Pith. sign in

REVIEW 3 major objections 6 minor 76 references

A single agent can replan radiotherapy for the day's anatomy in about two minutes, with target dose errors generally within 2 Gy of clinical plans.

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

2026-08-02 00:49 UTC pith:TK5GT3BF

load-bearing objection The real contribution is a unified multi-task foundation model plus agent orchestration—but the 'complete adaptive plan' claim does not match a system that predicts a dose map and never checks deliverability. the 3 major comments →

arxiv 2607.14870 v1 pith:TK5GT3BF submitted 2026-07-16 physics.med-ph

One-for-All Adaptive Radiotherapy Planning Agent: A Foundation Framework for Daily CBCT-guided Radiotherapy

Shaoyan Pan , Kirk Jon Luca , Yuan Gao , Shansong Wang , Mingzhe Hu , Ryan Sanford , Mojtaba Safari , Justin Roper
show 3 more authors
Zhen Tian Tonghe Wang Xiaofeng Yang
This is my paper
classification physics.med-ph
keywords adaptive radiotherapycone-beam CTfoundation modeldose predictionsynthetic CTimage registrationsegmentationagentic AI
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 tries to show that adaptive radiotherapy planning—normally a multi-hour, expert-driven process—can be automated end to end by one system. The One-for-All Adaptive Radiotherapy Planning Agent (A-RPA) takes the cone-beam CT acquired just before treatment, synthesizes a corrected CT, registers prior images, resegments tumors and organs, and predicts a new dose distribution in roughly two minutes. The authors report that across head-and-neck, lung, abdominal, and prostate cancers, and for both photon and proton therapy, the resulting plans match clinically delivered plans closely enough that target dose errors (D98) are generally within 2.0 Gy, and expert review found 93–99% of photon and 94–96% of proton cases clinically acceptable. If this holds, daily adaptation could become a routine clinical option rather than a resource-intensive exception.

Core claim

The paper's central claim is that a unified foundation model plus an agent orchestration layer can perform complete, treatment-specific online adaptive planning directly from daily CBCT, selecting the appropriate adaptation pathway (plan reuse, registration-based propagation, re-segmentation, or full synthetic-CT replanning) and producing a dose distribution comparable to the clinical reference. The evidence offered is dosimetric: D95/D98 errors of 0.5–2.0 Gy for photons and 0.5–1.6 Gy for protons, low voxel-wise mean absolute error, high gamma pass rates, organ-at-risk mean and max dose errors within a few Gy, and physicist-assigned acceptability ranks. A distinctive secondary claim is that

What carries the argument

The load-bearing mechanism is a latent-space cascade with two backbones: a conditional latent diffusion model generates the synthetic CT and later the 3D dose distribution, while a vision-transformer encoder with lightweight 3D adapters provides anatomical features for deformable registration and segmentation. The two are fused and the whole chain is post-trained jointly on longitudinal clinical data using progressive supervision, so each upstream module's loss includes the losses of the tasks that follow it. On top of this sits a large-language-model agent that reads the clinician's query, keeps a cache of intermediate results, and chooses the cheapest workflow that answers the day's clinic

Load-bearing premise

The central bet is that a predicted dose distribution—generated by a diffusion model from synthetic CT and contours—is a complete adaptive treatment plan; if a usable plan requires inverse optimization with deliverable machine parameters, the two-minute end-to-end claim is not yet established.

What would settle it

Take a set of A-RPA's predicted dose distributions and run them through the standard inverse-optimization and leaf-sequencing step in a treatment planning system; if the deliverable plans' D98 deviates from the prediction by more than about 2 Gy on a meaningful fraction of cases, the claim that A-RPA produces complete plans fails. A second check: evaluate on an external cohort with different CBCT scanners and reconstruction pipelines—severe artifacts or a limited field of view should degrade performance if the single-institution training limitation is real.

