The Push-Forward Transform for Continuous and Robust Comparison of Dynamic Shapes

Ivo F. Sbalzarini; Juan-Esteban Suarez Cardona; Roua Rouatbi

REVIEW 2 major objections 7 minor 28 references

The Push-Forward Transform for Continuous and Robust Comparison of Dynamic Shapes

T0 review · 2 major / 7 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that comparing shapes can be reduced to comparing scalar fields on one common reference ball, and that for signed distance functions this yields an invariant, training-free morphometric that is competitive with or better th

desk verdict Solid incremental work — the 3D extension and skeleton fusion are real, but the 'invariant by construction' claim is only as strong as a soft-enforced harmonic map, and the benchmarks need error bars. read the letter →

arxiv 2608.02306 v1 pith:CUFWNZVQ submitted 2026-08-03 cs.CV cs.LGcs.NAmath.NA

Roua Rouatbi , Juan-Esteban Suarez Cardona , Ivo F. Sbalzarini This is my paper

classification cs.CVcs.LGcs.NAmath.NA

keywords shapecomparisonpush-forwardtransformsigneddistancefunctionsmorphometricsdiffeomorphicmappingmedialaxisspectraldescriptorsinvariance

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

The paper is trying to establish that shape comparison can be made well-posed without landmarks, boundary parameterization, or learned representations: map any scalar field defined on a shape to a common reference ball through a smooth, non-folding deformation, and compare the pushed-forward fields pointwise. Applied to signed distance functions, this Push-Forward Transform yields a continuous, differentiable representation that encodes boundary and interior geometry together. By taking the magnitude of Fourier coefficients (in 2D) or degree-wise spherical-harmonic energies (in 3D), the resulting morphometric is invariant by construction to translation, rotation, reflection, re-parameterization, and uniform scaling, and robust to noise. The authors show on standard benchmarks that the descriptor matches or beats deep-learning classifiers while requiring no training and running orders of magnitude faster, and that it extends to time sequences and to additional intensity channels for joint shape-and-signal analysis. A sympathetic reader would care because this offers a reproducible, interpretable alternative in a field dominated by learned descriptors.

What carries the argument

The load-bearing object is the Push-Forward Transform (PF-T): for a scalar field f_S on a shape domain S and a diffeomorphism Ψ from the reference ball B_r onto S, it defines f_S*(x) = f_S(Ψ(x)). The boundary of Ψ is fixed by closest-point projection, and the interior is extended either radially (for star-convex shapes) or harmonically (in 2D, solving the Laplace equation with a positivity constraint on the Jacobian determinant enforced by a log-barrier term). The signed distance function is computed as a smooth polynomial surrogate of the viscous Eikonal equation, so the whole pipeline is differentiable; its normalized-gradient divergence locates the medial axis. Invariance to rotation and

What would settle it

Take a non-star-convex 2D shape such as a C-shape or dumbbell, compute the finite-dimensional harmonic extension by minimizing the paper's Eq. (5), and evaluate min det DΨ over the Legendre grid. If the minimum is negative, or if the resulting PF-SDF changes discontinuously as the polynomial degree is refined, the map folds and the descriptor is not a well-defined shape representation; the same check on a bean-shaped non-star-convex 3D object would test the 3D generalization.

Watch

Extended reading notes

Core claim

The central discovery is that composing a signed distance function with a diffeomorphism from a reference ball onto the shape domain normalizes shape geometry so that functions on different shapes become directly comparable. For topological spheres the authors construct such maps from closest-point boundary matching: a radial extension for star-convex shapes, and a harmonic extension in 2D that is a diffeomorphism exactly when the boundary Jacobian determinant stays positive. The pushed-forward SDF is represented in a polynomial spectral basis, and its normalized Fourier magnitude spectrum (2D) or spherical-harmonic degree energies (3D) form the Push-Forward Signed Distance Morphometric (PF-

Load-bearing premise

The whole construction assumes each shape can be smoothly and non-foldingly stretched onto a reference ball; the paper only proves this for star-convex shapes (radial map) and, in 2D, for harmonic maps whose boundary Jacobian stays positive, so general non-star-convex shapes—and all non-star-convex 3D shapes—rest on an unproven premise.

