Pith. sign in

REVIEW 4 major objections 5 minor 50 references

CT-PrepAgent: Bounded Policy and Controlled Execution for Adaptive CT Data Preparation

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

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

Pith's one-line read CT-PrepAgent claims that splitting adaptive CT data preparation into a bounded semantic decision plus deterministic, verifier-gated execution yields safer and more productive preparation than fixed or expert workflows.

desk verdict Worth reading for the bounded-policy architecture, but the empirical case for adaptive superiority collapses to tie-breaking and a single-seed CI that includes zero. read the letter →

arxiv 2608.01233 v1 pith:QALY2EQ4 submitted 2026-08-02 cs.AI

Xiaolin Fan , Yue Pei , Yingying Zhang , Haogang Zhu This is my paper
classification cs.AI
keywords CTdatapreparationboundedpolicyLLMagentDICOMseriesselectionpreprocessingprofileverifiableexecutionmedicalimagesegmentationregistration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

CT-PrepAgent tries to show that CT data preparation can be made adaptive without sacrificing safety: instead of letting an LLM generate code, parameters, or workflow structures, the agent is confined to choosing one eligible DICOM series or one of 12 predefined geometry–intensity preprocessing profiles, while deterministic inspection, execution, and postcondition verification handle everything else. The paper reports that this bounded design achieves the highest macro-average Dice (0.7609) across three public CT segmentation tasks, increases verified output yield on private raw-DICOM cohorts from 61.7% to 70.0%, and maintains similar registration metrics on the outputs both approaches can produce. It also reports that bounded recovery raises correct fault disposition from 50% to 100% on a controlled 60-instance fault taxonomy and that frozen decisions can be replayed without any LLM calls while drift is detected at the intended boundaries. If correct, this gives clinical pipelines a way to add adaptive preparation without exposing them to unrestricted code or parameter generation.

What carries the argument

The load-bearing mechanism is the bounded decision space plus the Guard–Resolve–Execute–Verify flow. The decision space is a finite catalog: for NIfTI preparation, the Cartesian product of two geometry skills (fixed 1.5-mm isotropic or planner-derived isotropic) and six intensity skills (global z-score plus five anatomy-specific clamped piecewise-linear HU mappings); for DICOM, the set of deterministically eligible series $E_j$ with a canonical acquisition phase, or the non-execution actions fallback/quarantine. The controlled execution flow first guards the candidate decision (schema, profile consistency, eligibility membership), resolves it into a versioned deterministic operator specifica

What would settle it

Run the public segmentation comparison with multiple seeds (e.g., 10) per policy-condition under the same protocol; if the adaptive policy's macro-average Dice advantage over the fixed policy (notably the +0.0619 on AMOS) does not persist across seeds, the segmentation claim is within training variability. Separately, have independent radiologists annotate the intended series and phase for the 60 private DICOM studies; if CT-PrepAgent's selections disagree with those annotations, or its 'verified' outputs fail a clinical readiness checklist, the claimed 70% verified yield does not establish cl

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that separating semantic choice from image manipulation makes adaptive CT preparation both safer and more productive than fixed or expert workflows. A deterministic inspector converts raw input and task into structured data–task profiles $(F_D, F_T)$, a bounded policy returns one typed candidate from a finite catalog — an eligible DICOM series with a canonical phase, or one of $|C_g| \times |C_i| = 2 \times 6 = 12$ NIfTI geometry–intensity profiles — and a controlled execution flow guards, resolves, executes, and verifies the decision against task-specific postconditions $\Gamma_T$ (Eq. 3). Only explicitly recoverable outcomes trigger bounded co

Load-bearing premise

The results stand only if the paper's own postcondition verifier and its hand-authored fault taxonomy are the right success criteria: 'verified' output means passing $\Gamma_T$ and 'correct disposition' means matching the authors' expected outcomes, with no independent clinical annotations of intended series or real-world fault prevalence.

