Pith. sign in

REVIEW 3 major objections 5 minor 40 references

A ranking of CT reconstruction methods measured on clean, idealized data does not predict how the same methods perform under mild input noise; the clean-order champion collapses while a physics-in-the-loop method rises to first.

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 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 04:23 UTC pith:2DMFSJYE

load-bearing objection Serious agentic-benchmark paper; the clean-to-noisy rank inversion is plausible but the flagship dd-supervised collapse may be an artifact of the agent's implementation — referee it, but ask for verification. the 3 major comments →

arxiv 2607.22824 v1 pith:2DMFSJYE submitted 2026-07-24 physics.med-ph cs.AIcs.CV

Agentic Autoresearch for CT Reconstruction

Andreas Maier , Lucas Kachelriess , Siming Bayer , Yixing Huang , Yan Xia , Amber Simpson , Moritz Zaiss This is my paper
classification physics.med-ph cs.AIcs.CV
keywords CT reconstructionagentic researchLLM agentbenchmarkingnoise robustnesssparse-view CTlow-dose CTknown operators
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.

This paper asks whether an autonomous large-language-model agent can take over the labor of comparing CT reconstruction methods, and whether rankings measured on clean, idealized data survive realistic noise. The agent independently implemented, tuned, and benchmarked 26 solvers under one frozen metric and a fixed compute budget, producing a top tier of statistically tied methods rather than a single winner. The decisive finding is negative: adding mild Poisson noise to test inputs, without retraining, nearly inverts the breast ranking; the clean-data champion falls to the bottom while a learned primal-dual method rises to the top, so the clean leaderboard carries almost no information about noisy performance (rank correlation near zero). Retraining on matched noise restores much of the original order, showing the collapse is a distribution-shift effect, not a permanent defect. The paper concludes that no single-factor benchmark can certify robustness, and argues for challenges that vary many realistic factors together.

Core claim

The paper's central claim is that an agentic research loop—an LLM that edits solver code, runs short cluster jobs, reads a frozen calibrated headroom score, and revises—can implement, tune, and benchmark 26 CT reconstruction methods fairly, and that doing so exposes a structural problem in how the field ranks methods. On noiseless and low-dose data, the benchmark yields a small tier of statistically indistinguishable top methods and a compact 969-parameter recombination that ties the top low-dose tier at the 1% level. When the same trained breast models are re-scored on inputs with mild Poisson noise (I0 = 1e5 photons, about 1–2% at the thickest ray), the noiseless ranking nearly inverts: th

What carries the argument

The machinery is a calibrated headroom metric, hr = max(0, 1 − RMSE / RMSE_FBP), computed inside the scan field of view and shared by all methods, together with a differentiable fan-beam projector that lets gradients flow through the forward model. Around these, the loop runs: the agent changes one knob, runs one short job, reads the frozen metric, and accepts or discards. The explanatory device is a two-axis taxonomy of the 26 solvers: how strongly the forward operator enters inference (none, one data-consistency step, in-loop unrolled, per-scene fit) and where the prior comes from (hand-crafted, supervised, self-supervised, generative, or implicit). The paper uses these axes to predict whi

Load-bearing premise

The ranking inversion rests on the assumption that the agent's reimplementations faithfully capture each published method; the paper concedes they reproduce each method's design but may differ in detail from official implementations, so the clean-to-noisy collapse could in principle be an artifact of unequal implementation quality rather than of the methods themselves.

What would settle it

Run the same no-retrain noise probe with the official, human-verified implementations of the clean-data champion (a supervised image-domain denoiser) and the learned primal-dual method on the same 200 test cases at I0 = 1e5. If the denoiser does not collapse to near-zero headroom while the primal-dual method rises to the top, the inversion is an implementation artifact, not a property of the methods.

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

If this is right

  • A clean-data leaderboard can reward methods that fail under a small realistic perturbation, so the strongest method on ideal data may be the least safe in practice.
  • Robustness is concentrated in two disjoint corners: methods that keep the forward operator in the loop and methods whose priors never overfit the clean distribution.
  • The best compact architecture depends on the bottleneck: a tiny denoiser on noise-limited data, a filtered data-consistency plus primal-dual unroll on sparse-view data, at 0.4–0.04% of the champion's parameters.
  • Retraining on matched noise brings most of the clean ranking back, so a collapse under distribution shift is recoverable rather than a permanent property of the architecture.
  • Because noise is one of many possible shifts, benchmarks should vary several realistic factors together instead of one at a time.