Editorial extensions

If this is right

If the central claim holds, shape descriptors can be made deterministic, reproducible, and interpretable without any training data, removing a major reproducibility barrier in shape analysis.
On the reported benchmarks, the method is competitive with or better than deep-learning baselines while running in minutes rather than hours, so it is practical for large-scale screening.
Because the spectral construction discards phase, rotation- and reflection-invariance are obtained by construction, and dropping the DC component yields a tunable descriptor that groups shapes by rotational symmetry.
The same push-forward machinery extends to time-varying shapes and to arbitrary scalar fields on the shape, so morphology and molecular or intensity signals can be compared jointly; the paper demonstrates this on a stem-cell-derived developmental time series.

Reading between the lines

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

If the diffeomorphism assumption were guaranteed for broader topologies (e.g., via topology-matched reference domains), the same construction would likely generalize to shapes with holes and branched skeletons, since the spectral-invariance step is topology-agnostic.
The full pipeline is differentiable, so one natural testable extension is to use PF-SDF as a differentiable geometric layer or regularizer inside a learned model; the paper only notes this as future work.
The skeleton channel helped on a topologically diverse 2D benchmark but not on a set of same-topology roundworms, suggesting the skeleton signal mostly carries topological/medial-axis variation; a testable prediction is that PF-SDM will excel precisely when classes differ in skeletal structure rather than boundary curvature.
The 3D results are restricted to star-convex shapes by the authors' own admission; a direct extension would be to validate fold-free harmonic or other diffeomorphic maps on non-star-convex topological spheres.

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.

Desk Editor's Note

read the letter

Quick read: this is a serious paper with a gap between theory and computation that the authors mostly acknowledge but don't fully close. The new content is real: the 3D spherical-harmonic PF-SDM, the divergence-based skeleton channel, and the multi-channel late-fusion pipeline for gastruloid fate prediction go beyond their earlier ISBI work. The benchmarks are also the most systematic I've seen for this class of descriptor: MPEG-7, BBBC010, several deep baselines, runtimes, and a code release for the 3D parts.

What's done well: the math for the radial extension is straightforward and correct; Proposition 1's spectral invariants are textbook but properly stated. The paper is honest about the restrictions — topological spheres, 3D star-convex only, scalar fields only. That honesty helps. The computational advantage is clear: minutes vs hours, with F1 essentially matching ShapeEmbed.

The soft spots: the invariance guarantee for non-star-convex 2D shapes (MPEG-7, BBBC010) relies on the harmonic-extension diffeomorphism, but the computed map is only approximately harmonic and the det>0 condition is enforced by a soft log-barrier at a finite set of collocation points. Theorem 1(ii) then doesn't directly apply to the computed object. The paper doesn't report a dense-boundary Jacobian check or Laplacian residual. So the claim 'invariant by construction' is stronger than what's verified for the actual shapes benchmarked. That's the main fix needed. Secondary: the MPEG-7/BBBC010 F1 numbers are from a single split, no error bars; a 0.91 vs 0.90 lead over ShapeEmbed may be noise. The free parameters (μ, polynomial degree, λ, truncation, fusion weights, quantile) are many, and sensitivity isn't reported.

None of this kills the method. On the demonstrated shape classes the descriptors work; the concerns are tightening claims, adding variance estimates, and verifying the Jacobian condition numerically. I'd send this to peer review — it's a real contribution for shape analysis and bioimage informatics, and the issues are addressable in revision.

Referee Report

2 major / 7 minor

Summary. The paper introduces the Push-Forward Transform (PF-T), a method for comparing scalar fields (e.g., signed distance functions, intensity images) defined over shape domains by mapping them to a common reference domain (unit ball/disk) via a diffeomorphism. The derived descriptors (PF-SDM), after Fourier (2D) or spherical-harmonic (3D) spectral analysis, are claimed to be invariant to translation, rotation, reflection, reparametrization, and uniform scaling, while being robust to noise. The authors present theoretical results (Theorem 1 for radial and harmonic extensions, Proposition 1 for spectral invariance), a polynomial-surrogate numerical implementation based on the viscous Eikonal equation, and experiments on synthetic 2D/3D data, MPEG-7 and BBBC010 shape benchmarks, and time-resolved mouse gastruloid data. The PF-SDM is benchmarked against classical descriptors and deep-learning baselines, reporting competitive or better F1 scores with much lower runtimes.

