Pith. sign in

REVIEW 3 major objections 4 minor 19 references

Predicting Brain Morphometry with MT-GNN: Mesh Evolution in Continuous Time with Graph-Based Metric Tensor Embeddings

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read MT-GNN forecasts brain-surface shape by predicting each vertex's intrinsic metric in continuous time and decoding it via as-rigid-as-possible, beating the no-change baseline by 2.29% and two learned baselines.

desk verdict A genuinely new continuous-time metric-tensor predictor for subcortical surfaces, with a solid but incomplete evaluation: the missing K-only ARAP baseline makes the headline gain over the temporal mean hard to attribute to prediction rather than representation. read the letter →

arxiv 2608.05132 v1 pith:MRH356TC submitted 2026-08-05 cs.CV cs.LG

classification cs.CVcs.LG
keywords longitudinalshapepredictionfirstfundamentalformmetrictensoras-rigid-as-possiblereconstructioncontinuous-timeforecastingsubcorticalbrainatrophygraphneuralnetworksoftrealizabilityconstraint
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

This paper claims that the right way to forecast how a subcortical brain structure will change in shape is to predict its intrinsic surface geometry, the per-vertex first fundamental form (metric tensor), a rigid-frame-invariant summary of local distances, rather than to extrapolate vertex positions or regress deformations directly. The proposed model, MT-GNN, is a single graph network that takes any causal multi-visit prefix and any lead time, and predicts a future metric as the sum of a log-Euclidean base mean, a learnable population shift, and a per-vertex subject residual. The predicted metric is decoded into a surface by a differentiable As-Rigid-As-Possible solver, and training end-to-end on the rigid-aligned vertex error of that decoded surface consistently beats supervising the metric directly, acting as a soft realizability constraint. On 14 ADNI subcortical structures, the model's mean vertex error is 2.29% below the strong temporal-mean baseline ($p=6.1\times10^{-5}$), it beats that baseline on all 14 structures, and it is ahead of both a geodesic-regression baseline (DCM) and a mesh-transformer baseline (TransforMesh) at every horizon, with the lead widening as the horizon grows.

What carries the argument

The central object is the per-vertex first fundamental form, the metric tensor $g_i \in \mathrm{SPD}(2)$, an intrinsic descriptor of local surface geometry written in a fixed tangent frame inherited from a common spherical parameterization. The load-bearing identity is the three-term prediction $\operatorname{Log}\hat{g}_i = K_i + \beta D_i(\Delta t) + r_i$, which decomposes the forecast into a log-Euclidean base metric, a scaled population-displacement direction frozen from training targets, and a per-vertex subject residual from a mesh graph network conditioned on a Fourier time encoding. The second mechanism is the differentiable As-Rigid-As-Possible (ARAP) decoder: it converts the predicted metric into target edge lengths through a symmetrized projection, then solves for vertex positions whose local edge geometry is as rigid as possible with respect to those lengths, initialized at the temporal-mean surface $S_0$ and unrolled for a fixed 15 iterations so the whole map is differentiable. Training through this reconstruction enforces a soft realizability constraint, because an arbitrary metric field need not be realizable by any embedded mesh, and the observation that generic triangulated surfaces are infinitesimally rigid explains both why the constraint helps and why the choice of starting surface selects the realization branch.

What would settle it

Re-run the identical training and evaluation protocol on the same ADNI scans processed through a different surfacing or registration pipeline (for example a different FreeSurfer version or spherical registration) and check whether MT-GNN still beats the temporal mean by roughly 2% and both baselines on held-out subjects at every horizon; if the advantage collapses or flips sign, the metric signal the model exploits is pipeline-specific rather than anatomical. A second, cheaper check is to repeat the evaluation with the per-sample outlier-exclusion step of Section 4.1 either removed or exactly specified, since its rule is not documented.

