Pith. sign in

REVIEW 3 major objections 7 minor 27 references

Bidirectional attention deforms a heart atlas into a full 3D mesh

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A bidirectional cross-attention point transformer deforms a heart atlas into a 3D four-chamber mesh from sparse cardiac MRI point clouds using locally affine diffeomorphic flows.

T0 review reviewed 2026-07-09 challenge →

load-bearing objection Solid architecture for cardiac mesh reconstruction, but the CMR claim rests on simulated data with thin margins the 3 major comments →

arxiv 2607.06923 v1 pith:RSAZJ7CX submitted 2026-07-08 cs.CV

Bi-PT: Bidirectional Cross-Attention Point Transformers for Four-Chamber Heart Reconstruction from Sparse Cardiac MRI Data

Chenchuhui Hu , Shaoming Pan , Leon Axel , Meng Ye This is my paper
classification cs.CV
keywords bi-ptdeformationheartpointcardiacfieldfour-chamberlearning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 proposes Bi-PT, a method for reconstructing a complete 3D four-chamber heart mesh from the sparse point cloud that clinical cardiac MRI protocols actually produce. Standard CMR acquisitions capture only a few 2D long-axis and short-axis slices, leaving large gaps in 3D geometry. Bi-PT starts from a shared atlas heart and deforms it onto the sparse observations using two mechanisms working together: a bidirectional cross-attention layer that lets atlas points and sparse-observation points exchange both local detail and global shape context, and a deformation model formulated as a neural ordinary differential equation whose per-point affine-plus-translation dynamics guarantee that the atlas bends and stretches without folding or tearing. A semantic-aware Chamfer distance loss ensures that atrium points match atrium points, ventricle points match ventricle points, and boundary points match boundary points, rather than letting geometry alone drive correspondences across anatomically wrong structures. On 100 held-out test cases, the method achieves 2.28 mm Chamfer distance and 1.41 mm point-to-surface error while producing zero self-intersecting faces, improving on five learning-based baselines and one classical registration method. The central claim is that bidirectional information flow between atlas and sparse observation, combined with topology-preserving deformation and label-aware matching, yields more accurate and anatomically valid whole-heart reconstruction than prior approaches.

Core claim

The paper establishes that letting information flow in both directions between a fixed atlas and a sparse point cloud, rather than only from observation to atlas or only from atlas to observation, produces substantially better correspondence estimation for 3D heart reconstruction. Removing the reverse pathway degrades Chamfer distance from 2.28 to 3.10 mm and introduces self-intersections, confirming that the global target-shape descriptor extracted via the sparse-point-cloud-to-atlas direction is load-bearing rather than decorative. Combined with a locally affine diffeomorphic deformation parameterized as a NODE, the atlas deforms to match patient-specific geometry while preserving topology

What carries the argument

Three components carry the argument. First, the bidirectional cross-attention layer: atlas tokens query sparse-point-cloud tokens to absorb local shape cues, while sparse-point-cloud tokens query atlas tokens to aggregate a global target-shape descriptor that is broadcast back to every atlas vertex. Second, the LADD (locally affine diffeomorphic deformation) dynamics: instead of predicting per-point translations alone, each point carries its own 3-by-3 affine matrix and translation vector, integrated through a NODE, which by the Cauchy-Lipschitz theorem guarantees a diffeomorphism and thus topology preservation. Third, the semantic-aware Chamfer distance: each point carries labels indicating

Load-bearing premise

The evaluation uses CT-derived data processed to mimic sparse CMR point clouds rather than actual clinical sparse cardiac MRI data. The claim of robust reconstruction from clinical CMR rests on the assumption that this simulation faithfully reproduces real CMR sparsity patterns, noise characteristics, and segmentation artifacts.

