Pith. sign in

REVIEW 3 major objections 6 minor 56 references

A single neural field fuses all heart MR views into one 3D+time volume.

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 →

T0 review · deepseek-v4-flash

2026-08-05 00:19 UTC pith:TZJ3CBJ7

load-bearing objection NISF++ is a real step forward for multi-view cardiac implicit fields, but the motion-correction and 'better than ground truth' claims need harder evidence. the 3 major comments →

arxiv 2608.00752 v1 pith:TZJ3CBJ7 submitted 2026-08-01 cs.CV

NISF++: Geometrically-grounded implicit representations of 3D+time cardiac function from 2D short- and long-axis MR views

classification cs.CV
keywords cardiac magnetic resonance imagingneural implicit functions3D+time segmentationmotion correctionsuper-resolution reconstructionmulti-view consistencyimplicit neural representations
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.

Standard cardiac MRI arrives as 2D cross-sectional slices in different orientations, and conventional deep networks process each slice independently, so their segmentations need not agree in 3D. This paper claims that one neural implicit function can learn a shared 3D+time representation of a subject's heart from every available short-axis and long-axis slice, regardless of orientation, making any imaging plane a cross-section of a single volume. Because all slices are explained by one field, intensity and segmentation predictions are spatially and temporally consistent, and the field can be sampled at any resolution. The paper further claims that learnable per-slice rotations and translations can correct inter-slice respiratory and patient motion, and that physics-informed point-spread-function supervision yields artifact-free super-resolution. If right, this replaces separate per-slice segmentation pipelines with a view-coherent 4D cardiac model built directly from standard clinical acquisitions.

Core claim

Central claim: one neural implicit function f_psi(x,t,h) maps a 4D spatiotemporal coordinate, plus a subject-specific latent vector h, to two continuous fields — intensity and four-class segmentation probabilities. All available 2D+time CMR slices are placed in a standardized world coordinate system and treated as a point cloud of voxels; cohort-wide training lets the network fill inter-slice regions from population patterns. During inference, a new subject's latent code and per-slice alignment parameters are optimized against image intensities only, yielding a shared 3D+time volume whose cross-sections along any plane reproduce the observed anatomy. The same field can be sampled at arbitrar

What carries the argument

The central object is a neural implicit field f_psi(x,t,h): a point-wise network that outputs field values at arbitrary continuous coordinates, implemented as a shared MLP body with two heads, one for intensity and one for segmentation. A sinusoidal positional encoder maps [x,y,z,sin(pi t),cos(pi t)], with time encoded cyclically to represent one heartbeat. Each subject is summarized by a learnable 128-dimensional conditioning vector h, following the DeepSDF-style conditioning design, while per-slice learnable rotation delta-theta and translation delta-tau parameters correct rigid motion in a standardized world coordinate system. A learnable per-slice intensity scaling factor reconciles inte

Load-bearing premise

The load-bearing premise is that each slice's misalignment can be captured by one learnable rigid rotation and translation, and that the network's learned population prior is accurate enough to separate true anatomy from motion-induced shifts — the paper measures alignment only through slice-intersection consistency, not against known motion.

