arxiv: 2603.22421 · v2 · submitted 2026-03-23 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

OsteoFlow: Lyapunov-Guided Flow Distillation for Predicting Bone Remodeling after Mandibular Reconstruction

Hamidreza Aftabi , Faye Yu , Brooke Switzer , Zachary Fishman , Eitan Prisman , Antony Hodgson , Cari Whyne , Sidney Fels

show 1 more author

Michael Hardisty

Authors on Pith no claims yet

Pith reviewed 2026-05-15 00:15 UTC · model grok-4.3

classification 💻 cs.CV

keywords bone remodelingmandibular reconstructionflow-based predictiontrajectory distillationLyapunov guidanceCT image forecastingmedical image registrationlong-term prediction

0 comments

The pith

OsteoFlow distills continuous flow trajectories from a registration velocity field using Lyapunov guidance to predict year-one bone remodeling from day-five CT scans after mandibular reconstruction.

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

The paper introduces OsteoFlow to forecast long-term bone changes following jaw reconstruction surgery. Standard generative models lose consistency when predicting far into the future, but this method extracts a full transport trajectory from an early scan to a later one. It treats a stationary velocity field obtained via image registration as a teacher and distills that field into a generative flow model while adding Lyapunov stability constraints and a resection-specific loss term. The result is a prediction that keeps anatomical landmarks aligned over many months without collapsing generative variety. If the approach holds, surgeons could obtain reliable estimates of final bone shape months ahead of the actual outcome.

Core claim

OsteoFlow predicts Year-1 post-operative CT scans from Day-5 scans by distilling a continuous trajectory over transport time from a registration-derived stationary velocity field teacher. Lyapunov guidance enforces stability along the trajectory, and a resection-aware image loss maintains geometric correspondence between predicted and observed anatomy. On 344 paired regions of interest the method reduces mean absolute error in the surgical resection zone by approximately 20 percent relative to prior baselines.

What carries the argument

Lyapunov-guided trajectory distillation, which converts a stationary velocity field from registration into a stable, continuous generative flow while preserving anatomical geometry via a resection-aware loss.

If this is right

Trajectory distillation produces lower error in the resection zone than one-step or standard generative baselines.
Lyapunov constraints keep predicted bone shapes consistent across the entire transport interval.
The resection-aware loss term improves geometric correspondence without eliminating the model's ability to generate varied but plausible outcomes.
The framework directly supports clinical visualization of expected year-one anatomy from an early post-operative scan.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same distillation pattern could be tested on other longitudinal medical imaging tasks such as tumor growth or joint degeneration where registration velocity fields are available.
Adding explicit stability penalties from dynamical systems may generalize to other long-horizon image forecasting problems where drift accumulates quickly.
If the teacher velocity field itself contains registration artifacts, the distilled model will inherit them; direct validation against independent biomechanical simulations would expose this inheritance.

Load-bearing premise

The registration-derived stationary velocity field provides an unbiased teacher signal whose distilled trajectories remain anatomically faithful over the full one-year horizon.

What would settle it

A new set of paired day-five and year-one CT scans in which OsteoFlow predictions show no reduction, or an increase, in mean absolute error within the resection zone compared with standard flow or diffusion baselines would falsify the central performance claim.

Figures

Figures reproduced from arXiv: 2603.22421 by Antony Hodgson, Brooke Switzer, Cari Whyne, Eitan Prisman, Faye Yu, Hamidreza Aftabi, Michael Hardisty, Sidney Fels, Zachary Fishman.

**Figure 2.** Figure 2: Model predictions for three representative cases (union, partial union, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

read the original abstract

Predicting long-term bone remodeling after mandibular reconstruction would be of great clinical benefit, yet standard generative models struggle to maintain trajectory-level consistency and anatomical fidelity over long horizons. We introduce OsteoFlow, a flow-based framework predicting Year-1 post-operative CT scans from Day-5 scans. Our core contribution is Lyapunov-guided trajectory distillation: Unlike one-step distillation, our method distills a continuous trajectory over transport time from a registration-derived stationary velocity field teacher. Combined with a resection-aware image loss, this enforces geometric correspondence without sacrificing generative capacity. Evaluated on 344 paired regions of interest, OsteoFlow significantly outperforms state of-the-art baselines, reducing mean absolute error in the surgical resection zone by ~20%. This highlights the promise of trajectory distillation for long-term prediction. Code is available on GitHub: OsteoFlow.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

