pith. machine review for the scientific record. sign in

arxiv: 2605.00901 · v1 · submitted 2026-04-28 · 💻 cs.CV · cs.AI

Recognition: unknown

RA-CMF: Region-Adaptive Conditional MeanFlow for CT Image Reconstruction

Md Shifatul Ahsan Apurba , Md Selim , Jin Chen
Authors on Pith no claims yet

Pith reviewed 2026-05-09 20:06 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords CT image reconstructionconditional MeanFlowreinforcement learningregion-adaptive enhancementlung cancer imagingradiomic featuresimage quality metricsflow-based models
0
0 comments X

The pith

A conditional MeanFlow model controlled by reinforcement learning reconstructs CT images by adaptively refining difficult regions.

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

The paper establishes a pipeline that models CT image enhancement as conditional flow fields predicted from intermediate states, trained jointly with consistency and reconstruction losses. A reinforcement learning policy network then reads the flow rollouts to decide tile-specific refinement budgets, stopping points, and total allocation. This setup is meant to concentrate effort on low-quality or tumor areas while leaving stable regions untouched, addressing the variability in noise, contrast, and texture that arises from different scanners and protocols. Such adaptive control matters for lung cancer workflows because reliable tumor features are needed for screening, planning, and prognosis. The authors report that the combined system reaches average tumor-region radiomic concordance of 0.96 together with global PSNR of 34.23 and SSIM of 0.95.

Core claim

The central claim is that a conditional MeanFlow network, which generates image-conditioned flow fields and is optimized with a MeanFlow consistency loss plus reconstruction loss, becomes spatially adaptive when paired with an RL policy network. The policy, trained by policy gradients to maximize enhancement gains while penalizing extra computation and instability, predicts per-tile budgets and stopping criteria directly from MeanFlow states. This integration lets the model allocate refinement where it is most needed, yielding high fidelity inside tumor ROIs and improved whole-image metrics.

What carries the argument

The reinforcement-learning policy network that consumes MeanFlow rollout information to output tile-wise refinement budgets, stopping criteria, and budget allocation.

If this is right

  • Tumor radiomic features remain highly concordant (average CCC 0.96) across varied input conditions.
  • Overall image quality rises to average PSNR 34.23 and SSIM 0.95 while avoiding unnecessary processing in good regions.
  • Enhancement effort is automatically concentrated on difficult tiles and withheld from already sufficient ones.
  • The same policy-gradient training can be reused on different MeanFlow backbones without redesigning the flow model itself.

Where Pith is reading between the lines

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

  • The adaptive stopping mechanism could translate to lower average inference time on high-quality inputs, an efficiency gain the paper does not quantify.
  • Similar RL controllers might be attached to other flow or diffusion models in medical imaging to achieve region-specific refinement.
  • The tile-wise formulation suggests a natural path to multi-resolution or hierarchical variants that refine only at needed scales.
  • If the policy generalizes across body sites, the same architecture could support reconstruction tasks beyond lung CT.

Load-bearing premise

The reinforcement learning policy can reliably predict tile-wise refinement budgets and stopping criteria from MeanFlow rollouts in a way that improves quality without creating instabilities or wasting computation.

What would settle it

New CT scans from unseen scanners and protocols where the full region-adaptive method produces lower average tumor-ROI radiomic CCC or lower global PSNR than a non-adaptive conditional MeanFlow baseline would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.00901 by Jin Chen, Md Selim, Md Shifatul Ahsan Apurba.

Figure 1
Figure 1. Figure 1: Conditional MeanFlow (CMF) backbone architecture. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Region adaptive conditional MeanFlow (RA-CMF) pipeline. The CMF backbone first predicts a coarse enhancement [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CT images reconstructed using all compared methods are displayed within the intensity window of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of regions to be refined (red tile). No [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Noise Power Spectrum (NPS) analysis computed from a homogeneous region (yellow box). The input image exhibits a [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

