Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Automated CT triage demands models that are simultaneously accurate across diverse pathologies and reliable under institutional shift. While Vision Transformers provide strong visual representations, many clinically significant findings are defined by quantitative imaging biomarkers rather than appearance alone. We introduce JANUS, a physiology-guided dual-stream architecture that conditions visual embeddings on macro-radiomic priors via Anatomically Guided Gating. On the MERLIN test set (N =5,082), JANUS attains macro-AUROC 0.88 and AUPRC 0.74, the best performance among reproduced baselines under our protocol. It generalizes to an external cohort (N =2,000; AUROC 0.87), with the largest gains on size and attenuation-defined findings and improved calibration on both cohorts. We further quantify prediction suppression using the Physiological Veto Rate (PVR), showing that under domain shift JANUS reduces high-confidence false positives substantially more often than true positives. Together, these results are consistent with physically grounded conditioning improving both discrimination and reliability in CT triage. Code is made publicly available at github repository https://github.com/lavsendahal/janus and model weights are at https://huggingface.co/lavsendahal/janus.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5443_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lavsendahal/janus https://huggingface.co/lavsendahal/janus

Link to the Dataset(s)

Merlin Dataset: https://stanfordaimi.azurewebsites.net/datasets/60b9c7ff-877b-48ce-96c3-0194c8205c40 Duke-Abdomen2026 Dataset: Private not available publicly

BibTex

@InProceedings{DahLav_JANUS_MICCAI2026,
        author = { Dahal, Lavsen AND Bhandari, Yubraj AND Rubin, Geoffrey AND Lo, Joseph Y.},
        title = { { JANUS: Anatomy-Conditioned Gating for Robust CT Triage Under Distribution Shift } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    JANUS is a dual-stream architecture for automated abdominal CT triage that addresses the “geometric gap” in vision transformers. Standard attention pooling discards quantitative information like organ volume and vessel diameter that defines many clinically urgent findings. It pairs a DINOv3 visual stream with a symbolic stream of macro-radiomic priors (organ sizes, densities, calcific burden) derived from TotalSegmentator, fusing them via disease-specific multiplicative gating rather than additive concatenation. The sigmoid-bounded gate acts as a “physiological veto,” structurally enabling the model to suppress visual predictions when quantitative measurements disagree. On MERLIN (N=5,082) and an external cohort (N=2,000), JANUS achieves macro-AUROC 0.88/0.87 with the best calibration, outperforming both a vanilla ViT baseline and an additive fusion approach that actually degrades under distribution shift. The gains are concentrated on measurement-defined findings (organomegaly, aneurysm, steatosis) while focal pathologies like appendicitis see minimal change, consistent with the priors carrying signal exactly where expected.

  • 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 core insight that quantitative measurements should constrain visual predictions via multiplicative gating rather than augment them additively is well-motivated and validated by the +OSF degradation under shift (0.84->0.81) versus JANUS maintaining performance (0.88->0.87)
    • Controlled experimental design: JANUS and ORACLE-CT+OSF share the same backbone, macro-radiomic prior bank, and training protocol, isolating the fusion mechanism as the variable.
    • The PVR metric and veto selectivity analysis (10.8×) directly test the claimed suppression mechanism rather than just reporting aggregate AUROC. -The pathology stratification by diagnostic mechanism (geometric, densitometric, focal, fluid) is clinically meaningful and shows the expected pattern: gains where priors carry signal, no degradation where they don’t -The gate behavior analysis (Fig. 2a) showing monotonic diameter-gate relationship within both positive and negative cases is a convincing mechanistic probe
  • 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.
    • JANUS relies on TotalSegmentator succeeding upstream. The scalar corruption experiment uses simulated noise rather than realistic segmentation failures (missed organs, boundary errors, etc). How does JANUS degrade when the segmentation mask itself is wrong?
    • The specific macro-radiomic features used per pathology are not well described. The paper describes the feature space broadly (organ volumes, diameters) but does not provide a per-disease/feature feature list or feature count. Without knowing which features feed into each gate, it is difficult to assess whether the gating mechanism is learning meaningful clinical associations or simply correlating with a few dominant features.
    • There is no ablation using macro-radiomic priors alone without the visual stream. A scalar-only baseline would clarify how much discriminative signal the priors carry independently and how much the visual stream contributes beyond what measurements already capture. This is important because for measurement-defined findings (where JANUS gains are largest), it is plausible that the priors alone approach the gated model’s performance.
  • 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

    LLMs were used solely for polishing review as per acceptable reviewer policy.