OsteoFlow distills continuous trajectories from a registration velocity field for year-one bone remodeling prediction, but the 20% MAE claim rests on thin evidence and a questionable teacher signal.

read the letter

The main thing to know is that OsteoFlow applies Lyapunov-guided distillation to pull full transport trajectories out of a stationary velocity field from image registration, then adds a resection-aware loss to keep things aligned in the surgical zone. This is a concrete step beyond one-step distillation for long-horizon medical prediction tasks. The approach is new in this domain and the GitHub release makes it easy to inspect or extend. The resection-aware term is a sensible way to protect geometric fidelity without killing generative capacity. That part is worth looking at if you work on flow models for medical sequences. The evaluation is the soft spot. The abstract reports a roughly 20% MAE drop on 344 paired ROIs in the resection zone, yet supplies no error bars, no statistical tests, and no information on how the ROIs were selected or how the train-test split was done. More critically, the teacher comes from standard registration, which optimizes for image overlap rather than biological remodeling. Bone changes over a year involve active resorption, formation, and vascular shifts that are not stationary transport. If those dynamics are not in the velocity field, the distilled predictions could simply reproduce registration biases instead of learning real remodeling. The stress-test concern lands here. This work is aimed at researchers in medical image synthesis or surgical planning who already use flow or generative models. A reader who wants to try trajectory distillation on their own longitudinal CT data could get ideas from it, but anyone expecting ready-to-use clinical numbers will be disappointed. The paper deserves peer review. The core idea is coherent enough that referees can pressure-test the validation and the biological assumptions, which is exactly what it needs.

Referee Report

2 major / 2 minor

Summary. The paper introduces OsteoFlow, a flow-based generative framework for predicting Year-1 post-operative CT scans from Day-5 scans after mandibular reconstruction. Its central contribution is Lyapunov-guided trajectory distillation from a registration-derived stationary velocity field teacher, combined with a resection-aware image loss to enforce geometric correspondence over long horizons. On 344 paired regions of interest, the method is reported to reduce mean absolute error in the surgical resection zone by approximately 20% relative to state-of-the-art baselines.

Significance. If the performance claims hold under rigorous validation, the work offers a promising direction for long-horizon medical image prediction by addressing trajectory consistency issues in generative models. The explicit code release supports reproducibility, and the focus on resection-aware losses could generalize to other post-surgical remodeling tasks.

major comments (2)

[Results] Results section (and abstract): The central claim of a ~20% MAE reduction on 344 paired ROIs is presented without error bars, statistical significance tests, details on train/validation/test splits, or exclusion criteria. This makes it impossible to determine whether the improvement is robust or could be explained by data partitioning choices.
[Method] Method section (trajectory distillation): The assumption that a registration-derived stationary velocity field provides a faithful teacher for biological bone remodeling over 1-year horizons is load-bearing but insufficiently justified. Registration optimizes for image alignment rather than active processes (resorption, formation, vascularization), so any systematic geometric bias in the teacher could be replicated by the distilled trajectories without reflecting true dynamics.

minor comments (2)

[Abstract] Abstract: 'state of-the-art' contains a spacing error and should read 'state-of-the-art'.
[Method] Notation: The precise definition of the Lyapunov function and its integration into the distillation objective should be stated explicitly with an equation number for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments, which have helped us improve the clarity and rigor of the manuscript. We address each major comment point by point below.

read point-by-point responses

Referee: [Results] Results section (and abstract): The central claim of a ~20% MAE reduction on 344 paired ROIs is presented without error bars, statistical significance tests, details on train/validation/test splits, or exclusion criteria. This makes it impossible to determine whether the improvement is robust or could be explained by data partitioning choices.

