pith. sign in

arxiv: 2509.13576 · v2 · submitted 2025-09-16 · 📡 eess.IV · cs.CV

Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction for Sparse-View CT

Haodong Li , Shuo Han , Haiyang Mao , Yu Shi , Changsheng Fang , Jianjia Zhang , Weiwen Wu , Hengyong Yu This is my paper

Pith reviewed 2026-05-18 16:23 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords Sparse-View CTDiffusion PriorsIterative ReconstructionOut-of-DistributionStochastic InterpolantClassifier-Free GuidanceDomain Invariance
0
0 comments X

The pith

Training a diffusion transformer across multiple CT datasets produces domain-invariant priors that stabilize iterative reconstruction of sparse-view scans from unseen scanners or anatomies.

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

The paper introduces a framework that trains one Scalable Interpolant Transformer on several CT datasets at once. By randomly replacing conditioning labels with a null token during training, the model acquires both dataset-specific details and features that remain stable across domains. These cross-distribution priors are then inserted into a model-based iterative loop that alternates between enforcing measurement consistency and drawing samples from the learned prior. A sympathetic reader would care because fewer X-ray views could cut patient dose and speed up scans, yet current methods lose detail or introduce artifacts precisely when the scanner or body region differs from the training set.

Core claim

A Scalable Interpolant Transformer (SiT) is trained under a unified stochastic interpolant framework with classifier-free guidance applied across multiple datasets. Randomly dropping the conditioning input to a null embedding during training yields both domain-specific and domain-invariant priors. In the reconstruction phase, the transformer exploits these priors inside alternating data-fidelity and sampling steps, permitting flexible control over multi-distribution-to-noise paths and decoupled sampling that together produce higher-fidelity images under out-of-distribution conditions.

What carries the argument

Cross-distribution diffusion priors generated by the SiT backbone inside a unified stochastic interpolant framework, which supply the regularizing signal during each sampling update of the iterative reconstruction loop.

If this is right

  • Alternating data-fidelity steps with sampling steps from the cross-distribution prior reduces streak artifacts while preserving fine anatomical detail.
  • The same trained model can be used without retraining on data from different scanner protocols or patient populations.
  • Decoupled sampling paths let the reconstruction adjust the strength of the prior relative to the measurements on a per-case basis.
  • Performance gains are largest precisely in the out-of-distribution regimes that currently limit clinical deployment of sparse-view CT.

Where Pith is reading between the lines

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

  • The same null-embedding trick might transfer to other medical inverse problems that suffer from scanner or protocol shifts.
  • Because the prior is learned from a continuous interpolant, it could support reconstruction at arbitrary view counts without retraining.
  • Extending the training set to include synthetic domain shifts might further strengthen invariance without altering the architecture.

Load-bearing premise

Randomly replacing conditioning labels with a null embedding during training is enough to create priors that remain useful when the reconstruction loop is applied to data from scanners or anatomies absent from the training set.

What would settle it

Apply the trained model to sparse-view sinograms acquired on a scanner model and anatomical site never seen in any training distribution and compare the resulting artifact level and structural similarity scores against both in-distribution baselines and competing methods.

Figures

Figures reproduced from arXiv: 2509.13576 by Changsheng Fang, Haiyang Mao, Haodong Li, Hengyong Yu, Jianjia Zhang, Shuo Han, Weiwen Wu, Yu Shi.

