule as TAP.

MethodDSC↑clDice↑Branch recall↑Gaps/case↓β0 err↓β1 err↓β2 err↓
Cubic masking80.8473.0457.0378.36170.4119.3214.39
Mixed masking81.4674.3059.1569.46148.8015.3314.26
SwinMM (Table 8)82.5575.2159.3873.42122.6614.4713.99
TAP85.0080.7263.7257.2483.4012.0413.17

TAP is best on every measure. Relative to cubic masking, TAP reduces internal gaps by 27%, Betti-0 error by 51% and Betti-1 error by 38%. Relative to SwinMM, the strongest controlled comparator, it reduces them by 22%, 32% and 17%. The clDice gain therefore reflects fewer breaks and fewer spurious connections. These measures also support your point that aggregate scores can understate structural differences. Mixed masking raises clDice by only 1.26 over cubic masking, yet removes 11% of the internal gaps and 13% of the Betti-0 errors. Stratified by caliber, TAP leaves fewer gaps than SwinMM in every diameter group. Below 2.5 mm, connected-branch recall rises from 42.78 to 47.40 and gaps fall from 51.63 to 41.18 per case, so the advantage over the strongest self-supervised comparator also holds for small vessels. The revision reports these branch-level measures in App. D.1 and refers to them in the ablation analysis of Sec. 4.3.

Grid and units. We compute ASD and NSD (App. A) on each dataset's native voxel grid, without resampling, in voxels and with a 1-voxel tolerance for NSD. App. B uses the same grid and differs only in units: caliber and false-positive volume are in millimetres, computed with each case's voxel spacing. As you note, a voxel has different physical sizes across datasets, so voxel-valued ASD and NSD are not a common physical distance. We therefore recomputed them in millimetres from the same predictions, with a 1 mm tolerance for NSD.

MethodASD (voxel)↓ASD (mm)↓NSD@1 voxel↑NSD@1 mm↑
SwinMM (Table 1)9.476.3870.8771.54
vesselFM6.983.9280.3380.35
TAP3.062.1584.4384.20

TAP remains best, and the ranking does not change. The revision states the grid in Sec. 4.1 and App. A, and App. A also gives the millimetre values.

Q2. What is transferred to ASOCA

  1. Comparators in Table 2. They start from their released checkpoints, as in Table 1, so Table 2 compares complete initialization pipelines. The revised caption states this, and point 3 gives the controlled comparison.
  2. TAP's settings were fixed before ASOCA. We pretrained TAP on unlabeled images of the four source datasets and chose all settings (tube radius, step, curvature, ρ, η, loss weights) on their validation splits. The resulting weights initialized fine-tuning on ASOCA unchanged. ASOCA played no part in pretraining or in choosing any setting.
  3. Isolating the pretraining objective. To compare objectives alone on this unseen dataset, we fine-tuned the controlled Table 8 checkpoints, which share corpus, backbone and schedule, on ASOCA under the same protocol.
Init. (same corpus and backbone)100%: DSC/clDice10%: DSC/clDice
Scratch80.71/79.0969.34/64.19
Cubic masking82.76/80.1371.92/69.30
VoCo83.58/82.4070.69/67.28
DAE82.83/81.5272.01/69.77
SwinMM84.75/83.0471.56/70.39
TAP86.13/84.5474.40/72.60

With corpus and backbone aligned, TAP leads on both metrics at both label budgets, by 1.38/1.50 over the best comparator with full labels and by 2.39/2.21 with 10% labels. On this unseen dataset, the gain can therefore be attributed to the pretraining objective.

Thank you again. Your questions have made the paper's claims more precise. We hope these results address your concerns, and if so, we would appreciate your considering an updated assessment.

Thank you for the detailed response and additional experiments. The new connectivity analyses, controlled ASOCA experiments, and clarification of physical-distance metrics have substantially addressed my major concerns. Although the specific contributions of individual tubular geometric properties remain unclear, I appreciate the authors' acknowledgment of this limitation and their more carefully qualified claims.

Overall, I consider my major concerns largely resolved and am updating my rating from Weak Rejection to Weak Acceptance.

AuthorsOpenReview ↗

Response to Reviewer jPY9

Thank you very much for your valuable review and for raising your rating. Your questions helped us make the paper clearer and more careful, and this means a lot to us. We really appreciate the time you spent on it. Thank you again.

