List of Papers Browse by Subject Areas Author List
Abstract
Multiphasic contrast-enhanced CT (CECT) is widely used for abdominal lesion characterization, yet it carries inherent risks of contrast-induced nephropathy, escalates acquisition burden, and heavily contributes to radiologist workload. To address these challenges, we introduce a novel multi-center benchmark for multi-organ abdominal disease diagnosis and automated radiology report generation, which learns to synthesize contrast-enhanced findings from single-phase non-contrast CT (NCCT). To support this, we curated a large-scale dataset of paired NCCT–CECT studies and their corresponding contrast-enhanced radiology reports from two centers, partitioned into internal sets and an external validation cohort. Under a unified evaluation protocol, we benchmarked five contemporary deep learning architectures encompassing chest-specific, abdomen-specific, and general-purpose multimodal domains. Extensive experiments demonstrate that NCCT retains diagnostic signals, achieving an average multi-organ AUC of 71.1% on the internal cohort and 66.2% on the external cohort, respectively. By releasing this dataset and standardized benchmark publicly, this study aims to catalyze future research into safer, resource-efficient, and globally accessible contrast-free abdominal imaging workflows. Code is available at: https://github.com/xmed-lab/TriALS-Report.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5454_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xmed-lab/TriALS-Report/
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ElbMar_AMulticenter_MICCAI2026,
author = { Elbakry, Mariam AND Sheha, Aliaa Sayed AND Tantawy, Salma Hassan AND Yassin, Aya AND Spampinato, Concetto AND Lekadir, Karim AND Li, Xiaomeng AND Elbatel, Marawan},
title = { { A Multi-center Benchmark for Abdominal Disease Diagnosis and Report Generation from Non-Contrast CT } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a multi-center benchmark for abdominal disease diagnosis and radiology report generation from non-contrast CT. The authors curated 1,254 paired cases from two Middle Eastern tertiary care centers, where each case contains a non-contrast CT volume and its corresponding triphasic contrast-enhanced radiology report. The benchmark includes two tasks: generating abdominal radiology findings from single-phase NCCT and performing multi-label disease classification across 15 organs and 53 abdominal pathologies. Under a unified evaluation protocol, the authors evaluate several existing 3D medical imaging and vision-language models, including CT2Rep, BTB3D, M3D, Merlin, and Pillar-0.They use GREEN, RadGraph-XL, AUROC, and macro-F1 to assess report generation quality and diagnostic performance. Overall, the main contribution is a clinically motivated NCCT-to-contrast-report benchmark and an initial feasibility study showing that NCCT may contain diagnostic signals exploitable by deep learning models.
- 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 main strength of this paper is the clinical relevance of the problem. Non-contrast CT is safer, cheaper, and more accessible than contrast-enhanced CT, especially for patients with renal impairment, contrast contraindications, or limited access to contrast-enhanced imaging. Therefore, studying whether NCCT contains useful diagnostic signals is clinically meaningful. Second, the paper provides a relatively systematic multi-center dataset and benchmark. The dataset contains 1,254 paired NCCT and contrast-enhanced report cases, with both internal testing and external validation, which is stronger than a purely single-center evaluation. Third, the task formulation is more comprehensive than single-disease classification. The paper evaluates both report generation and multi-label abdominal disease classification across 15 organs and 53 pathologies, making the setting closer to real abdominal radiology practice. Fourth, the report generation evaluation does not rely solely on text-overlap metrics such as BLEU and ROUGE. The authors also use GREEN and RadGraph-XL, which better capture clinical entity and relation correctness. This is more appropriate for medical report generation. Fifth, the paper compares several existing 3D medical imaging and vision-language models, including zero-shot and fine-tuned settings, providing useful baselines for future work.
- 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.