Figure 1
Figure 1. Figure 1: Overview of the proposed CDPIR framework. Sparse-view CT inputs from multiple domains are processed by a diffusion transformer that disentangles domain-invariant features to mitigate distribution shifts and domain-specific features with residual correction to suppress ar￾tifacts, yielding robust reconstructions across heterogeneous clinical settings. Note that noisy images are highlighted with red boxes, a… view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the proposed domain-robust SVCT reconstruction framework. The framework has two components. Feature disen￾tanglement during training: A diffusion transformer, guided by classifier-free guidance, learns domain-invariant features that capture anatomical structures and domain-specific features that encode texture and noise to suppress sparse-view artifacts. Residual-guided reconstruction durin… view at source ↗
Figure 3
Figure 3. Figure 3: OOD visualization of a representative image slice in XCAT dataset. The middle row are magnified region-of-interest images, and the bottom row are difference images with respect to ground-truth. For the 55-view ODD CT reconstruction task, the result of our model is clearer than the SOTA diffusion methods, including DDS and DiffusionMBIR. The display window is [-1000,750] HU [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 4
Figure 4. Figure 4: In-distribution visualization of a representative image slice in COCA dataset. The number of views and layout are the same as in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Zero-shot reconstruction on the Clinical GE dataset. Models are trained on AAPM and tested on 123-view clinical cardiac data. Iterative methods outperform U-Net diffusion models, but all baselines show severe sparse-view artifacts. Our method suppresses artifacts and achieves the highest SSIM (+0.07 over DDS), demonstrating strong generalization under extreme sparsity [PITH_FULL_IMAGE:figures/full_fig_p00… view at source ↗
Figure 6
Figure 6. Figure 6: Zero-shot reconstruction on MARS PCCT dataset. For the 74-views CT reconstruction task, our model shows strong performance on the details of the image compared with other methods. The second row are magnified ROIs, and third row are the SSIM map between the prediction and the ground truth. The display window is [-700, 1200] HU artifacts but also preserves intricate anatomical details in soft tissue and bon… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of S-2 and B-2 reconstructions on AAPM-COCA. Yellow boxes mark ROIs; bottom-right insets show magnified views. Display window: [-1000,1000] HU. using a representative OOD testing image from the COCA￾XCAT setup. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Quantitative comparison between DDS and CDPIR with respect to different sampling steps (100 to 1000) on a representative AAPM￾COCA testing case. CDPIR maintains superior quality with fewer steps. even lower inductive bias, such as Mamba [33], which may further enhance flexibility and performance. A central advantage of CDPIR is its ability to train across heterogeneous datasets, potentially spanning hundre… view at source ↗
read the original abstract

Sparse-View CT (SVCT) reconstruction enhances temporal resolution and reduces radiation dose, yet its clinical use is hindered by artifacts due to view reduction and domain shifts from scanner, protocol, or anatomical variations, leading to performance degradation in out-of-distribution (OOD) scenarios. In this work, we propose a Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR) framework to tackle the OOD problem in SVCT. CDPIR integrates cross-distribution diffusion priors, derived from a Scalable Interpolant Transformer (SiT), with model-based iterative reconstruction methods. Specifically, we train a SiT backbone, an extension of the Diffusion Transformer (DiT) architecture, to establish a unified stochastic interpolant framework, leveraging Classifier-Free Guidance (CFG) across multiple datasets. By randomly dropping the conditioning with a null embedding during training, the model learns both domain-specific and domain-invariant priors, enhancing generalizability. During sampling, the globally sensitive transformer-based diffusion model exploits the cross-distribution prior within the unified stochastic interpolant framework, enabling flexible and stable control over multi-distribution-to-noise interpolation paths and decoupled sampling strategies, thereby improving adaptation to OOD reconstruction. By alternating between data fidelity and sampling updates, our model achieves state-of-the-art performance with superior detail preservation in SVCT reconstructions. Extensive experiments demonstrate that CDPIR significantly outperforms existing approaches, particularly under OOD conditions, highlighting its robustness and potential clinical value in challenging imaging scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces the Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR) framework for sparse-view CT (SVCT). It trains a Scalable Interpolant Transformer (SiT) on multiple datasets using Classifier-Free Guidance with random null conditioning drops to learn cross-distribution priors, which are then integrated into an iterative reconstruction process alternating between data fidelity and sampling updates. The central claim is that this achieves state-of-the-art performance with improved detail preservation and robustness to out-of-distribution (OOD) scenarios arising from scanner, protocol, or anatomical variations.

Significance. If validated, the work offers a practical way to improve generalization in diffusion-based medical image reconstruction without per-domain retraining, which is valuable for clinical SVCT where scanner and protocol shifts are common. The unified stochastic interpolant framework with decoupled sampling extends DiT-style models to inverse problems in a flexible manner. Credit is due for the explicit multi-dataset training strategy and the alternating fidelity-sampling loop, which are clearly motivated.