The use of CT imaging is important for screening, diagnosis, therapy planning, and prognosis of lung cancers. Unfortunately, due to differences in imaging protocols and scanner models, CT images acquired by different means may show large differences in noise statistics, contrast, and texture. In this study, we develop a novel conditional MeanFlow pipeline for CT image reconstruction. We introduce a conditional MeanFlow network that models the enhancement trajectory by predicting image-conditioned flow fields given intermediate image states. The image enhancement network is trained with a MeanFlow consistency loss along with the image reconstruction loss. In order to provide an adaptive refinement process in terms of spatial location of enhancements, we integrate a regional reinforcement learning-driven policy network into our approach. The policy network receives information about the MeanFlow rollouts and provides predictions in terms of tile-wise refinement budgets, stopping criteria, and total budget allocation of enhancement processes. Our policy network is trained through reinforcement learning in a policy gradient framework, where the goal of the training reward is to maximize improvement of enhancements while minimizing unnecessary computations and avoiding instabilities. In this way, our approach combines conditional flow-based enhancement with reinforcement learning-based spatial enhancement control. This allows our approach to focus more attention on enhancing difficult areas while stabilizing areas already showing sufficient quality. Our results show high accuracy in the tumor ROI, with the average radiomic feature CCC being 0.96, an average PSNR of 31.30 $\pm$ 4.16, and average SSIM of 0.94 $\pm$ 0.07. Moreover, there is an improvement in the overall quality of images, with an average PSNR of 34.23 $\pm$ 1.71 and average SSIM of 0.95 $\pm$ 0.01.

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

2 major / 2 minor

Summary. The paper introduces RA-CMF, a region-adaptive conditional MeanFlow pipeline for CT image reconstruction in lung cancer applications. A conditional MeanFlow network models enhancement trajectories by predicting image-conditioned flow fields from intermediate states and is trained with a MeanFlow consistency loss plus image reconstruction loss. A reinforcement learning policy network is integrated to provide adaptive spatial control: it takes MeanFlow rollouts as input and outputs tile-wise refinement budgets, stopping criteria, and total budget allocation. The policy is trained via policy gradients with a reward that maximizes enhancement gains while minimizing unnecessary computations and instabilities. This combination is claimed to focus refinement on difficult regions while stabilizing high-quality areas. Quantitative results are given for tumor ROI (average radiomic feature CCC 0.96, PSNR 31.30 ± 4.16, SSIM 0.94 ± 0.07) and overall image quality (PSNR 34.23 ± 1.71, SSIM 0.95 ± 0.01).

Significance. If the adaptive RL mechanism can be shown through proper controls to deliver the reported gains, the work would offer a meaningful contribution to medical image reconstruction by demonstrating efficient, spatially varying refinement in flow-based models. This could improve handling of protocol and scanner variability in CT while preserving diagnostic features in tumor regions. The fusion of conditional flow modeling with RL-driven budget allocation is a plausible direction, though its practical impact hinges on reproducible evidence that the policy component outperforms non-adaptive baselines.

major comments (2)
  1. [Experimental Results] Experimental Results: the reported metrics (tumor-ROI CCC 0.96, overall PSNR 34.23 ± 1.71, SSIM 0.95 ± 0.01) are presented without any baseline comparisons to standard CT reconstruction methods or ablations that isolate the RL policy network, making it impossible to attribute improvements to the region-adaptive control rather than the base conditional MeanFlow.
  2. [Method] Method: the policy network is described as receiving MeanFlow rollouts and predicting tile-wise budgets via policy gradients, yet no concrete specification is given for state representation, action-space discretization, the exact reward terms (especially the instability penalty), or baseline subtraction, which are load-bearing for verifying that the policy reliably discovers spatially varying refinement strategies.
minor comments (2)
  1. [Abstract] The abstract introduces 'MeanFlow' and 'MeanFlow consistency loss' without a brief definition or reference on first use, which reduces accessibility for readers outside the immediate subfield.
  2. No mention is made of the dataset size, number of patients, or cross-validation procedure, which would help contextualize the reported standard deviations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important areas for strengthening the experimental validation and methodological details. We will revise the manuscript to address both major concerns directly.

read point-by-point responses

  1. Referee: [Experimental Results] Experimental Results: the reported metrics (tumor-ROI CCC 0.96, overall PSNR 34.23 ± 1.71, SSIM 0.95 ± 0.01) are presented without any baseline comparisons to standard CT reconstruction methods or ablations that isolate the RL policy network, making it impossible to attribute improvements to the region-adaptive control rather than the base conditional MeanFlow.

    Authors: We agree that the current results section does not sufficiently isolate the contribution of the RL-driven adaptive policy. In the revision we will add baseline comparisons against standard CT reconstruction techniques (FBP, MBIR) and non-adaptive conditional MeanFlow variants, together with an ablation that removes the policy network and uses uniform refinement budgets. These new experiments will quantify the incremental gains attributable to region-adaptive control on both tumor-ROI radiomic consistency and whole-image metrics. revision: yes

  2. Referee: [Method] Method: the policy network is described as receiving MeanFlow rollouts and predicting tile-wise budgets via policy gradients, yet no concrete specification is given for state representation, action-space discretization, the exact reward terms (especially the instability penalty), or baseline subtraction, which are load-bearing for verifying that the policy reliably discovers spatially varying refinement strategies.