The major weakness of this paper is its limited methodological novelty. The paper mainly applies existing models such as CT2Rep, BTB3D, M3D, Merlin, and Pillar-0 to a new abdominal NCCT benchmark. It does not propose a new model architecture, training objective, inference mechanism, or clinically grounded constraint. Therefore, the work is better positioned as a resource or benchmark paper rather than a methodologically novel contribution. Second, although the task is clinically meaningful, it also raises important medical risks. Many key findings in contrast-enhanced CT reports depend on enhancement patterns, such as arterial enhancement, portal venous washout, vascular invasion, and lesion enhancement characteristics. Generating such contrast-dependent descriptions from NCCT alone may encourage the model to infer or “guess” information that is not reliably visible in the input image. The paper does not sufficiently demonstrate that the generated contrast-enhanced findings are truly supported by image evidence rather than dataset bias or report-template memorization. Third, the reliability of the ground truth is limited. Disease labels are mainly extracted from contrast-enhanced reports using the RATE/LLM framework, and only a subset is audited by radiologists. The paper lacks comprehensive lesion-level manual annotations, localization validation, or lesion-level grounding. As a result, it is difficult to determine whether the model truly identifies lesion locations and imaging features. Fourth, the external validation set is relatively small. Center 2 contains only 169 cases, which is limited for evaluating 53 diseases across 15 organs. The performance of rare pathologies may therefore be unstable. The paper also lacks sufficient confidence intervals, statistical significance testing, or bootstrap analysis. Fifth, the classification performance is not particularly strong, especially in terms of macro-F1.Although the authors emphasize that NCCT retains a high proportion of diagnostic power, the external average AUC is only around 0.63 and the F1 scores are low. Thus, the results suggest that NCCT contains some usable diagnostic signal, but they do not support strong claims that NCCT can approach or replace contrast-enhanced CT. Sixth, the paper lacks systematic clinical error analysis. For medical report generation, the key issue is whether the model misses malignant findings, hallucinates non-existent lesions, misdescribes enhancement patterns, or reverses negation. The paper does not sufficiently analyze false positives, false negatives, hallucinations, or critical finding misses. Seventh, the baseline and training details are not sufficiently transparent. The input processing, preprocessing pipeline, prompts, fine-tuning settings, hyperparameters, number of epochs, and model selection criteria for each baseline are under-specified, which weakens the fairness and reproducibility of the benchmark.
- 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 provide sufficient information for 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?
I do not identify a clear ethical violation, but the information provided is insufficient. The main concern is that the manuscript does not adequately describe ethical approval, informed consent or consent waiver, data de-identification procedures, cross-center data sharing permissions, and how privacy leakage will be prevented if the dataset is released publicly. Since the study involves medical images and radiology report text, the authors should provide a detailed ethics and data governance statement.
- 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?
I assign a score of 3 because the paper has application value and dataset value, but it is not yet strong enough for acceptance. The paper proposes a clinically meaningful NCCT-to-contrast-report benchmark and evaluates several existing models on multi-center data, which are its main strengths. However, the work lacks substantial methodological novelty, with its main contribution lying in data collection and benchmark construction. The experimental results show that NCCT contains some usable diagnostic signal, but they do not sufficiently support strong claims about generating contrast-enhanced-equivalent reports or reducing reliance on contrast-enhanced CT. The external validation set is small, macro-F1 scores are low, and the paper lacks confidence intervals, statistical significance testing, lesion-level validation, and systematic clinical error analysis. In addition, the labels are mainly derived from reports using an LLM-based extraction pipeline, without comprehensive manual annotation or lesion-level grounding. Overall, the work is better viewed as a preliminary benchmark or resource paper, but the current version is limited in methodological novelty, clinical evidence, and reproducibility. Therefore, I recommend weak rejection.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 #2
- Please describe the contribution of the paper
First to construct and publicly release a multi center, paired non contrast CT (NCCT) and contrast enhanced CT (CECT) abdominal dataset with 1,254 patients, including internal and external validation splits. First to define and systematically evaluate the task of generating contrast enhanced radiology reports from single phase NCCT, covering 15 organs and 53 abdominal pathologies. First to benchmark five contemporary 3D medical imaging report generation models (CT2Rep, BTB3D, M3D, Merlin, Pillar 0) under a unified protocol using clinically relevant metrics (GREEN, RadGraph XL)
- 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.