What would settle it

Simulate known rigid motion on real cardiac MR slices by applying known rotations and translations to a subject's slice set, then run the inference-time optimization; the motion-correction claim stands only if the estimated per-slice transforms recover the applied motion to within a fraction of slice thickness, rather than merely improving intersection consistency.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Because every slice is a cross-section of one shared volume, predictions along short-axis and long-axis views cannot contradict each other; view consistency is a design property rather than a post-processing step.
  • The learned field can be sampled at arbitrary resolution at any cardiac phase, so super-resolution intensity volumes and smooth ventricular meshes can be generated directly from standard thick-slice clinical acquisitions.
  • Representations can be built from incomplete view sets: a subject's anatomy can be reconstructed from only short-axis slices or only long-axis slices, with held-out planes sampled afterwards.
  • Optimizing per-slice rotation and translation during inference improves slice-intersection consistency, suggesting that rigid inter-slice motion can be corrected without explicit atlases or a dedicated slice-intersection loss.
  • Structures that appear only partially in individual slices, such as papillary muscles and ventricular valves, emerge as continuous structures in the reconstructed 4D field.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: if the shared-representation claim holds under scrutiny, the same architecture could serve as a registration-free 4D cardiac atlas, since the latent code already separates subject-specific anatomy from a population prior.
  • Inference: because motion parameters are learned only against reconstruction fidelity and a population prior, a direct test with simulated known rigid motion would clarify how much of the reported alignment gain is attributable to geometry rather than prior strength.
  • Inference: the PSF-aware intensity model suggests a route to isotropic high-resolution cardiac volumes from clinical stacks; a natural extension would be to estimate slice-dependent PSF widths jointly rather than fixing them from resolution metadata.
  • Inference: the observation that long-axis-derived representations track through-plane apical motion more correctly than the 2D CNN labels implies the shared 3D+time field could itself generate cleaner training labels for slice-based segmentation networks.

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 / 6 minor

Summary. The paper proposes NISF++, a neural implicit function f_psi(x, t, h) that maps standardized world coordinates (x, y, z, t) and a subject-specific latent code h to intensity and segmentation fields. The method trains on all available short-axis (SA) and long-axis (LA) 2D+time CMR slices, treats slice coordinates as a point cloud in shared 3D+time space, and adds learnable per-slice rigid motion offsets, per-slice intensity scaling, and a PSF-based Monte-Carlo supervision loss. Experiments on 120 UK-Biobank subjects report in-plane segmentation performance described as on par with prior work (Table 1), improved slice-intersection consistency after motion correction (Table 2), 4D super-resolution, and reconstruction of held-out views (Appendix B). The paper's central claims are that the architecture yields view-consistent 3D+time representations, corrects inter-slice motion, and produces anatomically more accurate predictions than its own ground-truth labels.

Significance. If the central claims hold, the paper would be a meaningful step toward implicit, geometry-grounded models of cardiac function that natively integrate all CMR acquisition orientations, support super-resolution, and perform motion correction without an external atlas. Strengths of the work include the clear formulation of the shared implicit field, the physics-informed PSF supervision, the use of both SA and LA views, and the public code repository. However, the two most novel claims are currently supported by indirect evidence: motion correction is measured only by self-consistency along slice intersections, and the claim that the model exceeds its own training labels anatomically rests on qualitative inspection. The evaluation design is also entangled with the method's assumptions, as the automatic Bai et al. 2018 segmentations are used both as training targets and as the reference standard for segmentation quality.

