Pith. sign in

REVIEW 4 major objections 28 references

CT diagnosis can be treated as an iterative phase-sufficiency problem: start from plain CT, request contrast only when structured evidence is not enough.

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 · grok-4.5

2026-07-14 09:33 UTC pith:RWMDIJIR

load-bearing objection Solid systems paper that treats multi-phase CT as iterative evidence-sufficiency control; conversion gains look real but rest on small n and under-specified labels. the 4 major comments →

arxiv 2607.10748 v1 pith:RWMDIJIR submitted 2026-07-12 cs.LG

Policy-Driven CT-Agent: Modeling Phase-Aware Diagnostic Control for Clinically Consistent CT Reasoning

Yanmeng Dong , Han Li , Yujia Li , Jingsong Liu , Xun Ma , Yanzhu Hu , Zhengyang Xu , Zhicheng Li
show 2 more authors
Nassir Navab Shaohua Kevin Zhou
This is my paper
classification cs.LG
keywords multi-phase CTphase selectionmedical agentevidence sufficiencyvision-language modeldiagnostic policyclinical guidelines
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.

Most CT AI systems assume every imaging phase is already available and then run a static detection or classification task. Real radiology works differently: radiologists decide whether non-contrast CT is enough or whether arterial, venous, or excretory phases must still be acquired, balancing diagnostic certainty against radiation and cost. This paper argues that the missing piece is an agent that keeps an explicit diagnostic state, turns each phase into a structured evidence package, and uses guideline-derived rules to decide whether more phases are required. The proposed PD-CTAgent does exactly that. It claims the resulting loop improves both final diagnosis and correct phase-conversion decisions across liver, kidney, and lung CT, while remaining adaptable to different hospital protocols.

Core claim

PD-CTAgent shows that multi-phase CT diagnosis can be formulated as iterative evidence-sufficiency control rather than one-shot multi-phase analysis. By abstracting each phase into a phase-aware evidence package and letting a knowledge-guided controller escalate only when that package is insufficient, the agent outperforms general VLMs, medical VLMs, and a medical-agent baseline on both diagnostic accuracy and phase-conversion decisions on MCT-LTDiag, Kidney53, and LIDC.

What carries the argument

The paired CSAM–KDCM loop: CSAM builds a structured Evidence Package (findings, uncertainty, quality flags) from each CT phase and stores the trajectory; KDCM fuses retrieved guideline rules with VLM judgment under a rule-prioritized policy that decides whether to request the next phase or emit the final report.

Load-bearing premise

The labels that mark when another CT phase is truly needed, and the offline-compiled guideline rules used to make that call, are treated as valid clinical ground truth even though the paper notes that existing datasets do not supply phase-sufficiency supervision.

What would settle it

On a multi-phase cohort with independent radiologist phase-sufficiency labels and institutional protocols held fixed, check whether PD-CTAgent’s conversion F1 and diagnostic accuracy still beat the same baselines, and whether removing iterative re-entry into CSAM collapses the gains as claimed.

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

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

4 major / 0 minor

Summary. The paper proposes Policy-Driven CT-Agent (PD-CTAgent), an iterative agent that treats multi-phase CT diagnosis as evidence-sufficiency control rather than static single-shot analysis. CSAM converts each CT phase into a structured Evidence Package (organ findings, uncertainty metrics SDU/critical-region/MIC, quality flags) stored in a Case Memory Bank; KDCM then retrieves offline-compiled guideline rules, fuses rule-based and VLM decisions via a rule-prioritized policy (Eq. 7), and escalates phases until sufficiency is declared and a report is emitted. The agent can condition on hospital-specific rule sets. Experiments on MCT-LTDiag (liver multi-phase, n=60), Kidney53 (private kidney multi-phase, n=60; paired subset 46), and LIDC (plain lung CT) report gains over general VLMs, medical VLMs, and MedAgent-Pro on diagnostic Acc/F1 and phase-conversion Acc/F1 (Table II), with ablations (Table III) and controlled workflow analyses (Table IV) arguing for structured evidence, knowledge, and CSAM re-entry.