Highly clinically relevant problem – avoiding contrast‑induced nephropathy and allergic reactions, particularly important for patients with renal impairment and resource‑limited settings. Strong dataset contribution no publicly available paired NCCT‑CECT multi‑center abdominal dataset exists; this work fills a critical gap. Advanced evaluation metrics – moves beyond BLEU/ROUGE to GREEN and RadGraph‑XL, focusing on clinical entity correctness.
- 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.
Limited methodological novelty – the paper does not propose any new model or training strategy; it only fine‑tunes and evaluates existing models (CT2Rep, BTB3D, M3D, Merlin, Pillar‑0). The core contribution is task definition + dataset, not methodological innovation.
Insufficient clinical evaluation – despite using GREEN and RadGraph‑XL, no blinded radiologist assessment of generated reports is provided. It remains unclear whether the reports are clinically acceptable (e.g., missing small HCC or falsely reporting malignancy).
Significant performance drop on external validation without statistical testing – e.g., Merlin’s AUC drops from 0.691 to 0.631, F1 from 0.251 to 0.248.No p‑values or confidence intervals are reported, making it impossible to judge significance.
Fundamental information ceiling – certain enhancement kinetics (e.g., arterial hyperenhancement, portal venous washout) physically do not exist in NCCT. The paper does not discuss which lesions are in principle impossible to infer from NCCT, nor does it provide clinical risk disclaimers.
- 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.
(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 paper introduces a clinically important new task (NCCT → contrast‑enhanced report) and contributes a scarce multi‑center dataset. The experimental design is reasonably sound, with internal + external validation and metrics better than traditional NLG. However, methodological novelty is limited, clinical evaluation is incomplete, external performance drops noticeably without statistical testing, and the fundamental information ceiling of NCCT is not discussed. Overall, the paper is slightly above the acceptance threshold, but its rejection would not be surprising, depending on the rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
The paper proposes TriChat, a vision-language framework for automated report generation from triphasic abdominal CECT volumes. Its main contribution is to model cross-phase enhancement dynamics across arterial, portal venous, and delayed phases using Cross-Phase Spatiotemporal Interleaving, so that the model can better capture hemodynamic patterns such as arterial enhancement and washout. The paper also introduces the TriCECT dataset and an abdominal clinically unbiased reward engine to address long-tailed abdominal disease recognition during report generation.
- 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 paper has several strengths. (1) The writing logic and overall structure are clear, making the motivation, method, and experiments easy to follow. (2) The experimental comparison is relatively comprehensive, including several public or existing models as baselines. (3) The proposed method is interesting and clinically relevant, as it targets triphasic CECT report generation and attempts to model cross-phase information for abdominal diagnosis.
- 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.
The paper has some weaknesses. (1) There is a gap between the claimed goal and the evaluation design. The paper argues that TriChat captures cross-phase enhancement kinetics in triphasic CECT, but the main evidence is improvement in report generation metrics and disease-level macro-F1.These metrics evaluate whether the final report or disease label is correct, but they do not directly verify whether the model actually uses arterial, portal venous, and delayed-phase enhancement changes. A higher F1 score may come from single-phase appearance, lesion location, organ context, or dataset bias rather than true cross-phase reasoning. The paper would be stronger with phase-specific ablations, phase-order shuffling, single-phase vs triphasic comparisons, or subgroup analysis on diseases that explicitly depend on enhancement patterns, such as HCC. (2) The reward engine uses disease weighting and clinical keyword mappings, but the paper does not sufficiently analyze whether these rewards improve true clinical reasoning or mainly encourage matching predefined keywords. More detailed ablation and failure analysis would help support this component. (3) The clinical feasibility is not sufficiently validated. Although the paper targets automated radiology report generation, there is no radiologist reader study, clinical error analysis, or evaluation of whether the generated reports are acceptable for real clinical use. NLG metrics and disease F1 are useful, but they are not enough to prove clinical reliability.