Editorial extensions

If this is right

  • Adaptive preparation can be deployed without letting an LLM generate code, numerical parameters, or workflow topology; policy choice is the only adaptive step.
  • Fixed, expert-authored, statistical, and LLM policies are interchangeable under the same inspector, operators, verifier, and evaluator, so policy comparisons isolate selection effects.
  • On raw DICOM, verified output yield rises from 61.7% to 70.0% by resolving eligible candidate ties, while studies with no eligible series remain safely quarantined rather than forced through.
  • Bounded recovery — one correction, one replan, and configured expert-SOP fallback — raises correct disposition from 50% to 100% on the tested 60-instance fault taxonomy without weakening verification.
  • Frozen decisions can be replayed with zero LLM calls; input and plan drift are rejected before execution and semantic drift is caught afterward, supporting audit and reproducibility.

Reading between the lines

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

  • The paper's own controls show that the five additional DICOM outputs come from breaking expert-priority ties, and random or canonical tie-breaking reproduces the same five outputs; so the yield gain is not evidence of LLM-specific reasoning, only of resolving ties inside the eligible set.
  • Because 'verified' means passing the paper's verifier $\Gamma_T$ and no independent intended-series annotations exist, the 70% yield figure is a statement about internal postconditions, not clinical correctness; an external radiological annotation study would be the decisive test.
  • The public segmentation advantage rests on a single training run per condition, and several policy pairs resolve to identical prepared payloads; multi-seed retraining is needed to separate preprocessing effects from training variability.
  • The bounded-policy pattern — finite verifiable action catalog, deterministic executor, postcondition gate — is portable to other image-analysis pipelines (e.g., MRI, pathology) where an agent's freedom to improvise is the main safety risk.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CT-PrepAgent, a framework that separates a bounded semantic decision (selecting one eligible DICOM series or one of 12 predefined NIfTI geometry–intensity profiles) from deterministic, guard-resolved, verifier-checked execution. The authors report three main empirical results: (i) on three public CT segmentation tasks, the LLM-based bounded policy P-D attains the highest macro-average Dice (0.7609) among fixed (P-A), expert-SOP (P-B), statistical (P-C), and LLM (P-D) policies; (ii) on two private raw-DICOM cohorts, CT-PrepAgent increases verified output yield from 61.7% to 70.0% relative to an expert-SOP baseline, with similar registration metrics on common verified outputs; and (iii) controlled fault and replay tests show bounded recovery, safe quarantine, and policy-free replay. The supplementary material documents unusually careful protocol choices: test partitions are excluded from profiling, decisions are frozen before training, a hash-based rule selects the private 30-study cohorts, and an equivalence audit is used to avoid attributing differences to preprocessing when prepared payloads are identical.

Significance. If the empirical claims were fully supported, CT-PrepAgent would be a valuable architectural contribution: it shows a concrete way to introduce LLM-based adaptivity into CT preparation without permitting unrestricted code, parameter, or workflow generation, and its deterministic verification/replay machinery is a sensible template for auditable medical-image agents. The strengths of the paper are real: all policies share the same deterministic operators; the inspection layer excludes test information; the private-cohort selection is hash-based and frozen; the supplementary equivalence audit correctly identifies cases where policy differences cannot be attributed to preprocessing; and the replay/fault experiments are controlled and clearly specified. However, as detailed below, the two headline operational gains (aggregate segmentation superiority and improved DICOM yield) are substantially weaker, and partly policy-independent, than the abstract implies. The underlying framework is sound, but the central claims need re-analysis and recalibration.