Watch

Extended reading notes

Core claim

The central claim is that a single continuous-time predictor of the future per-vertex first fundamental form outperforms both trajectory extrapolation in shape space and direct vertex-deformation regression for longitudinal subcortical shape prediction. The forecast is the three-term identity $\operatorname{Log}\hat{g}_i = K_i + \beta D_i(\Delta t) + r_i$: the log-Euclidean mean of the prefix metrics, a learned scalar times a frozen population displacement estimated once from training targets, and a per-vertex subject residual produced by a mesh graph network conditioned on a Fourier encoding of the lead time. The predicted metric fixes target edge lengths, which a differentiable as-rigid-as-possible solve, initialized at the temporal-mean surface $S_0$, decodes into a valid embedded mesh; the model is trained on the mean vertex error of that decoded surface after rigid alignment. On 14 ADNI subcortical structures, MT-GNN reaches $-2.29\%$ mean vertex error relative to the temporal mean ($p=6.1\times10^{-5}$), improving on all 14 structures and on both learned baselines, DCM ($-0.19\%$) and TransforMesh ($-0.45\%$), at every horizon from 12 to 48 months. A geometric argument explains the design: for a generic triangulated genus-zero mesh the isometric-realization set is zero-dimensional, so the metric fixes the surface up to finitely many branches, which is why training through reconstruction matters and why adding a mean-curvature head helps only marginally (MT-GNN+H, $-2.48\%$).

Load-bearing premise

The method assumes that the per-vertex metric changes measured from FreeSurfer meshes under a shared spherical parameterization are a reproducible, subject-specific biological signal rather than segmentation and surfacing noise; the frozen population-displacement term and the temporal-mean initialization stand or fall with that assumption.

Editorial extensions

If this is right

  • One trained model answers forecasts at any continuous lead time from a causal prefix, so a clinical user can query a 17-month or 30-month horizon without retraining, although validation off the 12-month grid is left for future work.
  • Because the advantage over the temporal mean grows with horizon, intrinsic-metric prediction is specifically the regime where extrinsic trajectory extrapolation overshoots, and the gap between MT-GNN and DCM widens accordingly.
  • Supervising the decoded surface beats supervising the metric field itself ($-2.26\%$ vs $-1.78\%$), so reconstruction-constrained training functions as a practical realizability principle for any predict-a-tensor-then-solve pipeline.
  • The rigidity argument predicts that mean-curvature conditioning can only refine shape within a branch already chosen by the decoder, which matches the observed small, mostly localized gains of MT-GNN+H.
  • Beating the temporal mean on all 14 structures, significantly on 12 of 14 for MT-GNN and all 14 for MT-GNN+H, supports the claim that the metric residual carries reproducible subject-level signal above the segmentation noise floor.

Reading between the lines

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

  • If the metric-change signal is genuinely biological rather than pipeline noise, the same architecture should transfer to cortical surfaces or a unified multi-structure model, since the machinery only assumes a fixed-topology genus-zero atlas mesh with a shared spherical parameterization.
  • The branch-choice argument suggests the decoder's initialization at the temporal-mean surface acts as an implicit shape prior; a testable variant would initialize at a cohort-matched population-mean surface and check whether long-horizon forecasts change meaningfully.
  • A pipeline-transfer experiment (same subjects, different surfacing or registration software) would directly test the noise-floor assumption; the paper's per-sample outlier-exclusion rule, which is not fully specified, makes such a check important before clinical use.
  • The group-level significance results do not by themselves establish individual-level predictive power, but if per-subject residuals rank progression rates, the model could enrich clinical trials by selecting fast atrophiers from their first few visits.
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

3 major / 4 minor

Summary. The paper proposes MT-GNN, a continuous-time graph network that predicts the future per-vertex first fundamental form (metric tensor) of a subcortical surface from a causal multi-visit prefix, conditioned on a Fourier encoding of the lead time. The predicted metric is decoded into a surface by a differentiable as-rigid-as-possible (ARAP) solver, and the model is trained end-to-end on rigid-aligned vertex error. On 14 ADNI subcortical structures, the authors report a mean vertex error 2.29% below the temporal-mean baseline (p=6.1e-5), beating two learned baselines (DCM and TransforMesh) at every horizon, with the lead widening as the horizon grows. The paper also presents ablations supporting the log-Euclidean base, the base shift, and reconstruction-constrained training, plus a curvature-augmented variant MT-GNN+H.