Where Pith is reading between the lines

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

  • The inversion result implies a practical selection rule the paper does not state: if the noise level in deployment may drift, prefer in-loop-physics or hand-crafted-smoothing methods over pure image-domain supervised denoisers unless matched-noise retraining is possible.
  • The two-axis taxonomy can be tested prospectively: rank methods by physics engagement and prior source, then probe with beam hardening and scatter; the prediction is that brittleness tracks the supervised image-domain corner.
  • The compact-solver search suggests that the agent's chief value may lie in cheap parameter-efficient recombination, which could be extended by a multi-agent ensemble that deliberately explores beyond greedy metric improvement.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper describes an agentic loop in which a large language model writes, tunes, and benchmarks CT reconstruction solvers under a fixed differentiable fan-beam projector and a single frozen headroom metric. The authors benchmark 26 methods on two tasks (Mayo low-dose CT and a noiseless 128-view breast sparse-view challenge), build a compact recombination solver, and test robustness by adding Poisson noise to the breast test sinograms and re-scoring frozen models without retraining, then retraining on matched noise. They report that the noiseless ranking is essentially uncorrelated with the no-retrain noisy ranking (Spearman ρ≈0.04), that learned primal-dual and hand-crafted smoothing methods rise while supervised image-domain denoisers collapse, and that retraining on matched noise restores much of the ranking (ρ≈0.61). They conclude that an ideal-data leaderboard does not predict robustness.

Significance. If the empirical results hold, the paper provides a scalable, reproducible benchmarking infrastructure and a concrete demonstration that clean-data leaderboards can mislead about noisy robustness. The release of code, the fixed projector, the frozen metric, and the per-iteration provenance are notable strengths. The compact-solver result is interesting but currently undercut by a statistical inconsistency. The external validity of the central robustness claim depends on the faithfulness of the agent-written solver implementations, which the paper itself hedges; this needs either validation or a narrowed interpretation.

major comments (3)
  1. [Abstract; §3.1; §3.4] The paper repeatedly states that the compact Mayo solver 'ties the top tier at the 1% level (p=0.02–0.027)'. This is internally inconsistent: at the 1% significance level one requires p < 0.01. With p = 0.02–0.027 the correct statement is 'at the 5% level', or the test and p-values should be reported differently. This error appears in the Abstract and in the headline compact-solver claims, so it must be corrected.
  2. [§3.5, Table 4, §4.3] The central inversion claim treats the agent-written solvers as faithful implementations of the cited methods. Section 4.3 concedes that 'the agent-written solvers reproduce each method's design but may differ in detail from its official implementation.' The most acute case is the noiseless breast champion, dd-supervised, which §3.5 describes as 'a pure supervised image-domain denoiser trained only on clean FBP' despite being labelled dual-domain. If the projection-domain branch was not implemented, the collapse to hr 0.00 is an artifact of the agent's implementation, not a property of the cited dual-domain method. Since the inversion is the primary evidence for the headline claim, the authors should either validate implementation fidelity against official code for at least the two poles of the inversion (dd-supervised and learned-primal-dual), or explicitly restrict the robustness concl
  3. [§3.6] The 'framework' is presented as turning the reversal into 'a prediction: a method's position on the physics-engagement and prior-source axes tells us, before any noise is added, whether it is likely to be brittle or robust.' This is post-hoc: the two axes in Table 1 are a taxonomy constructed after the reversal was observed, and no prospective prediction was made or tested. The text should be reframed as a post-hoc explanation or a hypothesis-generating framework, not a validated prediction.
minor comments (5)
  1. [§2.6; Tables 2–4] Section 2.6 promises to report Cohen's d_z effect sizes for the breast comparisons, but no d_z values appear in the main tables or text. Please add them or point to the Supplement where they are reported.
  2. [§3.7] The statement 'a few solvers dip slightly (Hammernik-2017, 0.70→0.61)' is ambiguous: the noiseless score in Table 3 is 0.6265, the noisy no-retrain score is 0.7016, and the retrained score is 0.610. The 0.70 in the text apparently refers to the noisy no-retrain column, not the noiseless score. Please clarify.
  3. [Table 3] Table 3 orders rows by retrained rank but the noiseless and noisy no-retrain columns are not sorted, making it harder to visually compare the reordering. Consider sorting by the noiseless rank or adding rank columns.
  4. [§2.5; §3.5] The claim that I0=10^5 is 'mild' and corresponds to 'about 1–2% at the thickest ray' is not tied to the actual line-integral values of the breast phantom. A one-line derivation or reference for the attenuation range would help readers assess the perturbation size.
  5. [§3.4] The sentence 'The agent re-derived each optimum from its own runs; steered toward the Mayo answer on breast it did not transfer' is hard to parse. Please clarify whether the agent was explicitly instructed to transfer the Mayo architecture to the breast problem, and what the outcome was.