What would settle it

If the reverse SPC-to-atlas attention pathway is removed, the method should degrade significantly in geometric accuracy and produce self-intersections, which the ablation confirms. If the LADD formulation does not actually preserve topology, self-intersection ratios should be nonzero, which they are not. If the semantic-aware Chamfer distance does not improve correspondences, removing it should not degrade EMD or CD, which it does.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the bidirectional cross-attention principle generalizes, any atlas-to-observation deformation task with sparse or partial data, not just cardiac MRI, could benefit from letting the observation stream summarize global context back to the atlas stream.
  • The guarantee of zero self-intersections from the LADD formulation could make atlas-based reconstruction viable for downstream clinical tasks requiring topologically valid meshes, such as finite-element biomechanical simulation or electrophysiology modeling.
  • The semantic-aware matching hierarchy, where exact label-pair matches are preferred but relaxed to single-label matches when combinations are rare, offers a template for other multi-structure reconstruction problems where boundary and junction points are sparse.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The manuscript proposes Bi-PT, a pipeline for reconstructing 3D four-chamber heart meshes from sparse point clouds (SPCs) derived from clinical cardiac MRI (CMR) protocols. The method combines bidirectional cross-attention between an atlas and the SPC using Point Transformer architectures, a locally affine diffeomorphic deformation (LADD) model parameterized via Neural Ordinary Differential Equations (NODE), a semantic-aware Chamfer distance loss, and Laplacian regularization. The approach is evaluated against six baselines on 1,000 cardiac CT cases processed to mimic sparse CMR observations, reporting improvements in Chamfer distance and point-to-surface distance while achieving zero self-intersections.

Significance. The work addresses a clinically relevant problem: reconstructing dense 3D cardiac geometry from the sparse 2D views typical of routine CMR protocols. The combination of bidirectional cross-attention for atlas-SPC correspondence with topology-preserving deformation dynamics is a reasonable architectural contribution. The ablation study (Table 2) isolating each component is a strength, as is the reporting of mesh quality metrics (normal consistency, non-manifold faces, self-intersections) alongside geometric accuracy. The guarantee of diffeomorphic deformation via the Cauchy-Lipschitz theorem is a formal property that adds value. However, the significance is tempered by the evaluation being conducted on simulated data rather than real CMR, and by the relatively small improvement margins over existing baselines.

major comments (3)
  1. Title, abstract, and Sec. 3.1: The paper claims reconstruction 'from clinical sparsely sampled cardiac MRI data,' but all experiments use cardiac CT cases [25] processed to 'mimic CMR scans' (Sec. 3.1). The simulation procedure is described in a single sentence with no validation that the simulated SPCs reproduce real CMR sparsity patterns, through-plane resolution effects, or segmentation boundary noise. This is a gap between the central claim and the experimental evidence. The authors should either (a) reframe the title and abstract to accurately state that the method is evaluated on simulated sparse CMR-like data derived from CT, or (b) provide validation on actual clinical sparse CMR data, even on a small test set, to demonstrate transfer. As written, the claim of 'clinical' CMR reconstruction is not directly supported.
  2. Table 1 and Sec. 3.4: The improvement margins over the strongest baseline (LTN) are small. CD improves from 2.34 to 2.28 mm (2.6%), P2F is identical (1.41 mm), and EMD is actually worse (6.54 vs 6.36 mm). The abstract states 'accurate and robust performance compared to baselines,' but the mixed metric picture—particularly the EMD regression—should be discussed explicitly. The authors should clarify why EMD degrades while CD and P2F improve, and whether the Wilcoxon significance test (reported only for CD and P2F) also applies to EMD. Without this, the claim of overall superiority is not fully supported.
  3. Table 2 (ablation): Two of the four claimed contributions—LADD (CD 2.28 vs 2.32 for translation-only) and semantic-aware CD (CD 2.28 vs 2.32 for plain CD)—each contribute only 0.04 mm improvement. The bidirectional cross-attention is the only component with a substantial effect (CD 2.28 vs 3.10). Given that the LADD and semantic-aware CD are presented as core contributions alongside the cross-attention, the authors should discuss whether these small ablation effects justify the added architectural complexity, or whether the primary contribution is effectively the bidirectional cross-attention alone.
minor comments (7)
  1. Table 1: NDM reports NC = 0.04, which is anomalously low compared to all other methods (0.57–0.66). This may reflect a known issue with NDM's shape primitive approach, but a brief note explaining this would help readers interpret the table.
  2. Sec. 2.1: The label scheme is described as 'a 3-channel triplet (La=3) for the atlas and a 2-channel pair for the SPC (Ls=2),' but the exact encoding of these channels is not fully specified. Clarification of how the semantic labels in Fig. 2 map to the channel values would improve reproducibility.
  3. Sec. 2.4: The text mentions that 'we found the 3-label variant could degrade the shape reconstruction accuracy,' and Table 2 confirms this (CD 2.32 vs 2.28). A brief discussion of why adding a third label degrades performance would strengthen the design rationale.
  4. Sec. 3.2: The atlas is 'randomly selected a sample heart mesh from the training dataset.' It would be useful to know whether results are sensitive to atlas choice, e.g., by reporting performance with 2–3 different atlas selections.
  5. Fig. 1: The architecture diagram is dense. Expanding it or splitting into sub-figures (encoder, cross-attention, decoder, LADD) would improve readability.
  6. References [13] and [14] are arXiv preprints from 2025. If the manuscript is under review at a venue with citation deadlines, these may need updating; otherwise this is not a formatting error.
  7. Sec. 3.1: The evaluation reports 'mean result over the five heart structures,' but per-structure results are not provided. Given that the qualitative results (Fig. 3) show varying error patterns across chambers (e.g., RA errors in LTN), per-chamber quantitative results would be informative.