    Authors: We acknowledge that the method description is currently high-level and lacks the requested implementation details. The revised manuscript will expand this section to specify: (i) state representation as the concatenation of MeanFlow rollout features (flow fields, intermediate images, and per-tile statistics); (ii) discrete action space consisting of integer refinement budgets per tile (0–10 steps); (iii) the full reward as a weighted combination of enhancement gain (ΔPSNR + ΔSSIM), negative compute cost, and an instability penalty (variance of flow magnitude above a threshold); and (iv) REINFORCE with a learned value baseline for variance reduction. Pseudocode and hyper-parameter tables will be added for clarity. revision: yes

Circularity Check

0 steps flagged

No circularity; standard ML training objectives with no self-referential reductions.

full rationale

The paper presents a conditional MeanFlow network trained via a consistency loss plus reconstruction loss, plus an RL policy network trained by policy gradients on MeanFlow rollouts to output tile-wise budgets and stopping criteria. No equations or derivations appear in the abstract or described text. The consistency loss and RL reward are conventional training objectives that do not reduce to their own inputs by construction, nor do any predictions relabel fitted parameters. The approach is self-contained against external benchmarks (PSNR/SSIM/CCC metrics) with no load-bearing self-citation chains or uniqueness theorems invoked. The central claim rests on empirical results rather than a closed logical loop.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract does not specify any free parameters, axioms, or invented entities; the approach appears to build on standard deep learning and reinforcement learning components without additional postulates detailed here.

pith-pipeline@v0.9.0 · 5630 in / 1351 out tokens · 59407 ms · 2026-05-09T20:06:55.185342+00:00 · methodology

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

44 extracted references · 24 canonical work pages · 2 internal anchors

  1. [1]

    Cancer statistics, 2024,

    R. L. Siegel, A. N. Giaquinto, and A. Jemal, “Cancer statistics, 2024,” CA: A Cancer Journal for Clinicians, vol. 74, no. 1, pp. 12–49, 2024. [Online]. Available: https://acsjournals.onlinelibrary.wiley.com/doi/full/ 10.3322/caac.21820

    work page doi:10.3322/caac.21820 2024

  2. [2]

    Lung cancer,

    World Health Organization, “Lung cancer,” https://www.who.int/ news-room/fact-sheets/detail/lung-cancer, 2026, accessed: 2026-04-25

    2026

  3. [3]

    Cancer statistics, 2025,

    R. L. Siegel, A. N. Giaquinto, and A. Jemal, “Cancer statistics, 2025,” CA: A Cancer Journal for Clinicians, vol. 75, no. 1, pp. 10–45,

    2025

  4. [4]

    Cancer statistics, 2025,

    [Online]. Available: https://acsjournals.onlinelibrary.wiley.com/ doi/10.3322/caac.21871

    work page doi:10.3322/caac.21871

  5. [5]

    Cancer stat facts: Lung and bronchus can- cer,

    National Cancer Institute, “Cancer stat facts: Lung and bronchus can- cer,” https://seer.cancer.gov/statfacts/html/lungb.html, 2026, surveillance, Epidemiology, and End Results Program. Accessed: 2026-04-25

    2026

  6. [6]

    Reduced lung-cancer mortality with low-dose computed tomographic screening,

    D. R. Aberle, A. M. Adams, C. D. Berg, W. C. Black, J. D. Clapp, R. M. Fagerstrom, I. F. Gareen, C. Gatsonis, P. M. Marcus, and J. D. Sicks, “Reduced lung-cancer mortality with low-dose computed tomographic screening,”The New England Journal of Medicine, vol. 365, no. 5, pp. 395–409, 2011. [Online]. Available: https://www.nejm.org/doi/full/10.1056/NEJMoa1102873

    work page doi:10.1056/nejmoa1102873 2011

  7. [7]

    Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach,

    H. J. W. L. Aerts, E. R. Velazquez, R. T. H. Leijenaar, C. Parmar, P. Grossmann, S. Carvalho, J. Bussink, R. Monshouwer, B. Haibe-Kains, D. Rietveld, F. Hoebers, M. M. Rietbergen, C. R. Leemans, A. Dekker, J. Quackenbush, R. J. Gillies, and P. Lambin, “Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach,”Nature Communi...

