Pith. sign in

REVIEW 4 major objections 6 minor 48 references

A single frozen 3-D latent prior, decoded through coordinate queries, can both generate and reconstruct full-resolution CT and MRI volumes on one GPU.

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

2026-08-01 20:49 UTC pith:IYWE6EAE

load-bearing objection A genuinely useful LIIF-based 3D autoencoder for single-GPU 512^3 latent diffusion, with honest fidelity trade-offs but thin statistics and an undersupported 'continuous' claim. the 4 major comments →

arxiv 2607.16491 v1 pith:IYWE6EAE submitted 2026-07-17 physics.med-ph

Continuous 3-D Latent Diffusion for Medical Generation and Reconstruction

classification physics.med-ph
keywords latent diffusion modellocal implicit image functionimplicit neural representation3D medical image generationsparse-view CT reconstructionaccelerated MRI reconstructioncoordinate-based decodingsingle-GPU volumetric processing
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 paper tries to show that a 3-D autoencoder with a coordinate-conditioned implicit decoder can make latent diffusion practical for full clinical volumes on a single GPU. The key move is to run the expensive convolutional stage once on a compact latent grid and then render any voxel via lightweight local coordinate queries, avoiding overlapping patch decoding. The authors claim this makes a 512³ CT volume autoencode in about 9.55 seconds at 3.27 GB, roughly an order of magnitude faster than patch-based alternatives with comparable structural fidelity. They then freeze this autoencoder and train one diffusion prior on its latents, demonstrating that the same prior supports unconditional generation and, via hard data consistency, sparse-view CT and accelerated MRI reconstruction. The paper acknowledges that this loses voxel-level fidelity and trails pixel-domain reconstruction; the claim is a practical trade-off, not superiority.

Core claim

A continuous local-implicit decoder, LIIF-AE, represents a 3-D volume as a function of spatial coordinates: after a compact encoder maps the volume to a 4× downsampled latent grid, a single convolutional pass expands the latent to a feature grid, and a shared MLP evaluated at eight neighboring latent features predicts each queried voxel intensity with trilinear weights. Because the decoder is queried in memory-bounded chunks and is differentiable with respect to the latent, it can be embedded in latent-space data consistency without instantiating full-volume activations or stitching overlapping patches. The paper claims this design avoids overlapping sub-volume decoding, permits full 512³ CT

What carries the argument

LIIF-AE: a latent autoencoder whose decoder factors into a convolutional feature stage applied once on the coarse latent grid and a shared MLP implicit head that maps local feature vectors and relative coordinates to intensities, combined via normalized trilinear weights. It carries the argument by replacing dense or patched decoding with memory-bounded coordinate evaluation, keeping the composite map A∘Dψ differentiable for reconstruction.

Load-bearing premise

Training the implicit decoder with only 200,000 randomly sampled coordinate queries per iteration—less than 0.15% of the voxels in a 512³ volume—produces a continuous field smooth enough that full-volume rendering and gradient-based reconstruction inherit no hidden high-frequency artifacts.

What would settle it

Render a held-out 512³ CT volume at full resolution through the trained decoder and compare the decoded volume to the ground truth in high-frequency regions; if per-voxel error concentrates at vessel walls, bone edges, or lung fissures, or if randomly sampled query grids reveal interpolation error spikes, the smooth-field assumption fails. A quantitative continuity or interpolation-error metric over unseen coordinates would settle it.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Full 512³ CT volumes can be autoencoded on a single GPU in about 9.55 seconds at 3.27 GB peak memory, more than an order of magnitude faster than the evaluated sub-volume decoders.
  • The same frozen latent prior, without task-specific retraining, regularizes sparse-view CT and accelerated MRI reconstruction through hard data consistency.
  • Coordinate-query decoding avoids patch seams by construction, since no overlapping sub-volume predictions need blending.
  • A diffusion model trained on the compact latent inherits the autoencoder's fidelity ceiling; details not preserved at encoding cannot be recovered by better denoising.
  • Reconstruction through the differentiable decoder is less accurate than direct pixel-domain diffusion baselines, so the benefit is efficiency and practicality rather than raw fidelity.