major comments (4)
  1. [§3.2, Fig. 3; Supplementary §4.1, Table 13] The private-DICOM headline 'increased verified output yield from 61.7% to 70.0%' is not caused by the LLM policy. The five additional outputs all come from CCTA studies with expert-priority ties, and Supplementary Table 13 states that random and canonical tie-breaking produce the same five additional outputs and that canonical tie-breaking matches all five CT-PrepAgent selections. Thus the yield gain is a property of resolving eligibility ties, not of semantic LLM reasoning. The abstract and Section 3.2 should either attribute the gain to the deterministic bounded decision framework (with the LLM as one interchangeable tie-resolver) or provide evidence that the LLM's ranking adds value beyond a trivial tie-break.
  2. [Supplementary §2.2, Table 9; Supplementary Table 10; Table 1] The claim that CT-PrepAgent 'achieved the highest macro-average Dice' is not statistically supported relative to the expert-SOP baseline P-B. On AMOS, the only dataset where P-D and P-B resolve to different prepared payloads, the 95% CI for the paired difference is [−0.0012, +0.0172], which includes zero. On TsBone and TsLung, P-B and P-D are preprocessing-equivalent (Supplementary Table 10), so the reported differences are downstream training variability, not policy effects. Moreover, every segmentation condition is a single training run with seed 42 (Supplementary §2.2). The aggregate macro-average advantage (0.7609 vs. 0.7577) therefore rests on an uncertain AMOS difference plus equivalent-payload noise. The authors should either provide multiple seeds/error bars for the macro-average, or soften the claim to 'comparable to expert-SOP, with gains over fixed/statistical policies on anis
  3. [Abstract, §3.2; Supplementary §4.1, §5.1] The terms 'verified output' and 'correct disposition' are defined by the system's own verifier (Γ_T in Eq. 3) and by the authors' 60-instance fault taxonomy (Eq. 14; Supplementary §5.1). The supplementary material explicitly acknowledges that independent intended-series/phase annotations are unavailable and that the fault tests 'do not assess real-world fault prevalence or clinical safety.' Given these admissions, the abstract's unqualified statements about 'verified output yield' and 'safe quarantine' overstate what has been established. These qualifications should be moved into the main-text results and abstract, not confined to the supplement, because a reader could otherwise infer clinical selection accuracy or safety claims that the evaluation does not support.
  4. [§3.2, Table 2; Supplementary §4.1] The productivity value of the five additional DICOM outputs is not measured. Table 2 compares registration metrics only on the 37 studies that D-A and D-D resolve jointly; the authors acknowledge that the comparison 'does not assess registration utility for the five outputs.' Since the central claim is that CT-PrepAgent is more productive, the paper should either report registration utility (or another downstream metric) for the additional outputs, or explicitly frame the yield gain as 'more outputs passing the internal verifier, with unknown downstream utility.'
minor comments (5)
  1. [Table 2] Title contains a typo: 'private DIDOM' should be 'private DICOM'.
  2. [Figure 1] Label 'Bounded Sematic Decision' should be 'Bounded Semantic Decision'.
  3. [§2.3] The sentence 'the policy performs only context-dependent semantic decision space' appears to be missing a verb or intended word; likely '... performs only context-dependent semantic decision-making'.
  4. [Table 1 / Figure 2] The per-structure Dice differences in Figure 2 would be easier to interpret if the color scale were explicitly calibrated in the caption and if the values corresponding to each cell were printable at the figure resolution.
  5. [Supplementary §2.2] The paper should state explicitly whether the same downstream training seed is the only source of stochasticity or whether data loading/GPU nondeterminism was controlled; this would clarify the equivalence-pair differences in Supplementary Table 10.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central results are externally benchmarked; internal success metrics are explicitly scoped.