    work page doi:10.1038/ncomms5006 2014

  8. [8]

    Radiomics: the bridge between medical imaging and personalized medicine,

    P. Lambin, R. T. H. Leijenaar, T. M. Deist, J. Peerlings, E. E. C. de Jong, J. van Timmeren, S. Sanduleanu, R. T. H. M. Larue, A. J. G. Even, A. Jochems, Y . van Wijk, H. Woodruff, J. van Soest, T. Lustberg, E. Roelofs, W. van Elmpt, A. Dekker, F. M. Mottaghy, J. E. Wildberger, and S. Walsh, “Radiomics: the bridge between medical imaging and personalized ...

    work page arXiv 2017

  9. [9]

    Measuring computed tomography scanner variability of radiomics features,

    D. Mackin, X. Fave, L. Zhang, D. Fried, J. Yang, B. Taylor, E. Rodriguez-Rivera, C. Dodge, A. K. Jones, and L. Court, “Measuring computed tomography scanner variability of radiomics features,” Investigative Radiology, vol. 50, no. 11, pp. 757–765, 2015. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/26115366/

    work page arXiv 2015

  10. [10]

    A postreconstruction harmonization method for multicenter radiomic studies in PET,

    F. Orlhac, S. Boughdad, C. Philippe, H. Stalla-Bourdillon, C. Nioche, L. Champion, M. Soussan, F. Frouin, V . Frouin, and I. Buvat, “A postreconstruction harmonization method for multicenter radiomic studies in PET,”Journal of Nuclear Medicine, vol. 59, no. 8, pp. 1321–1328, 2018. [Online]. Available: https://jnm.snmjournals.org/ content/59/8/1321

    2018

  11. [11]

    Deep learning–based image conversion of CT reconstruction kernels improves radiomics reproducibility for pulmonary nodules or masses,

    J. Choe, S. M. Lee, K.-H. Do, G. Lee, J. B. Lee, S. M. Lee, and J. B. Seo, “Deep learning–based image conversion of CT reconstruction kernels improves radiomics reproducibility for pulmonary nodules or masses,”Radiology, vol. 292, no. 2, pp. 365–373, 2019. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/31210613/

    work page arXiv 2019

  12. [12]

    DiffusionCT: Latent diffusion model for CT image standardization,

    M. Selim, J. Zhang, M. A. Brooks, G. Wang, and J. Chen, “DiffusionCT: Latent diffusion model for CT image standardization,” inAMIA Annual Symposium Proceedings, 2024, pp. 624–633. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC10785850/

    2024

  13. [13]

    Mean Flows for One-step Generative Modeling

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He, “Mean flows for one-step generative modeling,”arXiv preprint arXiv:2505.13447, 2025. [Online]. Available: https://arxiv.org/abs/2505.13447

    work page internal anchor Pith review arXiv 2025

  14. [14]

    Radiomics: Extracting more information from medical images using advanced feature analysis,

    P. Lambin, E. Rios-Velazquez, R. Leijenaar, S. Carvalho, R. G. P. M. van Stiphout, P. Granton, C. M. L. Zegers, R. Gillies, R. Boellard, A. Dekker, and H. J. W. L. Aerts, “Radiomics: Extracting more information from medical images using advanced feature analysis,” European Journal of Cancer, vol. 48, no. 4, pp. 441–446, 2012. [Online]. Available: https://...

    work page arXiv 2012

  15. [15]

    Computational radiomics system to decode the radiographic phenotype,

    J. J. M. van Griethuysen, A. Fedorov, C. Parmar, A. Hosny, N. Aucoin, V . Narayan, R. G. H. Beets-Tan, J.-C. Fillion- Robin, S. Pieper, and H. J. W. L. Aerts, “Computational radiomics system to decode the radiographic phenotype,”Cancer Research, vol. 77, no. 21, pp. e104–e107, 2017. [Online]. Available: https://aacrjournals.org/cancerres/article/77/21/e10...