major comments (3)
  1. [§3.2 Motion correction; Table 2] The central claim that the learnable offsets δθ, δτ correct real patient/respiratory motion is not established. The only quantitative evidence is improved intersection agreement (NCC 0.521→0.706; Dice 0.756→0.862). This is a self-consistency metric, not an accuracy metric. Because the optimization simultaneously updates the subject latent h and per-slice intensity scales d, intersection consistency can be improved by reshaping the latent field or rescaling intensities rather than by recovering true slice poses. There is no validation against known or independently measured transformations (e.g., synthetic misalignment with known rigid transforms) and no comparison to an external registration method. The Discussion's caveat that ground-truth segmentations 'never perfectly align in 3D even if we were to assume perfect alignment' acknowledges label inconsistency but does not justify that th
  2. [Table 1; Appendix A] The segmentation evaluation is entangled with the method's own assumptions. The claim of being 'on-par with existing CMR segmentation methods' is made without a direct baseline comparison; no table reports Dice for Bai et al. 2018 or any other method on the same subjects. The ground truth for SA slices is the automatic Bai et al. output, which the paper itself argues is frequently erroneous in apical regions; using these labels both as training targets and as the reference standard for evaluating the model's anatomical accuracy is circular. The statement that the model produces predictions 'with better anatomical accuracy than its original training data' is supported only by qualitative figures in Appendix A, without blinded expert assessment or an independent reference standard. Given the low apical Dice values (0.69 LV BP, 0.58 LV Myo, 0.70 RV BP in Table 1), the current evidence does
  3. [§3.2 Motion correction, Eq. (rigid transform); §3.3 Eq. (loss)] The learnable motion-correction component is underdetermined in the current experimental design. The network has a per-subject latent h, per-slice intensity scale d, and per-slice rigid offsets; all are optimized simultaneously to minimize the same reconstruction objective. Without a known-pose ground truth or an ablation isolating δθ, δτ from the other parameters, improved slice-intersection metrics could be attributed to a combination of latent-code reshaping, intensity rescaling, and regularization rather than to accurate pose recovery. A minimal fix is to add a synthetic experiment where slices are displaced by known rigid transforms and quantify the recovery error of δθ, δτ. This directly tests the rigid assumption stated in §3.2.
minor comments (6)
  1. [§2 Background] Typo: 'to asses radial myocardial motion' should be 'to assess'.
  2. [§3.2 Motion correction] The definition of R(θ+δθ, λ) is ambiguous; clarify how voxel spacings λ enter the rotation matrix and define the coordinate spaces x_image, x_world explicitly.
  3. [Table 2] Report paired statistics and the number of slice-intersection pairs, not just subject-level means; with only 10 test subjects, confidence in the before/after difference is limited.
  4. [Appendix C] The statement in §3.3 that 'we empirically find no difference over using PSF sampling' at inference is hard to reconcile with Appendix C's claim that PSF provides a 'crucial form of regularization'. Clarify whether the PSF is used only during training or also during inference.
  5. [§3.1 Data overview] Manual LA segmentation details are missing: number of annotators, inter-observer variability, and how the three annotated time points were selected.
  6. [Figure 2a] Specify whether the reconstruction/segmentation curves are computed on training, validation, or test subjects and add error bars or confidence intervals.

Circularity Check

1 steps flagged

Motion-correction evaluation is a self-consistency check: the fitted per-slice transforms are optimized against the same reconstruction loss that implicitly measures intersection agreement.

specific steps
  1. fitted input called prediction [Section 3.2 'Motion correction' and Section 4, Table 2]
    "The overall objective for a given voxel can be summarized as follows: Ltotal(ψ,h_n, δθ_nj, δτ_nj, d_nj) = L_PSNR(î_v, i_v) + L_Dice(ŝ_v, s_v) + ... Because of the limited representational capacity of both the network fψ and the representation space H, motion-derived shape deviations ... can be resolved in an unsupervised manner by backpropagating gradients through the network. Table 2: Differences in ground-truth intensities and segmentations along slice intersection before and after optimization of rotation and translation parameters in test set."

    The per-slice rotation δθ and translation δτ are optimized by minimizing L_total, which includes per-voxel intensity and segmentation losses. For any world point on a slice intersection, the same network output î_v is supervised against observations from both slices; thus the loss directly penalizes inter-slice mismatch. Table 2 measures exactly this mismatch (NCC/Dice along intersections) before and after the optimization. The improvement is therefore an expected consequence of reducing L_total, not an independent validation against known poses. The paper's own caveat that ground-truth segmentations 'never perfectly align in 3D even if we were to assume perfect alignment' confirms that intersection consistency is not a gold standard for true alignment. The 'motion-correction' claim is thu

full rationale

The core view-consistency property is self-contained by construction: the architecture defines all plane predictions as cross-sections of one shared 3D+time field, and held-out view reconstruction (Appendix B) is a genuine external generalization test. The segmentation and super-resolution results are empirical and not circular. However, the motion-correction claim is supported only by Table 2, which reports improved intensity NCC and segmentation Dice along slice intersections. Because the shared-field loss supervises the same world point from multiple slices, optimizing the learnable rigid transforms to improve reconstruction fidelity necessarily reduces these intersection discrepancies. The paper does not provide any ground-truth motion benchmark or independent pose measurement; the evaluation metric is effectively a component of the optimized objective. This is a fitted-input-called-prediction circularity: the parameters are fit to a self-consistency target, and that same target is then reported as evidence of successful motion correction. This does not invalidate the other contributions, so the overall circularity is partial rather than total.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