- 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.
(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?
I recommend acceptance. The paper addresses a clinically meaningful problem and proposes an interesting framework for triphasic CECT report generation. The writing is clear, the structure is easy to follow, and the comparison with existing models is relatively comprehensive. However, the main limitation is that the evaluation does not directly verify the claimed cross-phase enhancement reasoning. The improvements in report-generation metrics and disease-level macro-F1 show better outputs, but do not prove that the model truly uses arterial, portal venous, and delayed-phase enhancement dynamics. The ACURE reward mechanism and clinical feasibility could also be further validated. Overall, despite these limitations, the paper presents a promising and clinically relevant direction.
- 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 #4
- Please describe the contribution of the paper
The paper investigates whether radiology reports for multiphasic contrast-enhanced abdominal CT can be generated from single-phase non-contrast CT volumes. Authors curate a multi-center dataset of paired non-contrast and contrast-enhanced abdominal CT studies with corresponding radiology reports, and introduce an expert-annotated benchmark for multi-organ abdominal disease diagnosis and report generation. The study evaluates several contemporary deep learning architectures under a unified protocol. The main finding is that non-contrast CT contains substantial diagnostic signal for predicting findings typically assessed using contrast-enhanced imaging. The proposed benchmark shows that models trained on non-contrast CT can achieve moderate multi-organ diagnostic performance and generate contrast-enhanced-style findings, suggesting a potential path toward reducing dependence on contrast agents in abdominal imaging workflows.
- 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 introduces a valuable benchmark dataset for the community, consisting of 1,085 cases from the internal center and 169 cases from an external center. The inclusion of expert annotations and multi-center validation makes the dataset a useful resource for future work on abdominal CT report generation and contrast-free imaging.
2- The study empirically demonstrates that latent representations learned from single-phase non-contrast CT, after task-specific fine-tuning, can capture clinically relevant information that partially overlaps with findings usually obtained from multiphasic contrast-enhanced CT. This is an important direction, given the cost, acquisition burden, and potential risks associated with contrast administration.
3- The paper provides a systematic evaluation of five contemporary foundation-model-based architectures under a unified protocol. The reported performance gains after fine-tuning are substantial, and the preservation of trends across two centers strengthens the empirical validity of the findings.The results also highlight an important limitation of current medical foundation models: large-scale generic pretraining does not necessarily generalize off-the-shelf to new anatomical regions or imaging tasks. However, the study shows that with a relatively modest fine-tuning dataset, these models can be adapted effectively to abdominal CT, suggesting that foundation models provide useful transferable priors rather than complete anatomical or clinical understanding out of the box.
- 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 report 5-fold cross-validation or confidence intervals obtained via bootstrapping. Although the gains over zero-shot baselines are substantial, stronger statistical validation would be expected to support the robustness of the results. The absence of 5-fold cross-validation weakens the generalization claims.
2- The qualitative report-generation analysis is supported by only a single example report. This is insufficient for evaluating whether the model consistently captures contrast-specific findings. The central claim of the paper is that a model trained on non-contrast CT can infer findings usually obtained from contrast-enhanced imaging, but this claim is not quantitatively assessed at the level of contrast-specific entities.It remains unclear whether the improvement after fine-tuning comes from learning abdomen-specific terminology and report style, or from truly capturing latent diagnostic information that is not easily visible to the human eye on non-contrast CT. This distinction is critical for the paper’s main claim. Without entity-level metrics, contrast-specific finding extraction, or systematic expert evaluation, the evidence is not sufficient to convincingly demonstrate that the model captures contrast-enhanced findings rather than simply generating more plausible abdominal CT reports.
