Pith. sign in

REVIEW 4 major objections 5 minor 78 references

Deformable registration and generative modelling of aortic anatomies by auto-decoders and neural ODEs

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces AD-SVFD, a single auto-decoder model that registers vascular anatomies to a reference by neural-ODE flows and generates new anatomies by sampling latent codes.

desk verdict Solid integration of auto-decoder and neural ODE registration for aortic shapes, but the headline accuracy claim rests on a test set mostly made of augmented copies of training patients. read the letter →

arxiv 2506.00947 v1 pith:LFEIOHXO submitted 2025-06-01 cs.CV cs.NAmath.NA

Riccardo Tenderini , Luca Pegolotti , Fanwei Kong , Stefano Pagani , Francesco Regazzoni , Alison L. Marsden , Simone Deparis This is my paper
classification cs.CVcs.NAmath.NA
keywords deformableregistrationneuralODEauto-decoderlatentshapecodesdiffeomorphicmappinggenerativemodellingaorticanatomypointcloud
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

AD-SVFD is a single deep-learning model that registers a whole cohort of vascular shapes to one reference and generates new anatomies from the same representation. Each shape becomes a weighted point cloud plus a low-dimensional latent code; the code conditions a neural network that defines a stationary velocity field, and integrating that field for unit time produces a smooth, invertible deformation. The paper reports that a model trained on roughly 880 synthetic plus 18 real aortas registers two held-out patients with maximal pointwise errors around 2-3 mm and average errors below 0.04 cm, at about 90 seconds per shape at inference. Because the inverse map is obtained by backward integration, the same model supplies both direct and reverse correspondences, and sampling new latent codes yields synthetic aortic geometries. If these results hold, AD-SVFD offers a fast, invertible, generative route to population-level vascular geometry analysis and simulation.

What carries the argument

The central object is the auto-decoder stationary vector field diffeomorphism (AD-SVFD). It names the combination of three mechanisms: (1) a per-shape latent code in $R^{256}$, reshaped into a 2x2x2x32 grid and trilinearly interpolated at each query point, which makes the network self-conditioning; (2) a fully connected velocity network with Fourier positional encoding, whose output is the time-independent right-hand side of the ODE d/dt $\varphi$(x;t) = v($\varphi$(x;t); Theta, z_i); and (3) bidirectional Chamfer-distance training with forward Euler integration for the direct map and a modified Euler step for the inverse map. The identity that carries the argument is that integrating the same ODE forward and backward yields a diffeomorphism and its inverse, so one network simultaneously solves both registration directions and can be repurposed for generation by sampling latent codes.

What would settle it

Hold out several patients whose anatomy contains features absent from the 18 training aortas, such as a subclavian artery bending toward the carotid or an unsegmented branch, run the frozen model, and measure pointwise error; if maximal error on these patients substantially exceeds the reported 0.2-0.28 cm range, the generalization claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that deformable registration and generative modelling can be unified in a single auto-decoder architecture, AD-SVFD. Each source anatomy carries a trainable 256-dimensional code; the code is reshaped into a low-resolution grid and interpolated at every spatial point to condition a fully connected network whose output is a stationary velocity field. The diffeomorphism between source and template is the unit-time flow of the ODE defined by that field, so the direct map is found by forward Euler integration and the inverse map by a modified Euler backward step. Both maps are trained jointly by minimising the Chamfer distance between the deformed point clouds, together with code and velocity regularisation. On the two held-out patients the baseline model achieves maximal forward local distances of 0.2777 cm and 0.2166 cm, average direct-test errors of 0.0387 cm and 0.0379 cm, and it improves on LDDMM for the inverse map while running in roughly one and a half minutes per shape. The same latent space, sampled with a fitted Gaussian, generates new anatomies by applying the inverse deformation to the reference.

Load-bearing premise

The load-bearing premise is that the thin-plate-spline-augmented dataset of synthetic aortas is a faithful proxy for real aortic variability, so accuracy on the synthetic test shapes and two held-out patients transfers to new patients; the paper's own cross-validation shows maximal test errors above 0.65 cm for patients with unique anatomy, so this premise is the fragile one.

Editorial extensions

If this is right

  • A single trained AD-SVFD model can register many unseen aortic shapes to the same reference in about 90 seconds each, without running an expensive optimisation from scratch.
  • Because the map is a diffeomorphism, the computed direct and inverse correspondences are guaranteed to be smooth, invertible, and topology-preserving, so registered geometries cannot tangle or tear.
  • The same latent space that drives registration doubles as a generative model: drawing codes from the fitted Gaussian and applying the inverse map to the reference yields new plausible aortic anatomies.
  • Weight sharing across the cohort makes the model substantially lighter and faster at inference than an auto-encoder, since only the low-dimensional code is fine-tuned for a new shape.
  • Training with plain Chamfer distance is sufficient; the paper finds that normals-based and Sinkhorn alternatives do not improve accuracy on this dataset.

