acceptodds
Under review as a conference paper at ICLR 2027

HemoPT: Flow-State Proxy-Supervised Pretraining on Heterogeneous Vascular Geometries

Abstract

Predicting vascular hemodynamic fields directly from geometry can reduce the computational cost of performing computational fluid dynamics (CFD) simulations across numerous vascular cases. However, training such surrogate models remains constrained by a substantial imbalance in supervision, with abundant unpaired vascular geometries but scarce and costly paired geometry-flow data. In this paper, we introduce HemoPT, a flow-state proxy-supervised vascular pretraining framework that requires no paired CFD fields. We construct a multi-fidelity, multi-anatomy pretraining corpus comprising 5,310 vascular geometries. By sampling 20 probe conditions per geometry, the corpus yields more than 100,000 probe-conditioned pretraining instances. HemoPT establishes spatial awareness of the vascular lumen through directional probes. This probe-derived spatial context parameterizes analytic responses associated with representative vascular flow states, producing geometry-specific hemodynamic proxies. The resulting geometry-proxy pairs are used to pretrain a Transolver encoder, which is subsequently fine-tuned for downstream velocity-field regression. A controlled study on the Aneumo dataset shows that pretraining with geometry-proxy pairs outperforms conditioning-only and geometry-oriented pretraining settings. Across four downstream tasks and varying data regimes, HemoPT achieves highly competitive performance among all compared training and pretraining methods. On the Vascular Model Repository (VMR) dataset, the gains from pretraining vary with vascular anatomy. These results indicate that HemoPT learns transferable geometry-flow representations from heterogeneous vascular geometries, with transfer benefits shaped by downstream anatomy. Code and preprocessing scripts are available at: https://anonymous.4open.science/r/HemoPT-66CF/

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific Machine Learning, Surrogate Models, Computational Fluid Dynamics, Heterogeneous Vascular Geometries, AI for Healthcare

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_ghMAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper constructed a multi-fidelity and multi-anatomy pretraining corpus comprising 5,310 vascular geometries that can yield more than 100,000 probe-conditioned pretraining instances. Building upon this, the authors introduced a flow-state proxy-supervised vascular pretraining framework (HemoPT) without the need for paired computational fluid dynamics (CFD) fields. HemoPT uses directional probes to learn the spatial information of the vascular lumen, and is later associated with representative vascular flow states, producing geometry-specific hemodynamic proxies. The authors then pre-trained a transolver encoder for downstream velocity-field regression across different data regimes.

Critical strengths
  1. The data-imbalance issue of abundant unpaired vascular geometries but scarce and costly paired geometry-flow data is well-targeted, and the exploration of how to leverage heterogeneous, unlabeled vascular geometries is well-motivated.

  2. The construction of the multi-fidelity corpus, the dataset information and the QC & processing pipeline, are described. The authors also included the code, which further details the implementation and promotes reproducibility.

  3. Multiple levels of low-data regime (i.e. 25%, 50%, 75%), multiple corpus scales are utilized, and different tier compositions are included.

Critical weaknesses
  1. The proposed flow proxy seems to have a fixed direction within each geometry. Equations A1-5 stated that each geometry is assigned one PCA-based frame, and the centerline used in Equations A7 and A8 is a straight line along the first principal axis, which doesn’t necessarily represent the centerline of all the branches in more complex models. Equations A14 to A18 further state that U1-U4 share the same direction vector e with different magnitude, and Table A11 shows that these four primitives receive most (86.5%, at epoch 200) of the mean routing weight on the pretraining corpus. The authors should discuss and clarify how the proxy could provide meaningful flow supervision on the branching or curved geometries.

  2. It’s unclear whether proxy design (i.e. dynamic routing and physical consistency losses) improves the transferred encoder. GeoPT-Vasc [1] has the same transolver backbone as HemoPT and the main design novelty is the flow-proxy branch. The authors showed in Table 2 the flow proxy and physical loss ablation, but the authors did not include a comparison between different encoders pre-trained with the proxy targets generated under each Table 2 configuration. Thus it is hard to conclude with the existing results that dynamic routing and physical consistency losses benefit the transfer.

  3. The authors mentioned in the related work session several applications adopted by other related literature, like wall shear stress estimation, and fractional flow reserve prediction, but only the velocity-field regression are adopted as the method’s down-streaming applications. More evaluation tasks should be used to highlight the benefit of pre-taining and bring more insights/evaluations of the model. In addition, experiments with linear probes and frozen encoder should be included to show and examine the rationality of the learned representation.

  4. Other readability issues: Certain figures like Figure 5, Figure A.1, Figure A.2 are hard to read, the texts in the figure are too small, and Figure A.2 is incomplete.

[1] Wu, H., Guo, M., Li, Z., Dou, Z., Long, M., He, K., & Matusik, W. (2026). Geopt: Scaling physics simulation via lifted geometric pre-training. arXiv preprint arXiv:2602.20399.

Questions for authors

So far the model’s input only considers the geometry and doesn’t seem to be able to adapt to different flow conditions. Could other factors potentially affect the flow that could potentially be considered (e.g., inlet flow rate)?

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

This manuscript proposes HemoPT, a vascular pretraining strategy that operates purely on geometry rather than additionally requiring flow data for pretraining. The authors assemble a pretraining dataset of 5,310 vascular geometries in total, which are subsequently augmented with 20 probing conditions to produce an augmented dataset of 106,200 pretraining instances. For each node, the input contains a combination of morphology information and probe condition, while the expected output is a vector containing a combination of the node’s spatial trajectory and a flow proxy. Decoding happens by predicting a combination of flow primitives that combine into the flow proxy. Training is done MAE-style: input geometry information is partially masked, but the model is still asked to output the full spatial + flow information; afterward, the decoder is discarded. Effectiveness is demonstrated through downstream hemodynamics estimation tasks on various geometries, where they show how their pretraining method substantially improves downstream performance, while showing a scaling effect with dataset size.