Significance. The clinical framing is valuable: most CT AI assumes all phases are already available, whereas real workflows decide whether additional contrast phases are needed under radiation and protocol constraints. Formulating phase selection as iterative, knowledge-guided sufficiency control, with protocol-conditioned reporting (Fig. 4) and explicit re-processing of newly acquired phases through CSAM (Table IV), is a clear systems contribution relative to static VLMs. Ablations and the controlled workflow study give more mechanistic support than typical agent papers. If conversion labels and knowledge assets are independently validated and results hold at larger scale, the work would be a useful bridge between multi-phase imaging and workflow-aware medical agents. Code release upon acceptance is a plus. Impact is currently tempered by very small multi-phase cohorts and under-specified ground truth for the phase-sufficiency task the paper itself identifies as missing from existing datasets.

major comments (4)
  1. Introduction and Abstract state that existing datasets lack supervision for phase-sufficiency decisions, yet Table II reports phase-conversion Acc/F1 as a primary claim (MCT-LTDiag 83.3/0.79; Kidney53 56.6/0.69). The manuscript never defines how conversion ground truth is obtained for each case (protocol-level labels, retrospective clinical orders, rule-derived labels from the same KDCM assets in §II.B.1, or expert re-annotation). Without an external gold standard independent of the offline-compiled Phase-Aware Diagnostic Knowledge Database and of Eq. 7’s rule-prioritized fusion, conversion gains risk measuring consistency with the compiled rules rather than clinical sufficiency. Please specify label provenance, inter-rater agreement if expert-labeled, and any separation between knowledge compilation and evaluation labels.
  2. Multi-phase evaluation cohorts are very small (MCT-LTDiag 60; Kidney53 60; paired kidney subset 46 in Table IV) and Table II reports point Acc/F1 with no confidence intervals, bootstrap, or significance tests. On Kidney53 diagnostic F1 the method is 0.40, below GPT-5.2-Pro (0.68), MedAgent-Pro (0.52), and Lingshu (0.54), so the abstract/contribution claim that PD-CTAgent “consistently outperforms” baselines is not uniformly supported. Either qualify the claim by task/dataset, add uncertainty estimates, or strengthen evaluation (larger hold-out, cross-validation, or external multi-phase set).
  3. §II.B.1 describes offline guideline-to-DSL compilation as keyword filtering, paragraph-to-rule conversion, and manual checking, but does not report the size of the rule set, organ/disease coverage, versioning, or how conflicts across guidelines are resolved. §II.C Eq. (7) forces escalation whenever d_r=1, so conversion performance can be driven by rule coverage rather than learned sufficiency. Please quantify rule coverage vs. VLM-only decisions, report how often d_r vs d_v determines D_t, and show sensitivity of Table II conversion metrics to rule-set composition.
  4. CSAM uncertainty (SDU boundary band r_b, critical radius r_c, MIC with k inferences and stability thresholds) and skill-memory calibration priors are free parameters that gate escalation (§II.A.2–3), yet no values, selection procedure, or sensitivity analysis appear in Experiments. Table III attributes large gains to “Uncertainty” and “Knowledge,” but without fixed thresholds or held-out tuning this confounds reproducibility and the causal interpretation of the ablation. Report the operating thresholds used for Table II/III and a brief sensitivity study.

Circularity Check

1 steps flagged

Mild evaluation circularity risk: phase-conversion labels lack independent gold standard and may be self-consistent with the offline knowledge assets that drive KDCM decisions.

specific steps
  1. other [Introduction; §II.B.1 Knowledge-Constrained Policy Database; Table II conversion columns]
    "the lack of supervision for phase-sufficiency decisions in existing datasets. ... we construct a Phase-Aware Diagnostic Knowledge Database, extracted from multi-phase radiology reports to encode structured diagnostic policies regarding phase sufficiency. ... PD-CT Agent (Ours) ... MCT-LTDiag (Conv) Acc 83.3 F1 0.79 ... Kidney53 (Conv) Acc 56.6 F1 0.69"

    Conversion Acc/F1 is presented as evidence that KDCM correctly decides when extra phases are required. Because the paper itself states that datasets supply no such supervision, the conversion ground truth must be derived (implicitly or via the same reports) from the guideline/report material that is also compiled into the offline knowledge assets used by the agent’s rule-prioritized policy (Eq. 7). High conversion scores are therefore partly self-consistent with the injected rules rather than an independent clinical-sufficiency signal; diagnostic Acc remains non-circular.

