List of Papers Browse by Subject Areas Author List
Abstract
Achieving tumor regression grade 0 (TRG0) after neoadjuvant therapy is critical for postoperative management in esophageal cancer. Recent studies indicate that preoperative CT, biopsy whole-slide images (WSIs), and their fusion can predict pCR or treatment response, suggesting that routine pre-treatment data carries efficacy-related signals. However, refining the endpoint to the decision-critical TRG0 vs TRG1–3 classification remains challenging: TRG is defined on post-treatment whole resection specimens, whereas preoperative CT and biopsy WSIs provide only partial and localized evidence, and false TRG0 predictions may lead to under-treatment or insufficient surveillance. This mismatch, compounded by multi-center shift, makes clinically required low-FPR operating point essential yet difficult. In this paper, we introduce DRGFuse, an alignment-guided CT–WSI framework for robust preoperative TRG0 vs TRG1–3 prediction. Specifically, DRGFuse leverages automatic segmentation to derive stable CT regions and encode tumor and peritumoral context, and employs a pathology foundation model to extract transferable histologic tokens from biopsy WSIs. Moreover, mismatch-aware reliability gating with alignment constraints suppresses irrelevant cross-modal interference and strengthens low-FPR discrimination. Extensive multicenter experiments demonstrate that DRGFuse consistently outperforms unimodal models and representative multimodal baselines, while exhibiting improved robustness under cross-center shift and missing-modality settings. Code is available at: https://github.com/Perasperaasastra/DRG-fuse
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1899_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{LiuZhe_DRGFuse_MICCAI2026,
author = { Liu, Zhenbing AND Liu, Hongzhi AND Yang, Nuo AND Zeng, Dong AND Min, Wenwen AND Li, Hongyu AND Liu, Zaiyi AND Wang, Changmiao AND Lu, Haoxiang},
title = { { DRGFuse: Alignment-Guided Dual-Layer Reliability-Gated CT–WSI Fusion for Preoperative TRG0 Vs TRG1–3 Prediction in Esophageal Cancer } },
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 studies preoperative prediction of TRG0 versus TRG1-3 in esophageal cancer using paired pre-treatment CT and biopsy WSIs. Its main contribution is DRGFuse, a multimodal fusion framework that explicitly models the structural mismatch between post-operative whole-specimen TRG labels and pre-operative, localized CT/WSI evidence, and uses a dual-layer reliability-gated design to regulate cross-modal interaction: a pre-gate suppresses potentially harmful token-level interaction when CT and WSI evidence disagree (via a mismatch-derived coefficient α = 1 - m), and a post-gate implements MoE-style routing with residual logit fusion over the fused and two unimodal experts. The method is paired with training-only regularizers (Align-Guard combining OT and MMD, co-teaching for label noise, and a pAUC surrogate targeting FPR ≤ 0.05) that are explicitly removed at inference, keeping deployment cost bounded. The paper further tailors evaluation to a clinically relevant low-FPR operating regime (pAUCn, TPR@Sp95) and reports five-fold internal plus two held-out external centers, showing consistent low-FPR gains over unimodal and multimodal baselines.
- 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.Rather than focusing on a broader pCR-style endpoint, it targets TRG0 vs. TRG1-3 and explicitly emphasizes low-FPR operation, which is well aligned with the stated clinical risk of false TRG0 predictions leading to under-treatment or insufficient surveillance.
2.The pre-treatment CT and biopsy WSI provide only partial and localized evidence for a post-operative whole-specimen label - is convincing, and the mismatch-aware dual-gate design is a sensible way to reduce harmful multimodal interaction under disagreement while preserving a principled fallback.
3.The separation between training-only regularizers (Align-Guard, co-teaching, low-FPR pAUC surrogate) and the inference-time forward path is explicitly documented, so deployment cost does not grow with the training-side alignment machinery. This is a useful engineering property that similar radiopathomics fusion papers often leave ambiguous.
4.The evaluation protocol is deployment-oriented and stronger than a standard AUC-only comparison. The use of patient-level stratified five-fold CV on the internal cohort combined with two held-out external centers, a Sp95 threshold fixed from validation only, and reporting of pAUCn and TPR@Sp95 better reflects the clinical operating regime the paper argues for.
5.Empirical gains are consistent and practically relevant in the low-FPR regime. Improvements over both unimodal models and strong multimodal baselines are not restricted to internal data, and the ablations support the claim that reliability-gated fallback is an important driver of external low-FPR robustness.
- 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 novelty is somewhat incremental at the component level. Many building blocks are established. The main novelty therefore lies in the integration, task formulation, and deployment-oriented design rather than in a fundamentally new modeling primitive.
2.The “ConventionalFusion” ablation in Table 3, which produces the single largest low-FPR drop and is therefore the strongest piece of evidence for the paper’s central design claim, is under-specified. It is described only as replacing the dual-gate reliability and fallback with “a conventional fusion (concat/single interaction)”. A more concrete specification - e. g. , plain concat + MLP vs. attention fusion vs. a specific prior baseline - is necessary for this ablation to be fully interpretable.