Significance. If the reported result holds, the paper makes a useful contribution to longitudinal shape prediction: it introduces an intrinsic metric-tensor representation for mesh evolution, a population-displacement plus subject-residual decomposition, and a differentiable ARAP decoder that enforces surface realizability during training. The evaluation is substantially stronger than typical in this area: a subject-disjoint split, paired per-subject t-tests, and Bonferroni correction across 70 comparisons are used, and the ablation table isolates several design choices. However, the central claim that the model 'predicts the future' is currently under-supported because the headline comparison to the temporal mean does not include a no-prediction baseline consisting of the ARAP reconstruction of the log-Euclidean mean metric alone. The contribution is therefore promising but needs additional evidence and clarification before the claims as stated can be accepted.

major comments (3)
  1. [§4.4, Table 4, Eq. (7)] The paper never reports the no-prediction baseline ARAP(S0, Exp K), obtained by setting β=0 and the residual r to zero in Eq. (7). This baseline is not the temporal-mean surface S0: S0 is a rigid-aligned average of vertex positions, whereas Exp K is the log-Euclidean mean of the prefix metric tensors. Because the ARAP decoder is initialized at S0 but can move vertices to fit the predicted edge lengths, the K-only reconstruction is a different and potentially better summary of the past than S0. The reported −2.29% improvement over the temporal mean may therefore conflate a better representation of the prefix with genuine time-conditioned prediction of future change. The ablations in Table 4 always include a learned residual r, so they do not isolate the K-only surface. This omission is load-bearing because the entire evaluation is built on beating the temporal mean; the authors should add this baseline and report its MVE relative to the temporal mean.
  2. [§4.2, Table 3] The headline p=6.1e-5 for MT-GNN vs. the temporal mean is not backed by any table. Table 3 reports paired tests among the learned methods (G, H, D, T), but it has no column comparing MT-GNN to the temporal mean, and the abstract's p-value does not appear in any table. Since the central claim is the improvement over the temporal mean, the authors must report the per-subject paired test against the temporal mean explicitly, including the Bonferroni correction details over the claimed 70 comparisons.
  3. [§4.1] The data filtering step is insufficiently documented. The text mentions a 'per-sample outlier flag based on Mean Absolute Distance and the persistence error' but does not specify the threshold, how the two criteria are combined, or how many subjects/visits were removed. This filtering can affect the difficulty of the prediction task and the comparability with baselines; the authors should provide the exact criterion or a sensitivity analysis over the threshold.
minor comments (4)
  1. [§4.1, §4.2, Table 1] Table 1 reports no error bars or confidence intervals for the per-horizon MVE values; the reader cannot assess the stability of the differences at individual horizons. The authors should add standard deviations or confidence intervals, as already done in Table 4 for the pooled averages.
  2. [§4.1] The DCM and TransforMesh baselines are described as 'adapted' or 'retrofitted' to the arbitrary-prefix protocol, but no hyperparameters, training schedules, or model selection details are given. Without this information the reader cannot judge whether the comparison is fair. Please provide tuning details and, ideally, multiple random seeds.
  3. [§3.2, Eq. (5)] The Fourier frequency set {ω_k} in Eq. (5) is never specified. Since this is a free parameter of the time encoding, the authors should state the number of frequencies and their values or how they were chosen.
  4. [§5, Limitations] The limitation paragraph notes that long-horizon validation samples are few and that off-grid validation is left for future work; this is appropriate, but the abstract and conclusions should temper the claim that the model is validated at 'arbitrary' horizons, since only the 12/24/36/48-month grid is tested.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MT-GNN is a supervised regressor with a subject-disjoint test split; the population displacement is a training-only statistic, and the Section 5 rigidity argument rests on external mathematics.

full rationale

The derivation chain is not circular. The predicted metric in Eq. (7) is a supervised composition: a log-Euclidean base K_i computed from the causal prefix (Eq. 2), a population displacement D_i(Δt) estimated once from training-set targets (Eq. 4), a learnable scalar β, and a graph-network residual r_i; all of θ and β are fitted on the training split and evaluated on a subject-disjoint test split with the same MVE metric. D(Δt) is explicitly frozen and depends only on training targets, so it is an ordinary fitted statistic, not a prediction renamed as a first-principles result. The temporal-mean baseline S0 (Eq. 3) is a different operation from the metric-mean reconstruction, and the paper does not report the K-only ARAP baseline; that is an experimental/representational limitation, but it does not make the prediction reduce to its inputs by construction. Self-citations ([7], [8]) are background or preprocessing tools and are not load-bearing. The realizability discussion in Section 5 invokes the external Gluck rigidity theorem [5], not an author-uniqueness claim, and is presented as an account of the small curvature gain, not as the source of the predictive result. No circular step can be exhibited from the paper's equations.