Authors: We agree that these statistical and methodological details are required to substantiate the performance claims. In the revised manuscript we have added: (i) error bars showing mean ± standard deviation across the 344 ROIs, (ii) paired Wilcoxon signed-rank tests with exact p-values comparing OsteoFlow against each baseline, (iii) explicit patient-level train/validation/test splits (60/20/20) chosen to prevent leakage, and (iv) exclusion criteria (incomplete 1-year follow-up, severe metal artifacts, or motion corruption). These additions appear in the Results section, a new supplementary table, and an updated abstract. revision: yes
Referee: [Method] Method section (trajectory distillation): The assumption that a registration-derived stationary velocity field provides a faithful teacher for biological bone remodeling over 1-year horizons is load-bearing but insufficiently justified. Registration optimizes for image alignment rather than active processes (resorption, formation, vascularization), so any systematic geometric bias in the teacher could be replicated by the distilled trajectories without reflecting true dynamics.

Authors: We acknowledge that the registration-derived SVF encodes geometric correspondence rather than explicit biological mechanisms. This choice is motivated by the practical absence of dense longitudinal data that would capture active remodeling dynamics at the required resolution. In the revision we have expanded the Method section with additional justification, referencing prior longitudinal registration studies in craniofacial surgery, and added a dedicated limitations paragraph that explicitly discusses the risk of propagating registration biases. We also include a new ablation that contrasts the Lyapunov-guided trajectories against a purely image-regression baseline to quantify the added value of the distillation step. revision: partial

Circularity Check

0 steps flagged

Derivation chain self-contained with independent evaluation

full rationale

The OsteoFlow framework introduces Lyapunov-guided trajectory distillation from a registration-derived stationary velocity field teacher to predict Year-1 post-operative CT scans from Day-5 scans, combined with a resection-aware image loss. Evaluation on 344 paired regions of interest measures mean absolute error reductions against actual follow-up data, without any equations or steps reducing the reported performance gains to quantities fitted inside the same training loop by construction. No self-definitional relations, fitted inputs renamed as predictions, or load-bearing self-citations appear in the derivation; the central claims rest on the distillation process and independent paired benchmarks rather than tautological reductions to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract provides insufficient detail to enumerate specific free parameters; the primary domain assumption is the reliability of the registration-derived velocity field as a teacher.

axioms (1)

domain assumption The registration-derived stationary velocity field accurately represents the underlying transport dynamics for bone remodeling trajectories.
Invoked as the teacher signal for trajectory distillation.

pith-pipeline@v0.9.0 · 5468 in / 1104 out tokens · 49248 ms · 2026-05-15T00:15:03.163638+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

registration-derived stationary velocity field teacher... diffeomorphic SVF registration

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 1 internal anchor

[1]

Computational models and their appli- cations in biomechanical analysis of mandibular reconstruction surgery.Comput

Hamidreza Aftabi, Katrina Zaraska, Atabak Eghbal, Sophie McGregor, Eitan Pris- man, Antony Hodgson, and Sidney Fels. Computational models and their appli- cations in biomechanical analysis of mandibular reconstruction surgery.Comput. Biol. Med., 169:107887, 2024

work page 2024
[2]

Gold- man, Howard Krein, Ryan N

Brian Swendseid, Ayan Kumar, Larissa Sweeny, Tingting Zhan, Richard A. Gold- man, Howard Krein, Ryan N. Heffelfinger, Adam J. Luginbuhl, and Joseph M. Curry. Natural history and consequences of nonunion in mandibular and maxil- lary free flaps.Otolaryngol. Head Neck Surg., 163(5):956–962, 2020

work page 2020
[3]

Lloyd, Amanda Ding, Benedikt Sagl, Eitan Prisman, Antony Hodgson, and Sidney Fels

Hamidreza Aftabi, John E. Lloyd, Amanda Ding, Benedikt Sagl, Eitan Prisman, Antony Hodgson, and Sidney Fels. Osteoopt: A bayesian optimization framework for enhancing bone union likelihood in mandibular reconstruction surgery. InMIC- CAI, pages 448–458, 2025

work page 2025
[4]

Durham, and Eitan Prisman

FarahnaSabiq,AbhiramCherukupalli,MohammadKhalil,LinhK.Tran,JamieJY Kwon, Thomas Milner, James S. Durham, and Eitan Prisman. Evaluating the ben- efit of virtual surgical planning on bony union rates in head and neck reconstructive surgery.Head Neck, 46(6):1322–1330, 2024

work page 2024
[5]

