Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Latent Space Consistency for Sparse-View CT Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read CLS-DM claims that adding contrastive alignment and an autoregressive constraint to latent diffusion makes one- or two-X-ray CT reconstruction outperform four baselines.

desk verdict A plausible new training recipe for sparse-view CT, but the reported evaluation is too inconsistent to support the 'outperforms' claim as written. read the letter →

arxiv 2507.11152 v1 pith:IIL6AXRV submitted 2025-07-15 eess.IV cs.AIcs.CV

classification eess.IVcs.AIcs.CV
keywords sparse-viewCTreconstructionlatentdiffusionmodelcontrastivelearningcross-modalalignmentX-raytotranslationautoregressiveconstraint3Dmedicalimaging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes CLS-DM, a three-stage latent diffusion model that reconstructs 3D CT volumes from one or two planar X-ray images. The central claim is that a plain latent diffusion model cannot align the 2D X-ray condition space with the 3D CT latent space, so the diffusion step has to do both denoising and cross-modal translation at once. CLS-DM adds a contrastive-learning stage that pulls X-ray features and CT latents of the same anatomy together, and an autoregressive reconstruction loss that keeps the X-ray encoder from losing anatomical detail. The authors report that CLS-DM outperforms CNN-, GAN-, INR-, and plain-LDM baselines on PSNR and SSIM in frontal, lateral, and two-view settings on two public CT datasets. If the comparison holds, sparse-view CT reconstruction becomes a more practical way to cut scan time and radiation dose.

What carries the argument

The load-bearing machinery is the three-stage training pipeline built around a 3D latent diffusion model. Stage one is a 3D VAE-GAN compressor that maps CT volumes $X_i \in \mathbb{R}^{1\times 128\times 128\times 128}$ to latents $Z_i^0 \in \mathbb{R}^{4\times 16\times 16\times 16}$, with Gaussian reparameterization for sampling. Stage two trains the Penetration-Aware Module and the conditional encoder: each 2D X-ray is expanded in depth, lifted to a coarse 3D reconstruction, fused across views, and compressed to a condition vector $c_i$; a contrastive loss aligns this vector with the CT latent of the same entity, and an autoregressive loss reconstructs the volume from $c_i$ to prevent feature degradation. Stage three concatenates the aligned condition $c_i$ to the noisy latent $z_t$ and trains a time-conditional UNet to predict the added noise $\epsilon_t$, yielding the LDM objective $\mathcal{L}_{\mathrm{LDM}} = \mathbb{E}_{t,z_0,\epsilon,c}[\|\epsilon_t - \hat{\epsilon}_t\|^2]$, with a fast ODE solver used at inference.

What would settle it

Re-run all four baselines under identical HU clipping, view generation, normalization, and 12-bit metric ranges; if CLS-DM's PSNR and SSIM lead disappears in either dataset's two-view setting, the claimed advantage is an artifact of protocol.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the failure of earlier latent-diffusion X-ray-to-CT models is not primarily a denoising problem but an alignment problem. The 3D perceptual compressor that maps CT volumes into latents and the conditional encoder that maps 2D X-rays into a condition vector create two spaces with no shared structure, so the UNet must compensate during training. CLS-DM solves this by inserting a contrastive alignment stage between compression and diffusion: same-entity X-ray and CT features are drawn together in latent space while different entities are pushed apart, using a temperature-scaled contrastive loss, and an autoregressive NLL loss simultaneously reconstructs a coarse 3D volume from the condition so the encoder retains fine anatomical cues. The result, as reported, is consistent gains in PSNR and SSIM over four baselines in all view configurations on both datasets, with ablations showing each added term contributes and that the dual-view setting benefits most from contrastive guidance.

Load-bearing premise

The central comparison assumes the baselines and CLS-DM were trained and evaluated under identical preprocessing, view selection, and metric ranges; if the baselines were disadvantaged by different HU clipping or evaluation ranges, the reported gaps could reflect protocol rather than method.