full rationale

This is an empirical agent/systems paper, not a first-principles derivation. CSAM evidence packages, uncertainty metrics (SDU/MIC), rule–LLM fusion (Eq. 7), and iterative re-entry are independent engineering choices whose diagnostic Acc/F1 on LIDC/MCT/Kidney53 are externally measurable against ordinary labels. No self-definitional equations, no fitted parameters renamed as predictions, no load-bearing uniqueness theorems imported from the authors, and no ansatz smuggled via self-citation. The sole mild issue is the conversion task itself: the Introduction explicitly states that existing datasets lack phase-sufficiency supervision, yet Table II reports conversion Acc/F1 as a primary claim of clinical consistency; the knowledge database that supplies the agent’s phase decisions is compiled offline from the same class of multi-phase reports/guidelines (II.B.1). Without a quoted external protocol-level gold standard independent of those assets, conversion gains can be partly self-consistent with the rules the model is designed to follow. That is a validity concern for the conversion metric, not a reduction of the whole result by construction; hence score 2 rather than 0 or 6+.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on domain assumptions about clinical phase policies, on engineered modules (Evidence Package, CSAM, KDCM), and on several hand-chosen control thresholds and fusion preferences rather than on a parameter-free theory. Free parameters are the uncertainty/quality thresholds and retrieval/fusion knobs that gate escalation. Axioms are standard ML/clinical-practice assumptions. Invented entities are the paper's named modules and memory structures used to operationalize phase control.

free parameters (6)
  • Boundary band radius r_b (SDU)
    Defines the adaptive boundary region over which entropy is averaged for Signed Distance Uncertainty; chosen design parameter that can trigger escalation.
  • Critical-region radius r_c
    Defines the decision-critical shell for critical-region uncertainty; hand-set geometric scale affecting phase requests.
  • MIC inference count k and stability threshold
    Number of stochastic segmentations and the low-MIC trigger for conservative escalation are implementation choices not derived from data-independent theory.
  • TopK retrieval size and rule-filter settings
    Structured RAG retrieves TopK evidence fragments after organ/phase filtering; K and filtering criteria are free system hyperparameters.
  • Rule-prioritized fusion policy (Eq. 7)
    Final decision forces escalation if rule says 1, else uses VLM; this conservative fusion preference is a design choice that shapes conversion metrics.
  • Uncertainty/quality escalation thresholds in skill memory
    Case/skill memory stores calibration priors such as uncertainty thresholds used online; values are calibrated rather than predicted from first principles.
axioms (5)
  • domain assumption Current imaging evidence can be judged sufficient or insufficient for diagnosis before all phases are acquired.
    Core clinical premise of the agent loop (Introduction; Eq. 1); without it phase conversion is undefined.
  • domain assumption Organ-level segmentation, uncertainty, and quality flags form a faithful enough Evidence Package for phase decisions in place of raw volumes.
    CSAM §II.A routes tools and suppresses raw pixels; performance depends on segmentation reliability.
  • domain assumption Offline-compiled guideline/report rules plus ontology-normalized retrieval adequately encode institutional phase-selection policies.
    KDCM knowledge database compilation and structured RAG (§II.B) supply the policy prior.
  • domain assumption A VLM conditioned on structured packages and retrieved rules can produce usable sufficiency decisions and reports.
    VLM-guided strategy generation and report writing are load-bearing for both conversion and diagnosis.
  • ad hoc to paper Standard supervised metrics Acc/F1 on the chosen datasets measure clinically consistent phase control.
    Paper itself notes missing phase-sufficiency supervision in existing datasets, yet reports conversion Acc/F1 as primary evidence of clinical consistency.
invented entities (3)
  • Clinical Structure Abstraction Module (CSAM) / Evidence Package no independent evidence
    purpose: Harmonize heterogeneous CT phases into structured findings, uncertainty, and quality for agent control.
    Named core module; independent evidence is only via ablations and downstream metrics in this paper, not external clinical validation of the package schema.
  • Knowledge-Guided Diagnostic Control Module (KDCM) with Phase-Aware Diagnostic Knowledge Database no independent evidence
    purpose: Evaluate phase sufficiency, retrieve rules/evidence, and decide escalation under hospital protocols.
    Central control invention; falsifiable only through the paper's conversion experiments and case studies.
  • Case Memory Bank and skill memory trajectory state S_t no independent evidence
    purpose: Store iterative diagnostic trajectory and reusable calibration priors without raw volumes.
    Agent-state construct enabling iterative re-entry; no external independent measurement of the memory design.