Where Pith is reading between the lines

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

  • If the learned implicit field is smooth beyond the 200k sampled training coordinates, the same decoder could render arbitrary output grids and voxel spacings, enabling heterogeneous-resolution generation from one latent prior.
  • The runtime advantage compounds in iterative reconstruction: decoder calls multiply across denoising steps, so converting each decode from minutes to seconds may matter more than the per-call numbers suggest.
  • A testable extension is to measure interpolation error on unseen coordinates or high-frequency regions; closing this gap without dense decoding would determine whether the trade-off can be improved.
  • The reconstruction gap to pixel-domain priors suggests a possible two-stage route: use the implicit decoder for coarse anatomy and a lightweight pixel-domain refinement for data consistency, if memory allows.

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

4 major / 6 minor

Summary. This paper proposes a 3-D latent diffusion framework for CT/MRI volume generation and reconstruction. The key component is LIIF-AE, an autoencoder whose decoder is a coordinate-conditioned local implicit function: a convolutional stage maps the latent grid to a feature grid, and an MLP renders intensities at arbitrary coordinates using trilinearly weighted neighbor features. The AE is trained with a per-coordinate L1 loss and KL regularization using N=200,000 random queries per iteration. A DDPM is then trained on the deterministic encoder means of the frozen AE. Reconstruction is performed by hard data consistency within the ReSample framework: no correction at high noise, image-space CG correction and re-encoding at intermediate noise, and latent-space backpropagation through A∘Dψ at low noise. Experiments on 512^3 CT and 256^3 MRI show that LIIF-AE is 12–32× faster and lower-memory than patch-based MAISI and 3D MedDiffusion autoencoders with modest PSNR loss. Unconditional generation yields CT FID slightly worse than baselines, and reconstruction is consistently worse than DDS but better than analytic/zero-filled references.

Significance. If the claims hold, the main value is computational: a frozen 3-D latent prior that can render full volumes on one GPU and be differentiated through for inverse problems is a practical engineering contribution. The paper's strengths are that the efficiency results are concrete and consistently ordered, the authors explicitly report the fidelity cost, they provide qualitative visual evidence, and they clearly state limitations (lossy AE ceiling, no external validation, no reader studies). The machine-checked aspects are limited; the code is promised but not yet released. However, the 'continuous, seam-free' claim is not quantitatively established, and the generation/reconstruction comparisons lack statistical support. The significance is therefore conditional on filling those gaps; if they are filled, the AE design is likely to be useful for memory-constrained volumetric diffusion.

major comments (4)
  1. [§II-B2, Eq. (10)-(11)] The claim that the LIIF decoder is 'continuous' and 'seam-free' and suitable for gradient-based data consistency rests entirely on a field trained with N=200,000 random coordinates per iteration. For a 512^3 CT volume this is 0.15% of voxels per iteration. The paper provides no quantitative continuity or interpolation-error metric, and the 'no visible patch seams' evidence is qualitative. Since Algorithm 1 Stage 3 differentiates through A∘Dψ, a decoder with hidden high-frequency errors or cell-boundary jumps could bias the latent-space data consistency. Please add dense-grid interpolation error vs. sampled-coordinate count, a quantitative continuity metric across latent cell boundaries, and an ablation on N.
  2. [§III-C, Table II] FID is computed from 40 reference volumes only, with no confidence intervals or repeated-run variability. The MRI comparison is explicitly contaminated by unmatched face masking between IXI references and MAISI samples; the text nevertheless reports 'lower FID on every plane' as a result before dismissing it. The CT ranking among Ours/MAISI/3D MedDiffusion is within a range that could easily change with more reference volumes. Please report bootstrap intervals or use a larger reference set, and remove the MRI FID claims from the abstract/results if the support mismatch is acknowledged.
  3. [§III-D, Table III] Reconstruction metrics are single point estimates with no error bars, standard deviations, or number of test volumes. At CT 60 views, the SSIM gap between Ours (0.9514) and DDS (0.9548) is 0.0034; without uncertainty quantification this cannot support the claim that 'our method stays close to DDS.' Likewise, the MRI gaps may be within noise. Please report per-volume metrics, means ± std, and number of subjects, and preferably perform paired significance tests.
  4. [§II-C, Eq. (12) and §II-B3, Eq. (11)] The diffusion prior is trained on deterministic posterior means z=µϕ(x), while Eq. (3) defines a stochastic posterior and Eq. (11) includes a KL term to N(0,I). Training on means is an additional assumption about the latent distribution not stated in the problem formulation. The KL weight λ_KLD is not reported or ablated. If λ_KLD is small, the mean encodings need not be close to the Gaussian prior used at sampling; if large, AE fidelity suffers. Please state λ_KLD and show either that the mean-latent distribution matches the prior or provide a comparison against training on sampled z.
minor comments (6)
  1. [Algorithm 1, lines 10-13] The variable names x̂0 and z^dc_{0|t} are used both as optimization variables and solutions; please clarify the initialization and whether the argmin is solved exactly or by CG/stopping tolerance.
  2. [§III-D] DDS is described as combining slice-wise denoising with conjugate-gradient data consistency, but axial total variation (TV) regularization is only mentioned in Sec. IV; specify where and how TV is incorporated.
  3. [§III-B] 'Free of patch boundaries and stitching artifacts by construction' is too strong; LIIF local ensemble does not by itself guarantee C0 continuity at latent-grid cell boundaries. Rephrase to match the empirical evidence.
  4. [Eq. (5)] The composition order notation for Eϕ could be confusing; write Eϕ(x) = E^{enc}_{ϕ2}(E^{down}_{ϕ1}(x)).
  5. [§II-A] 'We used a DPS algorithm inspired by [30] and [29]' — Algorithm 1 is closer to ReSample than DPS; align terminology.
  6. [Table I] Please state the hardware and software versions used for timing and memory (GPU model, PyTorch version) so the runtime comparison is reproducible.