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

If this is right

  • If A-RPA works as claimed, daily cone-beam CT becomes sufficient input for adaptive replanning, removing the need for repeated CT simulation.
  • The two-minute runtime puts adaptation inside the treatment slot, enabling per-fraction plan updates instead of weekly or one-time adaptation.
  • Joint training suggests that isolated task models underdeliver clinically; end-to-end optimization across the planning cascade is what makes intermediate outputs usable.
  • The human-in-the-loop design means the system can serve as decision support that escalates from simple plan check to full replanning based on clinician intent.
  • Because the framework spans multiple disease sites and both photon and proton therapy, the same model could standardize adaptive planning across diverse clinical scenarios.

Where Pith is reading between the lines

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

  • A natural next step the authors leave implicit: convert predicted dose into deliverable machine parameters (MLC sequences, monitor units) and verify with independent dose calculation; the paper's own dose-prediction caveat makes this the decisive test for the 'complete plan' claim.
  • The joint-optimization result implies that image-similarity benchmarks for CBCT-to-CT synthesis can be misleading; downstream clinical utility should be the primary evaluation criterion.
  • If the single-institution training limitation is addressed, the same architecture could be retrained on multi-institutional CBCT distributions with different scanners and artifact levels; a testable prediction is that such retraining would close the external-generalization gap without architectural change.
  • Because the agent caches intermediate results, temporal learning across fractions could be added to model anatomical trends, turning daily replanning into predictive replanning.

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 / 6 minor

Summary. The paper presents A-RPA, an agentic foundation-model framework for online adaptive radiotherapy. From a daily CBCT and clinician queries, the system orchestrates four modules: CBCT-to-CT synthesis (latent diffusion), multimodal deformable registration (DINOv2-based), structure segmentation, and 3D dose prediction (latent diffusion). All modules are first fine-tuned on public data and then jointly post-trained on an internal Emory cohort of 814 treatment courses spanning head-and-neck, brain, lung, abdomen, and pelvis, for both photon and proton therapy. The authors report dose errors (D95/D98 generally within 0.5–2.0 Gy), high gamma pass rates, improved segmentation on synthetic CT relative to baseline methods, and physicist-ranked acceptability in 93–99% of photon and 94–96% of proton cases. The central claim is that A-RPA performs 'complete, treatment-specific online adaptive planning' in under two minutes.

Significance. If the central claim were fully established, this would be a substantial contribution: it would demonstrate a single unified system performing the full adaptive replanning cascade, with human-in-the-loop control, across multiple sites and modalities. The manuscript has notable strengths: a large multi-task training corpus (public datasets plus 814 internal courses and 17,398 CBCTs), integration of synthesis-registration-segmentation-dose prediction in one framework, systematic comparison against several state-of-the-art baselines, and clinically oriented evaluation by physicists. These are genuine and should be credited. However, the paper's headline claim overreaches what is demonstrated: the final output is a predicted dose distribution, not a deliverable treatment plan with machine parameters and inverse optimization, and the evaluation largely measures agreement with the very clinical plans used as training labels. The significance is therefore currently more as an engineering and systems demonstration than as proof of a clinically deployable 'complete planning agent.'