pith-pipeline@v1.1.0-grok45 · 16468 in / 3834 out tokens · 57137 ms · 2026-07-14T09:33:20.753645+00:00 · methodology

0 comments
read the original abstract

Computed Tomography (CT) diagnosis often relies on dynamic selection of imaging phases, such as non-contrast, arterial, or venous phases, based on preliminary findings, clinical suspicion, and diagnostic guidelines. This phase-wise decision process is critical for reducing unnecessary radiation exposure while supporting timely staging and treatment planning. However, phase-selection protocols can vary across hospitals, regions, and guidelines, while most existing CT-based AI methods assume that all phases are available and focus on static tasks under a fixed imaging phase, failing to model whether additional phases are required. This limitation stems from heterogeneous multi-phase representations, the need for knowledge-guided phase control beyond visual cues, and the lack of supervision for phase-sufficiency decisions in existing datasets. To address these challenges, we propose Policy-Driven CT-Agent (PD-CTAgent) for clinically consistent CT phase selection and diagnostic reasoning. PD-CTAgent introduces a Clinical Structure Abstraction Module (CSAM) to harmonize heterogeneous CT phases into a unified, phase-aware evidence representation. Based on this representation, a Knowledge-Guided Diagnostic Control Model (KDCM) evaluates phase sufficiency and iteratively requests additional phases when necessary. The policy-driven agent design further allows PD-CTAgent to flexibly follow different institutional, regional, or guideline-specific diagnostic protocols. Together, PD-CTAgent bridges static CT analysis and real-world clinical workflows. Experiments on two public datasets, LIDC and MCT-LTDiag, and one private dataset demonstrate its effectiveness and clinical consistency. Code will be made public upon acceptance.

Figures

Figures reproduced from arXiv: 2607.10748 by Han Li, Jingsong Liu, Nassir Navab, Shaohua Kevin Zhou, Xun Ma, Yanmeng Dong, Yanzhu Hu, Yujia Li, Zhengyang Xu, Zhicheng Li.

Figure 1
Figure 1. Figure 1: This limitation is not merely a design choice, but stems [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Comparison between conventional CT diagnostic pipelines and our PD-CT Agent. Traditional methods rely on manually selected existing phase-specific [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall architecture of PD-CTAgent. The framework consists of a Clinical Structure Abstraction Module (CSAM) and a Knowledge-Guided [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative examples of PD-CTAgent on Kidney53. The left case shows sufficient plain-CT evidence without phase escalation, whereas the right case [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Case-level example of guideline-conditioned reasoning with four hospital-specific rules. The same CT case is evaluated under DNCCT-R2 for [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

28 extracted references · 6 linked inside Pith

  1. [1]

    Abn-blip: Abnormality-aligned bootstrapping language-image pre-training for pul- monary embolism diagnosis and report generation from ctpa,

    Z. Zhong, Y . Wang, L. Bi, Z. Ma, S. H. Ahn, C. J. Mullin, C. F. Greineder, M. K. Atalay, S. Collins, G. L. Bairdet al., “Abn-blip: Abnormality-aligned bootstrapping language-image pre-training for pul- monary embolism diagnosis and report generation from ctpa,”Medical image analysis, vol. 107, p. 103786, 2026

    2026

  2. [2]

    Mask-aware transformer with structure invariant loss for ct translation,

    W. Chen, W. Zhao, Z. Chen, T. Liu, L. Liu, J. Liu, and Y . Yuan, “Mask-aware transformer with structure invariant loss for ct translation,” Medical Image Analysis, vol. 96, p. 103205, 2024

    2024

  3. [3]

    Dct-net: Dual-branch ct recon- struction from orthogonal x-rays with diffusion model and contrastive learning,

    Z. Zhang, C. Shen, J. Tang, and Z. Liao, “Dct-net: Dual-branch ct recon- struction from orthogonal x-rays with diffusion model and contrastive learning,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 150–160

    2025

  4. [4]

    A multimodal contrastive learning for detecting aortic dissection on 3d non-contrast ct with anatomy simplification,

    D. Zhang, W. Xiao, C. Jiang, Y . Qiu, Z. Feng, H. Wang, Y . Zheng, and W. Zhu, “A multimodal contrastive learning for detecting aortic dissection on 3d non-contrast ct with anatomy simplification,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 3–12