Circularity Check

0 steps flagged

No significant circularity; the central claims are evaluated against external benchmarks and no prediction reduces to a fitted input.

full rationale

The paper's core derivation chain is not circular. The LIIF-AE is trained with the per-iteration coordinate-sampling objective (Eq. 11) on target volumes, and its fidelity is measured on held-out volumes in Table I against MAISI and 3D MedDiffusion, which are external baselines. The latent diffusion prior is trained on the frozen AE's latents via Eq. 12 and assessed by 2.5-D FID against real unseen volumes (Table II). Reconstruction uses the frozen prior and the hard-data-consistency procedure of ReSample (Algorithm 1) with CT/MRI forward operators, and is compared against DDS and analytic baselines (Table III). None of these evaluation quantities is identical to a fitted parameter or a training objective term. The claims that the decoder is 'continuous' and produces 'no visible patch seams' are structural properties of using a coordinate-conditioned MLP without overlapping patch decoding, not predictions obtained by fitting; the sufficiency of 200,000 sampled coordinate queries per iteration is a genuine empirical robustness risk, but that is a correctness concern rather than circularity. The only self-citation is [40], used for a peripheral remark about 256^3 tractability, and it is not load-bearing for the AE, generation, or reconstruction results. No step in the derivation reduces to its own input by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper relies on standard diffusion/INR background and on hand-chosen hyperparameters (r=4, Cz=4, N=200k, λ_KL, ReSample settings) that are not all reported. The most consequential unstated premise is that a diffusion prior trained on deterministic posterior means is a valid prior for latent-space reconstruction, and that sparse-coordinate LIIF training yields a globally faithful implicit volume. No invented entities are introduced.