full rationale

The paper's central segmentation and registration claims are anchored to external benchmarks, not to the system's own definitions. Public NIfTI segmentation is evaluated on held-out test labels from AMOS and TotalSegmentator, with test cases excluded from profiling (Supp. 1.1, 2.2); registration is evaluated with known rigid perturbations and true target registration error (Sec. 3.1, Supp. 2.3). These outcomes are not defined by CT-PrepAgent. The 'verified output yield' metric is indeed defined by the paper's own postcondition verifier (Eq. 3, Gamma_T), and 'correct disposition' by the authors' fault taxonomy (Eq. 14; Supp. 5.1). However, the paper explicitly disclaims external generalization in these settings: Supp. 4.1 states 'independent intended-series and phase annotations are unavailable, these results do not establish clinical selection accuracy,' and Supp. 5.1 states the fault tests 'do not assess real-world fault prevalence or clinical safety.' These are scope limitations, not circular derivations. The DICOM yield gain is decomposed in Supp. Table 13 and Supp. 4.1: all five added outputs come from resolving expert-priority ties, and random/canonical tie-breaking matches CT-PrepAgent's selections. This is a causal-attribution and ablation concern about the LLM's specific contribution, but the paper itself states the gain 'arises from resolving eligible ties rather than LLM-specific reasoning.' It is not a case of a prediction being equivalent to its inputs by construction. Similarly, the segmentation comparison is weakened by single-seed training (Supp. 2.2), by P-D and P-B sharing identical prepared payloads on TsBone and TsLung (Supp. Table 10), and by the AMOS P-D vs P-B confidence interval including zero (Supp. Table 9). These are statistical-strength and interpretation issues, not circularity. There are no load-bearing self-citations, no imported uniqueness theorems, and no parameter fitted to a target and then renamed as a prediction. Overall, the paper's derivation chain is self-contained against external benchmarks, and its internal metrics are honestly scoped.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The framework's central claims rest on hand-authored decision rules, self-defined verification and fault taxonomies, single-seed training runs, and an assumption of deterministic LLM behavior. These are disclosed in the main text and supplement and are enumerated here because each one bounds the generality of the reported yield, safety, and utility numbers. No physical or ontological entities are postulated: components such as the bounded policy view, the eligible set, the semantic signature, and linked run records are software constructs with specified behavior and carry no independent falsifiable handle beyond the paper's own implementation.

free parameters (4)
  • Planner geometry swap threshold and replacement percentile = axis ratio >= 3x; replacement at 10th percentile
    Hand-authored resolver rules (Supplementary Section 1.2) that fix isotropic spacing on anisotropic data; they produce the 1.1-mm AMOS spacing behind the main segmentation gain and are not validated against a broader corpus.
  • Anatomy-specific intensity profile breakpoints = lung (-1000,0)->(-200,1); bone (200,0)->(1500,1); mediastinum (-200,0)->(350,1); etc.
    Hand-specified HU mappings (Supplementary Eq. 2) define the six-intensity action space; their breakpoints shape whichever profile P-D selects and hence downstream Dice.
  • Shared executor settings (body crop, 144-mm field-of-view patch) = 144 mm
    Fixed resolver-controlled settings (Supplementary Section 1.2) assumed appropriate for all three public tasks without ablation.
  • Eligibility and QC thresholds (pass/review categories, hard-rejection rules) = categorical thresholds
    Hand-defined deterministic gates (Eq. 7) that decide which DICOM series are eligible; since the yield claim counts only eligible-series outputs, these thresholds directly bound the headline yield improvement.
assumptions (6)
  • ad hoc to paper The verifier postconditions (Gamma_T in Eq. 3) are a valid proxy for task-readiness and for correctness of DICOM series selection.
    All 'verified yield' numbers are outputs that pass the authors' own postcondition checks; Supplementary Section 4.1 concedes there are no independent intended-series annotations, so clinical correctness of selections is not established.
  • ad hoc to paper The 60 constructed fault instances represent real failure modes of clinical DICOM cohorts.
    Supplementary Section 5.1 states the tests 'do not assess real-world fault prevalence or clinical safety'; the 100% correct-disposition result is conditional on the authors' six-category fault taxonomy.
  • domain assumption Training-set statistics generalize to test cases for profile planning and intensity calibration.
    Standard ML assumption used in Section 2.2; mitigates leakage by excluding the test partition but assumes stable acquisition statistics across the cohort.
  • domain assumption The pinned LLM at temperature zero returns deterministic, reproducible decisions.
    P-D uses one temperature-zero call to Qwen3-Next-80B-A3B-Instruct-6bit (Supplementary Section 2.4); no sampling seed, decoding parameters, or model snapshot hash is given, so replay and replication rely on API determinism.
  • domain assumption One training run per condition isolates the effect of preprocessing choice.
    Each segmentation condition is a single seed-42 run (Supplementary Section 2.2); the bootstrap CIs in Table 9 are case-level only and explicitly exclude training variability, so preprocessing effects are confounded with training stochasticity.
  • standard math Standard background: Dice as defined, TRE under known rigid transforms, and standard interpolation behave as expected.
    Metric definitions in Supplementary Sections 2.2 and 2.3 follow standard practice and are not in dispute.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CT-PrepAgent: Bounded Policy and Controlled Execution for Adaptive CT Data Preparation." pith.science (2026). https://pith.science/paper/QALY2EQ4