Significance. If the claims hold, this is a meaningful contribution: a deterministic, training-free, interpretable shape descriptor with built-in invariances and spectral control (e.g., emphasizing rotational symmetry) that is competitive with learned representations, plus runtime advantages. The paper ships code, provides proofs of the core invariance theorems, and makes falsifiable benchmark comparisons. However, the scope is explicitly limited to topological spheres, and the load-bearing practical assumption of a computed diffeomorphism is not verified in the implementation.

major comments (2)

[Def. 11 / Eq. (8) and Sec. 5.2] The harmonic extension used on MPEG-7 and BBBC010 only enforces det DΨξ > 0 through the soft log-barrier L_r^n at NS boundary collocation points. Theorem 1(ii) requires det DΨ_h > 0 on the entire boundary, and Ψξ is only an approximate harmonic map (L_Δ^n minimizes the Laplacian residual, not exact ΔΨ=0). For non-star-convex shapes, no test of injectivity or Jacobian sign on a dense boundary grid is reported, so the diffeomorphism premise underlying the push-forward and the claimed invariance-by-construction is unverified exactly where the method is benchmarked. Please report the minimal/mean boundary Jacobian determinant and Laplacian residual for the benchmark shapes, or justify the invariance empirically.
[Sec. 4.1 / Algorithm A1] The spine extraction relies on PCA-based ordering of the skeleton point cloud (Algorithm A1, step 4). The paper acknowledges this restricts to elongated, single-axis shapes, which is an honest limitation. However, for the gastruloid application, the 'skeleton channel' is the raw divergence field (∇·η), not the parametric spine; this is fine, but the text should be clearer that the parametric spine is not used in the reported benchmarks.

minor comments (7)

[Abstract/Definition 1] Notation is inconsistent: ∂S is first called 'shape boundary' and S defined as interior, but later 'shape domain S' is used for the closed domain. Please define the convention once and use it consistently.
[Eq. (1)] The infimum in Definition 3 is over C^1 maps ψ_S, but the expression ∥ψ_S(x) − x_S(x)∥ uses x_S as the closest-point projection onto ∂S, which is not necessarily smooth or unique for non-convex shapes. This is acknowledged implicitly, but it would help to state the regularity assumptions on ∂S explicitly (smooth closed hypersurface) and cite standard results on the closest-point projection.
[Sec. 3.3, Eq. (6)] The notation L_n^Δ[ξ] appears to sum over multi-indices α of the squared residuals of the Laplacian, but the subscript ξ in Ψ_ξ^{x_i} is undefined; presumably it denotes the i-th component. Please define.
[Sec. 5.2 / Table 1] Pretrained MAE models underperform from-scratch MAE ViT-B on MPEG-7; this is surprising and should be briefly commented on (e.g., input modality mismatch with ImageNet natural images). Similarly, the PF-SDM runtime includes grid search over fusion weights; please clarify whether the reported PF-SDM+Int 0.86 on BBBC010 (Appendix G) is excluded from Table 1 for consistency.
[App. A / Theorem 1] The proof of Theorem 1(ii) states 'applying the Alessandrini–Nesi theorem' without stating the conditions of that theorem; at minimum, state that the boundary map must be orientation-preserving and the domain convex, and explain how the harmonic extension inherits C^1 regularity from ψ_S.
[Sec. 5.1] The caption of Fig. 7/8 says 'convariance' or similar typos in the text (e.g., 'convariance' in Sec. 5.1 text). Please proofread.
[Sec. 5.3] The balanced accuracy baseline from prior work is reported as 0.667±0.000, which is suspiciously constant; please clarify the discrepancy and whether the same train/test split protocol was used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PF-T/PF-SDM derivation is self-contained and benchmarked against external data; self-citations are provenance/baseline references, not load-bearing premises.