Circularity Check

0 steps flagged

No significant circularity; one minor self-citation for data generation procedure that is not load-bearing for the method's claims.

full rationale

The paper's derivation chain is self-contained. The bidirectional cross-attention (Sec. 2.2) computes features from atlas and SPC tokens via standard attention mechanics (Eqs. 1-2); no output is defined in terms of itself. The LADD dynamics (Sec. 2.3, Eq. 3) parameterize a velocity field via MLP-predicted affine A and translation b, with topology preservation guaranteed by the Cauchy-Lipschitz theorem cited from Brezis [4]—an external mathematics textbook, not a self-citation. The semantic-aware Chamfer distance (Sec. 2.4) is a straightforward modification of standard CD with label-based nearest-neighbor restriction. The Laplacian regularization is a standard mesh smoothing term from Nealen et al. [16]. The only self-citation is to NDM [23] (shared author: Meng Ye), used for (a) the SPC data generation procedure ('as in [23]') and (b) as a baseline method. Neither is circular: the SPC generation defines the input data, not a prediction, and NDM is one of six external baselines. The method is evaluated on a public dataset (ImageCas [25]) with 100 held-out test cases against five other independent baselines. No component of the evaluation reduces to a fitted quantity by construction. The minor self-citation for data generation does not raise the circularity score above 1.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities or postulated objects. The architectural components (bidirectional cross-attention layer, LADD dynamics, semantic-aware CD) are methodological constructs built from standard neural network operations. The free parameters are loss weights and architectural dimensions, all empirically set. The axioms are standard mathematical results (Cauchy-Lipschitz) and domain assumptions (CT-simulates-CMR, atlas generalizability) rather than ad hoc postulates.

free parameters (9)
  • lambda_plain = 0.5
    Weight for plain Chamfer distance loss, set empirically
  • lambda_sa = 0.5
    Weight for semantic-aware Chamfer distance loss, set empirically
  • lambda_1 = 0.3
    Weight for first NODE block CD loss, set empirically
  • lambda_2 = 0.7
    Weight for second NODE block CD loss, set empirically
  • lambda_lap = 1.0
    Weight for Laplacian regularization, set empirically
  • N0 (atlas points) = 5632
    Number of atlas vertices, chosen by design
  • M0 (SPC points) = 5632
    Number of SPC points, chosen by design
  • C (channel width) = 512
    Feature channel dimension, architectural choice
  • Atlas mesh selection = randomly selected training sample
    Atlas is a randomly selected sample from training set, not optimized
axioms (4)
  • standard math Cauchy-Lipschitz theorem guarantees that the LADD ODE (Eq. 3) produces a diffeomorphic (topology-preserving) deformation
    Invoked in Sec. 2.3 to justify topology preservation; standard ODE existence/uniqueness result applied to the per-point affine flow
  • domain assumption CT-derived point clouds processed to mimic CMR sparsity faithfully represent the clinical sparse CMR acquisition scenario
    Sec. 3.1: the entire evaluation rests on this simulation being representative of real clinical CMR data
  • domain assumption Point Transformer self-attention with k-nearest-neighbor aggregation produces meaningful local geometric features
    Sec. 2.1 adopts the Point Transformer architecture from [27] as the encoder backbone without modification
  • domain assumption A single fixed atlas mesh can be deformed to represent the full range of pathological and normal cardiac shapes
    The atlas-to-target deformation paradigm assumes one template suffices for all subjects; standard in atlas-based methods but unverified here for extreme pathologies