This is an empirical deep learning paper, so its ledger is dominated by fitted parameters and domain assumptions rather than invented entities. The central claim depends on three free parameter families per subject (latent code, slice motion offsets, intensity scalings) plus a PSF model and several hyperparameters. The assumptions that matter most are the rigid-motion simplification and the reliability of an implicit population prior as an anatomical anchor. These are reasonable but only weakly validated.

free parameters (5)
  • subject latent vector h_n (N x 128) = learned per subject, R128, initialized N(0,1e-4)
    Captures subject-specific anatomical deviations from the population prior; it is fit per subject and is central to modeling.
  • per-slice rotation offset delta_theta and translation delta_tau = learned per slice, L2-regularized with alpha 1e-4
    Motion correction parameters; they are fit to intensities and are the mechanism claimed to correct motion.
  • per-slice intensity scaling d_nj = learned per slice, L2-regularized with alpha 1e-2
    Accounts for inter-slice intensity variation; fit to intensities.
  • PSF Gaussian covariance sigma = kappa=2.355, diagonal entries from in-plane resolution and slice thickness
    A modeling choice from prior literature (Xu et al.) that determines how the intensity field is integrated; it is not fit to data but is a manually chosen assumption.
  • architecture and weighting hyperparameters = alpha_weights=1e-5, alpha_latent=1e-4, alpha_rot=1e-4, alpha_tra=1e-4, alpha_intens=1e-2, hidden 256, 16 residual blocks
    Chosen by hand; these shape the optimization and the representation.
axioms (5)
  • domain assumption A single rigid transform per slice can account for inter-slice patient and respiratory motion.
    Invoked in Section 3.2 Motion correction; the paper itself calls this a rigid assumption, and it fails for non-rigid or through-plane cardiac deformation.
  • domain assumption The population prior learned by f_psi in standardized world coordinates is a reliable anatomical anchor, strong enough to disambiguate motion from true anatomy.
    Crucial for unsupervised motion correction and for the claim that the model produces more accurate apical segmentations than its training labels; barely tested quantitatively.
  • ad hoc to paper The Bai et al. 2018 CNN segmentations, used as training ground truth, are errors that the implicit model can average away rather than encode.
    The paper relies on these synthetic labels for supervision yet also claims they are substantially wrong in apical regions (Appendix A); this is an unresolved tension.
  • domain assumption Voxel intensity can be modeled as a spatial average of a continuous field under an anisotropic Gaussian PSF, with K=16 Monte Carlo samples sufficient.
    Used in Section 3.2 Physics-informed intensity modeling; standard in the literature but an approximation.
  • domain assumption The standardized world coordinate system, built from DICOM header info and mid-ventricular landmarks, provides enough cross-subject alignment for a shared prior to generalize.
    Section 3.2 Standardized world coordinates; alignment errors propagate into the learned anatomy prior.
invented entities (2)
  • Shared 3D+time implicit intensity/segmentation field per subject independent evidence
    purpose: Represents cardiac anatomy and function continuously; enables cross-view consistent predictions and super-resolution.
    The model itself is the contribution, not a new physical entity. Its output is testable against withheld views and segmentations, as in Appendix B.
  • Per-slice rigid motion offsets (delta_theta, delta_tau) independent evidence
    purpose: Correct inter-slice motion without external registration.
    Falsifiable through slice-intersection consistency, though shown only indirectly in Table 2.

pith-pipeline@v1.3.0-alltime-deepseek · 12117 in / 8674 out tokens · 75312 ms · 2026-08-05T00:19:36.042059+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of NISF++: Geometrically-grounded implicit representations of 3D+time cardiac function from 2D short- and long-axis MR views." pith.science (2026). https://pith.science/paper/TZJ3CBJ7