Reading between the lines

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

  • The paper leaves implicit that the same code-sampling recipe could populate simulation cohorts with controlled geometric variability for haemodynamic studies, an application mentioned only as motivation.
  • The cross-validation results suggest the method's weakest spot is rare anatomy: several folds with unique patient features show maximal test errors above 0.5 cm, so a natural testable extension is to condition the latent space on clinical variables so that codes interpolate along disease-relevant directions rather than purely geometric ones.
  • Because the inverse map is trained on a bidirectional loss, the same model could in principle morph the template onto each source and then register sources to one another through the common reference, enabling population-level statistical shape analysis.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces AD-SVFD, a deep-learning model for deformable registration of aortic point clouds to a common template and for generative shape synthesis. Each anatomy is represented by a weighted point cloud and by a trainable latent code; the deformation is the unit-time flow of a stationary velocity field parameterized by a fully-connected neural network, integrated with K=10 Euler steps. Training minimizes a bidirectional Chamfer distance plus regularization on codes and velocities; at inference only the latent code is optimized. Experiments use 902 aortic geometries (882 generated by TPS-based augmentation from 20 real patients), compare several data attachment losses, and benchmark against CPD, TPS, LDDMM, ResNet-LDDMM, and SDF4CHD. The paper reports small average and maximal FLD/BLD values on the chosen test split, an average inference time of about 1m28s per shape, and a 10-fold cross-validation showing larger errors on patients with unique anatomy.

Significance. If the reported accuracy held on genuinely unseen anatomies, AD-SVFD would be a practically attractive combination of auto-decoder latent codes and neural-ODE diffeomorphisms: it offers fast inference, bidirectional registration, and a generative latent space, with a reasonable comparison against established baselines. The authors provide a public dataset link, perform a 10-fold cross-validation, and compare against LDDMM and several deep alternatives, which are concrete strengths. However, the central accuracy claim is not established for unseen anatomies: the main test split is dominated by TPS-augmented copies of training patients, and the paper's own cross-validation reports maximal test FLD values of 0.50-0.66 cm on several held-out patients, comparable to half the template inlet diameter. The abstract's unqualified 'extremely accurate' claim therefore overstates the evidence. The metric issue (FLD/BLD are components of the Chamfer training loss) and the lack of published code further limit verification.

major comments (4)
  1. [Results, Test 4 (Table 3)] The abstract's claim that AD-SVFD 'yields extremely accurate approximations' is not supported for genuinely unseen anatomies. Of the 38 test shapes, 36 are TPS-augmented versions of the 18 training patients, and the two original test patients (P#093, P#278) appear to be a favorable selection: in the 10-fold cross-validation, maximum direct test FLD reaches 0.5000 cm for P#275, 0.6429 cm for P#207, 0.6567 cm for P#188, and 0.5114 cm for P#205, against a template inlet diameter of 1.31 cm. These errors are of the same order as the vessel diameter and do not support the unqualified accuracy claim. The paper should either qualify the headline claim to the within-distribution/interpolation setting or provide evidence on genuinely held-out patients with unique anatomy.
  2. [Methods, Eq. (8); Results, Tables 1-2] The evaluation metrics FLD and BLD are the two per-point nearest-neighbor components of the Chamfer distance used as the training loss. Thus Tables 1 and 2 partly re-measure the training objective; they are valid for comparing methods trained with the same objective, but they are not an independent accuracy certificate. Please report at least one metric not minimized during training (for example volumetric overlap, surface-to-surface distance on the original meshes, or landmark-based distances if correspondences are available) to substantiate the absolute accuracy claim.
  3. [Appendix A.2, Algorithm 2] The data augmentation generates new anatomies by TPS interpolation between two existing patients with matching factors C_l uniformly drawn from [0.5,1]. The 36 synthetic test shapes used in the main test split are therefore convex combinations of training anatomies and lie close to the training manifold. This largely explains why test errors in Tables 1 and 2 are low while cross-validation on original anatomies shows much larger errors. The paper should state this limitation when interpreting the main results and should not present the 36 augmented test shapes as evidence of generalization to unseen anatomies.
  4. [Methods, Eqs. (5)-(7)] The inverse map used in all experiments is the modified-Euler approximation of an implicit backward Euler step, not the exact inverse of the discrete forward map. No experiment quantifies the inverse-consistency error (for example, mean and maximum of the composition error ||phi_i^{-1}(phi_i(x)) - x|| on the source points or its reverse). Since the abstract highlights invertibility and the method is explicitly discrete, please report this quantity and discuss the effect of K=10 on the inverse consistency.
minor comments (5)
  1. [Eq. (4)] The loss in Eq. (4) contains a weight-decay term w_Theta ||Theta||^2_2, but no value for w_Theta is reported in the text or in the hyperparameter tables; please state whether it was used and its value.
  2. [Abstract] The phrase 'extremely accurate approximations at competitive computational costs' should be qualified in light of the cross-validation results, or replaced by a claim limited to the tested distribution.
  3. [Results, Test 1] The sentence 'the training errors are computed only on the 18 original shapes' is useful, but Figure 3 also reports testing errors; please clarify in the caption that the testing set includes both original and augmented shapes.
  4. [Data and code availability] The code availability statement says the code is 'currently not available' and the text notes that exact reproducibility cannot be guaranteed due to non-deterministic PyTorch algorithms; releasing the code, trained weights, and random seeds would substantially strengthen the empirical contribution.
  5. [Appendix A] There is a typo 'seep learning' in the first sentence of Appendix A; also, the phrase 'quantifies' in the Table 4 caption should be 'quantified'.

Circularity Check

1 steps flagged · score 3.0 of 10

Reported accuracy partly re-measures the Chamfer training loss; no derivation-level circularity, but the generalization claim is further weakened by a test set that is mostly TPS-augmented copies of training patients.

