The DOI matches its Crossref record; none of the 27 sources its claims cite is checked for support yet.
PolyPose: Deformable 2D/3D Registration via Polyrigid Transformations
Open accessAbstract
Determining the 3D pose of a patient from a limited set of 2D X-ray images is a critical task in interventional settings. While preoperative volumetric imaging (e.g., CT and MRI) provides precise 3D localization and visualization of anatomical targets, these modalities cannot be acquired during procedures, where fast 2D imaging (X-ray) is used instead. To integrate volumetric guidance into intraoperative procedures, we present PolyPose, a simple and robust method for deformable 2D/3D registration. PolyPose parameterizes complex 3D deformation fields as a composition of rigid transforms, leveraging the biological constraint that individual bones do not bend in typical motion. Unlike existing methods that either assume no inter-joint movement or fail outright in this under-determined setting, our polyrigid formulation enforces anatomically plausible priors that respect the piecewise-rigid nature of human movement. This approach eliminates the need for expensive deformation regularizers that require patient- and procedure-specific hyperparameter optimization. Across extensive experiments on diverse datasets from orthopedic surgery and radiotherapy, we show that this strong inductive bias enables PolyPose to successfully align the patient's preoperative volume to as few as two X-rays, thereby providing crucial 3D guidance in challenging sparse-view and limited-angle settings where current registration methods fail. Additional visualizations, tutorials, and code are available at https://polypose.csail.mit.edu.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 of 27 cited sources checked · 0 supported · 0 flagged
- Findingown data
PolyPose surpasses both deep learning and optimization-based 2D/3D registration methods.
“Our method outperforms both deep learning and optimization-based 2D/3D registration methods”
Quote in the paperSubstantial claim - Findingown data
PolyPose remains robust even when only a few input views from limited angles are available.
“Empirically, across diverse datasets, PolyPose is robust even for a small number of input views from limited angles.”
Quote in the paperSubstantial claim - Findingown data
PolyPose outperforms both deep learning-based 2D/3D methods.
“PolyPose outperforms both deep learning-based 2D/3D methods [18, 19]”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [18]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [19]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingown data
In experiments on diverse orthopedic surgery and radiotherapy datasets, PolyPose aligns the preoperative volume to as few as two X-rays.
“Across extensive experiments on diverse datasets from orthopedic surgery and radiotherapy, we show that this strong inductive bias enables PolyPose to successfully align the patient’s preoperative volume to as few as two X-rays”
Quote in the paperSubstantial claim- Rests on an assumption: Despite non-rigid patient motion between acquisitions, there exists a global rigid transform for an individual articulated structure.
- Findingown data
Equation (5) yielded inaccurate deformation fields for volumes with articulated bodies of very different sizes.
“However, we found that Eq. (5) produced inaccurate deformation fields for volumes containing articulated bodies with very different sizes (Table 3).”
Quote in the paperSubstantial claim
- Findingown data
Among evaluated methods, PolyPose gives the most accurate deformation fields for every number of input X-rays used as registration targets.
“Of all evaluated methods, PolyPose estimates the most accurate deformation fields across all numbers of input X-rays available as registration targets.”
Quote in the paperSubstantial claim - Findingown data
PolyPose estimates deformation fields with minimal topological defects.
“PolyPose also estimates deformation fields with minimal topological defects.”
Quote in the paperSubstantial claim - Findingown data
Including the skull, cervical spine, and thoracic spine stabilizes the deformation fields estimated by PolyPose.
“Including the skull, cervical spine, and thoracic spine stabilizes the deformation fields estimated by PolyPose (Figure 7A).”
Quote in the paperSubstantial claim - Findingown data
The hyperparameter-free weighting function in Eq. (6) outperforms the widely used formulation in Eq. (5).
“Our hyperparameter-free weighting function in Eq. (6) outperforms the widely used formulation in Eq. (5).”
Quote in the paperSubstantial claim - Findingown data
With minimal supervision, only the low-dimensional deformation model enables localization of the misaligned femurs without topological defects.
“Given minimal supervision, only our low-dimensional deformation model enables the localization of the misaligned femurs without topological defects.”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [23]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [32]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [42]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Findingreasoning
Nearly all baselines exhibit high accuracy on the pelvis.
“As such, nearly all baselines (evaluated using our camera matrices) exhibit high accuracy on the pelvis.”
Quote in the paperNot a substantial claim - Findingown data
For the femurs, PolyPose produces the highest accuracy.
“However, for the femurs, PolyPose produces the highest accuracy.”
Quote in the paperSubstantial claim - Findingown data
Optimal performance for the left and right femurs is achieved for vastly different hyperparameter values when using Eq. (5).