free parameters (6)
  • Spatial downsampling factor r = 4
    Hand-chosen; maps 512^3 CT to 128^3 latent grid and 256^3 MRI to 64^3; determines the efficiency/fidelity trade-off; no ablation reported.
  • Latent channels C_z = 4
    Hand-chosen compact latent width; no sweep reported.
  • Coordinate query budget N per AE iteration = 200,000
    Hand-chosen; supervises <0.15% of CT voxels per iteration; no ablation on N or sampling strategy.
  • KL weight λ_KLD in Eq. (11) = not reported in text
    Value deferred to source code; sets VAE regularization strength and affects reconstruction fidelity.
  • ReSample reconstruction settings (t1=750, t2=300, γ=40, 50 CG iters, κ=0.9, DC every 10 steps) = listed in Sec. II-D/III-A
    Taken from ReSample's medical-image settings rather than tuned for this latent space; all reconstruction numbers depend on them.
  • Network architecture hyperparameters (C_h, C_u, residual block counts, MLP width/depth, U-Net channels, noise schedule) = unspecified; deferred to code
    Load-bearing for runtime/memory and quality figures, but not stated in the paper; code not yet available.
axioms (6)
  • standard math Tweedie's formula and DDIM sampling correctly estimate the clean latent z0|t from noisy zt (Algorithm 1, lines 3–5).
    Standard result from diffusion literature [1], used without proof; acceptable background.
  • domain assumption ReSample's hard data-consistency and StochasticResample steps preserve the target posterior when transplanted from 2D image latents to 3D volume latents with a lossy LIIF decoder.
    The paper imports ReSample's propositions [29, Prop. 2] but provides no formal argument that the guarantees survive the 3D LIIF autoencoder setting; reconstruction quality depends on this.
  • standard math The LIIF decoder (Eq. 10) is a continuous function over Ω, so chunked coordinate queries yield a seam-free volume.
    Trilinear normalized weights and a shared MLP imply continuity across cell boundaries; the paper asserts 'no patch seams by construction' without a formal proof or quantitative seam metric.
  • ad hoc to paper A diffusion prior trained on deterministic posterior means z=µϕ(x), rather than on sampled latents from qϕ(z|x), is a valid prior for downstream generation and reconstruction.
    Sec. II-C trains on z=µϕ(x); the paper does not justify this choice or analyze its effect on Tweedie estimates in ReSample. This is a design choice specific to this paper.
  • domain assumption The forward operators (parallel-beam CT projection via TorchRadon and slice-wise Fourier undersampling for MRI) are exact models of acquisition.
    All reconstruction experiments use simulated measurements; no cone-beam, multi-coil, or realistic noise models are used, so clinical generalization is not established.
  • domain assumption 700 CT and 540 MRI volumes constitute sufficient training data for a generalizable 3D latent prior.
    The paper acknowledges the 'limited scale and diversity of the modality-specific training cohorts'; if insufficient, FID and reconstruction numbers would not transfer.

pith-pipeline@v1.3.0-alltime-deepseek · 15191 in / 21578 out tokens · 224616 ms · 2026-08-01T20:49:16.320371+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Continuous 3-D Latent Diffusion for Medical Generation and Reconstruction." pith.science (2026). https://pith.science/paper/IYWE6EAE

@misc{pith2026260716491,
  author       = {Pith},
  title        = {Pith review of: Continuous 3-D Latent Diffusion for Medical Generation and Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IYWE6EAE}},
  note         = {Machine review of arXiv:2607.16491}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

High-resolution three-dimensional (3-D) medical diffusion models remain constrained by the cost of processing full volumes, even when denoising is performed in a compact latent space. We introduce a continuous 3-D latent diffusion model (LDM) framework for computed tomography (CT) and magnetic resonance imaging (MRI) generation and measurement-guided reconstruction. Its central component is a compact autoencoder (AE) with a coordinate-conditioned local implicit image function (LIIF) decoder that represents a volume as a continuous function of spatial coordinates. By evaluating the convolutional decoder once on the latent grid and restricting repeated computation to a lightweight implicit head, the proposed design avoids overlapping sub-volume decoding while remaining differentiable for inverse-problem optimization. We evaluate the framework on CT volumes of 512^3 voxels and MRI volumes of 256^3 voxels. On high-resolution CT, the proposed AE is approximately x12-32 faster than the evaluated reference autoencoders, achieves the lowest peak graphics processing unit (GPU) memory use, and retains comparable structural fidelity despite a moderate reduction in voxel-level accuracy. The resulting frozen 3-D latent prior generates coherent full volumes without visible patch seams and can be applied, without task-specific retraining, to sparse-view CT and accelerated MRI reconstruction through hard data consistency. Although direct pixel-domain reconstruction remains more accurate, the results demonstrate that a single volumetric latent prior can support both unconditional generation and measurement-conditioned reconstruction on one GPU. Overall, the framework provides a practical trade-off between continuous volumetric decoding, computational efficiency, and fine-detail preservation. Our code will be made available at https://github.com/mellak/.

