Abstract

Abstract: Endoscopic ultrasonography (EUS) interpretation for pancreatic cancer is highly operator-dependent, leading to significant diagnostic variability and limited reasoning transparency. We present PancCADx, an interpretable multimodal framework leveraging Qwen3-VL-8B-Thinking with a Chain-of-Thought (CoT) mechanism to align model reasoning with expert clinical logic. To ensure clinical alignment and minimize hallucinations, we utilized a two-stage alignment strategy: Supervised Fine-Tuning (SFT) followed by Direct Preference Optimization (DPO). Developed on an internal cohort (n=438) and validated on multi-center external cohorts (n=191), PancCADx achieved 95.74% external sensitivity, significantly outperforming state-of-the-art benchmarks (92.55%) with superior cross-institutional generalizability. In a prospective sequential validation involving 12 clinicians, AI assistance significantly enhanced diagnostic accuracy (+7.7%, p < 0.001) and decision confidence (p < 0.001). Notably, PancCADx effectively bridged the novice-expert gap and improved specificity for senior clinicians (up to 76%). By harmonizing high diagnostic performance with transparent, evidence-based reasoning, PancCADx provides a robust, clinically-aligned solution for objective and trustworthy pancreatic cancer diagnosis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1096_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{HuSha_PancCADx_MICCAI2026,
        author = { Hu, Shan AND Xiao, Changhong AND Qin, Xianzheng AND Mei, Bin AND Cheng, Bin AND Wang, Zhongyuan},
        title = { { PancCADx: A Multimodal Framework for Pancreatic Cancer Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    Authors proposes PancCADx, a multimodal reasoning framework for pancreatic cancer diagnosis on endoscopic ultrasonography. It employs a two-stage alignment strategy: Supervised Fine-Tuning to inject expert-level Chain-of-Thought reasoning, followed by Direct Preference Optimization to correct hallucinations without a separate reward model. Validated on multi-center external cohorts and a prospective study with 12 clinicians.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The SFT-DPO alignment strategy, reminiscent of RLHF which typically demands large-scale preference data, is effectively adapted to a low-resource medical setting through careful pipeline design — including expert-verified CoT distillation and targeted preference pair construction (n=386). The use of Low-Rank Adaptation further enables this data-efficient implementation on a single GPU. Moreover, the validation design is rigorous, incorporating multi-center external cohorts and a prospective clinical study with 12 stratified endoscopists, lending credibility to the reported performance gains.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The paper does not include a Limitations section, which is a notable omission.

    2.All p-values for the expert group and most of the senior groups are non-significant, yet the paper does not discuss whether this reflects genuinely limited room for improvement at high baselines or simply insufficient sample size. Explicit acknowledgment of this ambiguity is needed.

    3.PancCADx shows notably lower internal accuracy (83.72%) compared to Joint-AI (98.00%), yet the paper frames this solely as Joint-AI overfitting rather than analyzing potential weaknesses of PancCADx on in-distribution data. A more balanced discussion is warranted.