@misc{pith2026260801233,
  author       = {Pith},
  title        = {Pith review of: CT-PrepAgent: Bounded Policy and Controlled Execution for Adaptive CT Data Preparation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QALY2EQ4}},
  note         = {Machine review of arXiv:2608.01233}
}
read the original abstract

Heterogeneous computed tomography (CT) acquisitions and diverse downstream task requirements limit the transferability of fixed data preparation workflows across data sources and tasks. Existing approaches typically rely on manually designed or dataset-specific rules, making it difficult to accommodate changes in acquisition conditions and analytical objectives without manual intervention. Large language model (LLM)-based agents have shown promise for automating medical workflows, yet their potential for adaptive CT data preparation remains largely unexplored. To bridge this gap, we propose CT-PrepAgent, which enables adaptive CT data preparation through a bounded policy and controlled deterministic execution. Deterministic inspection constructs structured data--task profiles, from which a policy decides an eligible DICOM series or predefined preprocessing profile, while the controlled execution flow guards, resolves, executes, and verifies the decision with bounded recovery when enabled and safe quarantine otherwise. Across three public CT segmentation tasks, CT-PrepAgent derived data-task adaptive preprocessing decisions and achieved the highest macro-average Dice. On two private raw-DICOM cohorts, CT-PrepAgent increased verified output yield from 61.7\% to 70.0\% and yielded similar registration metrics on common verified outputs. Controlled fault and replay tests validate bounded recovery, safe quarantine, and policy-free replay under tested fault and drift settings.

Figures

Figures reproduced from arXiv: 2608.01233 by the authors.

Figure 1
Figure 1. Overview of CT-PrepAgent. The deterministic inspection constructs structured data-task profiles from public NIfTI [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. AMOS organ-level Dice differences for P-D relative to each baseline under the test protocol. Values are Dice percentage [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Private DICOM preparation results. (a) Verified outputs and safe quarantines for the Expert-SOP (D-A), one-shot [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Controlled execution ablation on 60 controlled [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [1]

    Computer Methods and Programs in Biomedicine , volume=

    Lazy Resampling: Fast and Information Preserving Preprocessing for Deep Learning , author=. Computer Methods and Programs in Biomedicine , volume=. 2024 , publisher=

    work page 2024

  2. [2]

    Computerized Medical Imaging and Graphics , volume=

    Improving Segmentation and Detection of Lesions in CT Scans Using Intensity Distribution Supervision , author=. Computerized Medical Imaging and Graphics , volume=. 2023 , publisher=

    work page 2023

  3. [3]

    and Gomez-Cardona, Daniel G

    Nelson, Brandon J. and Gomez-Cardona, Daniel G. and Thorne, Jamison E. and Huber, Nathan R. and Yu, Lifeng and Leng, Shuai and McCollough, Cynthia H. and Missert, Andrew D. , journal=. Multiple Kernel Synthesis of Head. 2024 , doi=

    work page 2024

  4. [4]

    and Ashburner, John and Smith, Jolinda and Rorden, Christopher , journal=

    Li, Xiangrui and Morgan, Paul S. and Ashburner, John and Smith, Jolinda and Rorden, Christopher , journal=. The First Step for Neuroimaging Data Analysis:. 2016 , doi=

    work page 2016

  5. [5]

    Cancers , volume=

    Impact of Preprocessing Parameters in Medical Imaging-Based Radiomic Studies: A Systematic Review , author=. Cancers , volume=. 2024 , doi=

    work page 2024

  6. [6]

    Journal of Machine Learning for Biomedical Imaging , volume=

    Buddenkotte, Thomas and Opfer, Roland and Kr. Journal of Machine Learning for Biomedical Imaging , volume=. 2024 , doi=

    work page 2024

  7. [7]