full rationale

The derivation chain is not circular. The Push-Forward Transform is defined as composition with a diffeomorphism (Def. 2), and the radial/harmonic extensions (Defs. 4–5) come with explicit diffeomorphism conditions (Thm. 1), including a proof for the radial case and an external theorem (Alessandrini–Nesi) for the 2D harmonic case. The spectral descriptors (Def. 6, Prop. 1) achieve rotation/reflection invariance by taking Fourier or spherical-harmonic magnitudes, which is a mathematical consequence rather than a fitted quantity. The finite-dimensional surrogates (Defs. 9–11) minimize residuals of the same PDEs, with the log-barrier term L_n^r in Eq. (8) softly enforcing the determinant condition from Eq. (3). The only label-informed quantities are the late-fusion weights alpha_c in Eq. (15), selected by grid search on the validation split, which is standard hyperparameter tuning. PF-SDM features are computed from geometry alone, and the benchmark results on MPEG-7, BBBC010, and the gastruloid data are evaluated on held-out test splits against external or independently implemented baselines. The paper's self-citations to [2] concern code provenance, the gastruloid evaluation protocol, and a previously reported baseline; they are not load-bearing because the present paper re-derives the representation and supplies its own theoretical statements and benchmark evidence. The main caveats are correctness risks, not circularity: the paper explicitly states there is no proof of the harmonic-extension diffeomorphism property in d=3, and the discrete solver only penalizes nonpositive Jacobian determinants at collocation points, so folding is not rigorously excluded for non-star-convex MPEG-7 shapes. These limitations are acknowledged in the text and are validity concerns, not cases where a prediction reduces to its input by definition.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The central derivations rely on standard PDE and diffeomorphism results, but the practical claims rest on several user-chosen parameters (μ, n, λ, truncation orders, fusion weights) and on domain restrictions (topological spheres, star-convexity for 3D) that are acknowledged only in the conclusion. No genuinely new physical or mathematical entities are introduced; the PF-T is a named composition of known objects.

free parameters (6)

Viscosity parameter μ (viscous Eikonal equation, Def. 13-14) = Not stated for benchmark descriptors; Fig. 5 uses 0.6 and 0.001
Controls the trade-off between exact SDF and smoothness; changes the diffuse skeleton signature. No sensitivity analysis is reported.
Polynomial degree n (Π_n(Ω) and T_n spaces, Def. 7-8) = Not stated
Discretization order for SDF surrogate and deformation map; a user-chosen accuracy/co st trade-off.
Regularization parameter λ (Eq. 8, Def. 11) = Not stated
Strength of the log-barrier term that enforces det DΨ>0 for the harmonic extension; chosen by hand.
Spectral truncation N_F (2D) / L (3D) = N_F=15 and 25, L=20
Truncation orders of Fourier/spherical-harmonic descriptors; user-selected; affects discriminative granularity.
Late-fusion weights α_c (Eq. 15) = Grid-searched per split; gastruloid example α_SDF=0.41, α_Skel=0.36
Fitted to validation labels to maximize target metric; a supervised tuning parameter.
Quantile level α (Algorithm A1) = Not stated
Threshold for extracting the skeleton point cloud in the parametric medial-axis algorithm; not used for the PF-SDM descriptor itself.

assumptions (7)

domain assumption The shape boundary ∂S is a smooth, closed co-dimension-one manifold embedded in a bounded Lipschitz domain Ω, and S is an open domain.
Required for SDF level-set definition, PDEs, and harmonic extension; excludes shapes with corners, holes, or non-manifold boundary.
domain assumption There exists a C^1 diffeomorphism Ψ: S_r → S between the reference domain and the shape domain.
Central to the PF-T; restricts the method to shapes topologically equivalent to the reference ball. The conclusion acknowledges: 'It is therefore restricted to shapes that are topologically equivalent to the reference domain.'
domain assumption For the radial extension, the shape S is star-convex with respect to the origin, and ν(θ)>0.
Needed for Theorem 1(i) and global injectivity of the radial map; all 3D benchmarks and the 2D synthetic set use this assumption.
standard math The 2D harmonic extension is a C^1 diffeomorphism if and only if det(DΨ_h)>0 on ∂B_r (Alessandrini–Nesi theorem).
Invoked in the Appendix A proof of Theorem 1(ii); the algorithm only enforces this condition softly via the penalty term L_r in Eq. (8).
ad hoc to paper The closest-point projection x_S(x) and the least-squares optimization in Def. 10 define a unique and stable boundary deformation map.
No injectivity or stability proof for the minimizer of Eq. (1); for non-convex or non-star-shaped boundaries, the closest-point correspondence can be ambiguous.
standard math The viscous Eikonal variational problem (Eq. 11) has a solution in H^2, and the polynomial surrogate converges to it.
Assumed well-posedness of the regularized Eikonal problem; the authors cite Refs. [11,12,17] for polynomial surrogates and Sobolev cubatures.
domain assumption The divergence of the normalized SDF gradient ∇·η is a faithful signature of the medial axis.
Motivated by prior work [13] and illustrated in Fig. 5; not proven for the viscous approximation, yet used as the skeleton channel in all experiments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Push-Forward Transform for Continuous and Robust Comparison of Dynamic Shapes." pith.science (2026). https://pith.science/paper/CUFWNZVQ