@misc{pith2026260800752,
  author       = {Pith},
  title        = {Pith review of: NISF++: Geometrically-grounded implicit representations of 3D+time cardiac function from 2D short- and long-axis MR views},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TZJ3CBJ7}},
  note         = {Machine review of arXiv:2608.00752}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Clinical acquisition in cardiac magnetic resonance (CMR) imaging involves obtaining cross-sectional planes of the heart along the radial and longitudinal directions. Despite these planes being 2D cross-sectional images of the heart, radiologists understand the 3D spatial and continuous temporal nature of the organ being imaged. The same can not be said about the conventional deep learning architectures used to process CMR images, which rely on in-plane and grid-based operations, and are hence unable to organically integrate information from all imaging planes. This paper builds upon previous work on neural implicit segmentation functions (NISF) to overcome unaddressed challenges in cardiac function modeling in the CMR domain. For a given subject, our architecture builds a shared 3D+time representations from all available acquisition planes regardless of orientation. By design, predictions along any imaging plane orientation are cross-sections of the same 3D representation, leading to spatio-temporal consistency across all slices. Moreover, our architecture makes the rotation and translation parameters of imaging planes learnable, allowing us to correct for the commonplace respiratory and patient motion between slice acquisitions under a rigid assumption. Furthermore, interpolation of intensities and segmentation can be performed in 4D at any desired resolution. We perform our study on a 120 subject sub-cohort of CMR imaging data from the UK-Biobank. Our in-plane segmentation performance is on-par with existing CMR segmentation methods and explore how the majority of failure cases arise from limitations in the ground-truth segmentation, for which our representations make predictions with better anatomical accuracy than its original training data. We also evaluate our motion-correction capabilities, displaying quantitative and qualitative improvements in slice alignment.

Figures

Figures reproduced from arXiv: 2608.00752 by Daniel Rueckert, Julian McGinnis, Maik Dannecker, Nil Stolt-Ans\'o, Steven Jia.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework: a) Images are transformed to a standard￾ized world-space. The NISF model performs rigid-body motion correction to account for patient and/or respiratory motion. b) Voxel coordinates are encoded via sinusoidal features and concatenated with (learnable) subject representation vector. c) The decoder predicts intensity and segmentation values for each input voxel. d) Intensi… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Segmentation and reconstruction metrics during inference-time optimization on image intensities. Segmentation metrics plateau soon after the initial rough anatomy is modeled. (b) Motion between acquisitions causes the anatomy in slices to not correctly overlap in scanner space. Ground-truth slice segmenta￾tions before alignment clearly show errors in overlap. After optimizing rotation and translation p… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Super-resolution intensity volume. For visualization purposes, highest inten￾sities have been made transparent, allowing us to look into the ventricular and atrial cavities. Despite sub-structures such as papillary muscles and ventricular valves being only partially visible at any given image, our representation ap￾pears to understand their the continuous nature. (b) Meshes of super-resolution segmenta… view at source ↗
Figure 4
Figure 4. Figure 4: Motion correction results displaying intensities along the intersection of two slices. Red line indicates intersection of slices prior to optimization, while the green line is the post-optimization intersection. Sampled intensities along both lines are displayed in the vertical section between each pair, showing how anatomical landmarks become spatially consistent across slice pairs. Letters A-E indicate l… view at source ↗
Figure 5
Figure 5. Figure 5: Top rows displays our synthetic ground-truth training data. These are predicted by a 2D CNN, resulting in frequent segmentation errors in apical regions. The LV myocardium appears to have holes, and detached segments appear inside other class labels. The bottom row shows cross-sectional slices of our model’s 3D predicted segmentation, which portrays a more realistic shape of the ventricles despite the erro… view at source ↗
Figure 6
Figure 6. Figure 6: Top rows displays our synthetic ground-truth training data. These are predicted by a 2D CNN, resulting in frequent segmentation errors in apical regions. The LV myocardium appears to be severely under-segmented. The bottom row shows cross-sectional slices of our model’s 3D predicted segmentation, which portrays a more realistic shape of the ventricles despite the erroneous training data. 17 [PITH_FULL_IMA… view at source ↗
Figure 7
Figure 7. Figure 7: Top rows displays our synthetic ground-truth training data. These are predicted by a 2D CNN, resulting in frequent segmentation errors in apical regions. As the apex moves through the imaging plane, the LV myocardium segmentation is spatially inconsistent across time. As the apex crosses the imaging plane again, the LV myocardium is undersegmented, failing to enclose the LV blood-pool. The bottom row shows… view at source ↗
Figure 8
Figure 8. Figure 8: The top row shows the original held-out ground-truth LA images. The second row displays the reconstructed LA images from the representation derived from optimizing on SA images. The third row shows the ground-truth segmentations overlayed on the orignal LA images. The last row shows the generated LA seg￾mentations from the representation derived from the SA images. 20 [PITH_FULL_IMAGE:figures/full_fig_p02… view at source ↗
Figure 9
Figure 9. Figure 9: The top row shows the original held-out ground-truth SA images. The second row displays the reconstructed SA images from the representation derived from optimizing on LA images. The third row shows the ground-truth segmentations overlayed on the original SA images. The last row shows the generated SA seg￾mentations from the representation derived from the LA images. 21 [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 10
Figure 10. Figure 10: Segmentation of an apical slice over multiple time points. Top row is the ground￾truth segmentation. Bottom row is the segmentation generated using a LA￾derived representation, which follows accuracte through-plane motion. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: In-plane reconstruction of a LA 3-chamber view (left) and a mid-ventricular SA view (right). When no PSF is used, banding artifacts occur along slice intersections. Using PSF removes these artifacts and improves reconstruction clarity. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_11.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