reviewed 2026-07-09 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Bi-PT: Bidirectional Cross-Attention Point Transformers for Four-Chamber Heart Reconstruction from Sparse Cardiac MRI Data." pith.science (2026). https://pith.science/paper/RSAZJ7CX

@misc{pith2026260706923,
  author       = {Pith},
  title        = {Pith review of: Bi-PT: Bidirectional Cross-Attention Point Transformers for Four-Chamber Heart Reconstruction from Sparse Cardiac MRI Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RSAZJ7CX}},
  note         = {Machine review of arXiv:2607.06923}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We propose Bi-PT, a pipeline for reconstructing 3D four-chamber human heart meshes from clinical sparsely sampled cardiac magnetic resonance imaging (CMR) data. This work addresses the error-prone generation of 3D cardiac shape from a sparse point cloud (SPC) extracted from 2D long-axis and short-axis views used in routine clinical CMR protocols. Bi-PT enables accurate inference of the four-chamber heart mesh from the SPC by learning robust point features via bidirectional point cross-attention between an atlas and the SPC, together with per-point semantic labels that improve correspondence estimation. We formulate the deformation field as a Neural Ordinary Differential Equation (NODE) parameterized by a per-point affine transformation and translation to deform the atlas toward the target heart shape. By learning such a NODE, we can guarantee the deformation field to be a locally affine diffeomorphic deformation. We also integrate a semantic label loss into the Chamfer distance to encourage label-consistent correspondences and add a smoothness regularization to stabilize and improve the learning of the deformation field. Extensive experiments demonstrate that Bi-PT achieves accurate and robust performance compared to baselines.

Figures

Figures reproduced from arXiv: 2607.06923 by Chenchuhui Hu, Leon Axel, Meng Ye, Shaoming Pan.

Figure 1
Figure 1. Figure 1: Architecture of Bi-PT. The bidirectional cross-attention Point Trans￾former (Bi-PT) learns fused atlas-SPC point features, which predict locally affine diffeomorphic deformation (LADD). Two NODE blocks then deform the atlas to produce an intermediate mesh and the final mesh. GAP: global average pooling. Point Transformer decoders, and (4) LADD through two NODE blocks. We give details of Bi-PT as follows. 2… view at source ↗
Figure 2
Figure 2. Figure 2: Each point is labeled by the set of chambers it touches: (a, a) interior, (a, b) boundary, (a, b, c) junction. Semantic-aware Chamfer Distance. Plain Cham￾fer distance (CD) [23] across the four chambers ignores chamber identity and can match points across structures. We therefore use a semantic￾aware CD (SA-CD) that restricts nearest￾neighbor search by the semantic labels. As shown in [PITH_FULL_IMAGE:fig… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of four-chamber whole-heart reconstruction on [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

  1. [1]

    Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences379(2212) (2021)

    Banerjee, A., Camps, J., Zacur, E., Andrews, C.M., Rudy, Y., Choudhury, R.P., Rodriguez, B., Grau, V.: A completely automated pipeline for 3d reconstruction of human heart from 2d cine magnetic resonance slices. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences379(2212) (2021)

    work page 2021

  2. [2]

    In: International Work- shop on Statistical Atlases and Computational Models of the Heart