“Note that the optimal performance for the left and right femurs is achieved for vastly different hyperparameter values (ϵ=100 vs. ϵ=10-3) when using Eq. (5).”
Quote in the paperSubstantial claim - Findingown data
PolyPose produces interpretable warps, for example modeling the external rotation of the femurs.
“Visualizations of the estimated deformation fields and warped CTs show that PolyPose produces interpretable warps, e.g., modeling the external rotation of the femurs (Figure 6A and B).”
Quote in the paperSubstantial claim - Premisecitation
Polyrigid transforms possess desirable properties by construction, including smoothness, invertibility, and coordinate frame invariance.
“polyrigid transforms have several desirable properties by construction, such as smoothness, invertibility, and coordinate frame invariance [26].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [26]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationreasoning
The polyrigid formulation imposes anatomically plausible priors that respect the piecewise-rigid nature of human movement.
“our polyrigid formulation enforces anatomically plausible priors that respect the piecewise-rigid nature of human movement.”
Quote in the paperSubstantial claim- The authors state this as an assumption: Bones are rigid bodies and individual bones do not bend during typical motion.
- Interpretationreasoning
PolyPose allows 3D localization of critical structures during medical interventions from intraoperative 2D images.
“enables the 3D localization of critical structures during medical interventions from intraoperative 2D images.”
Quote in the paperSubstantial claim - Interpretationreasoning
This method removes the need for costly deformation regularizers that require patient- and procedure-specific hyperparameter tuning.
“This approach eliminates the need for expensive deformation regularizers that require patient- and procedure-specific hyperparameter optimization.”
Quote in the paperSubstantial claim - Premisecitation
Rigid 2D/3D registration methods align global structure but ignore soft tissue deformation and articulated inter-joint motion during procedures.
“rigid 2D/3D registration methods align global structure [14–17], but do not account for the soft tissue deformation or articulated inter-joint motion that occurs during procedures”
Quote in the paperSubstantial claim0 of 4 sources checked- cites [14]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [15]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [16]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [17]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Premisecitation
Equation (5) has mostly been used when substructures have comparable volumes, such as certain brain regions or carpal bones.
“To our knowledge, Eq. (5) has largely only been used when the constituent substructures have comparable volumes, such as certain brain regions [51] or the carpal bones [52, 53].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [51]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [52]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [53]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Interpretationown data
PolyPose provides essential 3D guidance in difficult sparse-view and limited-angle scenarios where existing registration methods fail.
“thereby providing crucial 3D guidance in challenging sparse-view and limited-angle settings where current registration methods fail.”
Quote in the paperSubstantial claim - Interpretationreasoning
PolyPose cannot model separating a rigid body into two, such as opening the jaw, because a diffeomorphism cannot represent it.
“For example, separating a rigid body into two (e.g., opening the jaw) cannot be represented by a diffeomorphism and thus cannot be modeled by PolyPose.”
Quote in the paperNot a substantial claim- The authors state this as an assumption: Bones are rigid bodies and individual bones do not bend during typical motion.
- Interpretationreasoning
PolyPose may fail to generalize to settings with insufficient rigid bodies to constrain soft tissue deformation, such as within the abdomen.
“However, PolyPose may fail to generalize to settings where there are insufficient rigid bodies to constrain the deformation of soft tissues (e.g., within the abdomen).”
Quote in the paperNot a substantial claim- The authors state this as an assumption: Bones are rigid bodies and individual bones do not bend during typical motion.
- Interpretationnone given
PolyPose's lack of need for regularization and near-absence of hyperparameters make it generically applicable across a broad set of medical procedures.
“PolyPose’s lack of need for regularization and near-absence of hyperparameters make it generically applicable across a broad set of medical procedures.”
Quote in the paperNot a substantial claim - Interpretationreasoning
Training on the limited datasets available in interventional settings produces models that fail to generalize.
“suggesting that training on the limited datasets available in interventional settings produces models that fail to generalize.”
Quote in the paperSubstantial claim
- It may state two separate claims as one. 19
- The quote only partly states this claim. 1
It may state two separate claims as one.
FindingreasoningPolyPose needs no regularization and has no tunable hyperparameters besides the optimizer step size.
“PolyPose does not require any regularization and has no tunable hyperparameters other than the step size of the optimizer.”
Quote in the paperNot a substantial claimIt may state two separate claims as one.
Findingown dataPolyPose achieves the highest 3D Dice on both rigid structures and important soft tissue organs, even though the pose of these organs was not directly optimized.
“PolyPose achieves the highest 3D Dice on both rigid structures and important soft tissue organs, even though the pose of these organs was not directly estimated during optimization.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataPolyPose produces non-rigid deformations that are highly accurate, anatomically plausible, and topologically consistent.