    2017

  16. [16]

    The image biomarker standardization initiative: Standardized quantitative radiomics for high-throughput image-based phenotyping,

    A. Zwanenburg, M. Valli `eres, M. A. Abdalah, H. J. W. L. Aerts, V . Andrearczyk, A. Aditya, A. Saeed, S. Bakas, R. J. Beukinga, R. Boellaardet al., “The image biomarker standardization initiative: Standardized quantitative radiomics for high-throughput image-based phenotyping,”Radiology, vol. 295, no. 2, pp. 328–338, 2020. [Online]. Available: https://pu...

    work page arXiv 2020

  17. [17]

    Textural features for image classification,

    R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural features for image classification,”IEEE Transactions on Systems, Man, and Cybernetics, vol. SMC-3, no. 6, pp. 610–621, 1973. [Online]. Available: https://doi.org/10.1109/TSMC.1973.4309314

    work page doi:10.1109/tsmc.1973.4309314 1973

  18. [18]

    Neighboring gray level dependence matrix for texture classification,

    C. Sun and W. G. Wee, “Neighboring gray level dependence matrix for texture classification,”Computer Vision, Graphics, and Image Processing, vol. 23, no. 3, pp. 341–352, 1983. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0734189X83900324

    work page arXiv 1983

  19. [19]

    Texture analysis using gray level run lengths,

    M. M. Galloway, “Texture analysis using gray level run lengths,” Computer Graphics and Image Processing, vol. 4, no. 2, pp. 172– 179, 1975. [Online]. Available: https://doi.org/10.1016/S0146-664X(75) 80008-6

    work page doi:10.1016/s0146-664x(75 1975

  20. [20]

    Texture indexes and gray level size zone matrix: Application to cell nuclei classification,

    G. Thibault, B. Fertil, C. Navarro, S. Pereira, P. Cau, N. L ´evy, J. Sequeira, and J.-L. Mari, “Texture indexes and gray level size zone matrix: Application to cell nuclei classification,” inProceedings of the 10th International Conference on Pattern Recognition and Information Processing, 2009

    2009

  21. [21]

    Textural features corresponding to textural properties,

    M. Amadasun and R. King, “Textural features corresponding to textural properties,”IEEE Transactions on Systems, Man, and Cybernetics, vol. 19, no. 5, pp. 1264–1274, 1989. [Online]. Available: https://doi.org/10.1109/21.44046

    work page doi:10.1109/21.44046 1989

  22. [22]

    V oxel size and gray level normalization of CT radiomic features in lung cancer,

    M. S. ul Hassan, K. Latifi, G. Zhang, G. Ullah, R. Gillies, and E. Moros, “V oxel size and gray level normalization of CT radiomic features in lung cancer,”Scientific Reports, vol. 8, p. 10545, 2018. [Online]. Available: https://www.nature.com/articles/s41598-018-28895-9

    2018

  23. [23]

    A guide to ComBat harmonization of imaging biomarkers in multicenter studies,

    F. Orlhac, J. J. Eertink, A.-S. Cottereau, J. M. Zijlstra, C. Thieblemont, M. Meignan, R. Boellaard, and I. Buvat, “A guide to ComBat harmonization of imaging biomarkers in multicenter studies,”Journal of Nuclear Medicine, vol. 63, no. 2, pp. 172–179, 2022. [Online]. Available: https://jnm.snmjournals.org/content/63/2/172

    2022

  24. [24]

    STAN-CT: Standardizing CT image using generative adversarial networks,

    M. Selim, J. Zhang, B. Fei, G.-Q. Zhang, and J. Chen, “STAN-CT: Standardizing CT image using generative adversarial networks,”AMIA Annual Symposium Proceedings, vol. 2020, pp. 1100–1109, 2020. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/33936486/

    work page arXiv 2020

  25. [25]

    CT image harmonization for enhancing radiomics studies,

    ——, “CT image harmonization for enhancing radiomics studies,” in2021 IEEE International Conference on Bioinformatics and Biomedicine, 2021, pp. 1057–1062. [Online]. Available: https://arxiv. org/abs/2107.01337

    work page arXiv 2021

  26. [26]

    Wasserstein generative adversarial networks,

    M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adversarial networks,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 70. PMLR, 2017, pp. 214–223. [Online]. Available: https://proceedings.mlr.press/v70/arjovsky17a.html

    2017

  27. [27]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 6840–6851. [Online]. Available: https://proceedings.neurips. cc/paper/2020/hash/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html

    2020

  28. [28]

    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, 2022, pp. 10 684–10 695. [Online]. Available: https://openaccess.thecvf.com/content/CVPR2022/html/ Rombach High-Resolution Image Synthesis With La...