    Beetz, M., Banerjee, A., Grau, V.: Point2mesh-net: Combining point cloud and mesh-based deep learning for cardiac shape reconstruction. In: International Work- shop on Statistical Atlases and Computational Models of the Heart. pp. 280–290. Springer (2022)

    work page 2022

  3. [3]

    Medical image analysis90, 102975 (2023)

    Beetz, M., Banerjee, A., Ossenberg-Engels, J., Grau, V.: Multi-class point cloud completion networks for 3d cardiac anatomy reconstruction from cine magnetic resonance images. Medical image analysis90, 102975 (2023)

    work page 2023

  4. [4]

    Brezis, H., Brézis, H.: Functional analysis, Sobolev spaces and partial differential equations, vol. 2. Springer (2011)

    work page 2011

  5. [5]

    Advances in neural information processing systems31(2018)

    Chen, R.T., Rubanova, Y., Bettencourt, J., Duvenaud, D.K.: Neural ordinary dif- ferential equations. Advances in neural information processing systems31(2018)

    work page 2018

  6. [6]

    Medical image analysis74, 102228 (2021)

    Chen, X., Ravikumar, N., Xia, Y., Attar, R., Diaz-Pinto, A., Piechnik, S.K., Neubauer, S., Petersen, S.E., Frangi, A.F.: Shape registration with learned de- formations for 3d shape reconstruction from sparse and incomplete point clouds. Medical image analysis74, 102228 (2021)

    work page 2021

  7. [7]

    Medical Image Analysis74, 102228 (2021)

    Chen, X., Ravikumar, N., Xia, Y., Attar, R., Diaz-Pinto, A., Piechnik, S.K., Neubauer, S., Petersen, S.E., Frangi, A.F.: Shape registration with learned de- formations for 3d shape reconstruction from sparse and incomplete point clouds. Medical Image Analysis74, 102228 (2021)

    work page 2021

  8. [8]

    University of California, San Diego (2020)

    Gupta, K.: Neural mesh flow: 3d manifold mesh generation via diffeomorphic flows. University of California, San Diego (2020)

    work page 2020

  9. [9]

    Medical image analysis4(4), 335– 355 (2000)

    Haber, I., Metaxas, D.N., Axel, L.: Three-dimensional motion reconstruction and analysis of the right ventricle using tagged mri. Medical image analysis4(4), 335– 355 (2000)

    work page 2000

  10. [10]

    In: In- ternational conference on medical image computing and computer-assisted inter- vention

    He, X., Tan, C., Han, L., Liu, B., Axel, L., Li, K., Metaxas, D.N.: Dmcvr: Morphology-guided diffusion model for 3d cardiac volume reconstruction. In: In- ternational conference on medical image computing and computer-assisted inter- vention. pp. 132–142. Springer (2023)

    work page 2023

  11. [11]

    In: International workshop on shape in medical imaging

    Jayakumar, N., Hossain, T., Zhang, M.: Sadir: shape-aware diffusion models for 3d image reconstruction. In: International workshop on shape in medical imaging. pp. 287–300. Springer (2023)

    work page 2023

  12. [12]

    Medical image analysis74, 102222 (2021)

    Kong, F., Wilson, N., Shadden, S.: A deep-learning approach for direct whole-heart mesh reconstruction. Medical image analysis74, 102222 (2021)

    work page 2021

  13. [13]

    arXiv preprint arXiv:2512.00264 (2025) 10 Hu et al

    Ma, Z., Banerjee, A.: Heartformer: Semantic-aware dual-structure transform- ers for 3d four-chamber cardiac point cloud reconstruction. arXiv preprint arXiv:2512.00264 (2025) 10 Hu et al

    work page arXiv 2025

  14. [14]

    arXiv preprint arXiv:2510.01296 (2025)

    McMillian, E., Banerjee, A., Bueno-Orovio, A.: From 2d to 3d, deep learning-based shape reconstruction in magnetic resonance imaging: A review. arXiv preprint arXiv:2510.01296 (2025)

    work page arXiv 2025

  15. [15]

    IEEE trans- actions on pattern analysis and machine intelligence32(12), 2262–2275 (2010)