    Mixture-of-Experts Network via Frequency-Causal Reasoning for Spinal

    Song, Guoli and Ying, Yuhan and Zhao, Xingang and Zhao, Yiwen and Zhang, Lei , journal =. Mixture-of-Experts Network via Frequency-Causal Reasoning for Spinal. 2026 , doi =

    work page 2026

  8. [8]

    2026 , doi =

    Liu, Mengjie and Yao, Yuxin and Jia, Jinyong and Yao, Jiali and Huang, Zhengze and Zeng, Ziyang and Pu, Guangjin and Wu, Yan and Bai, Yuqi and Wang, Bin and Jiang, Lili , journal =. 2026 , doi =

    work page 2026

Show all 50 references
  1. [9]

    Jorge and Li, Wenqi and Brown, Richard and Ma, Nic and Kerfoot, Eric and Wang, Yiheng and others , journal =

    Cardoso, M. Jorge and Li, Wenqi and Brown, Richard and Ma, Nic and Kerfoot, Eric and Wang, Yiheng and others , journal =. 2022 , eprint =. doi:10.48550/arXiv.2211.02701 , url =

    2022 doi arXiv

  2. [10]

    Medical Image Analysis , volume =

    A Survey on Deep Learning in Medical Image Analysis , author =. Medical Image Analysis , volume =. 2017 , doi =

    2017

  3. [11]

    2019 , isbn =

    Automated Machine Learning: Methods, Systems, Challenges , editor =. 2019 , isbn =. doi:10.1007/978-3-030-05318-5 , url =

    2019 doi

  4. [12]

    2020 , eprint =

    Erickson, Nick and Mueller, Jonas and Shirkov, Alexander and Zhang, Hang and Larroy, Pedro and Li, Mu and Smola, Alexander , journal =. 2020 , eprint =. doi:10.48550/arXiv.2003.06505 , url =

    2020 doi arXiv

  5. [13]

    and Kohl, Simon A

    Isensee, Fabian and Jaeger, Paul F. and Kohl, Simon A. A. and Petersen, Jens and Maier-Hein, Klaus H. , journal =. 2021 , doi =

    2021

  6. [14]

    arXiv preprint arXiv:2604.21936 , year =

    An Artifact-Based Agent Framework for Adaptive and Reproducible Medical Image Processing , author =. arXiv preprint arXiv:2604.21936 , year =. doi:10.48550/arXiv.2604.21936 , url =. 2604.21936 , archivePrefix =

    doi arXiv

  7. [15]

    and Cao, Yuan , booktitle =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik R. and Cao, Yuan , booktitle =. 2023 , url =

    2023

  8. [16]

    Advances in Neural Information Processing Systems , volume =

    Schick, Timo and Dwivedi-Yu, Jane and Dess. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  9. [17]

    Advances in Neural Information Processing Systems , volume =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  10. [18]

    Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , series =

    Generative Agents: Interactive Simulacra of Human Behavior , author =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , series =. 2023 , doi =

    2023

  11. [19]

    2024 , publisher=

    Tang, Xiangru and Zou, Anni and Zhang, Zhuosheng and Li, Ziming and Zhao, Yilun and Zhang, Xingyao and Cohan, Arman and Gerstein, Mark , booktitle=. 2024 , publisher=. doi:10.18653/v1/2024.findings-acl.33 , url=

    2024 doi

  12. [20]

    2024 , doi=

    Kim, Yubin and Park, Chanwoo and Jeong, Hyewon and Chan, Yik Siu and Xu, Xuhai and McDuff, Daniel and Lee, Hyeonhoon and Ghassemi, Marzyeh and Breazeal, Cynthia and Park, Hae Won , booktitle=. 2024 , doi=

    2024

  13. [21]

    and Pierson, Emma and Koh, Pang Wei and Tsvetkov, Yulia , booktitle=

    Li, Shuyue Stella and Balachandran, Vidhisha and Feng, Shangbin and Ilgen, Jonathan S. and Pierson, Emma and Koh, Pang Wei and Tsvetkov, Yulia , booktitle=. 2024 , doi=