Figures

Figures reproduced from arXiv: 2607.16491 by Alexandre Bousse, Antoine De Paepe, Dimitris Visvikis, Youness Mellak.

Figure 1
Figure 1. Figure 1: Qualitative comparison of MRI autoencoding on a representative IXI volume. Rows show central sagittal, coronal, and [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of CT autoencoding on a representative held-out volume. Rows show central sagittal, coronal, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of unconditional CT volume generation. Rows show central sagittal, coronal, and axial slices. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of sparse-view CT reconstruction from 30 noisy projections. Rows show sagittal, coronal, and [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of sparse-view CT reconstruction from 60 noisy projections. Rows show sagittal, coronal, and [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of single-coil Cartesian MRI reconstruction at [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of single-coil Cartesian MRI reconstruction at [PITH_FULL_IMAGE:figures/full_fig_p012_7.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

48 extracted references · 4 canonical work pages · 2 internal anchors

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840–6851, 2020

  2. [2]

    Denoising diffusion probabilistic models for 3d medical image generation,

    F. Khader, G. Mueller-Franzes, S. Tayebi Arasteh, T. Han, C. Haar- burger, M. Schulze-Hagen, P. Schad, S. Engelhardt, B. Baessler, S. Foersch, J. Stegmaier, C. Kuhl, S. Nebelung, J. N. Kather, and D. Truhn, “Denoising diffusion probabilistic models for 3d medical image generation,” Scientific Reports , vol. 13, no. 1, p. 7303, 2023. DOI: 10.1038/s41598-02...

  3. [3]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022, pp. 10 684–10 695

  4. [4]

    Brain imaging generation with latent diffusion models,

    W. H. L. Pinaya, P.-D. Tudosiu, J. Dafflon, P. F. da Costa, V . Fernandez, P. Nachev, S. Ourselin, and M. J. Cardoso, “Brain imaging generation with latent diffusion models,” in Deep Generative Models, Springer Nature Switzerland, 2022, pp. 117–126. DOI: 10.1007/978- 3-031-18576-2 12

  5. [5]

    Memory-efficient 3d denoising diffusion models for medical image processing,

    F. Bieder, J. Wolleb, A. Durrer, R. Sandkuehler, and P. C. Cattin, “Memory-efficient 3d denoising diffusion models for medical image processing,” in Medical Imaging with Deep Learning , I. Oguz, J. Noble, X. Li, M. Styner, C. Baumgartner, M. Rusu, T. Heinmann, D. Kontos, B. Landman, and B. Dawant, Eds., ser. Proceedings of Machine Learning Research, vol. ...

  6. [6]

    Medsyn: Text- guided anatomy-aware synthesis of high-fidelity 3d ct images,

    Y . Xu, L. Sun, W. Peng, S. Jia, K. Morrison, A. Perer, A. Zandifar, S. Visweswaran, M. Eslami, and K. Batmanghelich, “Medsyn: Text- guided anatomy-aware synthesis of high-fidelity 3d ct images,” IEEE Transactions on Medical Imaging , 2024. DOI: 10.1109/TMI.2024. 3415032

  7. [7]

    Anatomically guided latent diffusion for high-resolution 3d chest ct synthesis,

    A. Oliveras, R. Mar ´ı, R. Redondo, O. Guardi `a, C. I. Ugwu, A. Tost, B. Nagarajan, C. Migliorelli, V . Ribas, and P. Radeva, “Anatomically guided latent diffusion for high-resolution 3d chest ct synthesis,” Scientific Reports, 2026. DOI: 10.1038/s41598-026-51634-4

  8. [8]

    Maisi: Medical ai for synthetic imaging,

    P. Guo, C. Zhao, D. Yang, Z. Xu, V . Nath, Y . Tang, B. Simon, M. Belue, S. Harmon, B. Turkbey, and D. Xu, “Maisi: Medical ai for synthetic imaging,” in Proceedings of the Winter Conference on Applications of Computer Vision (WACV), Feb. 2025, pp. 4430–4441

  9. [9]