Editorial extensions

If this is right

  • Two-view (frontal plus lateral) inputs stand to gain the most from contrastive alignment, since the condition space is richer and the alignment has more structure to enforce.
  • The three-stage design isolates each module's training, so adapting CLS-DM to a new dataset or modality requires retraining only the conditional encoder and projection heads, not the entire diffusion stack.
  • The autoregressive constraint is what prevents contrastive alignment from washing out anatomical detail; dropping it sharply reduces SSIM and harms structural fidelity.
  • CLS-DM inherits fast sampling from the latent diffusion formulation, so the extra contrastive stage does not add inference-time cost beyond a standard conditional diffusion pass.
  • The same alignment recipe is claimed to transfer to other cross-modal generation tasks, such as text-to-image synthesis, wherever a conditional encoder's space must meet a pre-trained latent space.

Reading between the lines

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

  • Editorial extension: if latent-space alignment is the active ingredient, CLS-DM should transfer to noisy or off-angle X-rays without retraining the diffusion stage; a testable prediction is that the contrastive gain grows as the condition distribution shifts further from the CT latent distribution.
  • Editorial extension: the view-count effect visible in the ablations suggests a measurable scaling law: PSNR improvement over plain LDM should increase monotonically as more views are fused, until the coarse reconstruction saturates.
  • Editorial extension: a clinically relevant evaluation the paper does not run is speed and memory per reconstruction; because inference uses a fast solver, the method may be deployable near real time, but that needs direct measurement.
  • Editorial extension: the same contrastive-alignment stage could be applied to other inverse problems with mismatched condition spaces, such as MRI from under-sampled k-space, where a latent diffusion prior already exists.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CLS-DM, a three-stage latent diffusion model for 3D CT reconstruction from one or two planar X-rays. Stage 1 trains a 3D VAE to compress CT volumes into a latent space; Stage 2 trains an X-ray conditional encoder using a SimCLR-style contrastive loss together with an autoregressive reconstruction constraint; Stage 3 trains a time-conditional UNet in the latent space, conditioned on the contrastively aligned encoder features. The method is evaluated on LIDC-IDRI and CTSpine1K for frontal, lateral, and two-view inputs, and compared with X2CT-CNN, X2CT-GAN, INRR3CT, and an LDM backbone using PSNR and SSIM. The paper claims that CLS-DM consistently outperforms these baselines, but also concedes in Section 4.4.1 that it has lower PSNR than X2CT-CNN for LIDC lateral and 2-view inputs. The ablation study separates the contributions of the contrastive module and the autoregressive constraint.

Significance. If the reported gains survive a controlled, single-protocol comparison, the idea of adding cross-modal contrastive alignment and an autoregressive constraint to a latent diffusion model is a useful contribution to sparse-view CT reconstruction. The paper's strengths include a released codebase, a clear three-stage training decomposition, and an ablation study that isolates the two proposed components. However, the current evidence is not sufficient to support the central superiority claim: the training and evaluation protocols are not shown to be identical across methods, the headline claim is internally contradicted by Table 1, and the diffusion objective in Equation 6 is malformed. These issues are load-bearing and must be resolved before the comparative conclusions can be accepted.