    2024

  14. [22]

    and Yang, Carl and Wang, May Dongmei , booktitle=

    Shi, Wenqi and Xu, Ran and Zhuang, Yuchen and Yu, Yue and Zhang, Jieyu and Wu, Hang and Zhu, Yuanda and Ho, Joyce C. and Yang, Carl and Wang, May Dongmei , booktitle=. 2024 , publisher=. doi:10.18653/v1/2024.emnlp-main.1245 , url=

    2024 doi

  15. [23]

    and Hashimoto, Tatsunori , booktitle=

    Ruan, Yangjun and Dong, Honghua and Wang, Andrew and Pitis, Silviu and Zhou, Yongchao and Ba, Jimmy and Dubois, Yann and Maddison, Chris J. and Hashimoto, Tatsunori , booktitle=. Identifying the Risks of. 2024 , url=

    2024

  16. [24]

    Advances in Neural Information Processing Systems , volume=

    Debenedetti, Edoardo and Zhang, Jie and Balunovi. Advances in Neural Information Processing Systems , volume=. 2024 , doi=

    2024

  17. [25]

    2025 , publisher=

    Fallahpour, Adibvafa and Ma, Jun and Munim, Alif and Lyu, Hongwei and Wang, Bo , booktitle=. 2025 , publisher=

    2025

  18. [26]

    2025 , url=

    Zhu, Yinghao and He, Ziyi and Hu, Haoran and Zheng, Xiaochen and Zhang, Xichen and Wang, Jiyao and Gao, Junyi and Ma, Liantao and Yu, Lequan , booktitle=. 2025 , url=

    2025

  19. [27]

    2025 , url=

    Yan, Weixiang and Liu, Haitian and Wu, Tengxiao and Chen, Qian and Wang, Wen and Chai, Haoyuan and Wang, Jiayi , booktitle=. 2025 , url=

    2025

  20. [28]

    2026 , doi=

    Yang, Cheng and Jin, Hui and Yu, Xinlei and Wang, Zhipeng and Liu, Yaoqun and Fan, Fenglei and Lei, Dajiang and Jia, Gangyong and Wang, Changmiao and Ge, Ruiquan , journal=. 2026 , doi=

    2026

  21. [29]

    2026 , doi=

    Ma, Siqi and Huang, Jiajie and Zhang, Fan and Wu, Jinlin and Shen, Yue and Fan, Guohui and Zhang, Zhu and Zang, Zelin , journal=. 2026 , doi=

    2026

  22. [30]

    Automated Segmentation of Organs and Tumors from Partially Labeled 3D

    Myronenko, Andriy and Yang, Dong and He, Yufan and Xu, Daguang , booktitle=. Automated Segmentation of Organs and Tumors from Partially Labeled 3D. 2023 , url=

    2023

  23. [31]

    and Bruss, Joel and Boes, Aaron and Jones, Randall and Johnson, Hans J

    Brzus, Michal and Riley, Cavan J. and Bruss, Joel and Boes, Aaron and Jones, Randall and Johnson, Hans J. , booktitle=. 2024 , publisher=. doi:10.1117/12.3006568 , url=

    2024 doi

  24. [32]

    The Neuroradiology Journal , volume=

    Automated Segmentation of Ventricular Volumes and Subarachnoid Hemorrhage from Computed Tomography Images: Evaluation of a Rule-Based Pipeline Approach , author=. The Neuroradiology Journal , volume=. 2025 , doi=

    2025

  25. [33]

    Pre-Processing and Quality Control of Large Clinical

    Jin, Benjamin and Vald. Pre-Processing and Quality Control of Large Clinical. Data Engineering in Medical Imaging , series=. 2024 , publisher=. doi:10.1007/978-3-031-73748-0_8 , url=

    2024 doi

  26. [34]

    and Pickhardt, Perry J

    Garrett, John W. and Pickhardt, Perry J. and Summers, Ronald M. , journal=. Methodology for a Fully Automated Pipeline of. 2025 , doi=

    2025

  27. [35]