Reviewer_pkYJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on 3D vascular image segmentation and addresses the challenge of data annotation, which is particularly complex and time-consuming. To that end, the authors propose a self-supervised strategy, denoted Tubularity-Aware Pretraining (TAP), that generates random vessel masks that are injected into patches of the images. A network is trained to reconstruct masked versions of such patches and to locate the implanted anatomies (vessels). The framework exploits 3D Haar wavelets to account for the multi-scale nature of vascular structures.

The method is evaluated on four datasets covering four body anatomies. All but one dataset are CT.

Critical strengths
  • The topic is timely and highly relevant. Vascular structure segmentation remains an open problem in the literature and this is largely due to the difficulty to annotate. This paper addresses the problem.
Critical weaknesses
  • Vascular generation process. There is a bunch of literature about vascular synthesis. The approach followed here seems very simple, especially when compared to recent approaches in the literature (see [1-9]). The paper does not discuss its reliability (i.e., are the generated structures "valid" vessels?), its generalizability (i.e., a generated patch may be "valid" for the heart, but not the lungs. Even more, a patch may be valid in a certain region of an organ, but not in another), nor their heterogeneity (i.e., how rich are the representations. do they offer enough variability? This is something that vesselFM addresses).
  • The method clearly states it is trained on CT patches. However, CAS (brain) is a TOF-MRA dataset. The two modalities have very different properties. For instance, early works on image synthesis [10] developed more complex strategies to achieve intensity transformation.s Therefore, the transferability proposed in Eq. 3, and the usage of 3D Haar wavelets become questionable.

references

  1. N. Rauch, M. Harders. “Methods for User-Controlled Synthesis of Blood Vessel Trees in Medical Applications: A Survey.” IEEE Access, vol. 13, pp. 22300–22313, 2025. doi:10.1109/ACCESS.2025.3536998.
  2. P. Feldman, M. Fainstein, V. Siless, C. Delrieux, E. Iarussi. “VesselVAE: Recursive Variational Autoencoders for 3D Blood Vessel Synthesis.” MICCAI 2023, LNCS 14220. doi:10.1007/978-3-031-43907-0_7.
  3. P. Feldman, M. Sinnona, C. Delrieux, V. Siless, E. Iarussi. “VesselGPT: Autoregressive Modeling of Vascular Geometry.” MICCAI 2025, LNCS 15975. doi:10.1007/978-3-032-05325-1_63.
  4. C. Prabhakar, S. Shit, F. Musio, K. Yang, T. Amiranashvili, J. C. Paetzold, H. B. Li, B. Menze. “3D Vessel Graph Generation Using Denoising Diffusion.” MICCAI 2024. doi:10.1007/978-3-031-72120-5_1.
  5. Y. Deo, F. Lin, H. Dou, N. Cheng, N. Ravikumar, A. F. Frangi, T. Lassila. “Few-shot learning in diffusion models for generating cerebral aneurysm geometries.” IEEE ISBI 2024. doi:10.1109/ISBI56570.2024.10635313.
  6. T. P. Kuipers, P. R. Konduri, H. Marquering, E. J. Bekkers. “Generating Cerebral Vessel Trees of Acute Ischemic Stroke Patients using Conditional Set-Diffusion.” MIDL 2024, PMLR 250, pp. 782–792.
  7. T. P. Kuipers, P. R. Konduri, E. J. Bekkers, H. Marquering. “Self-Supervised Synthetic Cerebral Vessel Tree Generation using Semantic Signed Distance Fields.” MIDL 2025, PMLR 301.
  8. S. Chen, G. Zhang, J. Lai, B. Shen, S. Zhang, C. Dong, X. Chen, Y. Li. “Hierarchical Part-based Generative Model for Realistic 3D Blood Vessel.” MICCAI 2025. doi:10.1007/978-3-032-04947-6_25.
  9. L. Kreitner, J. C. Paetzold, N. Rauch, et al., B. H. Menze, M. Harders, B. Knier, D. Rueckert, M. J. Menten. “Synthetic Optical Coherence Tomography Angiographs for Detailed Retinal Vessel Segmentation Without Human Annotations.” IEEE Transactions on Medical Imaging, 43(6), 2024.
  10. B. Wittmann, L. Glandorf, J. C. Paetzold, T. Amiranashvili, T. Wälchli, D. Razansky, B. Menze. “Simulation-Based Segmentation of Blood Vessels in Cerebral 3D OCTA Images.” MICCAI 2024.
  11. Z. Wang, R. Yi, X. Wen, C. Zhu, K. Xu. “VasTSD: Learning 3D Vascular Tree-State Space Diffusion Model for Angiography Synthesis.” CVPR 2025, pp. 15693–15702. doi:10.1109/CVPR52734.2025.01463.
  12. N. Burgos, M. J. Cardoso, K. Thielemans, M. Modat, S. Pedemonte, J. Dickson, A. Barnes, R. Ahmed, C. J. Mahoney, J. M. Schott, J. S. Duncan, D. Atkinson, S. R. Arridge, B. F. Hutton, S. Ourselin. “Attenuation Correction Synthesis for Hybrid PET-MR Scanners: Application to Brain Studies.” IEEE Transactions on Medical Imaging. vol. 33, no. 12, pp. 2332–2341, 2014
Questions for authors
  1. Some very recent works follow similar ideas [1,2]. While this does not invalidate the proposed work, I would like to understand how this work is positioned in this context. In [1], the authors' pipeline follows the same philosophy as this work. In [2], the authors consider controlled injections. In what ways is your strategy superior?
  2. The framework has been designed for the specifics of CT images. Because CT images are measured in Hounsfield units, using Eq. 3 is reliable. Moreover, the motivation behind the usage of 3D Haar wavelets is specific to CT. How can the method then work in the TOF-MRA (CAS) dataset? That is puzzling
  3. As a minor remark, the authors should consider adding references to all the baselines in their experiments. It is not the case for many of them. A paper should be readable and understandable for experts and beginners in a field. The latter ones would struggle with so many unidentified accronyms.