major comments (3)
  1. [Abstract / Results / Methods (Fine-tuning 4, Eq. (7))] The abstract and Results claim A-RPA 'executes the final clinical plan design' and produces 'complete, treatment-specific online adaptive planning.' However, the dose module in Methods (Fine-tuning 4, Eq. (7)) only predicts a 3D dose distribution. No monitor units, MLC leaf sequences, proton spot weights, or inverse optimization are generated, and no deliverability check is performed. The Discussion concedes that dose prediction 'does not fully replace inverse optimization.' Thus the output is a dose map, not a deliverable treatment plan, and the claimed 'complete plan' is not established. This mismatch should either be resolved by adding inverse-optimization/deliverable plan generation and verification, or the claims must be substantially revised to describe a dose-prediction and decision-support system.
  2. [Results (Fig. 2d,e) / Methods (Fine-tuning 4, Eq. (7))] The evaluation is partly circular. The dose prediction module is trained on clinical reference dose distributions (Eq. (7)), and the primary success metrics are agreement with those same reference plans: D95/D98 errors, MAE, gamma, and the rank definitions in Fig. 2d/e, where ranks 3–5 are substantially defined by dose differences <2 Gy (targets) and <3 Gy (OARs) from the reference. This rewards mimicry of the training distribution rather than independent clinical acceptability. The claim of clinical acceptability requires validation that does not use the reference dose as the target, e.g., independent recalculation of the predicted dose in a treatment planning system with machine parameters, constraint-based scoring without reference comparison, or blinded multi-institutional clinician review.
  3. [Methods (Datasets) / Discussion (Limitations)] All joint post-training and evaluation is performed on a single internal institution's cohort (814 Emory treatment courses). The limitations paragraph acknowledges that performance may differ for other scanners, reconstruction pipelines, and artifact levels, but the abstract's claim of generalization across 'diverse clinical scenarios' is not supported by the evidence. In addition, the physicist acceptability review appears to be non-blinded and single-institution; no inter-rater reliability or blinding procedure is reported. External multi-institutional validation, or at minimum a clearly blinded evaluation, is needed before the clinical-accuracy claims can be accepted.
minor comments (6)
  1. [Agent orchestration (Methods)] The paper inconsistently names the agent's language model: the Introduction says T5-XXL, while Methods 'Agent orchestration' states GPT-4o is the central controller. Please clarify which model is used and how it relates to the T5-XXL described earlier.
  2. [Results / Fig. 1d] The 'approximately 2 minutes' runtime is stated in the abstract, Fig. 1d, and Discussion, but no timing experiment, hardware specification, or variance is reported. Provide a concrete method for measuring end-to-end runtime, including the agent decision steps and all module inference times.
  3. [Fig. 3c] The comparison of 'dose calculation on the original CBCT' is not described. What plan was used for recalculation, and how was CBCT density converted for dose calculation? Specify the calculation method to make the comparison reproducible.
  4. [Methods (Fine-tuning 3)] The shared 76-channel label space is mentioned but not defined. Please list the structures in each site-specific group and state how targets/OARs are mapped to this common space.
  5. [Results (Clinical acceptability)] Rank 5 is defined as 'better than the original plan across multiple metrics,' but the specific metrics and the aggregation rule are not stated. Define the criteria for 'better' and how a case is assigned to rank 5 versus rank 4.
  6. [Data Availability] Public dataset URLs are provided, but no code or model weights are released. Providing executable code and trained checkpoints would substantially improve reproducibility and enable independent validation of the claimed results.

Circularity Check

0 steps flagged

No circularity found: the dose model is trained and evaluated on held-out clinical reference plans; the 'complete plan' overclaim is a scope/validity gap, not a circular derivation.

full rationale

The paper's derivation is a standard supervised pipeline: the dose module is trained with a denoising objective (Eq. 7) on reference dose latents, and evaluation compares generated dose distributions with clinically delivered reference plans on a held-out 30% of the internal cohort. There is no equation in which the output is defined by the evaluation metric, no parameter fitted to the test set, and no load-bearing self-citation or imported uniqueness theorem. The clinical acceptability ranks are thresholded agreement scores relative to reference plans; because those reference plans are the training labels, high agreement on held-out data is a legitimate empirical result, not a reduction-by-construction. The Discussion's concession that dose prediction 'does not fully replace inverse optimization' and that no MUs/MLC/deliverability are generated is a real limitation of the 'complete plan' claim, but it is a scope/validity gap, not a circular step: the predicted dose map is still produced by a model whose inference inputs (sCT, contours, prior dose, metadata) differ from its output, and whose held-out evaluation is independent. No pattern of renaming, ansatz-smuggling via self-citation, or fitted-input-called-prediction is present. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claim rests on a cascade of hand-weighted losses and unvalidated transfer assumptions. The most consequential assumption is that predicted dose is equivalent to a complete plan, and the most fragile methodological premise is that single-institution non-blinded review is sufficient evidence of clinical acceptability.