@misc{pith2026260802306,
  author       = {Pith},
  title        = {Pith review of: The Push-Forward Transform for Continuous and Robust Comparison of Dynamic Shapes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CUFWNZVQ}},
  note         = {Machine review of arXiv:2608.02306}
}

read the original abstract

We introduce a mathematical framework for shape comparison based on mapping functions from the shape domain to a common reference domain. This Push-Forward Transform enables invariant and robust comparison of shapes, preserving intrinsic geometric information. Quantitatively comparing shapes and their temporal evolution is a fundamental challenge in image analysis. Meaningful shape comparison requires representations that are invariant to transformations that do not alter shape itself, such as translation, rotation, reflection, re-parametrization, and uniform scaling, while remaining sensitive to intrinsic geometric variation. Existing approaches often rely on sensitive parameterizations, landmark correspondence, or learned representations that are difficult to interpret and reproduce. We show that the Push-Forward Transform (PF-T) applied to Signed Distance Functions (SDFs) yields a continuous representation that captures both boundary and interior geometry. We derive an interpretable morphometric that quantifies shape similarity and reveals features such as skeletal topology and rotational symmetries. The push-forward transform applies consistently to two- and three-dimensional shapes, extends to time-evolving geometries, and supports the joint analysis of shape and additional scalar fields defined over shapes, such as intensity or molecular signals. We present the mathematical formulation, describe an efficient algorithm, and benchmark the approach on 2D, 3D, and temporal data sets.

Figures

Figures reproduced from arXiv: 2608.02306 by the authors.

**Figure 2.** Harmonic mapping of a star-convex shape to the unit disk. Left: original shape. Middle: arc-length boundary [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Example of the signed distance function of a star-convex shape, showing the zero level-set [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: From left to right: Input binary mask of a star-convex shape, signed distance function (SDF) approximation [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Signature of a shape skeleton in the smooth polynomial SDF. Top row: solution from the viscous Eikonal [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Top: 2D synthetic data set of regular polygons and the approximately symmetric five-petaled flower with [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 7.** Figure 7: MDS embeddings of PF-SDM distance matrices computed using the full normlized Fourier descriptors [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: Hierarchical clustering of shapes based on PF-SDM FFT spectra. Left: full descriptor [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 9.** Figure 9: Mean normalized Fourier descriptors per shape class, illustrating the relationship between dominant fre [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: Top: the six 3D shape classes. Bottom: MDS embeddings of pairwise distance matrices computed from [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

**Figure 11.** Figure 11: Example binary masks from the benchmark data sets. (a) BBBC010 showing per-well masks of microscopy [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

**Figure 12.** Figure 12: Shape analysis of developing mouse gastruloids. (A) Representative gastruloid images from the single [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages

[1]

Osher and R

S. Osher and R. Fedkiw,Level Set Methods and Dynamic Implicit Surfaces. Applied Mathematical Sciences. New York, NY: Springer, 2003

2003
[2]

A continuous and interpretable morphometric for robust quantification of dynamic biological shapes,

R. Rouatbi, J.-E. Suarez Cardona, A. Villaronga-Luque, J. V . Veenvliet, and I. F. Sbalzarini, “A continuous and interpretable morphometric for robust quantification of dynamic biological shapes,” in2026 IEEE 23rd Int. Symp. Biomedical Imaging (ISBI), 2026, pp. 1–5

2026
[3]

Generalized Procrustes analysis,

J. C. Gower, “Generalized Procrustes analysis,”Psychometrika, vol. 40, no. 1, pp. 33–51, 1975

1975
[4]

Elliptic Fourier features of a closed contour,

F. P. Kuhl and C. R. Giardina, “Elliptic Fourier features of a closed contour,”Comput. Graph. Image Process., vol. 18, no. 3, pp. 236–258, 1982

1982
[5]

SHTools: Tools for working with spherical harmonics,

M. A. Wieczorek and M. Meschede, “SHTools: Tools for working with spherical harmonics,”Geochem. Geo- phys. Geosyst., vol. 19, no. 8, pp. 2574–2592, 2018

2018
[6]