“Our method, PolyPose, enables the estimation of highly accurate non-rigid deformations that are anatomically plausible and topologically consistent.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataThe proposed weight field removes challenging hyperparameter optimization while still producing topologically valid deformations.
“This formulation eliminates challenging hyperparameter optimization while still producing topologically valid deformations (Table 3).”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataAdding further rigid bodies yields only marginal improvements, showing PolyPose is expressive even when constrained to a subset of rigid bodies in an anatomical region.
“Adding further rigid bodies yields only marginal improvements, demonstrating that PolyPose is expressive even when constrained to a subset of rigid bodies in an anatomical region.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataPolyPose is memory-efficient, able to jointly optimize the poses of 26 rigid bodies in a large CT scan on a single 48 GB GPU.
“PolyPose is memory-efficient, capable of jointly optimizing the poses of 26 rigid bodies in a large CT scan on a single GPU with 48 GB of vRAM.”
Quote in the paperSubstantial claimIt may state two separate claims as one. The quote only partly states this claim.
Findingown dataIn a subject with only minimal head motion, PolyPose recovers the exact underlying deformation, while anatomix produces topologically-defective and irregular warps as measured by %Folds and σ(log|J|).
“For example, in a subject with only minimal head motion, PolyPose recovers the exact underlying deformation (Figure 4D)”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [38]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [63]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
Findingown dataConstructing the field from a small number of rigid components yields interpretable deformations that are more anatomically plausible than baselines.
“Our construction from a small number of rigid components yields interpretable deformation fields that are more anatomically plausible than baselines.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Findingown dataPolyPose estimated the most accurate and topologically correct warps in both sparse-view and limited-angle settings.
“PolyPose estimated the most accurate and topologically correct warps in both sparse-view and limited-angle settings.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
FindingcitationDense parameterizations yield anatomically implausible and unintelligible deformations because their objective prioritizes memorizing the appearance of the training views.
“In contrast, the dense parameterizations yield anatomically implausible and unintelligible deformations as their objective prioritizes memorizing the appearance of the training views.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [32]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [42]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
Findingown dataPolyPose is robust to extreme segmentation corruption, even outperforming baselines over erosion radii from 1mm to 3mm in Dice and Hausdorff Distance.
“We find that PolyPose is robust to extreme segmentation corruption, even outperforming the baselines in Table 1 over a range of erosion radii from 1mm to 3mm, both in terms of Dice and Hausdorff Distance.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningPolyPose models complex 3D deformation fields as a composition of rigid transforms, relying on the biological constraint that bones remain rigid in normal movement.
“PolyPose parameterizes complex 3D deformation fields as a composition of rigid transforms, leveraging the biological constraint that individual bones do not bend in typical motion.”
Quote in the paperSubstantial claim- The authors state this as an assumption: Bones are rigid bodies and individual bones do not bend during typical motion.
It may state two separate claims as one.
PremisereasoningDeformable models need extensive application-specific regularization to produce anatomically faithful warps, which introduces new modeling choices and hyperparameter tuning for each subject, procedure, and anatomical region.
“deformable models require extensive application-specific regularization to generate anatomically faithful warps, thereby introducing new modeling decisions and hyperparameter tuning for every subject, procedure, and anatomical region.”
Quote in the paperSubstantial claim0 of 8 sources checked- cites [18]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [19]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [20]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [21]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [22]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [23]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [24]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [25]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
InterpretationreasoningThe resulting polyrigid warp is diffeomorphic, anatomically constrained, and suitable for the ill-posed setting.
“By fusing log-transformed versions of the pose for each structure, as opposed to simply averaging their associated displacements, the resulting polyrigid warp is diffeomorphic, anatomically constrained, and well-suited to our ill-posed setting.”
Quote in the paperSubstantial claim- The authors state this as an assumption: Bones are rigid bodies and individual bones do not bend during typical motion.
It may state two separate claims as one.
Interpretationown dataPolyPose is an effective and reliable approach for deformable 2D/3D registration.
“we present PolyPose, a simple and robust method for deformable 2D/3D registration.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
InterpretationreasoningDense deformation models can only affect voxels with direct pixel supervision, whereas PolyPose extrapolates to unsupervised anatomy via piecewise-rigidity.
“Additionally, dense deformation models can only influence voxels on which they have direct pixel supervision (see the broken femurs in Figure 6B), whereas PolyPose extrapolates to unsupervised anatomy via piecewise-rigidity.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
PremisecitationLiftReg and 2D3D-RegNet cannot be trained on DeepFluoro because they require multiple CTs per patient, while each subject has only one volume.