major comments (1)
  1. [Abstract and §3 (SiT training with CFG)] Abstract and §3 (SiT training with CFG): The central OOD robustness claim rests on the assertion that random null-embedding drops during multi-dataset training produce transferable domain-invariant priors. No ablation (e.g., with vs. without null drops) or metric (e.g., cross-domain feature alignment or OOD reconstruction error on held-out scanner protocols) is provided to show that the learned component is invariant rather than an average of dataset-specific features. Without this, the flexible multi-distribution-to-noise paths cannot be guaranteed to deliver the claimed detail preservation under anatomical or protocol shifts in the iterative loop.
minor comments (2)
  1. [Abstract] Abstract: The SOTA and OOD superiority statements would be stronger if accompanied by at least one concrete metric (e.g., mean PSNR or SSIM gain on the primary OOD test set) rather than qualitative assertions.
  2. [Method (notation)] Notation: The unified stochastic interpolant path is described conceptually but would benefit from an explicit equation defining the multi-distribution conditioning variable and its interpolation schedule.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We are grateful to the referee for the thorough review and constructive suggestions. The feedback has helped us identify areas where the manuscript can be strengthened, particularly regarding the validation of the cross-distribution priors. We address the major comment in detail below and have made corresponding revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §3 (SiT training with CFG)] Abstract and §3 (SiT training with CFG): The central OOD robustness claim rests on the assertion that random null-embedding drops during multi-dataset training produce transferable domain-invariant priors. No ablation (e.g., with vs. without null drops) or metric (e.g., cross-domain feature alignment or OOD reconstruction error on held-out scanner protocols) is provided to show that the learned component is invariant rather than an average of dataset-specific features. Without this, the flexible multi-distribution-to-noise paths cannot be guaranteed to deliver the claimed detail preservation under anatomical or protocol shifts in the iterative loop.

    Authors: We thank the referee for this valuable observation. The use of random null-embedding drops in the multi-dataset training with Classifier-Free Guidance is intended to enable the model to learn domain-invariant priors by allowing it to operate in both conditioned and unconditioned modes. This design choice is motivated by the goal of creating flexible multi-distribution-to-noise paths that generalize across domains. We agree that direct evidence through ablations would strengthen this aspect of the paper. Accordingly, in the revised manuscript, we have incorporated an ablation study that compares the performance with and without the null conditioning drops. Furthermore, we have added metrics such as cross-domain feature alignment scores and specific OOD reconstruction errors on held-out scanner protocols. These results demonstrate that the null drops contribute to learning more invariant features rather than merely averaging dataset-specific ones, supporting the improved detail preservation observed in the iterative reconstruction under OOD conditions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on learned priors from external multi-dataset training and empirical validation

full rationale

The paper trains a SiT model on multiple datasets using classifier-free guidance with random null-embedding drops to produce cross-distribution priors, then alternates these priors with data-fidelity updates in an iterative reconstruction loop. No equation, parameter, or central claim reduces by construction to a fitted input or self-definition; the domain-invariant prior is an empirical outcome of the training procedure rather than a tautological renaming of the reconstruction target. Performance and OOD claims are presented as results of extensive experiments, not derived from internal self-citation chains or ansatzes smuggled via prior author work. The framework remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the untested transferability of CFG-trained domain-invariant priors and on the stability of alternating data-fidelity and diffusion sampling steps; no explicit free parameters or new entities are quantified in the abstract.

free parameters (1)
  • conditioning drop probability
    Random null-embedding drop rate used to learn domain-invariant priors; value is a training hyperparameter not reported in abstract.
axioms (1)
  • domain assumption Classifier-free guidance on a unified stochastic interpolant produces usable domain-invariant priors for OOD SVCT
    Invoked when the paper states that random conditioning drop enables both domain-specific and domain-invariant learning.

pith-pipeline@v0.9.0 · 5812 in / 1243 out tokens · 47280 ms · 2026-05-18T16:23:07.772293+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · 3 internal anchors

  1. [1]

    Photon-counting ct: technical principles and clinical prospects,

    M. J. Willemink, M. Persson, A. Pourmorteza, N. J. Pelc, and D. Fleischmann, “Photon-counting ct: technical principles and clinical prospects,”Radiology, vol. 289, no. 2, pp. 293–312, 2018

    work page 2018

  2. [2]