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

The model rests on several fitted parameters and data-derived terms (population displacement, base shift, Fourier features, outlier flag) and on domain assumptions about the reproducibility of the metric signal and the behavior of the ARAP solver. No new physical or geometric entities are postulated; the metric tensor is a standard object and the realizability constraint is a soft training objective rather than an invented entity.

free parameters (8)
  • base-shift scalar β = learned, not reported separately
    Scales the frozen population displacement D(∆t); trained end-to-end (Eq. 7).
  • population displacement D(∆t) = per-horizon mean log-metric shifts over training set
    Empirical per-horizon change of future metric relative to base, frozen after training (Eq. 4).
  • Fourier frequency set {ωk} = not reported
    Time encoding frequencies chosen by hand (Eq. 5); the set is unspecified and affects temporal conditioning.
  • outlier flag threshold = not reported
    Quality-control criterion for excluding segmentation/surfacing failures (Section 4.1); threshold and count not disclosed.
  • ARAP regularizer λ = small, not specified
    Keeps the global Laplacian system positive-definite (Eq. 9).
  • residual regularizer weight η = not specified
    Weights the L2 penalty on the subject residual in the loss (Eq. 10).
  • curvature head scale σH and bound κ = computed from training set / hand-chosen
    Normalize and bound the mean-curvature residual in the MT-GNN+H variant (Eq. 11).
  • curvature term weight µ = not specified
    Balances the curvature term in the global ARAP step for MT-GNN+H (Eq. 12).
assumptions (6)
  • standard math Gluck's rigidity theorem: almost all simply connected closed surfaces are infinitesimally rigid as bar-and-joint frameworks
    Used in Section 5 to argue that fixing the edge lengths (via the predicted metric) leaves only a zero-dimensional set of realizations, so the metric nearly determines the surface up to discrete branches.
  • domain assumption The log-Euclidean mean is a practical surrogate for the affine-invariant Karcher mean
    Used in Eq. 2 as the base metric; this approximates the true intrinsic mean and affects the reference.
  • domain assumption Per-vertex metric tensors computed from FreeSurfer meshes under a shared spherical parameterization capture reproducible biological change
    Underpins the entire prediction setup (Sections 3.1-3.2); if the metric is dominated by segmentation noise, the model cannot generalize.
  • domain assumption The ARAP solver converges to the branch containing the temporal-mean initialization S0 and this branch is the biologically relevant one
    Used in Section 5 to explain why mean curvature is only a second-order correction; this is an optimization and modeling assumption, not proven.
  • domain assumption Outlier exclusion removes only segmentation/surfacing failures
    Assumed by the data protocol in Section 4.1; if the flag removes systematic shape changes, the test-set estimates are biased.
  • domain assumption Subject-disjoint split and per-subject averaging yield valid inference for paired t-tests
    Standard experimental assumption; the paper aggregates multiple pairs per subject to one MVE, which limits but does not eliminate dependence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Brain Morphometry with MT-GNN: Mesh Evolution in Continuous Time with Graph-Based Metric Tensor Embeddings." pith.science (2026). https://pith.science/paper/MRH356TC

@misc{pith2026260805132,
  author       = {Pith},
  title        = {Pith review of: Predicting Brain Morphometry with MT-GNN: Mesh Evolution in Continuous Time with Graph-Based Metric Tensor Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRH356TC}},
  note         = {Machine review of arXiv:2608.05132}
}
abstract

Predicting how a subcortical structure's shape will evolve from a few prior scans could support prognosis and clinical-trial enrichment. Existing longitudinal mesh predictors either extrapolate shape trajectories via high-dimensional embeddings or regress vertex deformations directly. We instead predict the surface's intrinsic geometry in continuous time: a single per-structure graph network predicts the future per-vertex first fundamental form (metric tensor) for an arbitrary causal multiple-visit history and an arbitrary prediction horizon, conditioned on a Fourier encoding of the lead time. The predicted metric is decoded into a surface by a differentiable As-Rigid-As-Possible solver, and the model is trained end-to-end on the rigid-aligned vertex error. Training through the reconstruction keeps the decoded prediction a valid surface and consistently improves it. On 14 subcortical structures from the ADNI dataset, the proposed mesh evolution model (MT-GNN) predicts best among the evaluated methods at every horizon ($-2.29\%$ mean vertex error vs. the temporal mean, $p{=}6.1{\times}10^{-5}$, beating it on 14/14 structures), ahead of geodesic shape regression (DCM, $-0.19\%$) and a mesh transformer (TransforMesh, $-0.45\%$; $p{=}1.2{\times}10^{-4}$), with the lead widening as the horizon grows.