major comments (4)
  1. [§4.1, §4.2, Table 1] The training and evaluation HU ranges are inconsistent: Section 4.1 clips LIDC-IDRI training volumes to [0, 2500] and CTSpine1K volumes to [-1024, 1476], while Section 4.2 computes PSNR/SSIM over [0, 4095] and [-1024, 3071]. The paper does not state whether outputs are denormalized to the full metric ranges or whether all baselines were trained and evaluated with the same clipping, normalization, 12-bit conversion, and DRR geometry. Because several Table 1 gaps are small (e.g., LIDC frontal PSNR 23.02 vs. 23.01), the reported ranking could change under a different protocol. Please specify the exact preprocessing and metric pipeline for every method, or rerun all methods under one identical protocol.
  2. [§4.4.1, Table 1] The text states that 'CLS-DM consistently outperformed the baseline models in terms of PSNR and SSIM metrics across both datasets,' but the same section reports that CLS-DM has lower PSNR than X2CT-CNN on LIDC-IDRI lateral (24.96 vs. 25.71) and 2-view (27.36 vs. 27.49) inputs. This directly contradicts the abstract's unqualified claim that CLS-DM outperforms state-of-the-art generative models on PSNR and SSIM. Please revise the claims to match the actual numbers and provide significance testing or confidence intervals for the comparisons.
  3. [§3.1.2, Eq. (6)] Equation 6, written as L_LDM = E_{t,z0,eps,c}[eps_t - eps^2_t], is not a valid noise-prediction objective. The standard latent diffusion loss is E||epsilon - epsilon_theta(z_t, t, tau_theta(y))||^2, with the network output subtracted from the noise target and a squared norm. As written, the objective is dimensionally inconsistent and cannot be optimized as stated. Please correct the equation and confirm that the implementation matches the corrected objective.
  4. [§4.3.2, §4.1] The paper does not document whether the baseline methods X2CT-CNN, X2CT-GAN, and INRR3CT were retrained under the same data splits, HU clipping, normalization, DRR parameters, and metric computation, or whether the reported numbers are taken from previous publications. Since the central claim is comparative superiority, this information is essential. Please provide a complete training and evaluation protocol for each baseline, including hyperparameters and preprocessing.
minor comments (5)
  1. [§4.1] CTSpine1K is described as 'specifically tailored for colonography,' but this dataset is for spinal vertebrae segmentation; the description should be corrected.
  2. [References] Several references are duplicated: [3] and [4] are both the SimCLR paper, [29] and [30] are both the LDM paper, [36] and [37] are both the DIFR3CT paper, and [39] and [40] are both the X2CT-GAN paper. Please deduplicate the bibliography.
  3. [§3.3, Eq. (8)] The weights in Equation 8 are introduced as eλ1 and eλ2 but the equation uses λ1 and λ2, and these symbols are also used with different meanings in Equation 4. Please use distinct notation for the two loss decompositions.
  4. [§4.4.2] The subsection title 'Qualification' should be 'Qualitative evaluation' or similar; the current title appears to be a typo.
  5. [Abstract and §1] There are typos such as 'confronted with challenged' and 'Disscusion' in the conclusion heading; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CLS-DM is a trained pipeline whose central claims rest on held-out metrics and controlled ablations, not on equations that reduce to their inputs.

full rationale

CLS-DM is an empirical, trained system rather than a formal derivation chain. The three stages (VAE compression, contrastive/autoregressive conditional-encoder training, and latent diffusion) each have explicit loss functions (Eqs. 2-8) that are optimized on training data and then evaluated on test volumes. No equation in the paper defines the predicted metric in terms of a fitted parameter, and no result is imported from the authors' prior work; the references to LDM, SimCLR, DPM-Solver, and the baselines are external methods or controlled implementations. The contrastive loss (Eq. 7) directly optimizes cross-modal similarity, so the observation that contrastive learning aligns the features is a restatement of the training objective rather than a circular derivation of a novel prediction. The LDM baseline is the paper's own 'w/o CLIP, AR' ablation (the same numbers appear in Tables 1 and 2), which is a legitimate internal control, not a circular citation. The closest concern is benchmark fairness: Section 4.1 clips training CTs to [0, 2500] for LIDC-IDRI and [-1024, 1476] for CTSpine1K, while Section 4.2 computes PSNR/SSIM over [0, 4095] and [-1024, 3071], respectively; this could disadvantage some methods, but it is a protocol and comparability issue, not circularity. The paper itself concedes in Section 4.4.1 that CLS-DM has slightly lower PSNR than X2CT-CNN on LIDC lateral and 2-view reconstructions, and Section 5 concedes that contrastive learning somewhat compromises feature representational capacity; these concessions do not indicate circular reasoning. Overall, the central performance claim is self-contained against the reported experiments, and no load-bearing step reduces by construction to its inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a trained pipeline rather than a mathematical derivation. The main hand-chosen elements are the loss weights and InfoNCE temperature (values not reported), the HU clipping windows (reported but inconsistent between training and evaluation), and the architectural choices of the penetration-aware module. No physical entities are introduced; the only invented component is the PAM neural network module.