References

  1. E. Rainville, M. Ananian, T. Mirolla, H. Rivaz, Y. Xiao. “VesselSim: Learning 3D Blood Vessel Segmentation Without Expert Annotations.” MICCAI 2026. doi:10.1007/978-3-032-38085-2_58.
  2. M. Pentassuglia, X. Zhang, B. Billot, M. A. Zuluaga. “Anatomy-Conditioned Domain Randomization for Zero-Shot 3D Cerebrovascular Segmentation.” MICCAI 2026, doi:10.1007/978-3-032-38072-2_3.
7 comments
AuthorsOpenReview ↗

Response to Reviewer pkYJ (1/2)

Thank you for the expert review. We address each point below.

W1. Vascular generation: reliability, generalizability, heterogeneity

We first clarify the role of the tube generator in TAP, since it determines how these three questions apply. The random walk of Eq. (1) draws tube-shaped masks on real, unlabeled images. It does not generate vessel geometry or vessel images. Hide hides the real voxels inside a mask, and Implant fills a mask with real voxels taken from another crop of the batch. The inputs are derived from real images, and the targets are the real image or the drawn mask. Our phrases "synthetic tubular regions" and "tubular fragments" do not convey this clearly, and the revision replaces them with "tubular masks".

The revision also cites the vessel-synthesis literature [1–11] in Related Work and states the difference: these methods generate vessel geometry or images, whereas TAP generates no vessels and learns representations from real unlabeled images through self-supervised tasks.

Hide and Implant use this shape in different ways, Sharpen uses the scale of vessel detail, and a control in the paper supports each one. Hide mixes tubular masks with cubic ones, so the network must restore thin, elongated content along a path. At the same masking ratio, adding tubes raises DSC/clDice by 2.99/4.89 (Table 4b). Implant asks the network to find tube-shaped content that does not continue the surrounding anatomy. Tubular implants add 0.71/2.61, and volume-matched blobs 0.26/0.99 (Table 9). Sharpen asks the network to restore removed fine Haar detail and adds 1.00/2.47 (Table 10).

(a) Reliability. In TAP, the shape of a mask does not enter any target as a vessel label. The targets are the real image in Hide, the drawn mask in Implant and the removed real detail in Sharpen, so a mask that is not an anatomically valid vessel still yields a correct target. What TAP does need is that its masks are thin, elongated and connected like vessels, and the controls above show that this shape contributes to the gain. The generators you point to [1–8] could make the masks more realistic, for example with branches and varying radius, and the revision lists this as future work.

(b) Generalizability. For the same reason, a tube need not be valid for the organ or region in which it is drawn. In Hide, a tube in a lung image hides real lung, and the target is that real lung. In Implant, the target is the drawn mask, which is exact for any donor. All datasets therefore share one generator setting, and TAP gives the best DSC and clDice on coronary, cerebral, pulmonary and aortic trees alike (Table 1).

(c) Heterogeneity. TAP's variability comes from two sources. The image content is the real unlabeled corpus, with no synthetic appearance, so it varies as real data do. The masks are re-sampled at every iteration in number, shape and position. Their radius is currently fixed, and sampling radii and branches would add further variability. On ASOCA, which TAP never saw in pretraining, TAP exceeds vesselFM, which obtains variability through domain randomization, by 3.13/2.29 DSC/clDice under the same fine-tuning protocol (Table 2).