Circularity Check

0 steps flagged

No significant circularity: the central claims are empirical benchmark measurements, and the only same-author citations are interpretive rather than load-bearing.

full rationale

The paper's central claims—agentic benchmark feasibility, the clean/noisy ranking inversion, and retraining recovery—are direct measurements on a fixed metric (Eq. 5) with 200 shared test cases, not derivations from fitted parameters. The compact-solver result and the inversion are empirical endpoints; no equation reduces to a fitted parameter or to the conclusion. The known-operator theorem in Section 1 and its reuse in Section 3.4 ('Constraining a network with an exact operator provably cannot raise, and usually lowers, the maximum error bound...9') is a self-citation to prior work by the same authors, and ref. [11] is a same-group preprint, but the theorem is not used to compute any reported hr value, ranking, or Spearman correlation; it is interpretive context. Thus it is a minor self-citation, not a load-bearing circular step. The two-axis framework of Section 3.6 is a post-hoc qualitative explanation, not a definitional derivation: the axes are assigned from architecture (whether A is in the loop; prior source), not from the noisy outcomes, so the explanation is not equivalent to the result by construction. The limitation in Section 4.3 ('The agent-written solvers reproduce each method's design but may differ in detail from its official implementation') is a fidelity/validity caveat, not a circularity. Overall, no step reduces a claimed prediction to its own input.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities; the agent and compact solver are software artifacts, not postulated entities. The central claims rest on the chosen metric, the linear/Poisson forward model, the fidelity of agent-written implementations, and one hand-chosen noise level.

free parameters (3)
  • Poisson noise level I0 = 10^5 photons
    Chosen by hand; the entire robustness-inversion result is demonstrated at this single dose, with no sweep. The paper acknowledges a single dose in Section 4.3.
  • Per-iteration compute budget = 20 minutes
    Chosen to equalize comparisons; it determines which methods finish (per-scene methods DNF) and how deep the agent tunes each solver.
  • Autoresearch iterations per method = 20 for 26 methods; 40 for compact solver
    Protocol choice; rankings and the compact solver depend on this budget. The compact solver only reached hr 0.62 after human suggestions from iteration 20.
axioms (5)
  • domain assumption The CT scan is a linear operator A with additive noise (Eq. 1)
    Standard CT modeling; all solvers and the noise experiment inherit this.
  • domain assumption All 26 methods can be captured by regularized inversion Eq. (2) with data-consistency operator D and prior R
    Taxonomy in Table 1; some methods (generative/diffusion, implicit neural fields) do not literally minimize this objective, so the umbrella is approximate.
  • ad hoc to paper Poisson noise model Eq. (6) with I0=10^5 is representative of realistic noise
    Central to the transfer experiment; only one dose is tested, no sweep.
  • domain assumption Embedding a known operator cannot raise and usually lowers the maximum error bound (refs [9,11])
    Self-cited theoretical support for the compact-solver interpretation; not independently verified in this paper.
  • domain assumption The frozen headroom metric hr (Eq. 5) is a sufficient scalar for ranking reconstruction quality
    All conclusions about rankings and inversion depend on this metric; SSIM/PSNR are reported but not ranked.

pith-pipeline@v1.3.0-alltime-deepseek · 18474 in / 14941 out tokens · 145220 ms · 2026-08-01T04:23:39.478816+00:00 · methodology