56 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    International Conference on Medical Imaging with Deep Learning , pages=

    Learning shape reconstruction from sparse measurements with neural implicit functions , author=. International Conference on Medical Imaging with Deep Learning , pages=. 2022 , organization=

  2. [2]

    Patch-based Convolutional Neural Network for Whole Slide Tissue Image Classification

    Efficient multiple instance convolutional neural networks for gigapixel resolution image classification , author=. arXiv preprint arXiv:1504.07947 , volume=

  3. [3]

    2015 Fifth International Conference on Advanced Computing & Communication Technologies , pages=

    A review on the strategies and techniques of image segmentation , author=. 2015 Fifth International Conference on Advanced Computing & Communication Technologies , pages=. 2015 , organization=

  4. [4]

    Journal of Cardiovascular Magnetic Resonance , volume=

    Automated cardiovascular magnetic resonance image analysis with fully convolutional networks , author=. Journal of Cardiovascular Magnetic Resonance , volume=. 2018 , publisher=

  5. [5]

    Standardized cardiovascular magnetic resonance imaging (

    Kramer, Christopher M and Barkhausen, J. Standardized cardiovascular magnetic resonance imaging (. Journal of Cardiovascular Magnetic Resonance , volume=. 2020 , publisher=

  6. [6]

    2021 , publisher=

    Mildenhall, Ben and Srinivasan, Pratul P and Tancik, Matthew and Barron, Jonathan T and Ramamoorthi, Ravi and Ng, Ren , journal=. 2021 , publisher=

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Implicit neural representations with periodic activation functions , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    Saragadam, Vishwanath and LeJeune, Daniel and Tan, Jasper and Balakrishnan, Guha and Veeraraghavan, Ashok and Baraniuk, Richard G , journal=

  9. [9]

    Occupancy networks: Learning 3

    Mescheder, Lars and Oechsle, Michael and Niemeyer, Michael and Nowozin, Sebastian and Geiger, Andreas , booktitle=. Occupancy networks: Learning 3

  10. [10]

    Computer Vision – ECCV 2020 , volume =

    Convolutional Occupancy Networks , author =. Computer Vision – ECCV 2020 , volume =

  11. [11]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Implicit Neural Representations for Medical Imaging Segmentation , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2022 , organization=

  12. [12]

    Hypernetwork functional image representation , author=. Artificial Neural Networks and Machine Learning--ICANN 2019: Workshop and Special Sessions: 28th International Conference on Artificial Neural Networks, Munich, Germany, September 17--19, 2019, Proceedings 28 , pages=. 2019 , organization=

  13. [13]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Deepsdf: Learning continuous signed distance functions for shape representation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  14. [14]