    4.Some abbreviations in Table 1 (AIP, CP, NET, SPT) are left undefined. Providing full definitions would make the paper more accessible to early-stage readers who may not be familiar with these clinical terms.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The prospective validation with stratified clinicians is a clear strength over most AI-in-medical-imaging studies that remain purely retrospective. However, the very small expert subgroup (n=2), the unexplored confirmation bias risk in the Conclusion-First protocol, and inconsistencies in metric formatting prevent a stronger recommendation. These issues are addressable in revision — the authors are encouraged to discuss these limitations openly and transparently, which would strengthen the overall credibility of the work.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal addresses the main concerns from my review. The authors committed to adding a Limitations section covering the n=2 expert subgroup, Conclusion-First bias risk, and single-organ scope, and to including 95% CIs across all cohorts with unified metric formatting. They explicitly acknowledged that the non-significant p-values for the expert group reflect both small sample size and baseline ceiling effects, reframing this as AI reinforcing expert confidence rather than overclaiming improvement. The Conclusion-First bias defense grounded in recognition-primed decision-making, together with DPO-based mitigation, is reasonable though not fully resolving the underlying risk; appropriately deferred to the Limitations. The response to the internal vs external accuracy gap is partially satisfactory — the statistical instability argument (n=43, wide CI) is valid, but the Joint-AI overfitting framing remains one-sided. Overall, the prospective validation with stratified clinicians, the data-efficient SFT-DPO pipeline, and the authors willingness to incorporate the requested revisions justify an Accept. Remaining concerns are addressable in the camera-ready revision.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a multimodal framework for pancreatic cancer diagnosis that combines EUS images and structured clinical data, and further aligns a reasoning-capable vision-language model, aiming to improve both diagnostic performance and interpretability; the authors validate the model on multicenter cohorts and evaluate its value as an AI assistant for clinicians in a reader study.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The paper focuses on pancreatic cancer diagnosis from EUS, a high-impact task with substantial inter-operator variability, so improving consistency and decision support is clinically meaningful. 2.The method combines EUS images with 30 structured EMR variables, which better reflects how clinicians make diagnostic decisions than image-only approaches. 3.The model is evaluated not only on an internal cohort but also on data from three independent external centers, which is a stronger validation setting than single-center testing.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The claim of outperforming the SOTA is not fully supported by the reported results. PancCADx achieves higher sensitivity, but lower specificity, slightly lower accuracy, and slightly lower F1 than Joint-AI. 2.The paper does not present a strongly new technical method at the MICCAI level. 3.The paper does not sufficiently isolate the contribution of individual components, such as image-only vs. multimodal, SFT vs. SFT+DPO, or with vs. without CoT supervision, making it hard to determine which part of the pipeline drives performance gains. 4.Evaluation is incomplete and should include threshold-independent metrics such as AUC.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper addresses an important clinical problem, proposes a multimodal framework that combines EUS images with clinical variables, and includes both multi-center validation and a reader study, which increases its practical interest. However, the final score is limited by several important concerns: the claim of outperforming the state of the art is not fully supported by the reported results, key ablation studies are missing, the evidence for interpretability is still limited, and the evaluation would be more convincing if the authors reported threshold-independent metrics such as AUC in addition to fixed-threshold measures. Overall, while the work is promising and clinically relevant, the current experimental evidence is not yet strong enough to fully support its main claims.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper introduces PancCADx, an interpretable multimodal framework for AI-assisted pancreatic cancer diagnosis via endoscopic ultrasonography (EUS). The core contribution is a reasoning-driven clinical decision support system built on Qwen3-VL-8B-Thinking with a Chain-of-Thought mechanism, designed to align model reasoning with expert clinical logic rather than functioning as a black-box classifier. The authors employ a two-stage alignment strategy, Supervised Fine-Tuning followed by Direct Preference Optimization, to minimize hallucinations and ensure clinical relevance in model outputs. The framework is validated on both internal (n=438) and multi-center external (n=191) cohorts, achieving 95.74% sensitivity on single-frame inputs while outperforming multi-frame state-of-the-art baselines, a notable data efficiency advantage. A prospective sequential validation with 12 clinicians further demonstrates that AI assistance meaningfully improved diagnostic accuracy and decision confidence across experience levels, with particular impact in bridging the novice-expert gap and correcting specificity bias in senior endoscopists. Together, these contributions position PancCADx as a clinically grounded and transparent approach to objective pancreatic cancer diagnosis.