0 comments
read the original abstract

Comparing CT reconstruction methods fairly is labor-intensive and largely manual, and many benchmarks use idealized data. We ask whether a large language model (LLM) agent can do the labor of reconstruction research on its own, and whether a ranking measured on ideal data predicts behavior under realistic noise. We built an agentic loop: the agent edits a solver, runs a short cluster job, reads one frozen metric, and revises. The metric is a calibrated headroom score against the FBP baseline, inside the field of view; every method shares the same differentiable fan-beam projector. We benchmarked 26 methods on Mayo low-dose CT (noise-limited) and a 128-view sparse-view breast task from the noiseless DL-Sparse-View Challenge, with validation-selected iterations scored on a held-out test set. Every trained breast model was then re-scored on noisy inputs (I_0 = 10^5 photons) without retraining, and separately retrained on matched noise. The agent independently implemented, tuned, and benchmarked all 26 methods, and recombined them into a compact solver of 969 parameters that ties the top Mayo tier at the 1% level using 0.4% of the champion's parameters. Benchmarking gives a tier of statistically indistinguishable top methods, not one winner. Mild input noise nearly inverts the breast ranking: the noiseless champion (a supervised image denoiser, hr 0.89) collapses to 0.00, while a learned primal-dual method rises to champion (0.72 to 0.93). An ideal-data leaderboard therefore does not predict robustness. The inversion is a transfer effect, not a permanent deficit: retraining on matched noise restores much of the clean ranking (Spearman rho 0.04 to 0.61). Noise is only the easiest confounder in an open-ended set (beam hardening, scatter, anatomy, disease), so no single-factor challenge certifies generality. Benchmarks should model a broad spectrum of realistic factors at once.

Figures

Figures reproduced from arXiv: 2607.22824 by Amber Simpson, Andreas Maier, Lucas Kachelriess, Moritz Zaiss, Siming Bayer, Yan Xia, Yixing Huang.

Figure 1
Figure 1. Figure 1: Example slices from the two problems, each showing the FBP input the solvers start from and the ground truth. Left: Mayo low-dose CT, where quantum noise (visible graininess) is the bottleneck. Right: the 128-view sparse￾view breast phantom, where undersampling produces streak artifacts. Images are windowed for display, and the FBP is intensity-calibrated to the truth as in the metric. hashing: no test ima… view at source ↗
Figure 2
Figure 2. Figure 2: The agentic autoresearch loop and its provenance. The agent reads the previous result, names the failure, changes one thing, states a hypothesis, runs one short cluster job grounded in the same differentiable fan-beam projector, reads the frozen calibrated-headroom metric, and accepts or discards the change. Every iteration writes an immutable record (configuration, reconstruction, metric, comparison figur… view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy versus parameter count for both datasets (Mayo and breast noiseless). To overlay the two different hr scales, each solver’s hr is normalized to the percent of the best hr reached on its own board. The agent’s compact recombination solver is ringed: 969 parameters on Mayo, 195 on breast, i.e. 0.4 % and 0.04 % of the respective champion’s count; on Mayo it ties the top tier at the 1 % level. The div… view at source ↗
Figure 4
Figure 4. Figure 4: The reversal. Paired rank plot from the noiseless breast board to the noisy breast board (𝐼0 = 105 , same trained models, same 200 test cases, no retraining), with an arrow per method. The supervised image-domain champion collapses from rank 1 to last; the learned primal-dual method rises from mid-pack to champion; hand-crafted smoothing and in-loop physics priors rise. A small, realistic noise perturbatio… view at source ↗
Figure 5
Figure 5. Figure 5: Example reconstructions behind the reversal, for the two methods at opposite ends of it. Rows: the noiseless champion dual-domain-supervised (a supervised image-domain denoiser) versus learned-primal-dual (in-loop physics). Columns: ground truth, reconstruction from the clean input, reconstruction from the noisy input (𝐼0 = 105 , same trained model, no retraining), and the difference to truth. Per-case SSI… 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

40 extracted references · 7 canonical work pages

  1. [1]

    Sidky and Xiaochuan Pan

    Emil Y. Sidky and Xiaochuan Pan. Report on the AAPM deep-learning sparse-view CT (DL-sparse-view CT) grand challenge.MedicalPhysics,49(8):4935–4943,2022. doi: 10.1002/mp.15489

    doi:10.1002/mp.15489 2022

  2. [2]

    Learned primal-dual reconstruction.IEEE Transactions on Medical Imaging, 37(6): 1322–1332,2018

    Jonas Adler and Ozan Öktem. Learned primal-dual reconstruction.IEEE Transactions on Medical Imaging, 37(6): 1322–1332,2018. doi: 10.1109/TMI.2018.2799231

    arXiv 2018

  3. [3]