ShapeEmbed: A self-supervised learning framework for 2D contour quantification,

A. Foix Romero, C. Russell, A. Krull, and V . Uhlmann, “ShapeEmbed: A self-supervised learning framework for 2D contour quantification,”arXiv:2507.01009, 2025

arXiv 2025
[7]

Orientation-invariant au- toencoders learn robust representations for shape profiling of cells and organelles,

J. Burgess, J. J. Nirschl, M.-C. Zanellati, A. Lozano, S. Cohen, and S. Yeung-Levy, “Orientation-invariant au- toencoders learn robust representations for shape profiling of cells and organelles,”Nature Commun., vol. 15, no. 1, 2024

2024
[8]

A simple framework for contrastive learning of visual representations,

T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” inProc. 37th Int. Conf. Mach. Learn. (ICML), 2020

2020

Show all 28 references

[9]

Masked autoencoders are scalable vision learners,

K. He, X. Chen, S. Xie, Y . Li, P. Dollár, and R. Girshick, “Masked autoencoders are scalable vision learners,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 15979–15988

2022
[10]

Fourier descriptors for plane closed curves,

C. T. Zahn and R. Z. Roskies, “Fourier descriptors for plane closed curves,”IEEE Trans. Comput., vol. C-21, no. 3, pp. 269–281, 1972

1972
[11]

Negative order Sobolev cubatures: Preconditioners of partial differential equation learning tasks circumventing numerical stiffness,

J.-E. Suarez Cardona, P.-A. Hofmann, and M. Hecht, “Negative order Sobolev cubatures: Preconditioners of partial differential equation learning tasks circumventing numerical stiffness,”Mach. Learn.: Sci. Technol., vol. 5, no. 3, p. 035029, 2024

2024
[12]

Polynomial differentiation decreases the training time complexity of physics-informed neural networks and strengthens their approximation power,

J.-E. Suarez Cardona and M. Hecht, “Polynomial differentiation decreases the training time complexity of physics-informed neural networks and strengthens their approximation power,”Mach. Learn.: Sci. Technol., vol. 4, no. 4, p. 045005, 2023

2023
[13]

The Hamilton-Jacobi skeleton,

K. Siddiqi, S. Bouix, A. Tannenbaum, and S. W. Zucker, “The Hamilton-Jacobi skeleton,” inProc. 7th IEEE Int. Conf. Comput. Vis., 1999, pp. 828–834

1999
[14]

On Riemannian manifolds admitting a function whose gradient is of constant norm,

T. Sakai, “On Riemannian manifolds admitting a function whose gradient is of constant norm,”Kodai Math. J., vol. 19, no. 1, pp. 39–51, 1996

1996
[15]

Signed distance functions,

S. Osher and R. Fedkiw, “Signed distance functions,” inApplied Mathematical Sciences. New York, NY: Springer, 2003, pp. 17–22

2003
[16]

Viscosity solutions of Hamilton-Jacobi equations,

M. G. Crandall and P.-L. Lions, “Viscosity solutions of Hamilton-Jacobi equations,”Trans. Amer. Math. Soc., vol. 277, no. 1, p. 1, 1983

1983
[17]

A variational framework for the algorithmic complexity of PDE solutions,

J. E. Suarez Cardona, H. Boche, and G. Kutyniok, “A variational framework for the algorithmic complexity of PDE solutions,”arXiv:2510.21290, 2025

2025 arXiv
[18]

Distance solutions for medial axis transform,

H. Xia and P. G. Tucker, “Distance solutions for medial axis transform,” inProc. 18th Int. Meshing Roundtable. Berlin, Heidelberg: Springer, 2009, pp. 247–265

2009
[19]

Theλ-medial axis,

F. Chazal and A. Lieutier, “Theλ-medial axis,”Graphical Models, vol. 67, no. 4, pp. 304–331, 2005

2005
[20]

A skeletal measure of 2D shape similarity,

A. Torsello and E. R. Hancock, “A skeletal measure of 2D shape similarity,”Comput. Vis. Image Underst., vol. 95, no. 1, pp. 1–29, 2004

2004
[21]

scikit-image: Image processing in Python,

S. van der Walt, J. L. Schönberger, J. Nunez-Iglesias, F. Boulogne, J. D. Warner, N. Yager, E. Gouillart, and T. Yu, “scikit-image: Image processing in Python,”PeerJ, vol. 2, p. e453, 2014. 23 The Push-Forward Transform for Continuous and Robust Comparison of Dynamic ShapesA PREPRINT