    Journal of Imaging Informatics in Medicine , volume=

    Lemar. Journal of Imaging Informatics in Medicine , volume=. 2026 , doi=

    2026

  28. [36]

    and Yilmaz, Yasin and Rasool, Ghulam , journal=

    Tripathi, Aakash and Waqas, Asim and Schabath, Matthew B. and Yilmaz, Yasin and Rasool, Ghulam , journal=. 2025 , doi=

    2025

  29. [37]

    , journal =

    Bai, Zhe and Essiari, Abdelilah and Perciano, Talita and Bouchard, Kristofer E. , journal =. 2023 , eprint =. doi:10.48550/arXiv.2310.17780 , url =

    2023 doi arXiv

  30. [38]

    arXiv preprint arXiv:2309.17285 , year =

    Efficient Large Scale Medical Image Dataset Preparation for Machine Learning Applications , author =. arXiv preprint arXiv:2309.17285 , year =. doi:10.48550/arXiv.2309.17285 , url =. 2309.17285 , archiveprefix =

    doi arXiv

  31. [39]

    Harry and Gim, Nayoon and Harris, Nomi L

    Nikolov, Milen and Amorim, Edilberto and Caufield, J. Harry and Gim, Nayoon and Harris, Nomi L. and Houghtaling, Jared and Li, Xiang and Morrison, Danielle and Rameau, Ana. Toward. arXiv preprint arXiv:2512.03541 , year =. doi:10.48550/arXiv.2512.03541 , url =. 2512.03541 , ar...

    doi

  32. [40]

    How Well Can Modern

    Zheng, Qiaoyu and Wu, Chaoyi and Qiu, Pengcheng and Dai, Lisong and Zhang, Ya and Wang, Yanfeng and Xie, Weidi , journal =. How Well Can Modern. 2024 , eprint =. doi:10.48550/arXiv.2412.09529 , url =

    2024 doi arXiv

  33. [41]

    and Mileo, Alessandra , journal =

    Maksudov, Bulat and Kurenkov, Vladislav and Curran, Kathleen M. and Mileo, Alessandra , journal =. 2026 , eprint =. doi:10.48550/arXiv.2605.11224 , url =

    2026 doi arXiv

  34. [42]

    2022 , publisher =

    Ji, Yuanfeng and Bai, Haotian and Ge, Chongjian and Yang, Jie and Zhu, Ye and Zhang, Ruimao and Li, Zhen and Zhang, Lingyan and Ma, Wanling and Wan, Xiang and Luo, Ping , booktitle =. 2022 , publisher =. doi:10.52202/068431-2661 , url =

    2022 doi

  35. [43]

    and Pradella, Maurice and Hinck, Daniel and Sauter, Alexander W

    Wasserthal, Jakob and Breit, Hanns-Christian and Meyer, Manfred T. and Pradella, Maurice and Hinck, Daniel and Sauter, Alexander W. and Heye, Tobias and Boll, Daniel T. and Cyriac, Joshy and Yang, Shan and Bach, Michael and Segeroth, Martin , journal =. 2023 , doi =

    2023

  36. [44]

    2016 , publisher=

    Medical Image Computing and Computer-Assisted Intervention--MICCAI 2016 , series=. 2016 , publisher=

    2016

  37. [45]

    and Xu, Daguang , booktitle =

    Hatamizadeh, Ali and Tang, Yucheng and Nath, Vishwesh and Yang, Dong and Myronenko, Andriy and Landman, Bennett and Roth, Holger R. and Xu, Daguang , booktitle =

  38. [46]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=

  39. [47]

    Loshchilov, Ilya and Hutter, Frank , booktitle=

  40. [48]

    Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and others , booktitle =

  41. [49]

    IEEE Transactions on Medical Imaging , volume=

    The Distribution of Target Registration Error in Rigid-Body Point-Based Registration , author=. IEEE Transactions on Medical Imaging , volume=. 2001 , doi=

    2001

  42. [50]

    and Tustison, Nicholas J

    Avants, Brian B. and Tustison, Nicholas J. and Song, Gang and Cook, Philip A. and Klein, Arno and Gee, James C. , journal=. A Reproducible Evaluation of. 2011 , doi=

    2011

Pith tools

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