    LearningavariationalnetworkforreconstructionofacceleratedMRIdata.MagneticResonanceinMedicine, 79(6):3055–3071,2018

    KerstinHammernik,TeresaKlatzer,ErichKobler,MichaelP.Recht,DanielK.Sodickson,ThomasPock,andFlorian Knoll. LearningavariationalnetworkforreconstructionofacceleratedMRIdata.MagneticResonanceinMedicine, 79(6):3055–3071,2018. doi: 10.1002/mrm.26977

    doi:10.1002/mrm.26977 2018

  4. [4]

    Ghesu, Vincent Christlein, and Andreas Maier

    Tobias Würfl, Florin C. Ghesu, Vincent Christlein, and Andreas Maier. Deep learning computed tomography. In MedicalImageComputingandComputer-AssistedIntervention—MICCAI2016,LectureNotesinComputerScience, pages432–440.Springer,2016. doi: 10.1007/978-3-319-46726-9_50

    work page doi:10.1007/978-3-319-46726-9_50 2016

  5. [5]

    Deeplearningcomputedtomography: Learningprojection-domainweightsfromimagedomain 14 in limited angle problems.IEEE Transactions on Medical Imaging, 37(6):1454–1463, 2018

    Tobias Würfl, Mathis Hoffmann, Vincent Christlein, Katharina Breininger, Yixin Huang, Mathias Unberath, and AndreasK.Maier. Deeplearningcomputedtomography: Learningprojection-domainweightsfromimagedomain 14 in limited angle problems.IEEE Transactions on Medical Imaging, 37(6):1454–1463, 2018. doi: 10.1109/TMI.2018. 2833499

    doi:10.1109/tmi.2018 2018

  6. [6]

    Deeplearningtechniquesforinverseproblemsinimaging.IEEEJournalonSelectedAreasinInformationTheory,1 (1):39–56,2020

    GregoryOngie,AjilJalal,ChristopherA.Metzler,RichardG.Baraniuk,AlexandrosG.Dimakis,andRebeccaWillett. Deeplearningtechniquesforinverseproblemsinimaging.IEEEJournalonSelectedAreasinInformationTheory,1 (1):39–56,2020. doi: 10.1109/JSAIT.2020.2991563

    arXiv 2020

  7. [7]

    Some investi- gations on robustness of deep learning in limited angle tomography

    Yixing Huang, Tobias Würfl, Katharina Breininger, Ling Liu, Günter Lauritsch, and Andreas Maier. Some investi- gations on robustness of deep learning in limited angle tomography. InInternational Conference on Medical Image ComputingandComputer-AssistedIntervention(MICCAI),pages145–153.Springer,2018

    2018

  8. [8]

    A gentle introduction to deep learning in medicalimageprocessing.ZeitschriftfürMedizinischePhysik,29(2):86–101,2019.doi: 10.1016/j.zemedi.2018.12.003

    Andreas Maier, Christopher Syben, Tobias Lasser, and Christian Riess. A gentle introduction to deep learning in medicalimageprocessing.ZeitschriftfürMedizinischePhysik,29(2):86–101,2019.doi: 10.1016/j.zemedi.2018.12.003

    doi:10.1016/j.zemedi.2018.12.003 2019

  9. [9]

    Learning with known operators reduces maximum error bounds.NatureMachineIntelligence,1(8):373–380,2019

    AndreasK.Maier,ChristopherSyben,BernhardStimpel,TobiasWürfl,MathisHoffmann,FrankSchebesch,Weilin Fu, Leonid Mill, Lasse Kling, and Silke Christiansen. Learning with known operators reduces maximum error bounds.NatureMachineIntelligence,1(8):373–380,2019. doi: 10.1038/s42256-019-0077-5

    doi:10.1038/s42256-019-0077-5 2019

  10. [10]

    Known operator learning and hybridmachinelearninginmedicalimaging—areviewofthepast,thepresent,andthefuture.ProgressinBiomedical Engineering,4(2):022002,2022

    Andreas Maier, Harald Köstler, Marco Heisig, Patrick Krauss, and Seung Hee Yang. Known operator learning and hybridmachinelearninginmedicalimaging—areviewofthepast,thepresent,andthefuture.ProgressinBiomedical Engineering,4(2):022002,2022. doi: 10.1088/2516-1091/ac5b13

    doi:10.1088/2516-1091/ac5b13 2022

  11. [11]