3- Since this is a new dataset, data collection period, scanner brand and information regarding the availability of patient metadata (age, sex, etc.) would be very useful and should be reported.
4- Given the page limit, it is hard to include label distribution for each disease in the dataset. But this also makes it very hard to interpret the results especially for AUC and F1 scores. A random baseline can be included in Table 4 to make comparison easier.
5- Fig2 -> A sample ground truth and AI generated report was given. Given that many report generation metrics are not intuitively interpretable (not like sensitivity or specificity), it would be helpful to add the GREEN and RadGraph-XL metrics between ground truth and the AI generated report on the figure.
6- Discussion -> “our benchmark proves that AI-driven NCCT evaluation can retain over 90% of the diagnostic classification power of the clinical gold standard” This sentence might be misleading. AI-driven NCCT evaluation retains 90% of AI-driven contrast enhanced CT evaluation, not the clinical gold standard.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Although the paper aims to answer a clinically relevant question, the dataset size is relatively small (especially external validation) considering the large number of labels (53). The dataset metadata is largely missing except for acquisition settings.
- 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 #5
- Please describe the contribution of the paper
This paper introduces a multi-center benchmark for abdominal disease diagnosis and radiology report generation from non-contrast CT (NCCT). The key contribution lies in constructing a paired dataset of NCCT volumes and contrast-enhanced (CECT-derived) radiology reports, enabling the study of whether contrast-enhanced diagnostic information can be inferred from non-contrast imaging alone.
The authors further define two tasks: 1.multi-label disease classification across multiple abdominal organs and 2.free-text report generation, then evaluate a diverse set of existing 3D vision-language and CT-based models under a unified protocol. The work demonstrates that NCCT retains a substantial portion of diagnostic signal, achieving competitive performance compared to multiphasic CT, suggesting the feasibility of contrast-free diagnostic workflows.
Overall, the paper contributes a new dataset, a standardized benchmark, and an empirical study that highlights the potential of AI to reduce reliance on contrast-enhanced imaging in abdominal radiology.
- 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.Clinically meaningful problem formulation The paper addresses an important and underexplored problem: inferring contrast-enhanced diagnostic information from non-contrast CT. This has clear clinical relevance, particularly for patients with contraindications to contrast agents and in resource-limited settings.
2.Dataset and benchmark The construction of a multi-center dataset with paired NCCT scans and contrast-enhanced reports is a major contribution.
3.Multi-organ/multi-task setting Unlike prior works that focus on single organs or specific pathologies, this benchmark supports multi-organ (15 organs) and multi-disease (53 pathologies) evaluation, as well as both classification and report generation, which increases its practical relevance.
4.Evaluation protocol The authors evaluate SOTA models from multiple model families (CT-specific, multimodal, and general 3D architectures) under a unified setup.
5.The empirical finding that NCCT achieves a high fraction of the diagnostic performance of multiphasic CT is interesting and potentially impactful, suggesting that latent diagnostic information exists in non-contrast scans.
- 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.Limited population diversity Although the dataset is multi-center, all data are derived from two institutions within a single country. This raises concerns about the generalizability of the findings to other populations, imaging protocols, and healthcare systems.
2.Absence of true paired image-level supervision The framework relies on NCCT inputs and contrast-enhanced reports as supervision, without explicit pixel-level correspondence between NCCT and CECT images. This may limit the ability to fully capture enhancement-specific phenomena because the reports from CECT are information-filtered compared to original CECT images.
3.Potential bias from report-derived labels The use of automatically extracted labels from radiology reports (even with partial auditing) may introduce noise or bias.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 paper addresses a clinically important problem and provides a valuable multi-center dataset and benchmark for NCCT-based diagnosis and report generation. The experimental results are solid and suggest meaningful diagnostic potential from non-contrast CT. However, concerns remain regarding generalizability (single-country data) and label quality due to automated extraction.