    C. Zhao, P. Guo, D. Yang, Y . Tang, Y . He, B. Simon, M. Belue, S. Harmon, B. Turkbey, and D. Xu, Maisi-v2: Accelerated 3d high- resolution medical image synthesis with rectified flow and region- specific contrastive loss , 2025. arXiv: 2508.05772 [eess.IV]

  10. [10]

    3d meddiffusion: A 3d medical latent diffusion model for controllable and high-quality medical image generation,

    H. Wang, Z. Liu, K. Sun, X. Wang, D. Shen, and Z. Cui, “3d meddiffusion: A 3d medical latent diffusion model for controllable and high-quality medical image generation,” IEEE Transactions on Medical Imaging , vol. 44, no. 12, pp. 4960–4972, 2025. DOI: 10 . 1109/TMI.2025.3585372

  11. [11]

    Hierarchical amortized gan for 3d high resolution medical image synthesis,

    L. Sun, J. Chen, Y . Xu, M. Gong, K. Yu, and K. Batmanghelich, “Hierarchical amortized gan for 3d high resolution medical image synthesis,” IEEE Journal of Biomedical and Health Informatics , vol. 26, no. 8, pp. 3966–3975, 2022. DOI: 10 . 1109 / JBHI . 2022 . 3172976

  12. [12]

    Learning image priors through patch-based diffusion models for solving inverse problems,

    J. Hu, B. Song, X. Xu, L. Shen, and J. A. Fessler, “Learning image priors through patch-based diffusion models for solving inverse problems,” in Advances in Neural Information Processing Systems , vol. 37, 2024

  13. [13]

    Diffusionblend: Learning 3d image prior through position-aware diffusion score blending for 3d computed tomography reconstruction,

    B. Song, J. Hu, Z. Luo, J. A. Fessler, and L. Shen, “Diffusionblend: Learning 3d image prior through position-aware diffusion score blending for 3d computed tomography reconstruction,” in Advances in Neural Information Processing Systems , vol. 37, 2024

  14. [14]

    Nerp: Implicit neural representation learning with prior embedding for sparsely sampled image reconstruc- tion,

    L. Shen, J. Pauly, and L. Xing, “Nerp: Implicit neural representation learning with prior embedding for sparsely sampled image reconstruc- tion,” IEEE transactions on neural networks and learning systems , vol. 35, no. 1, pp. 770–782, 2022

  15. [15]

    Im- jense: Scan-specific implicit representation for joint coil sensitivity and image estimation in parallel mri,

    R. Feng, Q. Wu, J. Feng, H. She, C. Liu, Y . Zhang, and H. Wei, “Im- jense: Scan-specific implicit representation for joint coil sensitivity and image estimation in parallel mri,” IEEE Transactions on Medical Imaging, vol. 43, no. 4, pp. 1539–1553, 2023

  16. [16]

    Scan-specific unsupervised highly accelerated non-cartesian cest imaging using implicit neural repre- sentation and explicit sparse prior,

    B. Liu, H. She, and Y . P. Du, “Scan-specific unsupervised highly accelerated non-cartesian cest imaging using implicit neural repre- sentation and explicit sparse prior,” IEEE Transactions on Biomedical Engineering, vol. 71, no. 10, pp. 3032–3045, 2024

  17. [17]

    Spatiotemporal implicit neural representation for unsupervised dynamic mri reconstruction,

    J. Feng, R. Feng, Q. Wu, X. Shen, L. Chen, X. Li, L. Feng, J. Chen, Z. Zhang, C. Liu, et al., “Spatiotemporal implicit neural representation for unsupervised dynamic mri reconstruction,” IEEE Transactions on Medical Imaging, vol. 44, no. 5, pp. 2143–2156, 2025

  18. [18]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” Communications of the ACM , vol. 65, no. 1, pp. 99–106, 2021

  19. [19]

    Implicit neural representations with periodic activation functions,

    V . Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wetzstein, “Implicit neural representations with periodic activation functions,” Advances in neural information processing systems, vol. 33, pp. 7462– 7473, 2020

  20. [20]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,” ACM transactions on graphics (TOG) , vol. 41, no. 4, pp. 1–15, 2022

  21. [21]

    Bilevel optimized implicit neural representation for scan-specific accelerated mri reconstruction,