free parameters (4)
  • Joint post-training loss weights = λ1=0.4, λ2=0.4, λ3=0.7, λ4=1
    Hand-chosen weights in Eq 12; no sensitivity analysis reported. They determine how much each upstream module is shaped by downstream losses.
  • Registration loss weights = λmask=1, λgroup=0.4, λsmooth=0.15
    Hand-chosen in Eq 5; affect the trade-off between image similarity, mask consistency, groupwise agreement, and deformation smoothness.
  • Segmentation consistency weight = λreg=1
    Hand-chosen in Eq 6; balances segmentation Dice losses against deformation image consistency.
  • Prior-dose dropout rate = not reported
    Introduced during dose-module training to reduce overreliance on prior dose; the value is not given, so the training behavior cannot be reproduced.
axioms (6)
  • domain assumption A predicted dose distribution generated by a diffusion model, without inverse optimization or deliverable beam parameters, is a valid proxy for a complete adaptive treatment plan.
    The 'complete planning' claim depends on this; the Discussion acknowledges dose prediction 'does not fully replace inverse optimization.'
  • domain assumption Non-blinded medical physicist review of generated plans against the same institution's clinical reference is a valid clinical acceptability ground truth.
    No blinding or external audit is described; the rank definitions reward agreement with the reference dose distribution.
  • ad hoc to paper A frozen DINOv2 trained on natural images, with 3D adapters, provides suitable volumetric medical representations for registration and segmentation.
    No ablations are shown to validate this transfer; it is a core architectural premise.
  • ad hoc to paper Stable Diffusion 3.5's frozen VAE and text-conditioning generalize to 3D CT and dose latent spaces.
    The paper does not validate reconstruction fidelity of the frozen VAE for CT or dose volumes.
  • standard math Standard diffusion denoising L2 loss is an adequate training objective for dose prediction.
    Eq 7; standard practice in diffusion models, but dose distributions have clinical constraints that are not encoded in the loss.
  • ad hoc to paper Cascaded module-level supervision with hand-weighted downstream losses approximates end-to-end joint optimization.
    Equations 8-12 define the cascade; no evidence is provided that this weighting scheme is close to optimal.

pith-pipeline@v1.3.0-alltime-deepseek · 20331 in / 12549 out tokens · 117611 ms · 2026-08-02T00:49:11.499933+00:00 · methodology

0 comments
read the original abstract

In this work, we introduce the One-for-All Adaptive Radiotherapy Planning Agent, a unified foundation-model-based system that performs complete, treatment-specific online adaptive planning directly from daily cone-beam CT in under two minutes. The agent first autonomously predicts all essential planning components, including synthetic CT generation, multimodal alignment, and tumor/organ segmentation. It then intelligently leverages these outputs to execute the final clinical plan design, providing a comprehensive, automated solution for daily treatment. We also demonstrate that the agent enables clinicians to define planning with intent and intervene at critical decision points, ensuring a "human-in-the-loop" framework that generates acceptable plans before final approval. Evaluated on multiple datasets spanning head-and-neck, lung, abdominal, and prostate cancers with both photon and proton therapy, the proposed framework achieves clinically acceptable accuracy and plan quality comparable to clinically generated treatment plans, with target dose errors (D98) generally within 2.0 Gy of the reference plan. The strong performance of the One-for-All agent highlights the promise of a unified foundation-model approach and opens opportunities for fast, scalable, and fully automated online adaptive radiotherapy across diverse clinical scenarios.

Figures

Figures reproduced from arXiv: 2607.14870 by Justin Roper, Kirk Jon Luca, Mingzhe Hu, Mojtaba Safari, Ryan Sanford, Shansong Wang, Shaoyan Pan, Tonghe Wang, Xiaofeng Yang, Yuan Gao, Zhen Tian.