    First ct using medipix3 and the mars-ct-3 spectral scanner,

    M. Walsh, A. Opie, J. Ronaldson, R. Doesburg, S. Nik, J. Mohr, R. Ballabriga, A. Butler, and P. Butler, “First ct using medipix3 and the mars-ct-3 spectral scanner,”Journal of Instrumentation, vol. 6, no. 01, p. C01095, 2011

    work page 2011

  3. [3]

    Mars spectral molecular imaging of lamb tissue: data collection and image analysis,

    R. Aamir, A. Chernoglazov, C. J. Bateman, A. P. Butler, P. H. Butler, N. G. Anderson, S. Bell, R. Panta, J. Healy, J. Mohret al., “Mars spectral molecular imaging of lamb tissue: data collection and image analysis,” Journal of Instrumentation, vol. 9, no. 02, p. P02005, 2014

    work page 2014

  4. [4]

    Deep few-view high- resolution photon-counting extremity ct at halved dose for a clinical trial,

    M. Li, C. Niu, G. Wang, M. R. Amma, K. M. Chapagain, S. Gabrielson, A. Li, K. Jonker, N. de Ruiter, J. A. Clarket al., “Deep few-view high- resolution photon-counting extremity ct at halved dose for a clinical trial,”ArXiv, pp. arXiv–2403, 2024

    work page 2024

  5. [5]

    Dictionary learning,

    I. Toši ´c and P. Frossard, “Dictionary learning,”IEEE Signal Processing Magazine, vol. 28, no. 2, pp. 27–38, 2011

    work page 2011

  6. [6]

    Non-local means denoising,

    A. Buades, B. Coll, and J.-M. Morel, “Non-local means denoising,” Image Processing On Line, vol. 1, pp. 208–212, 2011

    work page 2011

  7. [7]

    Deep con- volutional neural network for inverse problems in imaging,

    K. H. Jin, M. T. McCann, E. Froustey, and M. Unser, “Deep con- volutional neural network for inverse problems in imaging,”IEEE transactions on image processing, vol. 26, no. 9, pp. 4509–4522, 2017

    work page 2017

  8. [8]

    Low-dose ct with a residual encoder-decoder convolutional neural network,

    H. Chen, Y . Zhang, M. K. Kalra, F. Lin, Y . Chen, P. Liao, J. Zhou, and G. Wang, “Low-dose ct with a residual encoder-decoder convolutional neural network,”IEEE transactions on medical imaging, vol. 36, no. 12, pp. 2524–2535, 2017

    work page 2017

  9. [9]

    A transformer-based iterative reconstruction model for sparse-view ct re- construction,

    W. Xia, Z. Yang, Q. Zhou, Z. Lu, Z. Wang, and Y . Zhang, “A transformer-based iterative reconstruction model for sparse-view ct re- construction,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2022, pp. 790–800

    work page 2022

  10. [10]

    Multi-domain integrative swin transformer network for sparse-view tomographic reconstruction,

    J. Pan, H. Zhang, W. Wu, Z. Gao, and W. Wu, “Multi-domain integrative swin transformer network for sparse-view tomographic reconstruction,” Patterns, vol. 3, no. 6, 2022

    work page 2022

  11. [11]

    Td-strans: Tri-domain sparse-view ct reconstruction based on sparse transformer,

    Y . Li, X. Sun, S. Wang, L. Guo, Y . Qin, J. Pan, and P. Chen, “Td-strans: Tri-domain sparse-view ct reconstruction based on sparse transformer,” Computer Methods and Programs in Biomedicine, vol. 260, p. 108575, 2025

    work page 2025

  12. [12]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inCVPR, 2022

    work page 2022

  13. [13]

    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

    work page 2020

  14. [14]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,”Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  15. [15]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,”arXiv preprint arXiv:2011.13456, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  16. [16]

    Improving diffusion inverse problem solving with decoupled noise an- nealing,

    B. Zhang, W. Chu, J. Berner, C. Meng, A. Anandkumar, and Y . Song, “Improving diffusion inverse problem solving with decoupled noise an- nealing,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 20 895–20 905

    work page 2025

  17. [17]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 22 542–22 551

    work page 2023

  18. [18]

    Decomposed diffusion sampler for accelerating large-scale inverse problems

    H. Chung, S. Lee, and J. C. Ye, “Decomposed diffusion sam- pler for accelerating large-scale inverse problems,”arXiv preprint arXiv:2303.05754, 2023

    work page arXiv 2023

  19. [19]