    2025

  5. [5]

    Incidental indeterminate renal lesions: Distinguishing non-enhancing from potential enhancing renal lesions using iodine quantification on portal venous dual-layer spectral ct,

    S. van der Star, P. A. de Jong, and M. Kok, “Incidental indeterminate renal lesions: Distinguishing non-enhancing from potential enhancing renal lesions using iodine quantification on portal venous dual-layer spectral ct,”Journal of Personalized Medicine, vol. 13, no. 11, p. 1546, 2023

    2023

  6. [6]

    Protocol optimization for renal mass detection and characterization,

    J. S. Chu and Z. J. Wang, “Protocol optimization for renal mass detection and characterization,”Radiologic Clinics, vol. 58, no. 5, pp. 851–873, 2020

    2020

  7. [7]

    Virtual non-contrast spectral ct in renal masses: is it time to discard conventional unenhanced phase?

    G. M. Bucolo, V . Ascenti, S. Barbera, F. Fontana, F. M. Aric `o, F. Piacentino, A. Coppola, G. Cicero, M. A. Marino, C. Boozet al., “Virtual non-contrast spectral ct in renal masses: is it time to discard conventional unenhanced phase?”Journal of Clinical Medicine, vol. 12, no. 14, p. 4718, 2023

    2023

  8. [8]

    Incidental finding of renal masses at unenhanced ct: prevalence and analysis of features for guiding management,

    S. D. O’Connor, P. J. Pickhardt, D. H. Kim, M. R. Oliva, and S. G. Sil- verman, “Incidental finding of renal masses at unenhanced ct: prevalence and analysis of features for guiding management,”American Journal of Roentgenology, vol. 197, no. 1, pp. 139–145, 2011

    2011

  9. [9]

    Mtanet: Multi- task attention network for automatic medical image segmentation and classification,

    Y . Ling, Y . Wang, W. Dai, J. Yu, P. Liang, and D. Kong, “Mtanet: Multi- task attention network for automatic medical image segmentation and classification,”IEEE Transactions on Medical Imaging, vol. 43, no. 2, pp. 674–685, 2023

    2023

  10. [10]

    Enhancing diagnostic images to improve the performance of the segment anything model in medical image segmentation,

    L. Kong, M. Huang, L. Zhang, and L. W. C. Chan, “Enhancing diagnostic images to improve the performance of the segment anything model in medical image segmentation,”Bioengineering, vol. 11, no. 3, p. 270, 2024

    2024

  11. [11]

    Revolutionizing healthcare: How deep learning is poised to change the landscape of medical diagnosis and treatment,

    A. Ahmad, A. Tariq, H. K. Hussain, and A. Y . Gill, “Revolutionizing healthcare: How deep learning is poised to change the landscape of medical diagnosis and treatment,”Journal of Computer Networks, Architecture and High Performance Computing, vol. 5, no. 2, pp. 458– 471, 2023

    2023

  12. [12]

    Deep learning for medical image-based cancer diagnosis,

    X. Jiang, Z. Hu, S. Wang, and Y . Zhang, “Deep learning for medical image-based cancer diagnosis,”Cancers, vol. 15, no. 14, p. 3608, 2023

    2023

  13. [13]

    Ct-agent: a multimodal- llm agent for 3d ct radiology question answering,

    Y . Mao, W. Xu, Y . Qin, and Y . Gao, “Ct-agent: a multimodal- llm agent for 3d ct radiology question answering,”arXiv preprint arXiv:2505.16229, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Cpllm: Clinical prediction with large language models,

    O. Ben Shoham and N. Rappoport, “Cpllm: Clinical prediction with large language models,”PLOS Digital Health, vol. 3, no. 12, p. e0000680, 2024

    2024

  15. [15]

    Enhancing medical image report generation through standard language models: leveraging the power of llms in healthcare,

    G. Leonardi, L. Portinale, A. Santomauroet al., “Enhancing medical image report generation through standard language models: leveraging the power of llms in healthcare,” inIn2nd AIxIA Workshop on Artificial Intelligence For Healthcare, 2023, pp. 41–45

    2023

  16. [16]