Figure 1
Figure 1. Figure 1: Overview of the One-for-All adaptive radiotherapy planning framework. a, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative clinician–agent interaction for head-and-neck adaptive radiotherapy [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Synthetic CT quality and downstream benefit for adaptive radiotherapy planning. a, [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Multimodal registration quality and downstream benefit for adaptive radiotherapy [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Segmentation quality and downstream benefit for adaptive radiotherapy planning. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

76 extracted references · 6 linked inside Pith

  1. [1]

    Sung et al

    H. Sung et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.CA: A Cancer Journal for Clinicians, 71:209–249, 2021

    2020

  2. [2]

    M. B. Barton et al. Estimating the demand for radiotherapy from the evidence: A review of changes from 2003 to 2012.Radiotherapy and Oncology, 112:140–144, 2014

    2003

  3. [3]

    S. S. Ahmad et al. Clinical development of novel drug-radiotherapy combinations.Clinical Cancer Research, 25:1455–1461, 2019

    2019

  4. [4]

    De Ruysscher et al

    D. De Ruysscher et al. Radiotherapy toxicity.Nature Reviews Disease Primers, 5:13, 2019

    2019

  5. [5]

    Atun et al

    R. Atun et al. Expanding global access to radiotherapy.The Lancet Oncology, 16:1153–1186, 2015

    2015

  6. [6]

    M. Goitein. The comparison of treatment plans.Seminars in Radiation Oncology, 2:246–256, 1992

    1992

  7. [7]

    L. B. Marks et al. Use of normal tissue complication probability models in the clinic.Interna- tional Journal of Radiation Oncology*Biology*Physics, 76:S10–S19, 2010

    2010

  8. [8]

    D. A. Jaffray. Image-guided radiotherapy: from current concept to future perspectives.Nature Reviews Clinical Oncology, 9:688–699, 2012

    2012

  9. [9]

    Beaton, S

    L. Beaton, S. Bandula, M. N. Gaze, and R. A. Sharma. How rapid advances in imaging are defining the future of precision radiation oncology.British Journal of Cancer, 120:779–790, 2019. 29

    2019

  10. [10]

    Bobić et al

    M. Bobić et al. Large anatomical changes in head-and-neck cancers – a dosimetric comparison of online and offline adaptive proton therapy.Clinical and Translational Radiation Oncology, 40, 2023

    2023

  11. [11]

    K. M. Langen and D. T. L. Jones. Organ motion and its management.International Journal of Radiation Oncology*Biology*Physics, 50:265–278, 2001

    2001

  12. [12]

    J. L. Barker et al. Quantification of volumetric and geometric changes occurring during fractionated radiotherapy for head-and-neck cancer using an integrated CT/linear accelerator system.International Journal of Radiation Oncology*Biology*Physics, 59:960–970, 2004

    2004

  13. [13]

    van der Horst, A

    A. van der Horst, A. C. Houweling, G. van Tienhoven, J. Visser, and A. Bel. Dosimetric effects of anatomical changes during fractionated photon radiation therapy in pancreatic cancer patients.Journal of Applied Clinical Medical Physics, 18:142–151, 2017

    2017

  14. [14]

    Hamming-Vrieze et al

    O. Hamming-Vrieze et al. Deterioration of intended target volume radiation dose due to anatomical changes in patients with head-and-neck cancer.Cancers (Basel), 13, 2021

    2021

  15. [15]

    Jafar and X

    A. Jafar and X. Jia. Towards human-centric intelligent treatment planning for radiation therapy. npj Digital Medicine, 9:155, 2026

    2026

  16. [16]

    D. L. Schwartz et al. Adaptive radiotherapy for head-and-neck cancer: Initial clinical outcomes from a prospective trial.International Journal of Radiation Oncology*Biology*Physics, 83:986– 993, 2012

    2012

  17. [17]

    J. M. Bryant et al. Structure-specific rigid dose accumulation dosimetric analysis of ablative stereotactic MRI-guided adaptive radiation therapy inultracentral lung lesions.Communications Medicine, 4:96, 2024

    2024

  18. [18]