    Low-dose ct for the detection and classification of metastatic liver lesions: results of the 2016 low dose ct grand challenge,

    C. H. McCollough, A. C. Bartley, R. E. Carter, B. Chenet al., “Low-dose ct for the detection and classification of metastatic liver lesions: results of the 2016 low dose ct grand challenge,”Medical physics, vol. 44, no. 10, pp. e339–e352, 2017

    work page 2016

  20. [20]

    Coca-coronary calcium and chest ct’s dataset,

    S. AIMI, “Coca-coronary calcium and chest ct’s dataset,”URL: https://stanfordaimi.azurewebsites.net/datasets/e8ca74dc-8dd4-4340- 815a-60b41f6cb2aa, 2022

    work page 2022

  21. [21]

    Steerable conditional diffusion for out-of- distribution adaptation in medical image reconstruction,

    R. Barbano, A. Denker, H. Chung, T. H. Roh, S. Arridge, P. Maass, B. Jin, and J. C. Ye, “Steerable conditional diffusion for out-of- distribution adaptation in medical image reconstruction,”IEEE Trans- actions on Medical Imaging, 2025

    work page 2025

  22. [22]

    Classifier-Free Diffusion Guidance

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” 2022. [Online]. Available: https://arxiv.org/abs/2207.12598

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  23. [23]

    Albergo, Nicholas M

    N. Ma, M. Goldstein, M. S. Albergo, N. M. Boffi, E. Vanden- Eijnden, and S. Xie, “Sit: Exploring flow and diffusion-based gen- erative models with scalable interpolant transformers,”arXiv preprint arXiv:2401.08740, 2024

    work page arXiv 2024

  24. [24]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” pp. 4195–4205, 2023

    work page 2023

  25. [25]

    Theory of ordinary differential equations,

    E. A. Coddington, N. Levinson, and T. Teichmann, “Theory of ordinary differential equations,” 1956

    work page 1956

  26. [26]

    Improved total variation regularized image reconstruction (itv) applied to clinical ct data,

    L. Ritschl and M. Kachelriess, “Improved total variation regularized image reconstruction (itv) applied to clinical ct data,” inMedical H LI, S HAN,et al.: SPARSE-VIEW DATA RECONSTRUCTION 11 Imaging 2011: Physics of Medical Imaging, vol. 7961. SPIE, 2011, pp. 786–798

    work page 2011

  27. [27]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    work page 2021

  28. [28]

    Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization,

    E. Y . Sidky and X. Pan, “Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization,” Physics in Medicine & Biology, vol. 53, no. 17, p. 4777, 2008

    work page 2008

  29. [29]

    Differentiable forward projector for x-ray computed tomography,

    H. Kim and K. Champley, “Differentiable forward projector for x-ray computed tomography,” 2023. [Online]. Available: https://arxiv.org/abs/2307.05801

    work page arXiv 2023

  30. [30]

    4d xcat phantom for multimodality imaging research,

    W. P. Segarset al., “4d xcat phantom for multimodality imaging research,”Medical physics, vol. 37, no. 9, pp. 4902–4915, 2010

    work page 2010

  31. [31]

    Improving diffusion models for inverse problems using manifold constraints,

    H. Chung, B. Sim, D. Ryu, and J. C. Ye, “Improving diffusion models for inverse problems using manifold constraints,”Advances in Neural Information Processing Systems, vol. 35, pp. 25 683–25 696, 2022

    work page 2022

  32. [32]

    Openood v1

    J. Zhang, J. Yang, P. Wang, H. Wang, Y . Lin, H. Zhang, Y . Sun, X. Du, Y . Li, Z. Liu, Y . Chen, and H. Li, “Openood v1.5: Enhanced benchmark for out-of-distribution detection,”arXiv preprint arXiv:2306.09301, 2023

    work page arXiv 2023

  33. [33]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  34. [34]

    Medical multimodal multitask foundation model for lung cancer screening,

    C. Niu, Q. Lyu, C. D. Carothers, P. Kaviani, J. Tan, P. Yan, M. K. Kalra, C. T. Whitlow, and G. Wang, “Medical multimodal multitask foundation model for lung cancer screening,”Nature Communications, vol. 16, no. 1, p. 1523, 2025

    work page 2025