Overall, I find the dataset contribution and clinical relevance sufficient for a weak accept.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Author Feedback
Dear Area Chairs and Reviewers,
We thank all reviewers and the Meta-Reviewer for constructive feedback. We are encouraged by the recognition of our “highly clinically relevant” (R2) and “important and underexplored problem” (R6); the “first multi-center, paired NCCT and CECT abdominal dataset with 1,254 patients” (R2), filling “a critical gap” (R2); a “valuable benchmark dataset” with “expert annotations and multi-center validation” (R5), under a “systematic” “unified protocol” (R5) with “advanced evaluation metrics” (R2). Our finding that NCCT retains “a high fraction of the diagnostic performance of multiphasic CT” was called “interesting and potentially impactful” (R6).
1.Methodological contributions (R1, R2): Our methodological novelty stems from task formulation and evaluation design rather than architectural changes specifically: (i) a novel task formulation: contrast-equivalent multi-organ (15 organs, 53 pathologies) report generation from single-phase NCCT; (ii) a unified clinical entity-level evaluation framework “beyond BLEU/ROUGE, focus on clinical entity correctness” (R2); (iii) the first quantitative evidence of latent NCCT imaging biomarkers across 53 abdominal pathologies as “latent diagnostic information not easily visible to the human eye” (R5).
2.NCCT image evidence and information ceiling (R1, R2, R5, R6): Disease classification is itself a falsifiable image-evidence test: AUROC against radiologist-audited labels on held-out external data cannot be achieved through template memorization. Decisive evidence: (a) on identical NCCT inputs, Merlin retains 90% of the multiphase upper bound (0.691 vs. 0.765 internal AUC, Table 4), a same-architecture, same-labels comparison isolating image content as the only variable; (b) zero-shot foundation models pretrained on contrast CT fail catastrophically (M3D zero-shot GREEN 1.695) while finetuning yields 459% relative GREEN improvement (Merlin 6.254→34.988); (c) gains are preserved across the external center with distinct reporting style, ruling out single-institution template bias. Per R5, we will rephrase as AI-driven NCCT retains ~90% of AI-driven CECT, not of the clinical gold standard. We agree with R2 that enhancement-kinetic findings (arterial hyperenhancement, portal venous washout) cannot be physically inferred from NCCT; we will add an information ceiling subsection delineating unattainable categories and reframe NCCT as a complementary screening modality with clinical risk disclaimers.
3.Statistical validation (R1, R2, R5): We will add 95% bootstrap CIs across all reported metrics, paired significance tests across baselines, and 5-fold cross-validation on the internal cohort in the camera-ready.
4.Label reliability and clinical error analysis (R1, R2, R5, R6): Labels combine the RATE pipeline (used in CT-RATE/Pillar-0) with radiologist auditing on a stratified subset, achieving 90% clinical relevance (§2.2). Our chosen GREEN and RadGraph-XL (Entity/Partial/Complete) provide systematic clinical error quantification at scale, penalizing hallucinated findings, missed findings, and incorrect entity–relation structure across all 388 test cases.
5.External scope, metadata, qualitative cases (R1, R5, R6): Our Egyptian multi-center cohort addresses an underserved population where contrast access is limited. We will add per-disease prevalence, a random baseline (R5), scanner brands, demographics, additional qualitative reports in supplementary, and explicit framing of regional scope as both strength and limitation.
6.Ethics and reproducibility (R1): Formal IRB approval was obtained. Full ethics statement, code, preprocessing, baseline prompts, hyperparameters, and the dataset will be released upon acceptance.
We thank the reviewers for their constructive feedback and suggestions and will incorporate them in the camera-ready version.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
Most reviewers accepted this paper, with only one reviewer recommending weak rejection due to limited methodological novelty, insufficient clinical evidence (small external validation set, low macro-F1 scores, no statistical testing or lesion-level validation), reliance on LLM-extracted labels without manual annotation, and an overall framing as a preliminary benchmark rather than a complete study. The authors should address the reviewers’ concerns in the revised manuscript.