free parameters (4)
  • lambda_1, lambda_2, lambda_3 = not reported
    Loss weighting coefficients in Equations 4 and 8; chosen by hand and values omitted from the paper.
  • InfoNCE temperature T = not reported
    Temperature in Equation 7 scales similarity scores; tuned implicitly, value not given.
  • learned uncertainty sigma and sigma_prime = learned during training, values not reported
    Gaussian NLL losses in Equations 2 and 8 treat sigma as learnable parameters.
  • CT HU clip windows = LIDC train [0,2500], CTSpine1K train [-1024,1476]; metrics [0,4095] and [-1024,3071]
    Hand-chosen preprocessing windows that define the reconstruction task and directly affect reported PSNR/SSIM values.
assumptions (5)
  • domain assumption DRR-simulated X-rays approximate clinical X-ray acquisitions.
    Section 4.1 generates training data with DRR at 0 and 90 degrees; the paper provides no validation on real X-rays.
  • domain assumption The 3D VAE latent space is a semantically coherent manifold where cosine similarity can align X-ray and CT features.
    Section 3.2 relies on this property for contrastive learning; no theoretical or empirical proof is given beyond the t-SNE illustration.
  • domain assumption Two orthogonal X-ray views (frontal and lateral) contain enough information for full 3D CT reconstruction.
    The method uses K=1 or 2 views, and clinical feasibility is asserted without measurement or clinical data.
  • ad hoc to paper Coarse reconstructions from each view can be combined by voxel-wise averaging in coronal space.
    Section 3.3; for orthogonal views, simple averaging is geometrically naive, though the network can partially compensate via learned transforms.
  • ad hoc to paper The reconstruction NLL loss can be called 'autoregressive guidance'.
    Equation 2 is a standard Gaussian reconstruction loss with learned variance, not an autoregressive model; the terminology is a re-labeling of standard practice.
invented entities (1)
  • Penetration-Aware Module (PAM)
    purpose: Maps 2D X-ray images to a 3D coarse reconstruction in voxel space before feature compression.
    Introduced in Section 3.3; there is no evidence outside this paper's experiments that the module provides a genuine 'penetration-aware' advantage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Space Consistency for Sparse-View CT Reconstruction." pith.science (2026). https://pith.science/paper/IIL6AXRV

@misc{pith2026250711152,
  author       = {Pith},
  title        = {Pith review of: Latent Space Consistency for Sparse-View CT Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IIL6AXRV}},
  note         = {Machine review of arXiv:2507.11152}
}
read the original abstract

Computed Tomography (CT) is a widely utilized imaging modality in clinical settings. Using densely acquired rotational X-ray arrays, CT can capture 3D spatial features. However, it is confronted with challenged such as significant time consumption and high radiation exposure. CT reconstruction methods based on sparse-view X-ray images have garnered substantial attention from researchers as they present a means to mitigate costs and risks. In recent years, diffusion models, particularly the Latent Diffusion Model (LDM), have demonstrated promising potential in the domain of 3D CT reconstruction. Nonetheless, due to the substantial differences between the 2D latent representation of X-ray modalities and the 3D latent representation of CT modalities, the vanilla LDM is incapable of achieving effective alignment within the latent space. To address this issue, we propose the Consistent Latent Space Diffusion Model (CLS-DM), which incorporates cross-modal feature contrastive learning to efficiently extract latent 3D information from 2D X-ray images and achieve latent space alignment between modalities. Experimental results indicate that CLS-DM outperforms classical and state-of-the-art generative models in terms of standard voxel-level metrics (PSNR, SSIM) on the LIDC-IDRI and CTSpine1K datasets. This methodology not only aids in enhancing the effectiveness and economic viability of sparse X-ray reconstructed CT but can also be generalized to other cross-modal transformation tasks, such as text-to-image synthesis. We have made our code publicly available at https://anonymous.4open.science/r/CLS-DM-50D6/ to facilitate further research and applications in other domains.