3.The experimental evidence for the deployment framing is still incomplete. Because the paper is explicitly built around low-FPR decision making, readers would reasonably expect explicit reporting of class prevalence per center. The external cohorts (n = 51 and n = 62) are small enough that Sp95-region estimates are statistically sensitive.
4.The abstract and introduction claim robustness under missing-modality settings, but I could not find corresponding quantitative results in the main experimental section instead of just single-modality comparison. Missing-modality robustness is listed as part of the stated contribution.
5.The binary endpoint definition is not clinically justified against alternatives. The paper fixes the cutoff as TRG0 vs. TRG1-3 with TRG0 as positive, motivated by the cost of false TRG0 predictions, but an alternative widely used cutoff is “major pathologic response” (e. g. , TRG0-1 vs. TRG2-3). A brief justification of why the TRG0-only positive class is preferred, and whether the conclusions would be qualitatively preserved under the alternative cutoff, would strengthen the clinical positioning.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
I lean weakly positive on this paper because it addresses a clinically important endpoint (TRG0 vs. TRG1-3), frames the problem in a deployment-relevant low-FPR regime, and presents a coherent multimodal method with reasonably strong multi-center evaluation. The strongest aspect of the submission is that the method design, the operating-point choice, and the evaluation protocol are aligned with the stated clinical motivation. The explicit separation of training-only regularization from the inference-time forward path is also an appealing engineering property.
At the same time, the technical novelty is integrative rather than fundamental, as most building blocks are established. The external cohorts remain modest in size (n = 51 and n = 62). The largest-gap ablation (“ConventionalFusion”) is under-specified, and the stated missing-modality robustness claim is not matched by quantitative results in the main text.
I would move toward Accept if the rebuttal convincingly addresses (a) paired statistical testing on external centers, (b) calibration at the high-specificity region, (c) an external validation that the mismatch score m is a faithful modality-disagreement signal rather than only a gating input, and (d) a more precise specification of the “ConventionalFusion” baseline. Conversely, I would move to Weak Reject if the largest-gap ablation cannot be specified, or if the external-center gains do not survive paired testing.
- 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
This paper proposes DRGFuse, a multimodal framework designed for the preoperative prediction of Tumor Regression Grade 0 (TRG0) versus TRG1-3 in esophageal cancer using 3D CT scans and biopsy WSI. Recognizing the inherent evidence mismatch between localized preoperative biopsies/CTs and whole-specimen postoperative TRG labels, this paper introduces an alignment-guided, dual-layer reliability-gated fusion core. This core explicitly estimates cross-modal mismatch to regulate feature interaction and control the final residual logit routing. The methodology is optimized for deployment safety, prioritizing low-FPR metrics using a specialized training objective.
- 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.Targeting the classification of TRG0 vs TRG1-3 is a critical clinical decision point for esophageal cancer surgery and surveillance. Furthermore, the explicit focus on optimizing for a low FPR aligns with the clinical reality that false TRG0 predictions carry severe risks of under-treatment.
2.The paper proposed dual-layer reliability-gated architecture addresses multimodal conflict. By measuring the mismatch and allowing the model to smoothly fallback to a dominant unimodal expert when modalities disagree, DRGFuse prevents the noise amplification in standard concatenation fusion.
3.The inclusion of a internal cohort and two external cohorts provides evidence of the model’s robustness and generalization capabilities against distribution shifts.
- 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.While the computational method accounts for multimodal mismatch, it cannot magically recover information lost due to physical biopsy sampling errors (e.g., tumor heterogeneity missing the primary aggressive clone). The upper bound of accuracy is fundamentally restricted by the physical evidence presented in the pre-op WSI.
2.The authors should provide a visualization analysis of cases where the WSI expert dominated the routing weight. It would be helpful to clarify what specific microscopic features (e.g., cell morphology, microenvironment) learned by the pathology foundation model are compensating for the macroscopic ambiguity of CT, allowing the model to overcome the localized nature of the biopsy.
3.While the clinical application is strong, the methodological novelty is somewhat incremental. The framework primarily integrates existing, well-established modules (e.g., nnU-Net, UNI, MoE).
- 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?
Please refer to the strengths and weaknesses.
- 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
(1) The study explicitly narrowed the task scope from general response/pCR prediction to a more clinically decision-oriented comparison of TRG0 versus TRG1-3: The paper does not engage in generalized ‘treatment response prediction,’ but rather focuses on the more sensitive and higher-risk decision-making classification of TRG0 versus non-TRG0, clearly stating that the low false positive rate (FPR) working point represents the true clinical deployment objective of concern. (2) A two-layer gated fusion framework is proposed to optimize weights based on the degree of modal conflict; (3) The loss function explicitly incorporates a low-FPR optimization objective for application-oriented training.
- 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 emphasizes that “false-positive TRG0 predictions may lead to inadequate treatment or insufficient follow-up,” thus prioritizing low False Positive Rate (FPR) as the primary objective. This argument holds validity and is more practical than solely pursuing overall Area Under the Curve (AUC). (2) The experimental setup is well-designed, incorporating both an external center and ablation. (3) The proposed fusion strategy demonstrates greater specificity compared to conventional multimodal approaches.