    H. Yu, J. A. Fessler, and Y . Jiang, “Bilevel optimized implicit neural representation for scan-specific accelerated mri reconstruction,” IEEE Transactions on Medical Imaging , 2026, Early access. DOI: 10.1109/ TMI.2026.3686724

  22. [22]

    Learning continuous image repre- sentation with local implicit image function,

    Y . Chen, S. Liu, and X. Wang, “Learning continuous image repre- sentation with local implicit image function,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2021, pp. 8628–8638

  23. [23]

    Arbitrary-scale image generation and upsam- pling using latent diffusion model and implicit neural decoder,

    J. Kim and T.-K. Kim, “Arbitrary-scale image generation and upsam- pling using latent diffusion model and implicit neural decoder,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2024, pp. 9202–9211

  24. [24]

    LoFi: Neural Local Fields for Scalable Image Reconstruction

    A. Khorashadizadeh, T. I. Liaudat, T. Liu, J. D. McEwen, and I. Dok- mani´c, “Lofi: Neural local fields for scalable image reconstruction,” arXiv preprint arXiv:2411.04995 , 2024. DOI: 10.48550/arXiv.2411. 04995

  25. [25]

    Diff-inr: Generative regularization for electrical impedance tomography,

    B. Tong, J. Wang, and D. Liu, “Diff-inr: Generative regularization for electrical impedance tomography,” arXiv preprint arXiv:2409.04494 ,

  26. [26]

    Highly accelerated mri via implicit neural representation guided posterior sampling of diffusion models,

    J. Chu, C. Du, X. Lin, X. Zhang, L. Wang, Y . Zhang, and H. Wei, “Highly accelerated mri via implicit neural representation guided posterior sampling of diffusion models,” Medical Image Analysis , vol. 100, p. 103 398, 2025. DOI: 10.1016/j.media.2024.103398

  27. [27]

    Medil: Implicit latent spaces for generating heterogeneous medical images at arbitrary resolutions,

    T. Spears, S. Zhu, Y . Jin, A. Shrivastava, and P. T. Fletcher, “Medil: Implicit latent spaces for generating heterogeneous medical images at arbitrary resolutions,” arXiv preprint arXiv:2504.09322 , 2025

  28. [28]

    Diffusion posterior sampling for general noisy inverse problems,

    H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye, “Diffusion posterior sampling for general noisy inverse problems,” arXiv preprint arXiv:2209.14687 , 2022

  29. [29]

    Solving inverse problems with latent diffusion models via hard data con- sistency,

    B. Song, S. M. Kwon, Z. Zhang, X. Hu, Q. Qu, and L. Shen, “Solving inverse problems with latent diffusion models via hard data con- sistency,” in International Conference on Learning Representations , vol. 2024, 2024, pp. 7624–7654

  30. [30]

    Denoising diffusion models for plug-and-play image restoration,

    Y . Zhu, K. Zhang, J. Liang, J. Cao, B. Wen, R. Timofte, and L. Van Gool, “Denoising diffusion models for plug-and-play image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 1219–1229

  31. [31]

    Prompt- tuning latent diffusion models for inverse problems,

    H. Chung, J. C. Ye, P. Milanfar, and M. Delbracio, “Prompt- tuning latent diffusion models for inverse problems,” arXiv preprint arXiv:2310.01110, 2023

  32. [32]

    The lung image database consortium (lidc) and image database resource initiative (idri): A completed reference database of lung nodules on ct scans,

    S. G. Armato, G. McLennan, et al. , “The lung image database consortium (lidc) and image database resource initiative (idri): A completed reference database of lung nodules on ct scans,” Medical Physics, vol. 38, no. 2, pp. 915–931, 2011. DOI: 10.1118/1.3528204

  33. [33]

    H. Roth, L. Lu, A. Seff, K. M. Cherry, J. Hoffman, S. Wang, J. Liu, E. B. Turkbey, and R. M. Summers, A new 2.5d representation for lymph node detection in ct , The Cancer Imaging Archive, Data set,

  34. [34]

    IXI, Ixi dataset: Information extraction from images , EPSRC GR/S21533/02, Accessed: 2026-06-21, 2006

  35. [35]