    Agent hospital: A simulacrum of hospital with evolvable medical agents,

    J. Li, Y . Lai, W. Li, J. Ren, M. Zhang, X. Kang, S. Wang, P. Li, Y .-Q. Zhang, W. Maet al., “Agent hospital: A simulacrum of hospital with evolvable medical agents,”arXiv preprint arXiv:2405.02957, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Medadapter: Efficient test-time adaptation of large language models towards medical reasoning,

    W. Shi, R. Xu, Y . Zhuang, Y . Yu, H. Sun, H. Wu, C. Yang, and M. D. Wang, “Medadapter: Efficient test-time adaptation of large language models towards medical reasoning,” inProceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Processing, 2024, pp. 22 294–22 314

    2024

  18. [18]

    Pathfinder: A multi-modal multi-agent system for medical diagnostic decision- making applied to histopathology,

    F. Ghezloo, M. S. Seyfioglu, R. Soraki, W. O. Ikezogwo, B. Li, T. Vivekanandan, J. G. Elmore, R. Krishna, and L. Shapiro, “Pathfinder: A multi-modal multi-agent system for medical diagnostic decision- making applied to histopathology,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 23 431–23 441

    2025

  19. [19]

    M 3builder: A multi-agent system,

    Y . Wang and W. Xie, “M 3builder: A multi-agent system,” inAI for Clinical Applications: First International Workshops, Agentic AI 2025, CREATE 2025, and Clinical MLLMs 2025, Held in Conjunction with MICCAI 2025, Daejeon, South Korea, September 23 and 27, 2025, Proceedings. Springer Nature, 2025, p. 115

    2025

  20. [20]

    Medchat: A multi-agent framework for multimodal diagnosis with large language models,

    P. R. Liu, S. Bansal, J. Dinh, A. Pawar, R. Satishkumar, S. Desai, N. Gupta, X. Wang, and S. Hu, “Medchat: A multi-agent framework for multimodal diagnosis with large language models,” in2025 IEEE 8th International Conference on Multimedia Information Processing and Retrieval (MIPR). IEEE, 2025, pp. 456–462

    2025

  21. [21]

    The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans,

    S. G. Armato III, G. McLennan, L. Bidaut, M. F. McNitt-Gray, C. R. Meyer, A. P. Reeves, B. Zhao, D. R. Aberle, C. I. Henschke, E. A. Hoffmanet al., “The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans,”Medical physics, vol. 38, no. 2, pp. 915–931, 2011

    2011

  22. [22]

    A multi-phase ct dataset for automated differential diagnosis of liver tumors,

    X. a. Wu, H. Su, Y . Hua, Y . Xu, L. Wang, X. Wang, S. Wang, B. Jin, X. Liu, X. Wanet al., “A multi-phase ct dataset for automated differential diagnosis of liver tumors,”Scientific Data, 2025

    2025

  23. [23]

    (2025) Introducing gpt-5.2

    OpenAI. (2025) Introducing gpt-5.2. https://openai.com/index/introducing-gpt-5-2/

    2025

  24. [24]

    Internvl3. 5: Advancing open-source multi- modal models in versatility, reasoning, and efficiency,

    W. Wang, Z. Gao, L. Gu, H. Pu, L. Cui, X. Wei, Z. Liu, L. Jing, S. Ye, J. Shaoet al., “Internvl3. 5: Advancing open-source multi- modal models in versatility, reasoning, and efficiency,”arXiv preprint arXiv:2508.18265, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Qwen2. 5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tanget al., “Qwen2. 5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Lingshu: A generalist foundation model for unified multimodal medical understanding and reasoning,

    W. Xu, H. P. Chan, L. Li, M. Aljunied, R. Yuan, J. Wang, C. Xiao, G. Chen, C. Liu, Z. Liet al., “Lingshu: A generalist foundation model for unified multimodal medical understanding and reasoning,”arXiv preprint arXiv:2506.07044, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning,

    J. Pan, C. Liu, J. Wu, F. Liu, J. Zhu, H. B. Li, C. Chen, C. Ouyang, and D. Rueckert, “Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 337–347

    2025

  28. [28]

    Medagent-pro: Towards evidence-based multi-modal medical diagnosis via reasoning agentic workflow,

    Z. Wang, J. Wu, L. Cai, C. H. Low, X. Yang, Q. Li, and Y . Jin, “Medagent-pro: Towards evidence-based multi-modal medical diagnosis via reasoning agentic workflow,”arXiv preprint arXiv:2503.18968, 2025

    Pith/arXiv arXiv 2025