    Myronenko, A., Song, X.: Point set registration: Coherent point drift. IEEE trans- actions on pattern analysis and machine intelligence32(12), 2262–2275 (2010)

    work page 2010

  16. [16]

    In: Proceedings of the 4th international conference on Computer graphics and interactive techniques in Australasia and Southeast Asia

    Nealen, A., Igarashi, T., Sorkine, O., Alexa, M.: Laplacian mesh optimization. In: Proceedings of the 4th international conference on Computer graphics and interactive techniques in Australasia and Southeast Asia. pp. 381–389 (2006)

    work page 2006

  17. [17]

    Park, J., Metaxas, D., Axel, L.: Analysis of left ventricular wall motion based on volumetric deformable models and mri-spamm (1996)

    work page 1996

  18. [18]

    Advances in neural information processing systems30(2017)

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. Advances in neural information processing systems30(2017)

    work page 2017

  19. [19]

    IEEE journal of biomedical and health informatics22(2), 503–515 (2017)

    Suinesiaputra, A., Ablin, P., Alba, X., Alessandrini, M., Allen, J., Bai, W., Cimen, S., Claes, P., Cowan, B.R., D’hooge, J., et al.: Statistical shape modeling of the left ventricle: myocardial infarct classification challenge. IEEE journal of biomedical and health informatics22(2), 503–515 (2017)

    work page 2017

  20. [20]

    International journal for numerical methods in biomedical engineering29(1), 83–103 (2013)

    Wang, H., Gao, H., Luo, X., Berry, C., Griffith, B., Ogden, R., Wang, T.: Structure- based finite strain modelling of the human left ventricle in diastole. International journal for numerical methods in biomedical engineering29(1), 83–103 (2013)

    work page 2013

  21. [21]

    Radiology: Artificial Intelligence 5(5), e230024 (2023)

    Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: robust segmen- tation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5), e230024 (2023)

    work page 2023

  22. [22]

    In: International Workshop on Statistical Atlases and Computational Models of the Heart

    Xu, Y., Xu, H., Sinclair, M., Puyol-Antón, E., Niederer, S.A., Chiribiri, A., Williams, S.E., Williams, M.C., Young, A.A.: Improved 3d whole heart geome- try from sparse cmr slices. In: International Workshop on Statistical Atlases and Computational Models of the Heart. pp. 43–52. Springer (2024)

    work page 2024

  23. [23]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Ye, M., Yang, D., Kanski, M., Axel, L., Metaxas, D.: Neural deformable models for 3d bi-ventricular heart shape reconstruction and modeling from 2d sparse car- diac magnetic resonance imaging. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14247–14256 (2023)

    work page 2023

  24. [24]

    Radiology216(2), 597–602 (2000)

    Young, A.A., Cowan, B.R., Thrupp, S.F., Hedley, W.J., Dell’Italia, L.J.: Left ven- tricular mass and volume: fast calculation with guide-point modeling on mr images. Radiology216(2), 597–602 (2000)

    work page 2000

  25. [25]

    Computerized Medical Imaging and Graphics109, 102287 (2023)

    Zeng,A.,Wu,C.,Lin,G.,Xie,W.,Hong,J.,Huang,M.,Zhuang,J.,Bi,S.,Pan,D., Ullah,N.,etal.:Imagecas:Alarge-scaledatasetandbenchmarkforcoronaryartery segmentation based on computed tomography angiography images. Computerized Medical Imaging and Graphics109, 102287 (2023)

    work page 2023

  26. [26]

    arXiv e-prints pp

    Zhang, C., Luo, Y., Wu, Y., Hwai Yap, C., Yang, G.: Topology-preserving loss for accurate and anatomically consistent cardiac mesh reconstruction. arXiv e-prints pp. arXiv–2503 (2025)

    work page 2025

  27. [27]

    In: Proceed- ings of the IEEE/CVF international conference on computer vision

    Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceed- ings of the IEEE/CVF international conference on computer vision. pp. 16259– 16268 (2021)

    work page 2021

This paper was first reviewed by glm-5.2 on July 9, 2026.