  1. self definitional [Results (Test 2, metric definition); Methods, 'Data attachment measures', Eq. (8); Algorithm 1 and Remark 2]
    "The model accuracy is quantified through the forward and backward local distances (FLD and BLD), expressed in cm. The former identifies the distance of each point in the mapped geometry from the closest one in the target, while the latter is the distance of each point in the target from the closest one in the mapped geometry. ... DCD (Y, Y′) := 1/M \sum_{i=1}^M \min_{c′∈Y ′} \|Y_i − c ′\|_2^2 + 1/M ′ \sum_{i′=1}^{M ′} \min_{c∈Y} \|c−Y ′_{i′}\|_2^2 ."

    FLD/BLD are the per-point nearest-neighbor distances whose squares are summed in the Chamfer distance of Eq. (8). The training loss in Eq. (4) is exactly E(S_i,T;φ) = D_CD(φ(S_i),T) + D_CD(φ^{−1}(T),S_i), and at inference Remark 2 states that 'only the latent code entries ... have to be optimized', using the same data-attachment loss. Thus the average FLD/BLD numbers reported in Tables 1-2 are essentially the two summands of the very objective being minimized, so the headline 'extremely accurate approximations' partly restates the training fit rather than an independent accuracy measure.

full rationale

The paper's methodological chain is self-contained: the neural-ODE/SVF formulation, the auto-decoder latent codes, and the backward integration for the inverse map are all stated as explicit model choices (Eqs. 2-7) rather than imported from a self-citation. The authors do cite their own prior SDF4CHD work [33] for the position-aware shape-encoding strategy and as a baseline, but nothing load-bearing is justified only by that self-citation, and no uniqueness theorem is invoked to forbid alternatives. The main circularity-adjacent issue is evaluative: FLD and BLD are the per-point components of the Chamfer distance used as the training loss, so the reported test errors partly re-measure the optimized objective. A second, non-circular but important confound is the test split: 36 of the 38 test shapes are TPS-augmented versions of the 18 training patients, so the statement that training and testing errors are comparable partly reflects near-training-manifold test data. The paper's own 10-fold cross-validation (Table 3) reports max direct FLD values up to 0.6567 cm on genuinely held-out patients with unique anatomy, and the discussion candidly attributes this to data paucity. Those honest limitations and the external comparisons to LDDMM, ResNet-LDDMM, and SDF4CHD give the central claim independent empirical content; they prevent a score above 3, while the loss/evaluation overlap prevents a score of 0.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

This is a fully data-driven deep learning paper; the central machinery consists of fitted network weights, fitted latent codes, and hand-calibrated hyperparameters, while the physical and geometric assumptions are the shared topology, the point-cloud surface representation, and the representativeness of the TPS augmentation. No new physical entities are introduced.

free parameters (8)
  • Neural network parameters Theta = approximately 278k parameters
    Learned by Adam from Eq. (4); parametrize the stationary velocity field v(x; Theta, z_i).
  • Latent shape codes z_i = 256-dimensional, one per training anatomy
    Auto-decoder conditioning; jointly optimized with Theta during training and fine-tuned at inference.
  • Latent dimension N_z = 256
    Selected in Test 1 as the best accuracy/efficiency trade-off; no prior justification.
  • Code regularization weight w_z = 1e-3
    Calibrated in Appendix B.2 to regularize the latent space without degrading registration.
  • Velocity regularization weight w_v = 1e-4
    Chosen by TPE hyperparameter search; controls kinetic energy of the deformation.
  • Learning rate lambda = 1e-3
    TPE-calibrated; used for both Theta and z_i.
  • Integration steps K = 10
    Adopted from ResNet-LDDMM [28]; not re-tuned for the multi-shape setting.
  • Adaptive sampling fraction a = 0.15
    TPE-calibrated; retains 15% of highest-loss points between epochs.
assumptions (6)
  • domain assumption All source and template anatomies share the same topology.
    Explicitly stated: 'Our approach is developed under the assumption that all shapes share the same topology.' If false, existence of a diffeomorphic map between shapes is not guaranteed.
  • standard math A fully-connected ANN with Leaky-ReLU activations is Lipschitz, ensuring a well-posed ODE.
    Invoked from [45] to justify existence/uniqueness of the flow in Eq. (3).
  • ad hoc to paper K=10 forward/modified Euler discretizations approximate the continuous flow and its inverse well enough.
    Methods Eqs. (5)-(7); no error bound or convergence analysis is provided for this step size.
  • domain assumption Cell-center point clouds with cell-area weights faithfully represent vessel surfaces for Chamfer-based registration.
    Used throughout; Eq. (1) and Methods; no sensitivity analysis with respect to sampling density.
  • domain assumption TPS-augmented shapes are physiologically plausible and representative of true anatomical variability.
    Appendix A.2; 118 of 1020 generated shapes were manually removed due to artifacts, and cross-validation shows poor performance on unique real anatomies.
  • domain assumption Gaussian sampling in the learned latent space produces realistic new anatomies.
    Test 5; only qualitative PCA and two interpolation examples, no quantitative or clinical validation of generated shapes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deformable registration and generative modelling of aortic anatomies by auto-decoders and neural ODEs." pith.science (2026). https://pith.science/paper/LFEIOHXO