    Kingma and Jimmy Ba , title =

    Diederik P. Kingma and Jimmy Ba , title =. 3rd International Conference on Learning Representations,

  15. [15]

    PLoS Med , volume =

    Cathie Sudlow and John Gallacher and Naomi Allen and Valerie Beral and Paul Burton and John Danesh and Paul Downey and Paul Elliott and Jane Green and Martin Landray and Bette Liu and Paul Matthews and Giok Ong and Jill Pell and Alan Silman and Alan Young and Tim Sprosen and Tim Peakman and Rory Collins , title =. PLoS Med , volume =

  16. [16]

    Data Showcase , howpublished =

  17. [17]

    Fully automatic segmentation of right and left ventricle on short-axis cardiac

    Budai, Adam and Suhai, Ferenc I and Csorba, Kristof and Toth, Attila and Szabo, Liliana and Vago, Hajnalka and Merkely, Bela , journal=. Fully automatic segmentation of right and left ventricle on short-axis cardiac. 2020 , publisher=

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    Fourier features let networks learn high frequency functions in low dimensional domains , author=. Advances in Neural Information Processing Systems , volume=

  19. [19]

    2021 , publisher=

    Gros, Charley and Lemay, Andreanne and Cohen-Adad, Julien , journal=. 2021 , publisher=

  20. [20]

    Isensee, Fabian and Jaeger, Paul F and Kohl, Simon AA and Petersen, Jens and Maier-Hein, Klaus H , journal=. nn. 2021 , publisher=

  21. [21]

    Sitzmann, Vincent and Chan, Eric and Tucker, Richard and Snavely, Noah and Wetzstein, Gordon , journal=

  22. [22]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Modulated periodic activations for generalizable local functional representations , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  23. [23]

    An Arbitrary Scale Super-Resolution Approach for 3

    Wu, Qing and Li, Yuwei and Sun, Yawen and Zhou, Yan and Wei, Hongjiang and Yu, Jingyi and Zhang, Yuyao , journal=. An Arbitrary Scale Super-Resolution Approach for 3. 2022 , publisher=

  24. [24]

    Shape completion using 3

    Dai, Angela and Ruizhongtai Qi, Charles and Nie. Shape completion using 3. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  25. [25]

    Learning 3

    Stutz, David and Geiger, Andreas , booktitle=. Learning 3

  26. [26]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  27. [27]

    Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas , booktitle=. U-net:. 2015 , organization=

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Data augmentation using learned transformations for one-shot medical image segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [29]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Attention to scale: Scale-aware semantic image segmentation , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  30. [30]

    Neural Implicit k-Space for Binning-free Non-Cartesian Cardiac

    Huang, Wenqi and Li, Hongwei and Cruz, Gastao and Pan, Jiazhen and Rueckert, Daniel and Hammernik, Kerstin , journal=. Neural Implicit k-Space for Binning-free Non-Cartesian Cardiac

  31. [31]

    International Conference on Medical Imaging with Deep Learning , pages=

    Implicit neural representations for deformable image registration , author=. International Conference on Medical Imaging with Deep Learning , pages=. 2022 , organization=

  32. [32]

    arXiv preprint arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

  33. [33]

    2023 , publisher=

    Xu, Junshen and Moyer, Daniel and Gagoski, Borjan and Iglesias, Juan Eugenio and Grant, P Ellen and Golland, Polina and Adalsteinsson, Elfar , journal=. 2023 , publisher=

  34. [34]

    Visualization in Biomedical Computing'92 , volume=

    Incremental volume reconstruction and rendering for 3-D ultrasound imaging , author=. Visualization in Biomedical Computing'92 , volume=. 1992 , organization=

  35. [35]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Stolt-Ans. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2023 , organization=

  36. [36]

    Gordaliza, Pedro and Georgantas, Costa and van Heeswijk, Ruud B and Richiardi, Jonas , booktitle=

    Banus, Jaume and Delaloye, Antoine and M. Gordaliza, Pedro and Georgantas, Costa and van Heeswijk, Ruud B and Richiardi, Jonas , booktitle=. 2025 , organization=

  37. [37]