    P. J. Keall et al. Integrated MRI-guided radiotherapy — opportunities and challenges.Nature Reviews Clinical Oncology, 19:458–470, 2022

    2022

  19. [19]

    Piperdi et al

    H. Piperdi et al. Adaptive radiation therapy in the treatment of lung cancer: An overview of the current state of the field.Frontiers in Oncology, 11:770382, 2021

    2021

  20. [20]

    O. M. Dona Lemus, M. Cao, B. Cai, M. Cummings, and D. Zheng. Adaptive radiotherapy: Next-generation radiotherapy.Cancers (Basel), 16, 2024

    2024

  21. [21]

    Yu et al

    L. Yu et al. Multicenter study on the versatility and adoption of AI-driven automated radiotherapy planning across cancer types.Nature Communications, 17:867, 2025

    2025

  22. [22]

    Hosny, C

    A. Hosny, C. Parmar, J. Quackenbush, L. H. Schwartz, and H. Aerts. Artificial intelligence in radiology.Nature Reviews Cancer, 18:500–510, 2018

    2018

  23. [23]

    K. N. McComas, A. Yock, K. Darrow, and E. T. Shinohara. Online adaptive radiation therapy and opportunity cost.Advances in Radiation Oncology, 8, 2023. 30

    2023

  24. [24]

    Zhang et al

    Y. Zhang et al. Breath-hold CBCT-guided CBCT-to-CT synthesis via multimodal unsupervised representation disentanglement learning.IEEE Transactions on Medical Imaging, 42:2313–2324, 2023

    2023

  25. [25]

    E. M. C. Huijben et al. Generating synthetic computed tomography for radiotherapy: Syn- thRAD2023 challenge report.Medical Image Analysis, 97:103276, 2024

    2024

  26. [26]

    Huang et al

    B. Huang et al. 3D lightweight network for simultaneous registration and segmentation of organs-at-risk in CT images of head and neck cancer.IEEE Transactions on Medical Imaging, 41:951–964, 2022

    2022

  27. [27]

    Nenoff et al

    L. Nenoff et al. Review and recommendations on deformable image registration uncertainties for radiotherapy applications.Physics in Medicine & Biology, 68, 2023

    2023

  28. [28]

    Cao et al

    X. Cao et al. Dual-core steered non-rigid registration for multi-modal images via bi-directional image synthesis.Medical Image Analysis, 41:18–31, 2017

    2017

  29. [29]

    Poel et al

    R. Poel et al. A comprehensive multifaceted technical evaluation framework for implementation of auto-segmentation models in radiotherapy.Communications Medicine, 5:319, 2025

    2025

  30. [30]

    Luo et al

    X. Luo et al. SegRap2023: A benchmark of organs-at-risk and gross tumor volume segmentation for radiotherapy planning of nasopharyngeal carcinoma.Medical Image Analysis, 101:103447, 2025

    2025

  31. [31]

    McIntosh et al

    C. McIntosh et al. Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer.Nature Medicine, 27:999–1005, 2021

    2021

  32. [32]

    Zhang et al

    Y. Zhang et al. DoseDiff: Distance-aware diffusion model for dose prediction in radiotherapy. IEEE Transactions on Medical Imaging, 43:3621–3633, 2024

    2024

  33. [33]

    J. Hu, Y. Song, Q. Wang, S. Bai, and Z. Yi. Incorporating historical sub-optimal deep neural networks for dose prediction in radiotherapy.Medical Image Analysis, 67:101886, 2021

    2021

  34. [34]

    Jiao et al

    Z. Jiao et al. TransDose: Transformer-based radiotherapy dose prediction from CT images guided by super-pixel-level GCN classification.Medical Image Analysis, 89:102902, 2023

    2023

  35. [35]

    Sapkota, K

    R. Sapkota, K. I. Roumeliotis, and M. Karkee. AI agents vs. agentic AI: A conceptual taxonomy, applications and challenges.Information Fusion, page 103599, 2025

    2025

  36. [36]