@misc{pith2026250600947,
  author       = {Pith},
  title        = {Pith review of: Deformable registration and generative modelling of aortic anatomies by auto-decoders and neural ODEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFEIOHXO}},
  note         = {Machine review of arXiv:2506.00947}
}
read the original abstract

This work introduces AD-SVFD, a deep learning model for the deformable registration of vascular shapes to a pre-defined reference and for the generation of synthetic anatomies. AD-SVFD operates by representing each geometry as a weighted point cloud and models ambient space deformations as solutions at unit time of ODEs, whose time-independent right-hand sides are expressed through artificial neural networks. The model parameters are optimized by minimizing the Chamfer Distance between the deformed and reference point clouds, while backward integration of the ODE defines the inverse transformation. A distinctive feature of AD-SVFD is its auto-decoder structure, that enables generalization across shape cohorts and favors efficient weight sharing. In particular, each anatomy is associated with a low-dimensional code that acts as a self-conditioning field and that is jointly optimized with the network parameters during training. At inference, only the latent codes are fine-tuned, substantially reducing computational overheads. Furthermore, the use of implicit shape representations enables generative applications: new anatomies can be synthesized by suitably sampling from the latent space and applying the corresponding inverse transformations to the reference geometry. Numerical experiments, conducted on healthy aortic anatomies, showcase the high-quality results of AD-SVFD, which yields extremely accurate approximations at competitive computational costs.

Figures

Figures reproduced from arXiv: 2506.00947 by the authors.

Figure 1
Figure 1. General structure of the AD–SVFD model. The proposed approach leverages deep learning techniques to perform the diffeomorphic registration of vascular anatomies to a reference. Invertible ambient space deforma￾tions are modeled as solutions at unit time of ODEs, whose right–hand sides are parametrized by neural networks. The source and template geometries, represented as point clouds, are provided as input to AD–SVF… view at source ↗
Figure 2
Figure 2. Healthy aortic shapes dataset overview. In particular: (a) original dataset of patient–specific anatomies; (b) topology of the considered geometries, with nomenclature of the different branches; (c) original shape and three synthetic samples, generated by deforming four anatomies with the implemented data augmentation pipeline. Test 2: Data attachment measures We analyse the AD–SVFD model performances considering th… view at source ↗
Figure 3
Figure 3. Deformable mapping results of the baseline AD–SVFD model. In particular, we report the average (a) and maximal (b) pointwise errors — quantified through the forward and backward local distances FLD and BLD, in cm — on training and testing datapoints, obtained for different shape code dimensions Nz; in (c), we show the geodesic paths between two source shapes (P#090 for training, P#093 for testing) and the reference … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Registration results of AD–SVFD considering different data attachment measures. In particular, we show the direct and inverse mapping pointwise errors, obtained on a training (P#090) and a testing (P#093) datapoint, for four different data adherence metrics. The errors…
Figure 5
Figure 5. Figure 5: Registration results obtained with AD–SVFD and three alternative approaches. In particular, we show the direct and inverse mapping pointwise errors, obtained with LDDMM [9], SDF4CHD [33], ResNet–LDDMM [28] and AD–SVFD on a training (P#090) and a testing (P#278) datapoi…
Figure 6
Figure 6. Figure 6: Representation of the latent space learned by the AD–SVFD model. In particular, we show the projection of the shape codes onto the two–dimensional subspace obtained through PCA on the whole set of training codes. We report the latent codes of the original patients (sta…
Figure 7
Figure 7. Figure 7: Visualization of the TPS interpolation results. In particular, for two patients in the dataset (P#090 and P#272) we show the locations of the interpolation points in the target and template geometries — color–coded so that corresponding points share the same value — an…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 71 canonical work pages

  1. [1]

    OUP Oxford, 2003

    Jan Modersitzki.Numerical methods for image registration. OUP Oxford, 2003

    work page 2003

  2. [2]

    SIAM, 2009

    Jan Modersitzki.FAIR: flexible algorithms for image registration. SIAM, 2009

    work page 2009

  3. [3]

    A fast diffeomorphic image registration algorithm.Neuroimage, 38(1):95–113, 2007

    John Ashburner. A fast diffeomorphic image registration algorithm.Neuroimage, 38(1):95–113, 2007

    work page 2007

  4. [4]

    Deformable medical image registration: a survey

    Aristeidis Sotiras, Christos Davatzikos, and Nikos Paragios. Deformable medical image registration: a survey. IEEE transactions on medical imaging, 32(7):1153–1190, 2013

    work page 2013

  5. [5]

    Deep learning in medical image registration: a survey.Machine Vision and Applications, 31(1):8, 2020

    Grant Haskins, Uwe Kruger, and Pingkun Yan. Deep learning in medical image registration: a survey.Machine Vision and Applications, 31(1):8, 2020

    work page 2020

  6. [6]

    Diffeomorphic demons: Efficient non–parametric image registration.NeuroImage, 45(1):S61–S72, 2009

    Tom Vercauteren, Xavier Pennec, Aymeric Perchant, and Nicholas Ayache. Diffeomorphic demons: Efficient non–parametric image registration.NeuroImage, 45(1):S61–S72, 2009

    work page 2009

  7. [7]

    Computing large deformation metric map- pings via geodesic flows of diffeomorphisms.International journal of computer vision, 61:139–157, 2005