Lloyd, Eitan Prisman, Antony Hodgson, and Sidney Fels

Hamidreza Aftabi, Benedikt Sagl, John E. Lloyd, Eitan Prisman, Antony Hodgson, and Sidney Fels. To what extent can mastication functionality be restored follow- ing mandibular reconstruction surgery? a computer modeling approach.Comput. Methods Programs Biomed., 250:108174, 2024

work page 2024
[6]

Lloyd, Benedikt Sagl, Amanda Ding, Eitan Prisman, Antony Hodgson, and Sidney Fels

Hamidreza Aftabi, John E. Lloyd, Benedikt Sagl, Amanda Ding, Eitan Prisman, Antony Hodgson, and Sidney Fels. Optimizing bone cuts enhances predicted bone union propensity in mandibular body reconstruction. InISBI, pages 1–4, 2025. 10 H. Aftabi et al

work page 2025
[7]

A machine learning-based multiscale model to predict bone formation in scaffolds.Nat

Chi Wu, Ali Entezari, Keke Zheng, Jianguang Fang, Hala Zreiqat, Grant P Steven, Michael V Swain, and Qing Li. A machine learning-based multiscale model to predict bone formation in scaffolds.Nat. Comput. Sci., 1:532–541, 2021

work page 2021
[8]

Schmon, and Chris G

Julian Wyatt, Adam Leach, Sebastian M. Schmon, and Chris G. Willcocks. Anod- dpm: Anomaly detection with denoising diffusion probabilistic models using sim- plex noise. InCVPR, pages 650–656, 2022

work page 2022
[9]

Adaptive latent diffusion model for 3d medical image-to-image translation: Multi-modal magnetic resonance imaging study

Jonghun Kim and Hyunjin Park. Adaptive latent diffusion model for 3d medical image-to-image translation: Multi-modal magnetic resonance imaging study. In W ACV, pages 7604–7613, 2024

work page 2024
[10]

Cross-conditioned diffusion model for medical image-to-image translation

Zhaohu Xing, Sicheng Yang, Sixiang Chen, Tian Ye, Yijun Yang, Jing Qin, and Lei Zhu. Cross-conditioned diffusion model for medical image-to-image translation. In MICCAI, pages 201–211, 2024

work page 2024
[11]

Clinicalfmamba: Advancing clinical assessment using mamba-based multimodal neuroimaging fusion

Meng Zhou and Farzad Khalvati. Clinicalfmamba: Advancing clinical assessment using mamba-based multimodal neuroimaging fusion. InMLMI, pages 245–255, 2025

work page 2025
[12]

Controllable flow matching for 3d contrast- enhanced brain MRI synthesis from non-contrast scans

Heng Chang, Yu Shang, Haifeng Wang, Yuxia Liang, Haoyu Wang, Fan Wang, Chen Niu, and Chunfeng Lian. Controllable flow matching for 3d contrast- enhanced brain MRI synthesis from non-contrast scans. InMICCAI, pages 119– 128, 2025

work page 2025
[13]

Maisi: Medical AI for synthetic imaging

Pengfei Guo, Can Zhao, Dong Yang, Ziyue Xu, Vishwesh Nath, Yucheng Tang, Benjamin Simon, Mason Belue, Stephanie Harmon, Baris Turkbey, et al. Maisi: Medical AI for synthetic imaging. InW ACV, pages 4430–4441, 2025

work page 2025
[14]

Difr3ct: Latent diffusion for probabilistic 3d CT reconstruction from few planar x-rays.arXiv preprint arXiv:2408.15118, 2024

Yiran Sun, Hana Baroudi, Tucker Netherton, Laurence Court, Osama Mawlawi, Ashok Veeraraghavan, and Guha Balakrishnan. Difr3ct: Latent diffusion for probabilistic 3d CT reconstruction from few planar x-rays.arXiv preprint arXiv:2408.15118, 2024

work page arXiv 2024
[15]

Anatomic-constrained medical image synthesis via physiological density sampling

Yuetan Chu, Changchun Yang, Gongning Luo, Zhaowen Qiu, and Xin Gao. Anatomic-constrained medical image synthesis via physiological density sampling. InMICCAI, pages 69–79, 2024

work page 2024
[16]

Mazurowski

Nicholas Konz, Yuwen Chen, Haoyu Dong, and Maciej A. Mazurowski. Anatomically-controllable medical image generation with segmentation-guided dif- fusion models. InMICCAI, pages 88–98, 2024

work page 2024
[17]

