List of Papers Browse by Subject Areas Author List
Abstract
Deploying medical Foundation Models (FMs) is often severely compromised by continuous clinical distribution shifts. While Test-Time Adaptation (TTA) offers a promising solution, existing paradigms rely on overly simplified, uni-directional optimization that blindly enforces confidence on flawed representations, inevitably precipitating error accumulation and model collapse. To address this critical safety gap, we propose Entropy-Consistency Flow Rectification (ECFR), which fundamentally redefines TTA as a spatial routing problem. Central to our approach are the Entropy-Consistency Quadrants, which explicitly disentangle streaming data into distinct diagnostic states. Guided by this spatial partitioning, ECFR pioneers a quadrant-wise flow rectification strategy: it consolidates potentially correct samples via minimization while actively sequestering high-risk erroneous inputs via maximization, strictly halting the propagation of erroneous gradients. Furthermore, a dynamic memory bank continuously calibrates these quadrant thresholds against distribution drift. Comprehensive evaluations on two state-of-the-art (SOTA) medical FMs across three diverse out-of-distribution (OOD) datasets demonstrate that ECFR successfully prevents model collapse, achieving superior accuracy and confidence calibration for safe clinical deployment. The code is available at https://github.com/EndoluminalSurgicalVision-IMR/ECFR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0402_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/EndoluminalSurgicalVision-IMR/ECFR
Link to the Dataset(s)
Kermany OCT/CXR dataset (Mendeley v2; CXR subset used): https://data.mendeley.com/datasets/rscbjbr9sj/2
Daffodil / Skin Disease Classification Dataset: https://data.mendeley.com/datasets/3hckgznc67/1
HAM10000: https://doi.org/10.1038/sdata.2018.161
BibTex
@InProceedings{LiSha_ECFR_MICCAI2026,
author = { Li, Shangkun AND Gu, Yun AND Zhang, Hanxiao},
title = { { ECFR: Entropy-Consistency Flow Rectification in Test-Time Adaptation for Medical Foundation Models } },
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
The paper proposes a test-time adaptation method that improves medical FMs under distribution shifts. The method utilizes entropy and prediction consistency and applies bidirectional optimization to consolidate reliable predictions or suppress high-risk ones.
- 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 work is well-motivated to address distribution shifts in medical AI deployment. 2.The entropy–consistency quadrants provide a clear and interpretable diagnostic view. 3.The results show consistent improvements across multiple datasets using multiple metrics.
- 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.Lack of clarity in notation and definitions in metric definition. This makes the method harder to follow.
- Streaming with batch size 1 makes the FIFO memory bank (N=128) the only source of distributional context. There are no sensitivity analyses for memory bank size N. 3.ECFR-0.5 uses q=0.5, is performance stable across other values?
- 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?
The paper is well-motivated. The results show the effectiveness of the proposed method. But the concerns needed to be addressed before acceptance.
- 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
The paper proposes a test-time adaptation method based on entropy-consistency performance within a Mean Teacher architecture. The entropy-consistency introduces 4 quadrants that define regions of well-adapted, high-risk, and transition quadrants. The loss, which combines a Jensen-Shannon divergence and entropy.
- 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 addresses a critical problem for model adaptation in a clinical setting with varying distributions but lacks some clarification on the methodology (see below).
- 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.
Several unclear methodology approaches require clarification - see additional questions below that require clarification from the authors to draw a conclusion on the contribution of the paper.
The paper has a section for “theoretical motivation,” while the section is just describing an intuition for the approach they are taking, as opposed to the theoretical backing of the method. Consider renaming this section.
- Please rate the clarity and organization of this paper
Poor
- 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
Major points:
- Why is the region defined as high risk is the one with high entropy and low consistency? A low consistency should be associated with high entropy, what would be a high risk situation would be a low consistency and low entropy (top left quadrant) since that means there is no indicator of the poor consistency between student/teacher (similar to an error in classification with high confidence). If I am misunderstanding this, please clarify.
- Why in the comparative methods, the authors should have included the lower and upper bound of performance on the datasets: lower bound is where no adaptation methods are introduced, just to see what the baseline performance of the model, and the upper bound is where the model is fully trained on the target data. Especially since the accuracy for Ark -> Mendeley-V2 is so low, it would have been important to see the upper / lower bounds for this dataset.
- Table 1 - evaluation lacks repeatability errors (running the same experiment several times with different seeds) - without error bounds, it is difficult to discern whether the gains in performance are significant.
Other comments:
- The first sentence in the introduction cites the paper “The clinician and dataset shift in artificial intelligence. New England Journal of Medicine 385(3), 283–286 (2021)” regarding the problem of distributional shifts in foundation models. The cited paper addresses medical data shift problem in general not specific to foundation models, and the paper pre-dates the wide adoption of the term and settings of foundation models. I suggest the authors supplement or replace this citation with a more recent study that specifically demonstrates the degradation of Foundation Models under multi-center medical distribution shifts to properly substantiate this claim
- Page 4 - Metric Definition - T_w / T_s are undefined.