    M Faisal Beg, Michael I Miller, Alain Trouv ´e, and Laurent Younes. Computing large deformation metric map- pings via geodesic flows of diffeomorphisms.International journal of computer vision, 61:139–157, 2005

    work page 2005

  8. [8]

    Surface matching via currents

    Marc Vaillant and Joan Glaunes. Surface matching via currents. InBiennial international conference on infor- mation processing in medical imaging, pages 381–392. Springer, 2005

    work page 2005

Show all 78 references
  1. [9]

    Deformetrica 4: an open-source soft- ware for statistical shape analysis

    Alexandre B ˆone, Maxime Louis, Benoˆıt Martin, and Stanley Durrleman. Deformetrica 4: an open-source soft- ware for statistical shape analysis. InShape in Medical Imaging: International Workshop, ShapeMI 2018, Held R.Tenderini et al.Deformable registration and generative mode...

    2018

  2. [10]

    Unsupervised probabilis- tic deformation modeling for robust diffeomorphic registration

    Julian Krebs, Tommaso Mansi, Boris Mailh ´e, Nicholas Ayache, and Herv´e Delingette. Unsupervised probabilis- tic deformation modeling for robust diffeomorphic registration. InDeep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th In...

    2018

  3. [11]

    Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces.Medical image analysis, 57:226–236, 2019

    Adrian V Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces.Medical image analysis, 57:226–236, 2019

    2019

  4. [12]

    V oxelmorph: a learning framework for deformable medical image registration.IEEE transactions on medical imaging, 38(8):1788–1800, 2019

    Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. V oxelmorph: a learning framework for deformable medical image registration.IEEE transactions on medical imaging, 38(8):1788–1800, 2019

    2019

  5. [13]

    Smooth shells: Multi–scale shape registration with functional maps

    Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi–scale shape registration with functional maps. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12265–12274, 2020

    2020

  6. [14]

    Neuromorph: Unsupervised shape interpolation and correspondence in one go

    Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2021

  7. [15]

    CycleMorph: cycle consistent unsupervised deformable image registration.Medical image analysis, 71:102036, 2021

    Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, and Jong Chul Ye. CycleMorph: cycle consistent unsupervised deformable image registration.Medical image analysis, 71:102036, 2021

    2021

  8. [16]

    Diffusemorph: Unsupervised deformable image registration using diffusion model

    Boah Kim, Inhwa Han, and Jong Chul Ye. Diffusemorph: Unsupervised deformable image registration using diffusion model. InEuropean conference on computer vision, pages 347–364. Springer, 2022

    2022

  9. [17]

    Transmorph: Transformer for unsupervised medical image registration.Medical image analysis, 82:102615, 2022

    Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration.Medical image analysis, 82:102615, 2022

    2022

  10. [18]

    A review of deep learning–based deformable medical image registration.Frontiers in Oncology, 12:1047215, 2022

    Jing Zou, Bingchen Gao, Youyi Song, and Jing Qin. A review of deep learning–based deformable medical image registration.Frontiers in Oncology, 12:1047215, 2022

    2022

  11. [19]

    A survey of non–rigid 3D registration.Computer Graphics Forum, 41(2):559–589, 2022

    Bailin Deng, Yuxin Yao, Roberto M Dyke, and Juyong Zhang. A survey of non–rigid 3D registration.Computer Graphics Forum, 41(2):559–589, 2022

    2022

  12. [20]

    Medical image registration in the era of Transformers: a recent review.Informatics in Medicine Unlocked, page 101540, 2024

    Hiba Ramadan, Dounia El Bourakadi, Ali Yahyaouy, and Hamid Tairi. Medical image registration in the era of Transformers: a recent review.Informatics in Medicine Unlocked, page 101540, 2024

    2024

  13. [21]

    In vivo preclinical photoacoustic imaging of tumor vasculature development and therapy.Journal of biomedical optics, 17(5):056016–056016, 2012

    Jan Laufer, Peter Johnson, Edward Zhang, Bradley Treeby, Ben Cox, Barbara Pedley, and Paul Beard. In vivo preclinical photoacoustic imaging of tumor vasculature development and therapy.Journal of biomedical optics, 17(5):056016–056016, 2012

    2012

  14. [22]

    Dual–scan photoacoustic tomography for the imaging of vascular structure on foot.IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 70(12):1703–1713, 2023

    Chuqin Huang, Yanda Cheng, Wenhan Zheng, Robert W Bing, Huijuan Zhang, Isabel Komornicki, Linda M Harris, Praveen R Arany, Saptarshi Chakraborty, Qifa Zhou, et al. Dual–scan photoacoustic tomography for the imaging of vascular structure on foot.IEEE Transactions on Ultrasonics...

    2023

  15. [23]

    A fast all–optical 3D photoacoustic scanner for clinical vascular imaging.Nature Biomedical Engineering, pages 1–18, 2024

    NT Huynh, E Zhang, O Francies, F Kuklis, T Allen, J Zhu, O Abeyakoon, F Lucka, M Betcke, J Jaros, et al. A fast all–optical 3D photoacoustic scanner for clinical vascular imaging.Nature Biomedical Engineering, pages 1–18, 2024

    2024

  16. [24]

    Registration of 3D point clouds and meshes: A survey from rigid to nonrigid

    Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, Xian-Fang Sun, and Paul L Rosin. Registration of 3D point clouds and meshes: A survey from rigid to nonrigid. IEEE transactions on visualization and computer graphics, 19(7...