Huaisheng Zhu, Teng Xiao, Shijie Zhou, Zhimeng Guo, Hangfan Zhang, Siyuan Xu, and Vasant G. Honavar. Simple distillation for one-step diffusion models. In NeurIPS, 2025

work page 2025
[18]

A new learning paradigm: Learning using privileged information.Neural Netw., 22(5–6):544–557, 2009

Vladimir Vapnik and Akshay Vashist. A new learning paradigm: Learning using privileged information.Neural Netw., 22(5–6):544–557, 2009

work page 2009
[19]

Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[20]

Kernel bundle diffeomorphic image registration using stationary ve- locity fields and wendland basis functions.IEEE Trans

Akshay Pai, Stefan Sommer, Lauge Sørensen, Sune Darkner, Jon Sporring, and Mads Nielsen. Kernel bundle diffeomorphic image registration using stationary ve- locity fields and wendland basis functions.IEEE Trans. Med. Imaging, 35(6):1369– 1380, 2015

work page 2015
[21]

Khalil and Jessy W

Hassan K. Khalil and Jessy W. Grizzle.Nonlinear Systems, volume 3. Prentice Hall, 2002

work page 2002
[22]

Lyapunov guidance: Stabilizing generative flows with one-line code

Jingdong Zhang, Xinze Li, Yize Jiang, Luan Yang, Minkai Xu, and Junhong Liu. Lyapunov guidance: Stabilizing generative flows with one-line code

work page
[23]

Abhay Datarkar, Bhavana Valvi, Suraj Parmar, and Jagadish Patil. Qualitative assessment of newly formed bone after distraction osteogenesis of mandible in Title Suppressed Due to Excessive Length 11 patients with facial asymmetry using 3-dimensional computed tomography.J. Oral Biol. Craniofac. Res., 11(3):410–414, 2021

work page 2021
[24]

Lienkamp, Thomas Brox, and Olaf Ronneberger

Özgün Çiçek, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning dense volumetric segmentation from sparse annotation. InMICCAI, pages 424–432, 2016

work page 2016
[25]

Film: Visual reasoning with a general conditioning layer

Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InAAAI, volume 32, 2018

work page 2018
[26]

Smedsrud, Michael A

Debesh Jha, Pia H. Smedsrud, Michael A. Riegler, Dag Johansen, Thomas De Lange, Pål Halvorsen, and Håvard D. Johansen. Resunet++: An advanced architecture for medical image segmentation. InISM, pages 225–2255, 2019

work page 2019
[27]

Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. InCVPR, pages 1125–1134, 2017

work page 2017
[28]

Davis, and Abhinav Shrivastava

Saksham Suri, Moustafa Meshry, Larry S. Davis, and Abhinav Shrivastava. Grit: Gan residuals for paired image-to-image translation. InW ACV, pages 4965–4975, 2024

work page 2024
[29]

Denoising diffusion probabilistic mod- els.Adv

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic mod- els.Adv. Neural Inf. Process. Syst., 33:6840–6851, 2020

work page 2020
[30]

CIResDiff: A clinically-informed residual diffusion model for predicting idiopathic pulmonary fibrosis progression

Caiwen Jiang, Xiaodan Xing, Zaixin Ou, Mianxin Liu, Walsh Simon, Guang Yang, and Dinggang Shen. CIResDiff: A clinically-informed residual diffusion model for predicting idiopathic pulmonary fibrosis progression. InMLMI, pages 83–93, 2024

work page 2024
[31]

Medvae: Efficient automated interpretation of medical im- ages with large-scale generalizable autoencoders.arXiv preprint arXiv:2502.14753, 2025

Maya Varma, Ashwin Kumar, Rogier Van der Sluijs, Sophie Ostmeier, Louis Blankemeier, Pierre Chambon, Christian Bluethgen, Jip Prince, Curtis Langlotz, and Akshay Chaudhari. Medvae: Efficient automated interpretation of medical im- ages with large-scale generalizable autoencoders.arXiv preprint arXiv:2502.14753, 2025

work page arXiv 2025
[32]

Roth, and Daguang Xu

Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R. Roth, and Daguang Xu. Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images. InMICCAI BrainLesion Workshop, pages 272–284, 2021

work page 2021