    2022

  29. [29]

    R. S. Sutton and A. G. Barto,Reinforcement Learning: An Introduction, 2nd ed. MIT Press, 2018. [Online]. Available: http://incompleteideas.net/book/the-book-2nd.html

    2018

  30. [30]

    Huynh-Thu, M

    Q. Huynh-Thu and M. Ghanbari, “Scope of validity of PSNR in image/video quality assessment,”Electronics Letters, vol. 44, no. 13, pp. 800–801, 2008. [Online]. Available: https://digital-library.theiet.org/ doi/10.1049/el:20080522

    work page doi:10.1049/el:20080522 2008

  31. [31]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004. [Online]. Available: https://ieeexplore.ieee.org/document/1284395

    work page arXiv 2004

  32. [32]

    Shape from focus,

    S. K. Nayar and Y . Nakagawa, “Shape from focus,” Columbia University, Tech. Rep., 1989. [Online]. Available: https://cave.cs. columbia.edu/Statics/publications/pdfs/Nayar TR89.pdf

    1989

  33. [33]

    Asynchronous methods for deep reinforcement learning,

    V . Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep reinforcement learning,” inProceedings of the 33rd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 48. PMLR, 2016, pp. 1928–1937. [Online]. Available: https://proceedings.mlr.p...

    2016

  34. [35]

    Proximal Policy Optimization Algorithms

    [Online]. Available: https://arxiv.org/abs/1707.06347

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [36]

    2013.doi: 10.7937/TCIA.HMQ8-J677

    National Lung Screening Trial Research Team, “Data from the national lung screening trial (NLST),” The Cancer Imaging Archive, 2013. [Online]. Available: https://doi.org/10.7937/TCIA.HMQ8-J677

    work page doi:10.7937/tcia.hmq8-j677 2013

  36. [37]

    The cancer imaging archive (TCIA): Maintaining and operating a public information repository,

    K. Clark, B. Vendt, K. Smith, J. Freymann, J. Kirby, P. Koppel, S. Moore, S. Phillips, D. Maffitt, M. Pringle, L. Tarbox, and F. Prior, “The cancer imaging archive (TCIA): Maintaining and operating a public information repository,”Journal of Digital Imaging, vol. 26, no. 6, pp. 1045–1057, 2013. [Online]. Available: https://doi.org/10.1007/s10278-013-9622-7

    work page doi:10.1007/s10278-013-9622-7 2013

  37. [38]

    NLST CT images,

    National Cancer Institute, “NLST CT images,” Cancer Data Access System, 2026, accessed: 2026-04-25. [Online]. Available: https: //cdas.cancer.gov/learn/nlst/images/

    2026

  38. [39]

    NLSTseg: A pixel-level lung cancer dataset based on NLST LDCT images,

    K.-H. Chenet al., “NLSTseg: A pixel-level lung cancer dataset based on NLST LDCT images,”Scientific Data, 2025. [Online]. Available: https://www.nature.com/articles/s41597-025-05742-x

    2025

  39. [40]

    NLSTseg: A pixel-level lung cancer dataset based on NLST LDCT images,

    Y . Lin, “NLSTseg: A pixel-level lung cancer dataset based on NLST LDCT images,” 2025. [Online]. Available: https://zenodo.org/records/ 14838349

    2025

  40. [41]

    A concordance correlation coefficient to evaluate reproducibility,

    L. I.-K. Lin, “A concordance correlation coefficient to evaluate reproducibility,”Biometrics, vol. 45, no. 1, pp. 255–268, 1989. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/2720055/

    work page arXiv 1989

  41. [42]

    The noise power spectrum of CT images,

    M. F. Kijewski and P. F. Judy, “The noise power spectrum of CT images,”Physics in Medicine and Biology, vol. 32, no. 5, pp. 565–575,

  42. [43]

    Available: https://doi.org/10.1088/0031-9155/32/5/003

    [Online]. Available: https://doi.org/10.1088/0031-9155/32/5/003

    work page doi:10.1088/0031-9155/32/5/003

  43. [44]

    Application of the noise power spectrum in modern diagnostic MDCT: Part i. measurement of noise power spectra and noise equivalent quanta,

    K. L. Boedeker, V . N. Cooper, and M. F. McNitt-Gray, “Application of the noise power spectrum in modern diagnostic MDCT: Part i. measurement of noise power spectra and noise equivalent quanta,” Physics in Medicine and Biology, vol. 52, no. 14, pp. 4027–4046,

  44. [45]

    Available: https://doi.org/10.1088/0031-9155/52/14/002

    [Online]. Available: https://doi.org/10.1088/0031-9155/52/14/002

    work page doi:10.1088/0031-9155/52/14/002