Assessment

This work provides an interesting direction for pretraining with a target more aligned with the downstream task compared to previous work, which mostly focuses on reconstructing geometry. Though the application to vascular flow pretraining is interesting, the dataset structuring and pruning make sense and are thorough. However, I do not think iit fits ICLR as a venue. (This is why I chose to rate it 1). It is a good contribution to the vascular flow dynamics field which contains interesting literature and is extensive, but as a general representation learning method it is under-evaluated (covers one downstream task) and may be a better fit for venues such as TMI, IPMI, MICCAI or MedIA. The work is also framed slightly larger than it is, with the 100K+ training dataset essentially being mostly augmentations of a 5K dataset of geometries (which is already a reasonably large set). Finally, it misses a large body of background literature extremely relevant to flow modelling, i.e. equivariant methods, geometric deep learning, and JEPA-like methods.

If the Area Chair is of the opinion that this work is in the scope of the conference and all points are addressed the work could merit publication at ICLR.

Critical strengths
  • The work identifies weaknesses in needing to rely on matching geometry-flow pairs with substantial online data available that is geometry-only, and exploits this by providing a matching pretraining strategy that optimizes for solving an encoder-decoder method that targets a flow surrogate, thus being more aligned with downstream task. They also augment the corpus of data in a meaningful way to simulate different flow conditions.

  • The idea of combining learned flow primitives is interesting and is a natural fit that forces the method to be constrained to flow priors that it has to rely on in order to learn accurate decoding.

Critical weaknesses
  • The authors point out that classical geometry reconstruction-based pre-training optimizes for the wrong goal (C.1: “biases model attention toward high-frequency boundary noise rather than volumetric convective momentum.”), as the learned representation should be tuned for the downstream task. It therefore makes sense to learn something more analogous to flow during decoding, which the authors do. However, these targets are still hand-designed, and an answer to this would be to not specify a target at all and let it be discovered. A comparison with Point-JEPA [1] or an I-JEPA-style [2] vascular baseline would therefore be fair, as these would address representation learning in an unsupervised fashion and could allow for fair results. Finally, the work could benefit from applying the method to another PDE-like downstream task where they could pretrain on ShapeNet and test for downstream performance on e.g. DrivAerNet++ [3]. Though this would be an extensive revision, it would show the purpose of the method as a more general-purpose physics-simulation-based auto-encoder.

  • The authors miss an entire literature on geometric and equivariant deep learning [4], which may play a large role in the gains of the proposed method (which may come from supplying orientation consistency rather than hemodynamic structure). Could authors report the gain of HemoPT when compared with an encoder that handles orientation natively (e.g. LaB-GATr [5])? If a backbone swap is out of scope for the rebuttal, a cheap control could be to retrain Scratch and HemoPT with rotation augmentation.

[1] Saito, A., Kudeshia, P., & Poovvancheri, J. (2025, February). Point-jepa: A joint embedding predictive architecture for self-supervised learning on point cloud. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (pp. 7348-7357). IEEE. [2] Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., ... & Ballas, N. (2023, June). Self-supervised learning from images with a joint-embedding predictive architecture. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 15619-15629). IEEE. [3] Elrefaie, M., Morar, F., Dai, A., & Ahmed, F. (2024). Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks. Advances in Neural Information Processing Systems, 37, 499-536. [4] Suk, J., Nannini, G., Rygiel, P., Brune, C., Pontone, G., Redaelli, A., & Wolterink, J. M. (2025). Deep vectorised operators for pulsatile hemodynamics estimation in coronary arteries from a steady-state prior. Computer methods and programs in biomedicine, 108958. [5] Suk, J., Imre, B., & Wolterink, J. M. (2024, October). LaB-GATr: geometric algebra transformers for large biomedical surface and volume meshes. In International conference on medical image computing and computer-assisted intervention (pp. 185-195). Cham: Springer Nature Switzerland.

Questions for authors
  • Could the authors compare to JEPA-style training (see [1,2])?
  • Compare to an equivariant baseline, or retrain with rotation augmentation (see [5])?
  • The corpus of data is overstated in size. Since a dataset of 5K geometries is already impressive, I see no reason to overstate it by sampling 20 probe conditions. Furthermore, sampling these conditions with ray casting could likely be made efficient enough to essentially simulate any probe condition on the fly, resulting in an “infinite” corpus of data. Could the authors frame this appropriately?
  • At a fixed training budget, is it better to spend that budget on more geometries or more views per geometry? The authors could vary K vs R, which could show whether geometries are the limiting factor or the number of probe conditions.

Comments on the paper

1 comment

I thank the authors for the additional experiments and clarifications. The paired runs strengthen the evidence that the flow proxy provides benefits beyond spatial pretraining, particularly on Aneumo and ICA-Sim. I also appreciate the more qualified interpretation of the flow templates and the case-level uncertainty analysis. Nevertheless, the incremental gains remain modest, the 4DFlow confidence intervals include zero, and VMR-AOCOA results remain mixed with high absolute errors. The boundary-condition families and their variation across cases also remain insufficiently specified. Overall, the rebuttal meaningfully addresses several concerns, but does not fully establish the strength and generality of the central flow-proxy contribution. I therefore maintain my recommendation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.