  • 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?

    JANUS presents a well-motivated architectural contribution — multiplicative gating of macro-radiomic priors to constrain visual predictions — with a carefully controlled experimental design that isolates the fusion mechanism as the variable. The results are convincing where expected: measurement-defined findings show clear gains, the veto selectivity analysis (10.8×) directly validates the suppression mechanism, and the model maintains performance under distribution shift while the additive baseline degrades.

    The concerns are around completeness. The macro-radiomic features per pathology are not described, there is no scalar-only ablation to assess how much signal the priors carry independently, focal findings show slight degradation with no mitigation explored. These are addressable in revision without undermining the core contribution.

  • Reviewer confidence

    Somewhat confident (2)

  • [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

    The paper introduces JANUS, a dual-stream architecture designed for robust 3D CT triage. Its primary contribution is the Anatomically Guided Gate, a multiplicative bottleneck that conditions visual representations (from a Vision Transformer) on macro-radiomic priors (quantitative biomarkers like volume, diameter, and Hounsfield Unit density).This approach addresses the pooling-induced geometric gap in standard transformers, where structural measurements are lost during feature aggregation. Additionally, the authors propose a new metric, the Physiological Veto Rate (PVR), which quantifies the model’s ability to selectively suppress high-confidence false positives that contradict physical reality.The framework demonstrates superior calibration and robustness under distribution shift across internal (N = 5,082) and external (N = 2,000) cohorts.

  • 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.Novel Methodological Inductive Bias - The shift from additive fusion to multiplicative gating is a significant innovation. It allows the symbolic stream to act as a physiological veto, attenuating visual features that are physically implausible (e.g., suppressing an aneurysm prediction if the vessel diameter is measured at a normal 2.0 cm). 2.State-of-the-Art Visual Backbone - The implementation leverages DINOv3, incorporating advanced features like register tokens to stabilize attention maps and Gram anchoring to maintain feature consistency. 3.Rigorous Validation of Robustness - The study goes beyond standard AUROC reporting by including noise-stress tests (corrupting scalar priors by up to 50%) and demonstrating that JANUS maintains performance (AUROC 0.84) where additive models fail. 4.Clinically Relevant Metrics - The introduction of PVR and veto selectivity (achieving 10.8x on external data) provides a transparent way to measure how effectively the model reduces alarm fatigue caused by overconfident false positives. 5.Superior Calibration - JANUS significantly reduces Expected Calibration Error (ECE) to 0.09 (internal) and 0.15 (external), which is critical for clinical trust and triage reliability.

  • 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.Sequential Pipeline Dependency - The model relies on a pre-processing segmentation step (TotalSegmentator). While robust, the paper does not fully quantify how the system handles cases of catastrophic segmentation failure beyond a uniform-weight fallback mechanism. 2.Dimensionality Constraint - The visual stream uses a 2.5D tri-slice encoding. While computationally efficient, it may miss subtle out-of-plane features compared to native 3D backbones or causal convolutional architectures like BTB3D. 3.Pathology-Specific Prior Mapping - The current framework requires a predefined mapping of pathologies to specific macro-radiomic features. This may limit the discovery of complex, non-obvious multi-organ biomarkers (e.g., hepatorenal syndromes) that a more dynamic cross-attention mechanism could capture. 4.Limited Baseline Comparison - While the comparison against ORACLE-CT is thorough, the paper lacks a direct benchmark against frozen 3D foundation models like CT-CLIP or MedSigLIP under the same evaluation protocol .

  • 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.

    (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?

    JANUS addresses a fundamental limitation in the application of Vision Transformers to volumetric medical imaging, which is the disconnect between visual textures and the quantitative thresholds that define clinical diagnosis. The methodological innovation is technically sound and provides a clear mechanism for enhancing model reliability under institutional distribution shift. The experimental design is rigorous, utilizing a large-scale internal dataset and a sizable external cohort with a high-quality label consensus. The results show consistent gains in both discriminative power (AUROC) and calibration (ECE), particularly for finding categories where physical dimensions are paramount.

  • 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 presents an automated method for CT triage through Guided Gating guided by macro-radiomic priors. The method outperforms the state-of-the-art and under domain shift, it reduces false positives and improve discrimination and reliability in CT triage.