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    This paper has several compelling strengths that make it a strong contribution to the MICCAI community. First, the clinical motivation is sharp and well-executed. EUS interpretation for pancreatic cancer is genuinely operator-dependent, and the authors don’t just acknowledge this, they build their entire framework around solving it. The decision to address both diagnostic variability and reasoning transparency in a single system is thoughtful and clinically grounded. Second, the two-stage alignment strategy is a standout methodological contribution. The combination of Supervised Fine-Tuning for reasoning injection followed by Direct Preference Optimization for clinical alignment is novel in this context. What makes it particularly interesting is the construction of the preference dataset, capturing discrepancies between SFT outputs and expert consensus to generate 386 high-quality preference pairs is a rigorous and principled approach to minimizing hallucinations in a high-stakes clinical setting. Third, the Chain-of-Thought distillation pipeline is well-designed and clinically motivated. The Conclusion-First protocol, anchoring reasoning to the diagnosis before retrospective morphological analysis, is a deliberate design choice that mirrors expert recognition-primed decision-making. Having three senior experts with over 15 years of experience refine the reasoning chains adds meaningful credibility to the interpretability claims. Fourth, the evaluation is genuinely rigorous. The multicenter external validation across three independent institutions (n=191) is a meaningful test of generalizability, and the results hold up well, PancCADx achieves 95.74% pooled sensitivity against a SOTA baseline trained on richer multi-frame inputs, using only single-frame data. The direct comparison against Joint-AI is fair and transparent, and the authors are commendably honest about where PancCADx trails on specificity. Fifth, the prospective sequential validation study is one of the strongest aspects of the paper. The self-controlled two-phase design with a washout period is methodologically sound, and the results are meaningful across all experience levels. The finding that AI assistance improved senior clinician specificity from 40% to 76% is striking and clinically significant. The dual benefit of CoT reasoning, reinforcing expert confidence while serving as an educational scaffold for novices, is a genuinely interesting finding that extends the contribution beyond pure diagnostic performance. Finally, the privacy and data handling pipeline is thorough. The de-identification protocol, IRB approval across participating centers, and STARD guideline adherence reflect a level of clinical rigor that is not always present in AI diagnostic papers.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    While this is a strong paper overall, there are a few areas worth addressing that would further solidify the contribution. The most notable concern is the internal validation performance. On the Center A holdout (n=43), PancCADx reports 83.72% accuracy and 68.75% specificity, a meaningful drop compared to its external validation results. This is somewhat counterintuitive, and the paper does not fully account for it. The authors make a convincing case that Joint-AI’s near-perfect internal performance reflects overfitting, but the inverse pattern in PancCADx warrants its own explanation. Understanding what drives this discrepancy would strengthen confidence in the framework’s consistency across evaluation settings. Relatedly, the internal validation set of n=43 is quite small. At this sample size, performance estimates are statistically fragile, and the absence of confidence intervals for this cohort makes it difficult to interpret the results with confidence. Reporting confidence intervals across all cohorts, not just the sequential validation study, would add important statistical transparency. On the human-AI collaboration findings, the results for expert clinicians specifically deserve closer attention. While the overall narrative around bridging the novice-expert gap is well supported, none of the expert-level improvements in Table 3 reach statistical significance, sensitivity (p=.215), specificity (p>.99), and accuracy (p=.154) all fall short. This does not undermine the broader findings, but the paper should be careful not to overstate AI benefit uniformly across experience levels. The story for novices and seniors is compelling; the expert story is more nuanced. The framework is validated exclusively on pancreatic EUS, which is appropriate for a focused contribution but limits the generalizability claims. The authors gesture toward broader applicability across GI cancers and imaging modalities, but this remains undemonstrated. Framing the scope more precisely, as a validated solution for pancreatic EUS specifically, with potential for broader application, would be more accurate and still compelling. Finally, the CoT reasoning quality evaluation relies primarily on subjective Likert-scale assessment across five dimensions. While this captures clinician perception well, there is no quantitative measure of reasoning correctness or factual accuracy independent of the diagnostic outcome. A structured evaluation of the reasoning chains themselves, even on a subset, would meaningfully strengthen the interpretability claims that are central to this work.