2014
[22]

Identification of approximate symmetries in biological development,

P. Gandhi, M.-V . Ciocanel, K. Niklas, and A. T. Dawes, “Identification of approximate symmetries in biological development,”Phil. Trans. R. Soc. A, vol. 379, no. 2213, 2021

2021
[23]

MPEG-7 Core Experiment CE-Shape-1 Test Set: Benchmarking image database for shape recognition techniques,

“MPEG-7 Core Experiment CE-Shape-1 Test Set: Benchmarking image database for shape recognition techniques,” 2009. [Online]. Available:https://www.ehu.eus/ccwintco/index.php/MPEG-7_Core_ Experiment_CE-Shape-1_Test_Set._Benchmarking_image_database_for_shape_recognition_ techniques

2009
[24]

Annotated high-throughput microscopy image sets for valida- tion,

V . Ljosa, K. L. Sokolnicki, and A. E. Carpenter, “Annotated high-throughput microscopy image sets for valida- tion,”Nature Methods, vol. 9, no. 7, p. 637, 2012

2012
[25]

The OpenCV library,

G. Bradski, “The OpenCV library,”Dr. Dobb’s J. Softw. Tools, 2000

2000
[26]

SpinePy enables automated 3D spatiotemporal quantification of multicellular in vitro systems,

R. G. Savill, A. Villaronga-Luque, M. Trani Bustos, Y . Maroudas-Sacks, J. Batki, A. Meissner, A. Q. Ryan, C. D. Modes, O. Campàs, and J. V . Veenvliet, “SpinePy enables automated 3D spatiotemporal quantification of multicellular in vitro systems,”bioRxiv, doi:10.1101/2025.09....

2025 doi
[27]

Cartan,Differential Forms, Dover Books on Mathematics

H. Cartan,Differential Forms, Dover Books on Mathematics. Mineola, NY: Dover Publications, 2006

2006
[28]

Invertible harmonic mappings, beyond Kneser,

G. Alessandrini and V . Nesi, “Invertible harmonic mappings, beyond Kneser,”Ann. Sc. Norm. Super. Pisa Cl. Sci., pp. 451–468, 2009. 24

2009

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Osher and R

S. Osher and R. Fedkiw,Level Set Methods and Dynamic Implicit Surfaces. Applied Mathematical Sciences. New York, NY: Springer, 2003

2003

[2] [2]

A continuous and interpretable morphometric for robust quantification of dynamic biological shapes,

R. Rouatbi, J.-E. Suarez Cardona, A. Villaronga-Luque, J. V . Veenvliet, and I. F. Sbalzarini, “A continuous and interpretable morphometric for robust quantification of dynamic biological shapes,” in2026 IEEE 23rd Int. Symp. Biomedical Imaging (ISBI), 2026, pp. 1–5

2026

[3] [3]

Generalized Procrustes analysis,

J. C. Gower, “Generalized Procrustes analysis,”Psychometrika, vol. 40, no. 1, pp. 33–51, 1975

1975

[4] [4]

Elliptic Fourier features of a closed contour,

F. P. Kuhl and C. R. Giardina, “Elliptic Fourier features of a closed contour,”Comput. Graph. Image Process., vol. 18, no. 3, pp. 236–258, 1982

1982

[5] [5]

SHTools: Tools for working with spherical harmonics,

M. A. Wieczorek and M. Meschede, “SHTools: Tools for working with spherical harmonics,”Geochem. Geo- phys. Geosyst., vol. 19, no. 8, pp. 2574–2592, 2018

2018

[6] [6]

ShapeEmbed: A self-supervised learning framework for 2D contour quantification,

A. Foix Romero, C. Russell, A. Krull, and V . Uhlmann, “ShapeEmbed: A self-supervised learning framework for 2D contour quantification,”arXiv:2507.01009, 2025

arXiv 2025

[7] [7]

Orientation-invariant au- toencoders learn robust representations for shape profiling of cells and organelles,

J. Burgess, J. J. Nirschl, M.-C. Zanellati, A. Lozano, S. Cohen, and S. Yeung-Levy, “Orientation-invariant au- toencoders learn robust representations for shape profiling of cells and organelles,”Nature Commun., vol. 15, no. 1, 2024

2024

[8] [8]

A simple framework for contrastive learning of visual representations,

T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” inProc. 37th Int. Conf. Mach. Learn. (ICML), 2020

2020