    Adeepriskestimatorforknownoperator learning.arXivpreprintarXiv:2605.08517,2026

    AndreasMaier,MdHasan,PaulinaConrad,andPaulaAndreaPerez-Toro. Adeepriskestimatorforknownoperator learning.arXivpreprintarXiv:2605.08517,2026

    Pith/arXiv arXiv 2026

  12. [12]

    autoresearch: AI agents running research automatically

    Andrej Karpathy. autoresearch: AI agents running research automatically. GitHub repository, 2026.https:// github.com/karpathy/autoresearch

    2026

  13. [13]

    The AI scientist: Towards fully automatedopen-endedscientificdiscovery.arXivpreprintarXiv:2408.06292,2024

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist: Towards fully automatedopen-endedscientificdiscovery.arXivpreprintarXiv:2408.06292,2024. arXiv:2408.06292

    Pith/arXiv arXiv 2024

  14. [14]

    Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press

    John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. SWE-agent: Agent-computerinterfacesenableautomatedsoftwareengineering. InAdvancesinNeuralInformation ProcessingSystems37(NeurIPS2024),pages50528–50652,2024. arXiv:2405.15793

    Pith/arXiv arXiv 2024

  15. [15]

    SWE- bench: Can language models resolve real-world GitHub issues? InInternational Conference on Learning Represen- tations(ICLR),pages54107–54157,2024

    CarlosE.Jimenez,JohnYang,AlexanderWettig,ShunyuYao,KexinPei,OfirPress,andKarthikNarasimhan. SWE- bench: Can language models resolve real-world GitHub issues? InInternational Conference on Learning Represen- tations(ICLR),pages54107–54157,2024. arXiv:2310.06770

    Pith/arXiv arXiv 2024

  16. [16]

    LectureNotesinComputerScience.Springer,2026

    AndreasMaier,ChristianBergler,SimingBayer,andMoritzZaiss,editors.VibeCoding: SoftwareEngineeringinthe AgeofAI. LectureNotesinComputerScience.Springer,2026. inprint

    2026

  17. [17]

    Agentic MR sequence development: Leveraging LLMs with MR skills for automatic physics-informed sequence development

    Moritz Zaiss, Amr Aly, Jonathan Endres, Tobias Dornstetter, Simon Weinmüller, and Andreas Maier. Agentic MR sequence development: Leveraging LLMs with MR skills for automatic physics-informed sequence development. arXivpreprintarXiv:2604.13282,2026. arXiv:2604.13282;Agent4MR

    Pith/arXiv arXiv 2026

  18. [18]

    HolmesIII,AliceE.Huang,FarhanaKhan,ShuaiLeng,KyleL.McMillan,GregoryJ.Michalak,KristinaM.Nunez, LifengYu,andJoelG.Fletcher

    CynthiaH.McCollough,AdamC.Bartley,RickeyE.Carter,BaiyuChen,TammyA.Drees,PhillipEdwards,DavidR. HolmesIII,AliceE.Huang,FarhanaKhan,ShuaiLeng,KyleL.McMillan,GregoryJ.Michalak,KristinaM.Nunez, LifengYu,andJoelG.Fletcher. Low-doseCTforthedetectionandclassificationofmetastaticliverlesions: Results ofthe2016lowdoseCTgrandchallenge.MedicalPhysics,44(10):e339–e35...

    work page doi:10.1002/mp.12345 2017

  19. [19]

    Techni- calnote: PYRO-NN:Pythonreconstructionoperatorsinneuralnetworks.MedicalPhysics,46(11):5110–5115,2019

    ChristopherSyben,MarkusMichen,BernhardStimpel,StephanSeitz,StefanPloner,andAndreasK.Maier. Techni- calnote: PYRO-NN:Pythonreconstructionoperatorsinneuralnetworks.MedicalPhysics,46(11):5110–5115,2019. doi: 10.1002/mp.13753

    doi:10.1002/mp.13753 2019

  20. [20]

    Reconstruct anything model: A lightweight general model for computational imaging

    Matthieu Terris, Samuel Hurault, Maxime Song, and Julián Tachella. Reconstruct anything model: A lightweight general model for computational imaging. InInternational Conference on Learning Representations (ICLR), 2026. arXiv:2503.08915

    arXiv 2026

  21. [21]