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    This is a well-executed and clinically meaningful paper that makes a genuine contribution to AI-assisted endoscopic diagnosis. The two-stage alignment strategy, the prospective sequential validation design, and the honest benchmarking against a stronger baseline are all commendable. The work is clearly motivated, methodologically sound, and the results are largely convincing. The authors should be encouraged, this is the kind of human-AI collaboration study the MICCAI community needs more of. A few additional thoughts for the authors to consider: The internal validation performance dip is the one finding that lingers. Even a brief discussion acknowledging and contextualizing this, whether attributable to the small holdout size, cohort composition, or the training-validation split strategy, would go a long way in preempting reviewer concern and demonstrating full transparency. On the expert clinician findings, it may be worth reframing the narrative slightly. The AI assistance story is strongest for novices and seniors, and those results are genuinely impressive. Presenting the expert findings as a nuanced outcome, where CoT reasoning reinforced confidence rather than changing decisions, is actually an interesting insight in itself and worth developing rather than grouping under the general improvement narrative. The CoT reasoning evaluation would benefit from even a small-scale structured audit of reasoning chain accuracy independent of diagnostic outcome. Given that interpretability is a central claim of the framework, grounding it in something more quantitative than clinician perception would make the contribution significantly more defensible. Finally, the privacy and de-identification pipeline is thorough and well-documented, this is worth highlighting more prominently, as it directly addresses one of the key barriers to real-world clinical deployment of AI diagnostic tools. Overall, this is a strong submission and a worthy addition to the MICCAI program.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper is recommended for acceptance. PancCADx addresses a clinically significant and well-scoped problem, the operator-dependency and reasoning opacity of EUS-based pancreatic cancer diagnosis and does so with a methodologically rigorous and technically innovative framework. The two-stage SFT-DPO alignment strategy is a novel and principled contribution, and the prospective sequential validation study with stratified clinicians elevates this well above a typical diagnostic AI paper. The major factors driving this recommendation are the strength and fairness of the evaluation, the clinical grounding of the design choices, and the significance of the human-AI collaboration findings. Achieving 95.74% pooled sensitivity on multicenter external data using only single-frame inputs, against a baseline trained on richer multi-frame data, is a compelling demonstration of both diagnostic performance and data efficiency. The finding that AI assistance improved senior clinician specificity from 40% to 76% is clinically striking and independently meaningful. The weaknesses identified, particularly around the internal validation performance gap, the small holdout size, and the subjectivity of the CoT quality evaluation, are real but do not undermine the core contribution. They are addressable refinements rather than fundamental flaws, and the overall rigor of the work more than compensates. The honest benchmarking approach, IRB compliance across four centers, and STARD adherence further reflect the maturity of this submission. PancCADx represents a meaningful step toward transparent, clinically aligned AI decision support in gastrointestinal oncology, and it deserves a place in the MICCAI program. A rebuttal is not needed to reach this conclusion.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    My recommendation remains Accept. The authors’ rejoinder was well-structured, targeted, and effectively addressed the primary points mentioned by all three reviewers without going overboard or promising results beyond the published work.Concerning the internal validation performance gap, the authors clarified that the lower accuracy (83.72%, n=43) reflects statistical instability at small sample sizes, with an approximate 95% confidence interval of [72%, 92%], and a higher proportion of diagnostically challenging cases in the Center A holdout. This is a reasonable and honest explanation, and the commitment to publishing confidence intervals across all cohorts in revision represents a significant advance. Regarding the expert clinician findings, the rebuttal appropriately reframes the null results; n=2 restricts statistical power; and the authors now expressly characterize this as a confidence reinforcement finding rather than a performance increase. This is exactly the nuance the study required, and I appreciate that they dealt with it immediately. The SOTA comparability concern identified by R#2 is convincingly addressed. The sensitivity-specificity trade-off is clinically grounded; in a cancer with a 5-year survival rate of about 10%, missed diagnoses are far more risky than false positives. The rebuttal’s clarification that AUC does not apply to reasoning-based generative models is technically correct and should be explicitly mentioned in the final version. The article contains specific and actionable improvements, such as the addition of a Limitations section, confidence intervals across all cohorts, the definition of Table 1 abbreviations, and structured CoT quality auditing. These are the types of adjustments that boost an existing strong contribution without altering its core assertions. The flaws noted in my previous review remain, but the rebuttal shows that the writers understand them well and have credible replies. Nothing in the response undermines my trust in the fundamental contribution. PancCADx is still a rigorous, clinically informed, and innovative framework that deserves a position in the MICCAI 2026 program.