Figures

Figures reproduced from arXiv: 2608.05132 by the authors.

Figure 1
Figure 1. MT-GNN pipeline. From an m-visit causal prefix, the per-vertex input met￾ric g, a Fourier time embedding γ(∆t) of the lead time, and the population-displacement term Dpop(∆t) feed a mesh graph network (alternating Linear and MeshConv layers) that predicts a per-vertex log-metric residual. Combined (⊕) with the log-Euclidean base metric and the base shift β Dpop(∆t) (skip connections), this yields the predicted first… view at source ↗
Figure 2
Figure 2. Accuracy vs. horizon (14 structures). Mean MVE relative to the temporal mean (%, lower is better; y-axis inverted) at each lead time ∆t. MT-GNN and MT￾GNN+H are the top two lowest-error curves, with MT-GNN improving monotonically and its lead over the temporal mean widening with ∆t, while DCM rises above the temporal mean at ∆t=36; the MT-GNN+H gain over MT-GNN is small and significant per structure in only 4/14 str… view at source ↗
Figure 3
Figure 3. Absolute accuracy vs. horizon (14 structures). [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Per-structure error difference, averaged over the test set. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Qualitative forecasts on test subjects. For the right caudate, putamen and hippocampus (rows), the left block shows the three-visit causal input prefix (screen￾ing, 12 and 24 months; shaded) and the right block the forecasts of MT-GNN, DCM and TransforMesh at the targe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 8 canonical work pages

  1. [1]

    ACM Transactions on Graphics (Proc

    Aigerman, N., Gupta, K., Kim, V.G., Chaudhuri, S., Saito, J., Groueix, T.: Neural Jacobian fields: Learning intrinsic mappings of arbitrary meshes. ACM Transactions on Graphics (Proc. SIGGRAPH)41(4) (2022). https://doi.org/10.1145/3528223.3530141

  2. [2]

    Magnetic Resonance in Medicine56(2), 411–421 (2006)

    Arsigny, V., Fillard, P., Pennec, X., Ayache, N.: Log-Euclidean metrics for fast and simple calculus on diffusion tensors. Magnetic Resonance in Medicine56(2), 411–421 (2006). https://doi.org/10.1002/mrm.20965

  3. [3]

    In: Global Analysis (Proc

    Ebin, D.G.: The manifold of Riemannian metrics. In: Global Analysis (Proc. Sym- pos. Pure Math., Vol. XV, Berkeley, Calif., 1968), pp. 11–40. American Mathemat- ical Society, Providence, RI (1970)

  4. [4]

    NeuroImage62(2), 774–781 (2012)

    Fischl, B.: FreeSurfer. NeuroImage62(2), 774–781 (2012). https://doi.org/10.1016/j.neuroimage.2012.01.021

  5. [5]

    In: Ge- ometric Topology (Proc

    Gluck, H.: Almost all simply connected closed surfaces are rigid. In: Ge- ometric Topology (Proc. Conf., Park City, Utah, 1974), Lecture Notes in Mathematics, vol. 438, pp. 225–239. Springer, Berlin, Heidelberg (1975). https://doi.org/10.1007/BFb0066118

  6. [6]

    In: IEEE/CVF International Confer- ence on Computer Vision Workshops (ICCVW)

    Gong, S., Chen, L., Bronstein, M., Zafeiriou, S.: SpiralNet++: A fast and highly efficient mesh convolution operator. In: IEEE/CVF International Confer- ence on Computer Vision Workshops (ICCVW). pp. 4141–4148. IEEE (2019). https://doi.org/10.1109/ICCVW.2019.00509

  7. [7]

    In: Information Processing in Medical Imaging (IPMI 2015)