- 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 total sample size was 525 cases, with only 51/62 cases in the internal and external datasets, resulting in a relatively small number of external test sets. (2) It would be preferable to include error case analyses, particularly the imaging/pathological patterns of false-positive TRG0 cases. (3) The abstract claims robustness in handling missing-modality settings, but the main text’s experimental tables lack systematic presentation of missing-modality experiments. This claim currently appears insufficiently substantiated. (4) The correspondence between formulas and flowcharts lacks directness. Please conduct a thorough review of all formulas and variables in the text to ensure clear specification of variable dimensions or domains.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
(1) The research question posed in the paper is valuable, focusing not on generalized multimodal classification but on TRG0 high-specificity prediction driven by clinical safety considerations; (2) The methodological design aligns with the research question, particularly the core approach of mismatch-aware dual gating + fallback, which is well-founded. (3) The experiments are robust, incorporating external validation centers, ablation studies, and comprehensive mechanism analysis.
- 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
We thank the reviewers and AC for their constructive feedback and recognition of our work. We first address three shared concerns, followed by reviewer/AC-specific responses. G1: Methodological novelty. We understand that several components have prior foundations. Our contribution is not an isolated module, but a mismatch-aware dual-layer reliability-gated fusion design for CT-biopsy WSI prediction of TRG0 vs. TRG1-3 under postoperative-label/preoperative-evidence mismatch and low-FPR deployment needs. Since false-positive TRG0 is clinically risky, DRGFuse is designed to remain reliable in the high-specificity region: the mismatch score identifies CT-WSI conflict, the pre-gate suppresses unreliable token interaction, and the post-gate routes among fusion/CT/WSI experts while preserving unimodal fallback. These modules jointly serve “conflict detection - controlled fusion - reliable fallback.” Table 3 supports this: replacing it with ConventionalFusion reduces External A/B TPR@Sp95 from 0.456/0.433 to 0.302/0.280.G2: External validation and calibration. We agree that External A/B are modest and Sp95 estimates can be sensitive. However, the study does not rely only on these cohorts: the Internal cohort contains 525 cases with patient-level stratified five-fold validation. Validation data are used for early stopping, model selection, and Sp95 threshold determination; the fixed threshold is then applied to Internal test and held-out External A/B. External centers are never used for training, tuning, model selection, or threshold determination. All methods use identical splits and the fixed-Sp95 protocol, with mean±SD reporting. We also performed Brier score and ECE analyses to evaluate calibration. G3: Missing-modality robustness. We agree this point was not presented in a sufficiently concentrated way. Table 1 contains complete-missing-modality boundary tests: DRGFuse is degraded into CT-only and WSI-only variants by retaining only the corresponding modality branch and patient-level classifier, removing the other modality path and cross-modal fusion components, while keeping the same training pipeline, splits, and fixed-Sp95 protocol. CT-only / WSI-only retain some discriminative ability, but low-FPR performance is limited; on External A/B, TPR@Sp95 is 0.252/0.242 and 0.270/0.251, compared with 0.456/0.433 for full DRGFuse. R1Q2: ConventionalFusion definition. Due to space limitations, the manuscript only summarized it as conventional fusion. In implementation, ConventionalFusion replaces only the DRGFuse fusion core, while keeping CT/WSI encoders, training settings, splits, and evaluation unchanged. It removes the mismatch-controlled pre-gate, three-expert post-gate routing, and residual logit fallback; projected CT/WSI global features are concatenated and fed into an MLP classifier. Thus, this ablation tests the necessity of reliability gating and fallback. AC: Mismatch score validity. The mismatch score m is not post-hoc, but a reliability signal involved in fusion control. It is computed from CT/WSI representations, their difference, and element-wise interaction, and directly controls alpha=1-m and post-gate routing. Training/validation logs show m is sample-dependent rather than constant; high-m samples have lower alpha and fusion-expert weight, consistent with Fig. 2.The ConventionalFusion ablation further supports m-driven reliability control. R2Q1-Q2/R3Q2: Error cases, interpretability and biopsy limitation. We agree that false-positive TRG0 and WSI-dominant routing analyses help interpret model behavior. The manuscript provides case-level routing information through mismatch score, alpha, and expert weights, which can help analyze weak CT appearance, insufficient WSI residual-risk cues, biopsy locality, heterogeneity, or macro-micro mismatch. DRGFuse does not recover information absent from biopsy; it mitigates unreliable fusion through mismatch-aware gating and fallback under limited or inconsistent CT-WSI evidence.
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.
All reviewers accepted this paper, but they also noted its incremental novelty, pointing out that no new modules were proposed and that the validation set was too small to demonstrate robust results. Specifically, R1 requested paired statistical testing on external centers, calibration at the high-specificity region, validation of the mismatch score as a faithful modality-disagreement signal, and a precise specification of the baseline. The authors should address these concerns in the rebuttal and explain how they plan to revise the paper accordingly.