    Oquab et al

    M. Oquab et al. DINOv2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Raffel et al

    C. Raffel et al. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1–67, 2020

    2020

  38. [38]

    R. Gao, B. Lou, Z. Xu, D. Comaniciu, and A. Kamen. Flexible-Cm GAN: Towards precise 3D dose prediction in radiotherapy. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 715–725, 2023. 31

    2023

  39. [39]

    Miften et al

    M. Miften et al. Tolerance limits and methodologies for IMRT measurement-based verification QA: Recommendations of AAPM task group no. 218.Medical Physics, 45:e53–e83, 2018

    2018

  40. [40]

    Z. Wang, E. P. Simoncelli, and A. C. Bovik. Multiscale structural similarity for image quality assessment. InThe Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, volume 2, pages 1398–1402, 2003

    2003

  41. [41]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018

    2018

  42. [42]

    Özbey et al

    M. Özbey et al. Unsupervised medical image translation with adversarial diffusion models. IEEE Transactions on Medical Imaging, 42:3524–3539, 2023

    2023

  43. [43]

    Shi et al

    K. Shi et al. Anatomy-conserving unpaired CBCT-to-CT translation via schrödinger bridge. InMedical Image Computing and Computer Assisted Intervention – MICCAI 2025, Part IV, pages 46–55, Daejeon, Korea (Republic of), 2025. Springer-Verlag

    2025

  44. [44]

    Wasserthal et al

    J. Wasserthal et al. TotalSegmentator: robust segmentation of 104 anatomic structures in CT images.Radiology: Artificial Intelligence, 5:e230024, 2023

    2023

  45. [45]

    Isensee et al

    F. Isensee et al. nnU-Net revisited: A call for rigorous validation in 3D medical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 488–498. Springer, 2024

    2024

  46. [46]

    Hatamizadeh et al

    A. Hatamizadeh et al. Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images. InInternational MICCAI Brainlesion Workshop, pages 272–284. Springer, 2021

    2021

  47. [47]

    Zhou et al

    H.-Y. Zhou et al. nnFormer: Interleaved transformer for volumetric segmentation.arXiv preprint arXiv:2109.03201, 2021

    Pith/arXiv arXiv 2021

  48. [48]

    Roy et al

    S. Roy et al. MedNeXt: transformer-driven scaling of convnets for medical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 405–415. Springer, 2023

    2023

  49. [49]

    Jiang and H

    J. Jiang and H. Veeraraghavan. One shot PACS: Patient specific anatomic context and shape prior aware recurrent registration-segmentation of longitudinal thoracic cone beam CTs.IEEE Transactions on Medical Imaging, 2022

    2022

  50. [50]

    Xu et al

    S. Xu et al. Volumetric computed tomography with carbon nanotube X-ray source array for improved image quality and accuracy.Communications Engineering, 2:71, 2023

    2023

  51. [51]

    L. J. Wilson et al. Cone beam CT dose optimisation: A review and expert consensus by the 2022 ESTRO physics workshop IGRT working group.Radiotherapy and Oncology, 209:110958, 2025. 32

    2022

  52. [52]

    Buchanan, S

    L. Buchanan, S. Hamdan, Y. Zhang, X. Chen, and X. A. Li. Deep learning-based prediction of deliverable adaptive plans for MR-guided adaptive radiotherapy: A feasibility study.Frontiers in Oncology, 13:939951, 2023

    2023

  53. [53]

    Sun et al

    Z. Sun et al. A hybrid optimization strategy for deliverable intensity-modulated radiotherapy plan generation using deep learning-based dose prediction.Medical Physics, 49:1344–1356, 2022

    2022

  54. [54]

    Vorontsov et al

    E. Vorontsov et al. A foundation model for clinical-grade computational pathology and rare cancers detection.Nature Medicine, 30:2924–2935, 2024

    2024

  55. [55]

    Gao et al

    C. Gao et al. Synthetic data accelerates the development of generalizable learning-based algorithms for X-ray image analysis.Nature Machine Intelligence, 5:294–308, 2023