    Radimagenet: An open radiologic deep learning research dataset for effective transfer learning,

    X. Mei, Z. Liu, P. M. Robson, B. Marinelli, M. Huang, A. Doshi, A. Jacobi, C. Cao, K. E. Link, T. Yang, et al. , “Radimagenet: An open radiologic deep learning research dataset for effective transfer learning,” Radiology: Artificial Intelligence , vol. 4, no. 5, e210315, 2022

  36. [36]

    Torchradon: Fast differentiable routines for computed tomography,

    M. Ronchetti, “Torchradon: Fast differentiable routines for computed tomography,” arXiv preprint arXiv:2009.14788 , 2020. eprint: arXiv: 2009.14788

  37. [37]

    Solving 3d inverse problems using pre-trained 2d diffusion models,

    H. Chung, D. Ryu, M. T. McCann, M. L. Klasky, and J. C. Ye, “Solving 3d inverse problems using pre-trained 2d diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 542–22 551

  38. [38]

    Improving 3d imaging with pre-trained perpendicular 2d diffusion models,

    S. Lee, H. Chung, M. Park, J. Park, W.-S. Ryu, and J. C. Ye, “Improving 3d imaging with pre-trained perpendicular 2d diffusion models,” in Proceedings of the IEEE/CVF international conference on computer vision , 2023, pp. 10 710–10 720

  39. [39]

    Decomposed diffusion sampler for accelerating large-scale inverse problems,

    H. Chung, S. Lee, and J. C. Ye, “Decomposed diffusion sampler for accelerating large-scale inverse problems,” inInternational conference on learning representations , vol. 2024, 2024, pp. 38 922–38 949

  40. [40]

    Adaptive diffusion models for sparse-view motion-corrected head cone-beam ct,

    A. De Paepe, A. Bousse, C. Phung-Ngoc, Y . Mellak, and D. Visvikis, “Adaptive diffusion models for sparse-view motion-corrected head cone-beam ct,” IEEE Transactions on Radiation and Plasma Medical Sciences, 2025

  41. [41]

    Improving 2d diffusion models for 3d medical imaging with inter-slice consistent stochasticity,

    C. Du, Q. Wu, X. Tian, J. Yu, H. Wei, and Y . Zhang, “Improving 2d diffusion models for 3d medical imaging with inter-slice consistent stochasticity,” arXiv preprint arXiv:2602.04162 , 2026

  42. [42]

    Solving linear inverse problems provably via posterior sampling with latent diffusion models,

    L. Rout, N. Raoof, G. Daras, C. Caramanis, A. Dimakis, and S. Shakkottai, “Solving linear inverse problems provably via posterior sampling with latent diffusion models,” Advances in Neural Informa- tion Processing Systems , vol. 36, pp. 49 960–49 990, 2023

  43. [43]

    Image neural field diffusion models,

    Y . Chen, O. Wang, R. Zhang, E. Shechtman, X. Wang, and M. Gharbi, “Image neural field diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 8007–8017

  44. [44]

    Towards lossless implicit neural representation via bit plane decomposition,

    W. K. Han, B. Lee, H. Cho, S. Im, and K. H. Jin, “Towards lossless implicit neural representation via bit plane decomposition,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 2269–2278. 14

  45. [45]

    Sana: Efficient high-resolution image synthesis with linear diffusion transformers,

    E. Xie, J. Chen, J. Chen, H. Cai, H. Tang, Y . Lin, Z. Zhang, M. Li, L. Zhu, Y . Lu,et al., “Sana: Efficient high-resolution image synthesis with linear diffusion transformers,” arXiv preprint arXiv:2410.10629, 2024

  46. [46]

    Sana-video: Efficient video generation with block linear diffusion transformer,

    J. Chen, Y . Zhao, J. Yu, R. Chu, J. Chen, S. Yang, X. Wang, Y . Pan, D. Zhou, H. Ling, et al., “Sana-video: Efficient video generation with block linear diffusion transformer,” arXiv preprint arXiv:2509.24695, 2025

  47. [2015]

    DOI: 10.7937/K9/TCIA.2015.AQIIDCNM

  48. [2024]

    DOI: 10.48550/arXiv.2409.04494