    Geometric Deep Learning in Medical Image Analysis , pages=

    Reconstructing 3D cardiac anatomies from misaligned multi-view magnetic resonance images with mesh deformation U-Nets , author=. Geometric Deep Learning in Medical Image Analysis , pages=. 2022 , organization=

  38. [38]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Fit Pixels, Get Labels: Meta-learned Implicit Networks for Image Segmentation , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2025 , organization=

  39. [39]

    Reconstruction of fetal brain

    Kuklisova-Murgasova, Maria and Quaghebeur, Gerardine and Rutherford, Mary A and Hajnal, Joseph V and Schnabel, Julia A , journal=. Reconstruction of fetal brain. 2012 , publisher=

  40. [40]

    and Glenn, O

    Rousseau, F. and Glenn, O. and Iordanova, B. and Rodriguez-Carranza, C. and Vigneron, D. and Barkovich, J. and Studholme, C. Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2005. 2005

  41. [41]

    2019 , organization=

    Chen, Chen and Biffi, Carlo and Tarroni, Giacomo and Petersen, Steffen and Bai, Wenjia and Rueckert, Daniel , booktitle=. 2019 , organization=

  42. [42]

    2024 , organization=

    Zhang, Yundi and Chen, Chen and Shit, Suprosanna and Starck, Sophie and Rueckert, Daniel and Pan, Jiazhen , booktitle=. 2024 , organization=

  43. [43]

    2017 , organization=

    Patravali, Jay and Jain, Shubham and Chilamkurthy, Sasank , booktitle=. 2017 , organization=

  44. [44]

    2003 , publisher=

    Mitchell, Steven C and Bosch, Johan G and Lelieveldt, Boudewijn PF and Van der Geest, Rob J and Reiber, Johan HC and Sonka, Milan , journal=. 2003 , publisher=

  45. [45]

    2022 , organization=

    Beetz, Marcel and Banerjee, Abhirup and Grau, Vicente , booktitle=. 2022 , organization=

  46. [46]

    2023 , publisher=

    Meng, Qingjie and Bai, Wenjia and O’Regan, Declan P and Rueckert, Daniel , journal=. 2023 , publisher=

  47. [47]

    2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=

    Intensity-based 3D motion correction for cardiac MR images , author=. 2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=. 2024 , organization=

  48. [48]

    2008 , publisher=

    Chandler, Adam G and Pinder, Richard J and Netsch, Thomas and Schnabel, Julia A and Hawkes, David J and Hill, Derek LG and Razavi, Reza , journal=. 2008 , publisher=

  49. [49]

    Advances in Neural Information Processing Systems , volume=

    Learning large-scale neural fields via context pruned meta-learning , author=. Advances in Neural Information Processing Systems , volume=

  50. [50]

    Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

    Image-gs: Content-adaptive image representation via 2d gaussians , author=. Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

  51. [51]

    International Conference on Machine Learning , pages=

    Implicit Geometric Regularization for Learning Shapes , author=. International Conference on Machine Learning , pages=. 2020 , organization=

  52. [52]

    Medical Imaging with Deep Learning , year=

    Self-Supervised Synthetic Cerebral Vessel Tree Generation using Semantic Signed Distance Fields , author=. Medical Imaging with Deep Learning , year=

  53. [53]

    Computers in Biology and Medicine , volume=

    Reconstruction and completion of high-resolution 3D cardiac shapes using anisotropic CMRI segmentations and continuous implicit neural representations , author=. Computers in Biology and Medicine , volume=. 2023 , publisher=

  54. [54]

    ACM transactions on graphics (TOG) , volume=

    Instant neural graphics primitives with a multiresolution hash encoding , author=. ACM transactions on graphics (TOG) , volume=. 2022 , publisher=

  55. [55]

    arXiv preprint arXiv:2302.03130 , year=

    Spatial functa: Scaling functa to imagenet classification and generation , author=. arXiv preprint arXiv:2302.03130 , year=

  56. [56]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Film: Visual reasoning with a general conditioning layer , author=. Proceedings of the AAAI conference on artificial intelligence , volume=