“The deep learning methods LiftReg [18] and 2D3D-RegNet [19] cannot be trained on this dataset since they require multiple CTs from each patient”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [18]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [19]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
It may state two separate claims as one.
PremisereasoningPolyPose has O(K) optimizable parameters and is well suited for ill-posed settings, whereas dense representations have O(M) parameters with K much smaller than M.
“PolyPose has only 𝒪(K) optimizable parameters and is thus well suited for ill-posed settings, whereas the under-constrained dense representations have 𝒪(M) parameters with K≪M.”
Quote in the paperSubstantial claimIt may state two separate claims as one.
Interpretationown dataEquation (5) has a large hyperparameter search space, requiring a different epsilon for every rigid body.
“Thus, Eq. (5) has a large hyperparameter search space, requiring a different ϵ for every rigid body.”
Quote in the paperSubstantial claim
The authors state this as an assumption.
Bones are rigid bodies and individual bones do not bend during typical motion.
“Our approach is instead motivated by a generic anatomical prior: bones are rigid bodies.”
The paper relies on this without support.
Despite non-rigid patient motion between acquisitions, there exists a global rigid transform for an individual articulated structure.
“While patients move non-rigidly between the acquisitions of V and I, there exists a global rigid transform for an individual articulated structure.”
The paper relies on this without support.
X-ray imaging systems used in clinical practice are calibrated, so intrinsic parameters can be obtained from image metadata.
“as X-ray imaging systems used in clinical practice are calibrated, the intrinsic parameters K1,…,KN can easily be obtained from each image’s metadata”
The authors state this as an assumption.
All images in the set I are acquired simultaneously.
“we assume all images in I are acquired simultaneously”
The authors state this as an assumption.
The patient is in different positions for the acquisitions of the volume V and the images I.
“assume the patient is in different positions for the acquisitions of V and I”
The authors state this as an assumption.
All bones have constant density when estimating mass from volume.
“assuming a constant density for all bones”
The authors state this as an assumption.
X-ray projection can be approximated without scattering, beam hardening, and similar effects.
“i.e., no scattering, beam hardening, etc.”
The paper relies on this without support.
The anchor structure is reliably visible across all views in the image set I.
“a structure Sanchor that is reliably visible across all views in I”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Reported weight parameter exceeds its stated bound
MinorThe weight-function definition restricts ε to at most 1, but the ablation reports using ε=100 for that function.
“ϵ≤1 is a hyperparameter controlling the rate of decay of wk as x moves further away from Sk.”
“the optimal performance for the left and right femurs is achieved for vastly different hyperparameter values (ϵ=100 vs. ϵ=10-3) when using Eq. (5).”
Citations
Cited by
Sorted by citations
Adaptive Riemannian optimization for multi-scale diffeomorphic matching
Jena, R., Chaudhari, P., & Gee, J. C. (2026) · Nature Communications · doi:10.1038/s41467-026-72508-3 · Cited by 1
FlowReg: Reconstruction-Guided Flow Matching for Universal Spinal CT/X-Ray Registration
Shen, A., Jiang, J., Fu, X., Zeng, Q., Tang, Y., Chen, Z., Nong, L., Wang, F., & Zhou, S. K. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-38186-6_8
Jiang, J., Shen, A., Wang, Z., Fu, X., ZhengMing, C. H. E. N., Nong, L., Wang, F., & Zhou, S. K. (2026) · Medical Image Analysis · doi:10.1016/j.media.2026.104319
Metadata
- Authors
- Gopalakrishnan, V., Dey, N., & Golland, P.
- Affiliations
- Massachusetts Institute of Technology (US)
- Harvard University (US)
- Venue
- Advances in Neural Information Processing Systems 38
- Published
- 2025
- Pages
- 67454-67480
- ISSN
- 1049-5258 (ISSN-L)
- PMC
- PMC13464490
- OpenAlex
- W4415314062
- Cited by
- 3
- References
- 0
- Citations by year
- 2026: 3
- FWCI
- 2.26
- Type
- conference paper
- Language
- en
- Open access
- green
- pmc.ncbi.nlm.nih.gov
- Licence
- cc-by
- Locations
- neural information processing systems (publishedVersion)
- PubMed Central (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
- Physical Sciences › Engineering › Biomedical Engineering › Medical Imaging and Analysis
- Physical Sciences › Physics and Astronomy › Radiation › Advanced Radiotherapy Techniques
- Keywords
- 2D-3D registration, deformable image registration, X-ray images, intraoperative guidance, limited-angle CT, orthopedic surgery, radiotherapy, anatomical priors, 3D pose estimation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