Author Feedback

We thank all reviewers for their thorough and constructive feedback.

1.Technical Contributions (R#2) Our novelty lies in three innovations: (1) Error-driven preference mining: instead of generic RLHF, we mine preference pairs from the SFT model’s failure modes against expert consensus (386 pairs), making DPO data-efficient in low-resource settings; (2) Conclusion-First CoT grounded in dual-process cognitive theory: prior medical CoT deduces forward (findings to diagnosis), whereas we anchor diagnosis first to prevent logic drift; (3) First evidence that single-frame VLM reasoning matches multi-frame CNN ensembles on EUS, suggesting reasoning quality can compensate for data quantity.

2.SOTA Comparison (R#2) On comparable multi-center external cohorts (3 centers, ~190 patients), PancCADx achieves 95.74% sensitivity with single-frame input, outperforming Joint-AI’s 92.55% with multi-frame voting. The lower specificity (80.41% vs 84.54%) reflects a deliberate trade-off: in pancreatic cancer (5-year survival ~10%), missed diagnosis carries far greater risk than false positives requiring EUS-FNA. We will revise to “superior sensitivity with stronger interpretability and data efficiency” rather than unconditional superiority. Regarding AUC: PancCADx outputs categorical conclusions with reasoning chains, not continuous probability scores. AUC requires adjustable thresholds, which is inapplicable to reasoning-based generative models. We will state this explicitly.

3.Ablation and Component Contributions (R#2) CoT contributes interpretability, not classification gain. In Table 3, CoT reasoning improved senior specificity from 40% to 76% (p=0.003) by providing auditable diagnostic logic. DPO corrects hallucinations. As described in Section 2.2, error-driven discrepancy mining identified 386 cases where SFT outputs diverged from expert consensus. These preference pairs target systematic biases, e.g., over-reliance on “clear boundary = benign” heuristics and anchoring on normal biomarkers while ignoring indirect imaging signs. DPO penalizes these patterns (Eq. 2), aligning reasoning with expert logic. Ablating generative VLMs differs from modular CNNs: removing CoT transforms the output structure entirely, requiring different optimization. Per rebuttal guidelines, we address this based on existing manuscript content.

4.Conclusion-First Bias (R#1) This design mirrors expert recognition-primed decision-making (Ref. 22): experts form rapid diagnostic impressions then retrospectively verify morphological evidence. Moreover, our DPO stage partially mitigates this risk: the 386 preference pairs include cases where the model reached correct conclusions but through selective reasoning, and DPO penalizes such incomplete evidence chains. We will further discuss potential bias risks in the Limitations section.

5.Internal Validation (R#1, R#3) The lower internal accuracy (83.72%, n=43) reflects statistical instability (95% CI approximately [72%, 92%]) and a higher proportion of diagnostically challenging cases in Center A’s holdout. This reflects fragility of point estimates at small sample sizes, not model weakness. We will add CIs across all cohorts.

6.Expert Group Results (R#1, R#3) We stratify findings: novices 69% to 90% accuracy (p<0.001); seniors 40% to 76% specificity (p=0.003); experts no significant change (n=2, high baseline). We acknowledge n=2 limits statistical power. For experts, AI reinforces confidence rather than changing decisions, a clinically meaningful finding we will emphasize rather than grouping under general improvement.

7.Response to R#3 We appreciate R#3’s insightful evaluation and will incorporate structured CoT quality auditing in the revision.

8.Planned Revisions Limitations section (n=2, Conclusion-First bias, single-organ scope); Table 1 abbreviations (AIP/CP/NET/SPT); 95% CIs for all metrics; unified metric formatting; code and model weights upon acceptance.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The manuscript was reviewed by three expert reviewers. While acknowledging merits in the proposed work, the reviewers also raised some concerns about methodological novelty and validation. Such concerns should be addressed to strengthen the value of the work.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns MIC research.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper presents a multimodal framework for pancreatic cancer diagnosis from EUS images and structured clinical data, using SFT and DPO to align a reasoning-capable vision-language model with expert clinical logic. Reviewers appreciated the clinical relevance of the task, the multimodal design, the data-efficient alignment strategy, the multicenter external validation, and especially the prospective reader study with stratified clinicians. Concerns were raised regarding limited ablation analysis, overstatement of state-of-the-art claims, the small expert subgroup, non-significant expert-level improvements, the internal validation performance gap, and the need for clearer limitations and confidence intervals. The rebuttal addressed most of these concerns constructively by agreeing to revise the claims, add limitations, provide confidence intervals, define unclear abbreviations, and better contextualize the expert results and internal cohort instability. While some concerns about component-level attribution remain only partially resolved, the overall clinical validation and human-AI study are strong. I therefore recommend acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper has strong clinical translation value. It combines multicenter external validation with a prospective self-controlled clinician study, which is very rare among the submissions. PancCADx achieves a high pooled external sensitivity of 95.74% and shows meaningful human-AI benefit, especially for novices and senior clinicians, where accuracy/specificity improves significantly. Personally I like this paper much and recommend the acceptance.



back to top