    Gutman, B.A., Fletcher, P.T., Cardoso, M.J., Fleishman, G.M., Lorenzi, M., Thompson, P.M., Ourselin, S.: A Riemannian framework for intrinsic compari- son of closed genus-zero shapes. In: Information Processing in Medical Imaging (IPMI 2015). Lecture Notes in Computer Science, vol. 9123, pp. 205–218. Springer (2015). https://doi.org/10.1007/978-3-319-19992-4_16

  8. [8]

    8159, book section 24, pp

    Gutman, B.A., Madsen, S.K., Toga, A.W., Thompson, P.M.: A Family of Fast Spherical Registration Algorithms for Cortical Shapes, Lecture Notes in Computer Science, vol. 8159, book section 24, pp. 246–257. Springer International Publish- ing (2013). https://doi.org/10.1007/978-3-319-02126-3_24,http://dx.doi.org/ 10.1007/978-3-319-02126-3_24

Show all 19 references
  1. [9]

    Journal of Magnetic Res- onance Imaging27(4), 685–691 (2008)

    Jack Jr., C.R., Bernstein, M.A., Fox, N.C., Thompson, P., Alexander, G., Harvey, D., Borowski, B., Britson, P.J., Whitwell, J.L., Ward, C., et al.: The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. Journal of Magnetic Res- onance Imaging27(4), 685–691 (2008)...

  2. [10]

    SIAM Journal on Matrix Analysis and Applications40(4), 1353– 1370 (2019)

    Lin, Z.: Riemannian geometry of symmetric positive definite matrices via Cholesky decomposition. SIAM Journal on Matrix Analysis and Applications40(4), 1353– 1370 (2019). https://doi.org/10.1137/18M1221084

  3. [11]

    International Journal of Computer Vision66(1), 41–66 (2006)

    Pennec, X., Fillard, P., Ayache, N.: A Riemannian framework for tensor computing. International Journal of Computer Vision66(1), 41–66 (2006). https://doi.org/10.1007/s11263-005-3222-z

  4. [12]

    In: International Conference on Learning Representations (ICLR) (2021)

    Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., Battaglia, P.W.: Learning mesh- based simulation with graph networks. In: International Conference on Learning Representations (ICLR) (2021)

  5. [13]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Rubanova, Y., Chen, R.T.Q., Duvenaud, D.: Latent ODEs for irregularly-sampled time series. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019)

  6. [14]

    In: Machine Learning in Medi- cal Imaging (MLMI 2021, held with MICCAI 2021)

    Sarasua, I., Pölsterl, S., Wachinger, C.: TransforMesh: A transformer network for longitudinal modeling of anatomical meshes. In: Machine Learning in Medi- cal Imaging (MLMI 2021, held with MICCAI 2021). Lecture Notes in Computer Predicting Subcortical Surfaces via the Metric ...

  7. [15]

    In: Proceedings of the Eurographics Symposium on Geometry Processing (SGP)

    Sorkine, O., Alexa, M.: As-rigid-as-possible surface modeling. In: Proceedings of the Eurographics Symposium on Geometry Processing (SGP). pp. 109–116. Euro- graphics Association (2007). https://doi.org/10.2312/SGP/SGP07/109-116

  8. [16]

    In: Advances in Neural In- formation Processing Systems (NeurIPS)

    Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Sing- hal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: Advances in Neural In- formation Processing System...

  9. [17]

    In: Shape in Medical Imaging (ShapeMI 2023, held with MICCAI 2023)

    Türkseven, D., Rekik, I., von Tycowicz, C., Hanik, M.: Predicting shape develop- ment: A Riemannian method. In: Shape in Medical Imaging (ShapeMI 2023, held with MICCAI 2023). Lecture Notes in Computer Science, vol. 14350, pp. 211–222. Springer (2023). https://doi.org/10.1007/...

  10. [18]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 30 (2017)

  11. [19]

    ACM Transactions on Graphics (Proc

    Yu, Y., Zhou, K., Xu, D., Shi, X., Bao, H., Guo, B., Shum, H.Y.: Mesh editing with Poisson-based gradient field manipulation. ACM Transactions on Graphics (Proc. SIGGRAPH)23(3), 644–651 (2004). https://doi.org/10.1145/1015706.1015774

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.