    2012

  17. [25]

    A review of non–rigid transformations and learning–based 3D point cloud registration methods.ISPRS journal of photogrammetry and remote sensing, 196:58–72, 2023

    Sara Monji-Azad, J ¨urgen Hesser, and Nikolas L ¨ow. A review of non–rigid transformations and learning–based 3D point cloud registration methods.ISPRS journal of photogrammetry and remote sensing, 196:58–72, 2023

    2023

  18. [26]

    Flownet3D: Learning scene flow in 3D point clouds

    Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3D: Learning scene flow in 3D point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 529–537, 2019

    2019

  19. [27]

    Unsupervised diffeomorphic surface registration and nonlinear modelling

    Balder Croquet, Daan Christiaens, Seth M Weinberg, Michael Bronstein, Dirk Vandermeulen, and Peter Claes. Unsupervised diffeomorphic surface registration and nonlinear modelling. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Confe...

    2021

  20. [28]

    ResNet–LDDMM: advancing the LDDMM framework using deep residual networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3707– 3720, 2022

    Boulbaba Ben Amor, Sylvain Arguill `ere, and Ling Shao. ResNet–LDDMM: advancing the LDDMM framework using deep residual networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3707– 3720, 2022

    2022

  21. [29]

    Learning implicit fields for generative shape modeling

    Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5939–5948, 2019

    2019

  22. [30]

    DISN: Deep implicit surface network for high–quality single–view 3D reconstruction.Advances in neural information processing systems, 32, 2019

    Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. DISN: Deep implicit surface network for high–quality single–view 3D reconstruction.Advances in neural information processing systems, 32, 2019

    2019

  23. [31]

    DeepSDF: learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: learning continuous signed distance functions for shape representation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165–174, 2019

    2019

  24. [32]

    Topology–preserving shape reconstruction and registration via neural diffeomorphic flow

    Shanlin Sun, Kun Han, Deying Kong, Hao Tang, Xiangyi Yan, and Xiaohui Xie. Topology–preserving shape reconstruction and registration via neural diffeomorphic flow. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20845–20855, 2022

    2022

  25. [33]

    SDF4CHD: Generative modeling of cardiac anatomies with congenital heart defects.Medical Image Analysis, 97:103293, 2024

    Fanwei Kong, Sascha Stocker, Perry S Choi, Michael Ma, Daniel B Ennis, and Alison L Marsden. SDF4CHD: Generative modeling of cardiac anatomies with congenital heart defects.Medical Image Analysis, 97:103293, 2024

    2024

  26. [34]

    Neural ordinary differential equa- tions.Advances in neural information processing systems, 31, 2018

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equa- tions.Advances in neural information processing systems, 31, 2018

    2018

  27. [35]

    Conditional neural fields.Advances in neural information processing systems, 22, 2009

    Jian Peng, Liefeng Bo, and Jinbo Xu. Conditional neural fields.Advances in neural information processing systems, 22, 2009

    2009

  28. [36]

    Auto–encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma and Max Welling. Auto–encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  29. [37]

    InInternational conference on learning representations, 2016

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mo- hamed, and Alexander Lerchner.β–V AE: learning basic visual concepts with a constrained variational frame- work. InInternational conference on learning representations, 2016

    2016

  30. [38]

    An introduction to variational autoencoders.Foundations and Trends in Machine Learning, 12(4):307–392, 2019

    Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders.Foundations and Trends in Machine Learning, 12(4):307–392, 2019

    2019

  31. [39]

    Sam Bond-Taylor, Adam Leach, Yang Long, and Chris G Willcocks. Deep generative modelling: a comparative review of V AEs, GANs, normalizing flows, energy–based and autoregressive models.IEEE transactions on pattern analysis and machine intelligence, 2021

    2021

  32. [40]

    Point set registration: coherent point drift.IEEE transactions on pattern analysis and machine intelligence, 32(12):2262–2275, 2010

    Andriy Myronenko and Xubo Song. Point set registration: coherent point drift.IEEE transactions on pattern analysis and machine intelligence, 32(12):2262–2275, 2010

    2010

  33. [41]

    Morphometry of anatomical shape complexes with dense deformations and sparse parameters.Neu- roImage, 101:35–49, 2014

    Stanley Durrleman, Marcel Prastawa, Nicolas Charon, Julie R Korenberg, Sarang Joshi, Guido Gerig, and Alain Trouv´e. Morphometry of anatomical shape complexes with dense deformations and sparse parameters.Neu- roImage, 101:35–49, 2014

    2014

  34. [42]

    A Bayesian framework for joint morphometry of surface and curve meshes in multi–object complexes.Medical image analysis, 35:458–474, 2017

    Pietro Gori, Olivier Colliot, Linda Marrakchi-Kacem, Yulia Worbe, Cyril Poupon, Andreas Hartmann, Nicholas Ayache, and Stanley Durrleman. A Bayesian framework for joint morphometry of surface and curve meshes in multi–object complexes.Medical image analysis, 35:458–474, 2017

    2017

  35. [43]

    ImplicitAtlas: learning deformable shape templates in medical imaging

    Jiancheng Yang, Udaranga Wickramasinghe, Bingbing Ni, and Pascal Fua. ImplicitAtlas: learning deformable shape templates in medical imaging. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15861–15871, 2022