AuthorsOpenReview ↗

Response to Reviewer pkYJ (2/2)

W2 / Q2. CAS2023 is TOF-MRA, and TAP is not CT-specific

Yes, CAS2023 is TOF-MRA, and we have corrected the descriptions of our data as CT in the revision. TAP, however, is not designed for CT. Its tasks are defined by the shape of the masks and the scale of the removed detail, and every operation acts on intensities standardized within each crop. No operation depends on the modality, so TAP treats CT and MRA in the same way, and it has been pretrained on both together from the start.

The two steps you ask about follow this principle.

  • Eq. (3) acts on standardized intensities, not on HU. TAP standardizes every crop to zero mean and unit variance, and Eq. (3) then rescales the implanted content to the mean and standard deviation of the host region it replaces. Both steps use only statistics of the crop itself, so they work in the same way for CT and MRA. Eq. (3) also does not translate intensities between modalities, the harder problem addressed by the PET-MR synthesis work you cite (Burgos et al.). It only keeps the implant from standing out by brightness, so that the network must detect it by its texture and structure. Without Eq. (3), DSC/clDice drop from 85.00/80.72 to 84.40/78.65 on average (Table 9), and on CAS2023 from 87.33/83.47 to 86.96/83.21.
  • The Haar wavelet is motivated by vessel anatomy, not by CT. Sharpen supervises the fine-scale morphology of vessels (Sec. 1), a property of the anatomy rather than of the modality. The Haar wavelet splits any image into coarse content and fine detail at each scale, so it serves this purpose in CT and MRA alike. To check that the target emphasizes vessels in both modalities, we compared the per-voxel energy of level-1/2 Haar detail inside annotated vessels with that in the background. Vessels carry 2.81 times the background energy in CTA and 24.27 times in TOF-MRA, so in both modalities the target emphasizes vessels over the background.

The results on TOF-MRA agree with this. On CAS2023, every TAP component improves DSC and clDice (Table 3), and with 10% of the labels TAP exceeds training from scratch by 3.35 DSC and the strongest competing method by 1.21 (Table 2). On COSTA-IXI-HH, a brain TOF-MRA dataset excluded from pretraining (48 training and 12 test images), TAP exceeds SwinMM, the strongest self-supervised baseline in Table 8, by 0.51/1.09 DSC/clDice with all labels and by 0.84/1.10 with 10% of the labels.

Q1. Positioning relative to VesselSim [1] and ACDR [2]

Thank you for pointing us to these two works. Both appeared at MICCAI 2026, close to our submission, and the revision cites them in Related Work.

The three methods share ingredients, namely random curves, masked reconstruction and vessel injection, but use them in different settings. VesselSim and ACDR train a segmentation network only on synthetic images with vessel masks, so they need no labels in the target domain. TAP pretrains on real unlabeled images and is then fine-tuned with the available labels. The shared ingredients therefore play different roles. In VesselSim and ACDR, the vessel mask is the segmentation label, and the image is synthesized around it. In TAP, a tubular mask only marks which real voxels are hidden or replaced, and the targets come from the real image. VesselSim trains masked reconstruction alongside segmentation on synthetic data and uses it at test time to adapt to real scans, whereas TAP uses it to pretrain on real images. ACDR adds vessels only in plausible intracranial regions, since they become labels, whereas TAP's Implant injects real content from another image, and its label is simply the set of replaced voxels.

Because the settings differ, we see the approaches as complementary. VesselSim and ACDR are the natural choice when no target-domain labels exist. TAP suits the common case in which unlabeled target-domain images and a small number of labels are available. It learns from real images, so it has no synthetic-to-real gap and needs no test-time adaptation, and it makes good use of few labels. On ASOCA, unseen in pretraining, TAP exceeds vesselFM, which is trained on annotated, domain-randomized and generated vessel data, by 3.13/2.29 DSC/clDice with all labels and by 4.00/1.31 with 10% of the labels (Table 2). The two directions could also be combined, for example by drawing TAP's masks with a vessel generator, or by pretraining with TAP before training on synthetic vessels.

Q3. Baseline references. Thank you for the suggestion. The revision cites the 8 supervised architectures and 8 pretraining methods in the experiment tables, wherever they appear in Tables 1, 2 and 8.

We hope these clarifications address your concerns. If they do, we would appreciate your considering an updated assessment, and we are happy to answer further questions.