  • 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 methodology is sound and well presented. The paper presents an architecture named JANUS that incorporates the aforementioned guided gating. Such gating is learrned through two streams. The first one uses visual information through DINOv3 transformer, thus obtaining ROI with masked attention which are transformed into weights that applied to the tokens, obtaining a score (equation 1). Such score passes depending on the information coming from macro-radiomics information (second stream), i.e. a binary anatomically guided gate which is based in a macro-radiomic prior through a simple model of type Ws+b (equation 2). The experimental evaluation seems to be exhaustive considering I am no expert in this field of CT triage. It evaluates against the very recent ORACLE-CT in two datasets, one public (MERLIN) and one private from an US hospital. JANUS seem to outperform that recent work on both datasets on different evaluation metrics as well as reducing false positives (table 1 and figure 2) which is consistent across (almost) all pathologies (table 2).

  • 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.

    Some aspects of the methodology where unclear to me:

    • Specifically how are the priors obtained, I am assuming it is through segmentation (assumption is that s_l belongs to S), in which case, which is the segmentation being applied. As this prior is applying a binary gate over all the visual procedure (ViT+ROI-masked attention), knowing the segmentation procedure is particularly relevant.
    • Other details could be giving more detail as well. Section 2.3 was too brief for me, i.e. why the choice for the the epochs, what are E_ignore, and E_ramp?, why w_max is set to 0.3.- For the experimental setup, since I am no expert in CT-Triage, I did not understand what are Qwen-3 and MedGemma. Also, I think the authors should not have mentioned the country of the private-data hospital (US) for a better blind process.
    • Some insight on why JANUS underperforms for Focal/Control (table 2) would have been of interest.
  • 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.

    (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 is a well presented paper. There are some weaknesses, but the interest of the methodology and the experimental evaluation overcome those.

  • 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 #4

  • Please describe the contribution of the paper

    The authors propose JANUS, a physiology-guided dual-stream architecture that leverages conditioned gating to encode quantitative imaging biomarkers. Their central argument is that traditional visual representations learned through standard objective functions may not adequately capture defined quantitative imaging biomarkers from appearance alone, and could therefore create spurious associations between visual presentations, outcomes, and diagnostic biomarkers. The authors evaluate this anatomically guided gating on the MERLIN test set (N=5,082) and an external private cohort (N=2,000). Compared against baselines including DINOv3 ViT-B/16 and ORACLE-CT+OSF, JANUS outperforms these models across almost all tasks.

  • 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.

    i. Encoding quantitative biomarkers already associated with disease prognostics is clinically relevant and leverages years of established clinical knowledge. ii. Strong comparison models — including DINOv3 ViT-B/16, ORACLE-CT, and ORACLE-CT+OSF, provide well-defined benchmarks for evaluating the performance improvement. iii. Stratified performance across multiple targets on an external dataset shows no signs of performance degradation and highlights the model’s ability to generalize under distribution shift, spanning cohort, disease, and potentially population differences.

  • 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.

    i. Demographic details for both cohorts would allow for a better understanding of how well JANUS responds to distribution shift. As currently presented, we are left to assume the two cohorts are distinct, but they could share similar race, gender, and age demographics that could bias the assessment of model generalization. ii. Corruption of biomarker signals is limited in scope and represents only a preliminary attempt at exploring the impact of synthetic segmentation errors or inter-reader variability. The evaluation uses random uniform noise rather than realistic segmentation failure modes, which may underestimate vulnerability to segmentation errors on unusual anatomy or post-surgical cases. iv. Additional ablation experiments would have been informative to better understand the individual contribution of the symbolic stream. It remains unclear whether the model is substantially benefiting from the additional information or receiving only a marginal boost in performance, particularly for findings where scalar priors are weakly informative.

  • 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

    i. Including a demographic comparison table between the MERLIN and external cohort datasets would substantiate any claims of true distribution shift across demographic variables such as age, sex, and race. ii. The external cohort labels are derived from a dual-LLM consensus (Qwen-3 and MedGemma) rather than radiologist ground truth. The reliability and systematic biases of this labeling approach relative to expert annotation are not discussed in the limitations sections, and limits confidence in the external validation results.

  • 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 proposed work is well grounded clinically and addresses an important problem in medical vision modeling: imaging biomarkers associated with disease prognosis are not always well embedded in learned representations, and there is limited methodology to verify that they are being leveraged. The JANUS framework reintroduces clinically meaningful quantitative information through a principled gating mechanism, with demonstrated improvements in both discrimination and calibration. The external cohort evaluation supports generalization, though demographic characterization of the shift would strengthen these claims.

  • Reviewer confidence

    Somewhat confident (2)

  • [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

N/A



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.

    This paper received four consistent reviews to recommend the decision of accept.



back to top