    2022

  36. [44]

    Registration of anatomical images using paths of diffeomorphisms parameterized with stationary vector field flows.International Journal of Computer Vision, 85: 291–306, 2009

    Monica Hernandez, Matias N Bossa, and Salvador Olmos. Registration of anatomical images using paths of diffeomorphisms parameterized with stationary vector field flows.International Journal of Computer Vision, 85: 291–306, 2009

    2009

  37. [45]

    CortexODE: Learning cortical surface reconstruction by neural ODEs.IEEE Transactions on Medical Imaging, 42(2):430– 443, 2022

    Qiang Ma, Liu Li, Emma C Robinson, Bernhard Kainz, Daniel Rueckert, and Amir Alansary. CortexODE: Learning cortical surface reconstruction by neural ODEs.IEEE Transactions on Medical Imaging, 42(2):430– 443, 2022

    2022

  38. [46]

    Recent advances in shape correspondence.The Visual Computer, 36(8):1705–1721, 2020

    Yusuf Sahillio ˘glu. Recent advances in shape correspondence.The Visual Computer, 36(8):1705–1721, 2020

    2020

  39. [47]

    Optimal transport for diffeomorphic registration

    Jean Feydy, Benjamin Charlier, Franc ¸ois-Xavier Vialard, and Gabriel Peyr´e. Optimal transport for diffeomorphic registration. InMedical Image Computing and Computer Assisted Intervention- MICCAI 2017: 20th Interna- tional Conference, Quebec City, QC, Canada, September 11-13,...

    2017

  40. [48]

    Geometric distortion metrics for point cloud compression

    Dong Tian, Hideaki Ochimizu, Chen Feng, Robert Cohen, and Anthony Vetro. Geometric distortion metrics for point cloud compression. In2017 IEEE International Conference on Image Processing (ICIP), pages 3460–3464. IEEE, 2017

    2017

  41. [49]

    Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

    2013

  42. [50]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  43. [51]

    Optimization methods for large–scale machine learning.SIAM review, 60(2):223–311, 2018

    L ´eon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large–scale machine learning.SIAM review, 60(2):223–311, 2018

    2018

  44. [52]

    SimVascular: an open source pipeline for cardiovascular simulation.Annals of biomedical engineering, 45:525– 541, 2017

    Adam Updegrove, Nathan M Wilson, Jameson Merkow, Hongzhi Lan, Alison L Marsden, and Shawn C Shadden. SimVascular: an open source pipeline for cardiovascular simulation.Annals of biomedical engineering, 45:525– 541, 2017

    2017

  45. [53]

    The vascular model repository: a public resource of medical imaging data and blood flow simulation results.Journal of medical devices, 7(4), 2013

    Nathan M Wilson, Ana K Ortiz, and Allison B Johnson. The vascular model repository: a public resource of medical imaging data and blood flow simulation results.Journal of medical devices, 7(4), 2013

    2013

  46. [54]

    Splines minimizing rotation–invariant semi–norms in Sobolev spaces

    Jean Duchon. Splines minimizing rotation–invariant semi–norms in Sobolev spaces. InConstructive Theory of Functions of Several Variables: Proceedings of a Conference Held at Oberwolfach April 25–May 1, 1976, pages 85–100. Springer, 1977

    1976

  47. [55]

    Algorithms for hyper–parameter optimization

    James Bergstra, R ´emi Bardenet, Yoshua Bengio, and Bal´azs K´egl. Algorithms for hyper–parameter optimization. Advances in neural information processing systems, 24, 2011

    2011

  48. [56]

    Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures

    James Bergstra, Daniel Yamins, and David Cox. Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures. InInternational conference on machine learning, pages 115–

  49. [57]

    Variational auto–decoder: A method for neural generative modeling from incomplete data.arXiv preprint arXiv:1903.00840, 2019

    Amir Zadeh, Yao-Chong Lim, Paul Pu Liang, and Louis-Philippe Morency. Variational auto–decoder: A method for neural generative modeling from incomplete data.arXiv preprint arXiv:1903.00840, 2019

    1903 arXiv

  50. [58]

    Surrogate modeling for fluid flows based on physics– constrained deep learning without simulation data.Computer Methods in Applied Mechanics and Engineering, 361:112732, 2020

    Luning Sun, Han Gao, Shaowu Pan, and Jian-Xun Wang. Surrogate modeling for fluid flows based on physics– constrained deep learning without simulation data.Computer Methods in Applied Mechanics and Engineering, 361:112732, 2020

    2020

  51. [59]

    Geometry–aware physics–informed neural network surrogate for solving Navier–Stokes equation (GAPINN).Advanced Mod- eling and Simulation in Engineering Sciences, 9(1):8, 2022

    Jan Oldenburg, Finja Borowski, Alper ¨Oner, Klaus-Peter Schmitz, and Michael Stiehm. Geometry–aware physics–informed neural network surrogate for solving Navier–Stokes equation (GAPINN).Advanced Mod- eling and Simulation in Engineering Sciences, 9(1):8, 2022

    2022

  52. [60]

    Francesco Regazzoni, Stefano Pagani, and Alfio Quarteroni. Universal Solution Manifold Networks (USM– Nets): non–intrusive mesh–free surrogate models for problems in variable domains.Journal of Biomechanical Engineering, 144(12):121004, 2022