I appreciate the detailed response from the authors.

As the use of TOF-MRA has been confirmed, I think it would have been good to use it as an out-of-distribution dataset to see how well the technique generalizes. It is a more solid OOO test than using Parse, which, although it is a different organ, is also CT.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their careful reviews. We are encouraged that the reviewers found the problem timely and relevant (pkYJ), the motivation clear and practically relevant and the experimental evaluation reasonably comprehensive (DEjg), and the central result supported by meaningful controls (jPY9).

Below we summarize our answers to the five main questions. Each reviewer's reply gives the full answers and the new experiments. We have uploaded a revised PDF, with all changes marked in blue.

How TAP works. TAP turns vessel morphology into pretext tasks on real unlabeled images. Vessels are thin, elongated and connected. Hide masks tubular regions, mixed with cubic ones, and asks the network to restore the real content. Implant replaces tubular regions with real content from another image and asks the network to find them. Sharpen removes fine-scale detail and asks the network to restore it. The vessel prior lies in the shape and scale of these tasks, so TAP needs no annotations or vessel detector.

Q1. Does the vessel-oriented design contribute to the gain, and does it improve connectivity? (jPY9, DEjg) Yes. The paper's controls keep the localized corruption and change only the mask shape. Tubular masks outperform cubic masks in Hide (Table 4b) and volume-matched blobs in Implant (Table 9), so the tubular design improves on generic localized corruption. We added branch-level connectivity measures, stratified by vessel diameter. Averaged over the four test sets, TAP leaves 27% fewer internal gaps than cubic masking and 22% fewer than SwinMM, the strongest Table 8 baseline. It leaves the fewest gaps and has the highest connected-branch recall in every diameter group. These results support the paper's premise that vessel morphology can guide self-supervised tasks even when vessel locations are unknown.

Q2. What does TAP add beyond VasoMIM and WaMIM? (DEjg) VasoMIM masks more often where a vesselness filter finds vessels, and WaMIM reconstructs wavelet coefficients. TAP gives part of its masks the shape of vessels, adds Implant, a localization task, and asks for the fine detail removed from the input. With the same 3D backbone, data and budget, we replaced TAP's masks by VasoMIM-style masking and Sharpen by WaMIM-style targets. TAP's designs remain better, by 2.66/3.24 and 0.61/1.85 DSC/clDice.

Q3. How does TAP relate to vessel synthesis? (pkYJ) TAP uses vessel shape to define pretext tasks on real images, not to generate vessels. The random walk draws tubular masks on real, unlabeled images. Hide hides the real voxels inside a mask, and Implant fills a mask with real voxels from another crop of the batch. The targets are the real image or the drawn mask, so a mask that is not an anatomically valid vessel still yields a correct target.

Q4. Does TAP transfer beyond its pretraining data? (DEjg, jPY9) We fine-tuned the Table 8 checkpoints of TAP and SwinMM on three datasets excluded from pretraining, ASOCA (coronary CTA), COSTA-IXI-HH (brain TOF-MRA) and AirRC-GE (pulmonary arteries, CT). TAP leads SwinMM in all six comparisons (three datasets, full and 10% labels), by 1.93/1.91 DSC/clDice on average with 10% labels. Pretrained without Parse2022, TAP still exceeds on Parse2022 every Table 8 baseline, all pretrained with it (85.17/75.46 against at best 84.17/73.32).

Q5. CAS2023 is TOF-MRA. Is TAP CT-specific? (pkYJ) Yes, CAS2023 is TOF-MRA, and we have corrected the descriptions of our data as CT. TAP itself is not CT-specific. CAS2023 has always been part of pretraining and evaluation, so the reported results already include TOF-MRA, and no step of TAP uses a CT property. Eq. (3) acts on per-crop standardized intensities rather than HU, the Haar wavelet in Sharpen is a standard decomposition for images of any modality, and on CAS2023 every TAP component improves DSC and clDice. TAP also leads SwinMM on COSTA-IXI-HH, an unseen TOF-MRA dataset (Q4).

Revisions. The revision (i) corrects the modality description throughout, (ii) cites all 16 baselines in the experiment tables, (iii) revises the few sentences that went beyond the premise in Q1 and replaces wording such as "synthetic tubular regions" that suggests vessel generation, (iv) cites the vessel-synthesis literature, including VesselSim and ACDR, and WaMIM in Related Work, and (v) adds the branch-level connectivity results, the transfer results and surface distances in millimetres to the appendix.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.