Figures

Figures reproduced from arXiv: 2507.11152 by the authors.

Figure 1
Figure 1. The diagram illustrates two logically opposite imag [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The complete workflow of the CLS-DM algorithm, encompassing both the conditional encoder (left portion) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The modality consistency optimization process [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Visual comparison between the results of various models and the Ground Truth under the LIDC-IDRI dataset [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison between the results of various [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: A detailed comparison of the results obtained from [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: The results illustrate the distribution of X-ray fea [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 23 canonical work pages

  1. [1]

    Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al . 2011. The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans. Medical physics ...

  2. [2]

    Alexandre Cafaro, Quentin Spinat, Amaury Leroy, Pauline Maury, Alexandre Munoz, Guillaume Beldjoudi, Charlotte Robert, Eric Deutsch, Vincent Grégoire, Vincent Lepetit, et al. 2023. X2Vision: 3D CT reconstruction from biplanar X-rays with deep structure prior. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springe...

  3. [4]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Interna- tional conference on machine learning . PmLR, 1597–1607

  4. [5]

    Hyungjin Chung, Dohoon Ryu, Michael T McCann, Marc L Klasky, and Jong Chul Ye. 2023. Solving 3d inverse problems using pre-trained 2d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22542–22551

  5. [6]

    Bin Dai and David Wipf. 2019. Diagnosing and enhancing VAE models. arXiv preprint arXiv:1903.05789 (2019)

  6. [7]

    Yang Deng, Ce Wang, Yuan Hui, Qian Li, Jun Li, Shiwei Luo, Mengke Sun, Quan Quan, Shuxin Yang, You Hao, et al. 2021. Ctspine1k: A large-scale dataset for spinal vertebrae segmentation in computed tomography. arXiv preprint arXiv:2105.14711 (2021)

  7. [8]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

  8. [9]

    Patrick Esser, Robin Rombach, and Bjorn Ommer. 2021. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 12873–12883

Show all 41 references
  1. [10]

    Davide Evangelista, Elena Morotti, and Elena Loli Piccolomini. 2023. RISING: A new framework for model-based few-view CT image reconstruction with deep learning. Computerized Medical Imaging and Graphics 103 (2023), 102156

  2. [11]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks. Commun. ACM 63, 11 (2020), 139–144

  3. [12]

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. 2020. Bootstrap your own latent-a new approach to self-supervised learning. Advances in ne...

  4. [13]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 9729–9738

  5. [14]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

  6. [15]

    Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. 2022. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research 23, 47 (2022), 1–33

  7. [16]

    Dianlin Hu, Jin Liu, Tianling Lv, Qianlong Zhao, Yikun Zhang, Guotao Quan, Juan Feng, Yang Chen, and Limin Luo. 2020. Hybrid-domain neural network processing for sparse-view CT reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences 5, 1 (2020), 88–98

  8. [17]

    Dongwei Jiang, Wubo Li, Miao Cao, Wei Zou, and Xiangang Li. 2020. Speech simclr: Combining contrastive and reconstruction objective for self-supervised speech representation learning. arXiv preprint arXiv:2010.13991 (2020)

  9. [18]

    Diederik P Kingma, Max Welling, et al. 2013. Auto-encoding variational bayes

  10. [19]

    Zhifeng Kong and Wei Ping. 2021. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132 (2021)

  11. [20]

    Daeun Kyung, Kyungmin Jo, Jaegul Choo, Joonseok Lee, and Edward Choi. 2023. Perspective projection-based 3d ct reconstruction from biplanar x-rays. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

  12. [21]

    S Laporte, W Skalli, JA De Guise, F Lavaste, and D Mitton. 2003. A biplanar reconstruction method based on 2D and 3D contours: application to the distal femur. Computer Methods in Biomechanics & Biomedical Engineering 6, 1 (2003), 1–6