    2022

  53. [61]

    Francisco Sahli Costabal, Simone Pezzuto, and Paris Perdikaris.∆–PINNs: Physics–informed neural networks on complex geometries.Engineering Applications of Artificial Intelligence, 127:107324, 2024

    2024

  54. [62]

    Simone Brivio, Stefania Fresca, and Andrea Manzoni. Handling geometrical variability in nonlinear reduced order modeling through continuous geometry–aware DL-ROMs.Computer Methods in Applied Mechanics and Engineering, 442:117989, 2025

    2025

  55. [63]

    UNIST: unpaired neural implicit shape translation network

    Qimin Chen, Johannes Merz, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. UNIST: unpaired neural implicit shape translation network. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18614–18622, 2022

    2022

  56. [64]

    Fourier features let networks learn high frequency functions in low–dimensional domains.Advances in Neural Information Processing Systems, 33:7537–7547, 2020

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low–dimensional domains.Advances in Neural Information Proce...

    2020

  57. [65]

    On the spectral bias of neural networks

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. InInternational conference on machine learning, pages 5301–5310. PMLR, 2019

  58. [66]

    The earth mover’s distance as a metric for image retrieval

    Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover’s distance as a metric for image retrieval. International journal of computer vision, 40:99–121, 2000

    2000

  59. [67]

    DeepEMD: A transformer–based fast estimation of the earth mover’s distance

    Atul Kumar Sinha and Franc ¸ois Fleuret. DeepEMD: A transformer–based fast estimation of the earth mover’s distance. InInternational Conference on Pattern Recognition, pages 1–15. Springer, 2024. R.Tenderini et al.Deformable registration and generative modelling of aortic anat...

    2024

  60. [68]

    Faster Wasserstein distance estimation with the Sinkhorn divergence.Advances in Neural Information Processing Systems, 33: 2257–2269, 2020

    Lenaic Chizat, Pierre Roussillon, Flavien L ´eger, Franc ¸ois-Xavier Vialard, and Gabriel Peyr´e. Faster Wasserstein distance estimation with the Sinkhorn divergence.Advances in Neural Information Processing Systems, 33: 2257–2269, 2020

    2020

  61. [69]

    Diffeomorphic registration using Sinkhorn divergences.SIAM Journal on Imaging Sciences, 16(1):250–279, 2023

    Lucas De Lara, Alberto Gonz ´alez-Sanz, and Jean-Michel Loubes. Diffeomorphic registration using Sinkhorn divergences.SIAM Journal on Imaging Sciences, 16(1):250–279, 2023

    2023

  62. [70]

    Inter- polating between optimal transport and mmd using Sinkhorn divergences

    Jean Feydy, Thibault S´ejourn´e, Franc ¸ois-Xavier Vialard, Shunichi Amari, Alain Trouv´e, and Gabriel Peyr´e. Inter- polating between optimal transport and mmd using Sinkhorn divergences. InThe 22nd international conference on artificial intelligence and statistics, pages 268...

    2019

  63. [71]

    Density–aware chamfer distance as a comprehensive metric for point cloud completion.arXiv preprint arXiv:2111.12702, 2021

    Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Density–aware chamfer distance as a comprehensive metric for point cloud completion.arXiv preprint arXiv:2111.12702, 2021

    2021 arXiv

  64. [72]

    Delving deep into rectifiers: Surpassing human– level performance on ImageNet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human– level performance on ImageNet classification. InProceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015

    2015

  65. [73]

    There is more than one way to frame a curve.The American Mathematical Monthly, 82(3): 246–251, 1975

    Richard L Bishop. There is more than one way to frame a curve.The American Mathematical Monthly, 82(3): 246–251, 1975

    1975

  66. [74]

    A low order, torsion deformable spatial beam element based on the absolute nodal coordinate formulation and Bishop frame.Multibody System Dynamics, 51(3):247– 278, 2021

    Mehran Ebrahimi, Adrian Butscher, and Hyunmin Cheong. A low order, torsion deformable spatial beam element based on the absolute nodal coordinate formulation and Bishop frame.Multibody System Dynamics, 51(3):247– 278, 2021

    2021

  67. [75]

    Method for registration of 3D shapes

    Paul J Besl and Neil D McKay. Method for registration of 3D shapes. InSensor fusion IV: control paradigms and data structures, volume 1611, pages 586–606. Spie, 1992

    1992

  68. [76]

    Robust registration of 2D and 3D point sets.Image and vision computing, 21(13-14): 1145–1153, 2003

    Andrew W Fitzgibbon. Robust registration of 2D and 3D point sets.Image and vision computing, 21(13-14): 1145–1153, 2003

    2003

  69. [77]

    New algorithms for 2D and 3D point matching: pose estimation and correspondence.Pattern recognition, 31(8):1019–1031, 1998

    Steven Gold, Anand Rangarajan, Chien-Ping Lu, Suguna Pappu, and Eric Mjolsness. New algorithms for 2D and 3D point matching: pose estimation and correspondence.Pattern recognition, 31(8):1019–1031, 1998

    1998

  70. [78]

    extended

    Haili Chui and Anand Rangarajan. A new algorithm for non–rigid point matching. InProceedings IEEE Con- ference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 2, pages 44–51. IEEE, 2000. R.Tenderini et al.Deformable registration and generative ...

    2000

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.