    Sidky and Xiaochuan Pan

    Emil Y. Sidky and Xiaochuan Pan. Image reconstruction in circular cone-beam computed tomography by con- strained, total-variation minimization.Physics in Medicine & Biology, 53(17):4777–4807, 2008. doi: 10.1088/ 0031-9155/53/17/021

    2008

  22. [22]

    A novel filtered backprojection-based algorithm for sparse-view CT image reconstruction

    Meng Wu, Andreas Maier, Qiao Yang, and Rebecca Fahrig. A novel filtered backprojection-based algorithm for sparse-view CT image reconstruction. InThe 13th International Meeting on Fully Three-Dimensional Image Recon- structioninRadiologyandNuclearMedicine(Fully3D),pages202–205,2015

    2015

  23. [23]

    ProjectionspacedenoisingwithbilateralfilteringandCTnoisemodelingfordosereductioninCT

    ArmandoManduca,LifengYu,JoshuaD.Trzasko,NataliaKhaylova,JamesM.Kofler,CynthiaM.McCollough,and JoelG.Fletcher. ProjectionspacedenoisingwithbilateralfilteringandCTnoisemodelingfordosereductioninCT. MedicalPhysics,36(11):4911–4919,2009. doi: 10.1118/1.3232004

    doi:10.1118/1.3232004 2009

  24. [24]

    Ultralow-parameterdenoising: Trainable bilateral filter layers in computed tomography.MedicalPhysics, 49(8):5107–5120, 2022

    Fabian Wagner, Mareike Thies, Mingxuan Gu, Yixing Huang, Sabrina Pechmann, Mayank Patwari, Stefan Ploner, OliverAust,StefanUderhardt,GeorgSchett,SilkeChristiansen,andAndreasMaier. Ultralow-parameterdenoising: Trainable bilateral filter layers in computed tomography.MedicalPhysics, 49(8):5107–5120, 2022. doi: 10.1002/mp. 15718. 15

    doi:10.1002/mp 2022

  25. [25]

    Maltz, and Hengyong Yu

    Yongshun Xu, Shuo Han, Dayang Wang, Ge Wang, Jonathan S. Maltz, and Hengyong Yu. Hybrid U-Net and Swin- transformernetworkforlimited-anglecardiaccomputedtomography.PhysicsinMedicine&Biology,69(10):105012,

  26. [26]

    SwinIR: Image restora- tion using Swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restora- tion using Swin transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops,pages1833–1844,2021. doi: 10.1109/ICCVW54120.2021.00210. arXiv:2108.10257

    arXiv 2021

  27. [27]

    Swintransformer: Hierarchicalvisiontransformerusingshiftedwindows

    ZeLiu,YutongLin,YueCao,HanHu,YixuanWei,ZhengZhang,StephenLin,andBainingGuo. Swintransformer: Hierarchicalvisiontransformerusingshiftedwindows. InProceedingsoftheIEEE/CVFInternationalConferenceon ComputerVision(ICCV),pages9992–10002,2021. doi: 10.1109/ICCV48922.2021.00986. arXiv:2103.14030

    arXiv 2021

  28. [28]

    TransCT:Dual-pathtransformerforlow-dose computed tomography

    ZhichengZhang,LequanYu,XiaokunLiang,WeiZhao,andLeiXing. TransCT:Dual-pathtransformerforlow-dose computed tomography. InMedicalImageComputingandComputer-AssistedIntervention—MICCAI2021, Lecture NotesinComputerScience,pages55–64.Springer,2021. doi: 10.1007/978-3-030-87231-1_6. arXiv:2103.00634

    Pith/arXiv arXiv 2021

  29. [29]

    A deep learning architecture for limited- angle computed tomography reconstruction

    Kerstin Hammernik, Tobias Würfl, Thomas Pock, and Andreas Maier. A deep learning architecture for limited- angle computed tomography reconstruction. InBildverarbeitung für die Medizin 2017, Informatik aktuell, pages 92–97.Springer,2017. doi: 10.1007/978-3-662-54345-0_25

    work page doi:10.1007/978-3-662-54345-0_25 2017

  30. [30]

    On the benefit of dual-domain denoisinginaself-supervisedlow-doseCTsetting

    Fabian Wagner, Mareike Thies, Laura Pfaff, Oliver Aust, Sabrina Pechmann, Daniela Weidner, Noah Maul, Max- imilian Rohleder, Mingxuan Gu, Jonas Utz, Felix Denzinger, and Andreas Maier. On the benefit of dual-domain denoisinginaself-supervisedlow-doseCTsetting. In2023IEEE20thInternationalSymposiumonBiomedicalImag- ing(ISBI),pages1–5.IEEE,2023. doi: 10.1109...