  13. [22]

    Hoyeon Lee, Jongha Lee, Hyeongseok Kim, Byungchul Cho, and Seungryong Cho. 2018. Deep-neural-network-based sinogram synthesis for sparse-view CT image reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences 3, 2 (2018), 109–119

  14. [23]

    Suhyeon Lee, Hyungjin Chung, Minyoung Park, Jonghyuk Park, Wi-Sun Ryu, and Jong Chul Ye. 2023. Improving 3D imaging with pre-trained perpendicular 2D diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 10710–10720

  15. [24]

    Yiqun Lin, Zhongjin Luo, Wei Zhao, and Xiaomeng Li. 2023. Learning deep intensity field for extremely sparse-view cbct reconstruction. In International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 13–23

  16. [25]

    Wen-Chih Liu, Jayanth Pratap, Abhiram R Bhashyam, Neal C Chen, Quanzheng Li, and Xiang Li. 2023. Recent advances in sparse and ultra-sparse reconstruction for medical imaging. Digital Holography and Three-Dimensional Imaging (2023), HTh2C–1

  17. [26]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu

  18. [27]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning . Pmlr, 8821–8831

  19. [28]

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. 2019. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)

  20. [30]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

  21. [31]

    Robin Rombach, Patrick Esser, and Bjorn Ommer. 2020. Network-to-network translation with conditional invertible neural networks. Advances in Neural Information Processing Systems 33 (2020), 2784–2797

  22. [32]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceeding...

  23. [33]

    Julian Schön, Raghavendra Selvan, and Jens Petersen. 2022. Interpreting latent spaces of generative models for medical images using unsupervised methods. In MICCAI Workshop on Deep Generative Models . Springer, 24–33

  24. [34]

    Liyue Shen, Wei Zhao, and Lei Xing. 2019. Patient-specific reconstruction of volumetric computed tomography images from a single projection view via deep learning. Nature biomedical engineering 3, 11 (2019), 880–888

  25. [35]

    Kihyuk Sohn. 2016. Improved deep metric learning with multi-class n-pair loss objective. Advances in neural information processing systems 29 (2016)

  26. [37]

    Yiran Sun, Hana Baroudi, Tucker Netherton, Laurence Court, Osama Mawlawi, Ashok Veeraraghavan, and Guha Balakrishnan. 2024. DIFR3CT: Latent Diffusion for Probabilistic 3D CT Reconstruction from Few Planar X-Rays. arXiv preprint arXiv:2408.15118 (2024)

  27. [38]

    Yiran Sun, Tucker Netherton, Laurence Court, Ashok Veeraraghavan, and Guha Balakrishnan. 2023. Ct reconstruction from few planar x-rays with application towards low-resource radiotherapy. In International Conference on Medical Image Computing and Computer-Assisted Intervention...

  28. [40]

    Xingde Ying, Heng Guo, Kai Ma, Jian Wu, Zhengxin Weng, and Yefeng Zheng

  29. [41]

    Guangming Zang, Mohamed Aly, Ramzi Idoughi, Peter Wonka, and Wolfgang Hei- drich. 2018. Super-resolution and sparse view CT reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV) . 137–153

  30. [42]

    Han Zhang and Yuan Cao. 2024. Understanding the Benefits of SimCLR Pre-Training in Two-Layer Convolutional Neural Networks. arXiv preprint arXiv:2409.18685 (2024)

  31. [43]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    X2CT-GAN: reconstructing CT from biplanar X-rays with generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10619–10628

  32. [46]

    Zhicheng Zhang, Xiaokun Liang, Xu Dong, Yaoqin Xie, and Guohua Cao. 2018. A sparse-view CT reconstruction method based on combination of DenseNet and deconvolution. IEEE transactions on medical imaging 37, 6 (2018), 1407–1417

  33. [2022]

    Advances in Neural Information Processing Systems 35 (2022), 5775–5787

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35 (2022), 5775–5787

Pith tools

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