    2023

  56. [56]

    Y. C. Tham et al. Building the world’s first truly global medical foundation model.Nature Medicine, 31:3580–3585, 2025

    2025

  57. [57]

    Ding et al

    T. Ding et al. A multimodal whole-slide foundation model for pathology.Nature Medicine, 31:3749–3761, 2025

    2025

  58. [58]

    Thapa et al

    R. Thapa et al. A multimodal sleep foundation model for disease prediction.Nature Medicine, 32:752–762, 2026

    2026

  59. [59]

    Fan et al

    X. Fan et al. Bi-level learning of task-specific decoders for joint registration and one-shot medical image segmentation. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11726–11735, 2024

    2024

  60. [60]

    Gupta, A

    T. Gupta, A. Kamath, A. Kembhavi, and D. Hoiem. Towards general purpose vision systems: An end-to-end task-agnostic vision-language architecture. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16399–16409, 2022

    2022

  61. [61]

    Y. Ai, X. Zhou, H. Huang, L. Zhang, and R. He. Uncertainty-aware source-free adaptive image super-resolution with wavelet augmentation transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8142–8152, 2024

    2024

  62. [62]

    Liu et al

    J. Liu et al. Difflow3D: Toward robust uncertainty-aware scene flow estimation with iterative diffusion-based refinement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15109–15119, 2024

    2024

  63. [63]

    Tang et al

    H. Tang et al. Code repair with LLMs gives an exploration-exploitation tradeoff.Advances in Neural Information Processing Systems, 37:117954–117996, 2024

    2024

  64. [64]

    J. Hong, A. Dragan, and S. Levine. Planning without search: Refining frontier LLMs with offline goal-conditioned RL.arXiv preprint arXiv:2505.18098, 2025

    arXiv 2025

  65. [65]

    Uehara et al

    M. Uehara et al. Feedback efficient online fine-tuning of diffusion models.arXiv preprint arXiv:2402.16359, 2024. 33

    Pith/arXiv arXiv 2024

  66. [66]

    Fan et al

    Y. Fan et al. DPOK: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:79858–79885, 2023

    2023

  67. [67]

    Hiranaka et al

    A. Hiranaka et al. HERO: Human-feedback efficient reinforcement learning for online diffusion model finetuning.arXiv preprint arXiv:2410.05116, 2024

    Pith/arXiv arXiv 2024

  68. [68]

    S. Li, K. Kallidromitis, A. Gokul, Y. Kato, and K. Kozuka. Aligning diffusion models by optimizing human utility.Advances in Neural Information Processing Systems, 37:24897–24925, 2024

    2024

  69. [69]

    Pandya and M

    K. Pandya and M. Holia. Automating customer service using LangChain: Building custom open-source GPT chatbot for organizations.arXiv preprint arXiv:2310.05421, 2023

    Pith/arXiv arXiv 2023

  70. [70]

    Hurst et al

    A. Hurst et al. GPT-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  71. [71]

    Xu et al

    T. Xu et al. A generalizable 3D framework and model for self-supervised learning in medical imaging.npj Digital Medicine, 8:639, 2025

    2025

  72. [72]

    Esser et al

    P. Esser et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024

    2024

  73. [73]

    E. J. Hu et al. LoRA: Low-rank adaptation of large language models.International Conference on Learning Representations, 1:3, 2022

    2022

  74. [74]

    Chen et al

    J. Chen et al. TransMorph: Transformer for unsupervised medical image registration.Medical Image Analysis, 82:102615, 2022

    2022

  75. [75]

    M. P. Heinrich et al. MIND: Modality independent neighbourhood descriptor for multi-modal deformable registration.Medical Image Analysis, 16:1423–1435, 2012

    2012

  76. [76]

    Zhang, X

    Y. Zhang, X. Wu, H. M. Gach, H. Li, and D. Yang. GroupRegNet: a groupwise one-shot deep learning-based 4D image registration method.Physics in Medicine & Biology, 66:045030, 2021. 34

    2021