    arXiv 2023

  31. [31]

    Hendriksen, Daniël M

    Allard A. Hendriksen, Daniël M. Pelt, and K. Joost Batenburg. Noise2inverse: Self-supervised deep convolutional denoisingfortomography.IEEETransactionsonComputationalImaging,6:1320–1335,2020. doi: 10.1109/TCI.2020. 3019647

    work page doi:10.1109/tci.2020 2020

  32. [32]

    DM4CT:Benchmarkingdiffusionmodelsforcomputedtomog- raphyreconstruction

    JiayangShi,DaniëlM.Pelt,andK.JoostBatenburg. DM4CT:Benchmarkingdiffusionmodelsforcomputedtomog- raphyreconstruction. InInternationalConferenceonLearningRepresentations(ICLR),2026. arXiv:2602.18589

    arXiv 2026

  33. [33]

    Paul Friedrich, Julia Wolleb, Florentin Bieder, Alicia Durrer, and Philippe C. Cattin. WDM: 3d wavelet diffusion models for high-resolution medical image synthesis. InProc. 4th MICCAI Workshop on Deep Generative Models (DGM4MICCAI),LNCS,pages11–21.Springer,2024. doi: 10.1007/978-3-031-72744-3. arXiv:2402.19043

    Pith/arXiv arXiv 2024

  34. [34]

    Solvinginverseproblemsinmedicalimagingwithscore-based generativemodels

    YangSong,LiyueShen,LeiXing,andStefanoErmon. Solvinginverseproblemsinmedicalimagingwithscore-based generativemodels. InInternationalConferenceonLearningRepresentations(ICLR),2022. arXiv:2111.08005

    Pith/arXiv arXiv 2022

  35. [35]

    McCann, Marc L

    Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, and Jong Chul Ye. Diffusion posterior sam- pling for general noisy inverse problems. InInternational Conference on Learning Representations (ICLR), 2023. arXiv:2209.14687

    Pith/arXiv arXiv 2023

  36. [36]

    doi: 10.1007/978-3-031-16446-0_42

    RuyiZha,YanhaoZhang,andHongdongLi.NAF:Neuralattenuationfieldsforsparse-viewCBCTreconstruction.In MedicalImageComputingandComputer-AssistedIntervention—MICCAI2022,LectureNotesinComputerScience, pages442–452.Springer,2022. doi: 10.1007/978-3-031-16446-0_42. arXiv:2209.14540

    Pith/arXiv arXiv 2022

  37. [37]

    R2-Gaussian: Rectifying radia- tive Gaussian splatting for tomographic reconstruction

    Ruyi Zha, Tao Jun Lin, Yuanhao Cai, Jiwen Cao, Yanhao Zhang, and Hongdong Li. R2-Gaussian: Rectifying radia- tive Gaussian splatting for tomographic reconstruction. InAdvances in Neural Information Processing Systems 37 (NeurIPS2024),2024. arXiv:2405.20693

    Pith/arXiv arXiv 2024

  38. [38]

    A review on 3D Gaussian splattingforsparseviewreconstruction.ArtificialIntelligenceReview,58:215,2025.doi: 10.1007/s10462-025-11171-4

    Haitian Liu, Binglin Liu, Qianchao Hu, Peilun Du, Jing Li, Yang Bao, and Feng Wang. A review on 3D Gaussian splattingforsparseviewreconstruction.ArtificialIntelligenceReview,58:215,2025.doi: 10.1007/s10462-025-11171-4

    work page doi:10.1007/s10462-025-11171-4 2025

  39. [39]

    Single-slicerebinningmethodforhelicalcone-beamCT.Physics inMedicine&Biology,44(2):561–570,1999

    FrédéricNoo,MichelDefrise,andRolfClackdoyle. Single-slicerebinningmethodforhelicalcone-beamCT.Physics inMedicine&Biology,44(2):561–570,1999. doi: 10.1088/0031-9155/44/2/019. 16

    work page doi:10.1088/0031-9155/44/2/019 1999

  40. [2024]

    doi: 10.1088/1361-6560/ad3db9

    work page doi:10.1088/1361-6560/ad3db9