- The sentence “to mitigate catastrophic forgetting we integrate stochastic restoration” several times in both section 2.3 and 3.1.Questions:
- Experimental Setup: Implementation: The authors state that ECFR-Acc is where q is set to the estimated target accuracy, how is the target accuracy estimated?
- In equation 2, why is lambda set to -1 or 1? What is the explanation for setting a loss term to -1?
- In Equation 3, how are the threshold established?
- In section 2.3, the authors mention the use of the method stochastic restoration to mitigate catastrophic forgetting, how would this impact the learning and adaptation of the model? How would the hyperparameter p (rate of restoration) impact the learning or forgetting?
- 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?
Several points / questions that require clarification in order to make a better informed decision.
- 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 authors adequately address most of my concerns. Given this additional information, I believe the paper meets the MICCAI standards, and I thank the authors for their clear answers.
Review #3
- Please describe the contribution of the paper
The authors address a limitation in test-time adaptation (TTA), arguing that existing methods treat entropy and consistency in a nascent and fragmented manner. To overcome this, they propose a new TTA approach that explicitly models the interaction between prediction entropy and consistency
- 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.
Prior TTA methods either relied on either entropy minimization or consistency regularization separately. The present study introduces a quadrant formulation to simultaneously handle both. The method performs quadrant-wise, bidirectional optimization, which allows it to selectively reinforce reliable predictions while actively suppressing error-prone ones. The inclusion of a dynamic memory bank further stabilizes this process. This is a novel formulation and the results are promising.
- 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 theortical motivation didn’t appear rigorously formulated. It appears largely heuristic to me.
- 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.
(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 because the paper presents a novel and well-motivated framework that unifies entropy and consistency in TTA
- Reviewer confidence
Somewhat confident (2)
- [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.
I stay with my original recommendation.
Author Feedback
We thank all reviewers. For bounds/sensitivity/repeatability, we had already run these checks. They did not affect the main conclusion, but were omitted due to the page limit. Since rebuttal rules bar new experimental results, we report no new numbers here. If accepted, we will add bounds, q/N sensitivity, and repeatability metrics in the revision, while clarifying design choices below.
Q: High-risk quadrant (R2) A: R2 suggested that low-entropy & low-consistency samples may look risky because they resemble confident but unstable predictions. It is intuitive, but ECFR defines high-risk differently: the quadrant with the largest proportion/number of incorrect predictions. This is supported by the zero-shot preliminary study, not intuition alone. In Fig.1(b), samples were split by median entropy and median consistency for comparable quadrant statistics. LCHE had the lowest accuracy and the densest wrong predictions, while LCLE did not. Thus ECFR names LCHE the high-risk zone. We will clarify this definition and the preliminary study.
Q: Metric definitions, λ, and thresholds (R1&R2) A: We agree that metric definitions were too compressed due to space. T_w and T_s denote weak and strong stochastic transformations of the same test image. We will define them clearly. In Eq.(2), λ=±1 is not a loss-weight hyperparameter. It only controls direction: +1 minimizes the entropy/JS term, while -1 makes minimizing the signed objective equivalent to maximizing that term. This gives the two opposite flows in Fig.2.For Eq.(3), as stated in Sec.2.3(pp.5-6), thresholds are online FIFO-bank quantiles: τ_H=Q(B_H,q) is the q-quantile of recent entropy values, and τ_D=Q(B_D,q) is the q-quantile of recent JS divergence values. In Sec.3.1, q is 0.5 for ECFR-0.5 or an accuracy prior for ECFR-Acc.
Q: Selection and stability of q (R1&R2) A: ECFR-Acc is designed for a prior-informed setting with an overall target-domain reliability estimate from deployment logs, validation records, or a small audit cohort. Since public benchmarks lack such logs, our experiments used zero-shot accuracy as its substitute to set q. This does not use sample labels during online adaptation. It only maps an overall accuracy to the routing quantile. ECFR can do this because its thresholds are adjustable: ECFR-0.5 is the no-prior version and already outperforms other baselines in Table 1, while ECFR-Acc shows how ECFR can use an available reliability prior that most TTA baselines cannot exploit. Our q-sensitivity check supports stable performance near the accuracy range. Prior-free adaptive q selection is future work.
Q: Bounds, N sensitivity, and repeatability (R1&R2) A: For bounds, Table 1 reports zero-shot, which is the no-adaptation lower bound. We also checked the label-using upper-bound setting where the model is fully trained on target data. It provides ceiling context but is not a fair comparator to unsupervised online TTA. For the FIFO memory bank, N-sensitivity shows low sensitivity across a broad range, so N=128 is not fragile. For repeatability, different-seed streams show variation much smaller than the Table 1 gaps. These checks support, rather than alter, the submitted conclusion. If accepted, we will add the corresponding metrics to the revision.
Q: Rationale, p restoration, and citation issues (R2&R3) A: We agree that “theoretical motivation” is too strong and mainly empirical/heuristic. We will rename it “heuristic motivation” and soften the claims. For stochastic restoration, the exact sentence appears once, but the idea is restated once in Sec.3.1.We will remove this redundancy. p controls reset frequency toward source weights: larger p limits drift/forgetting but weakens target adaptation, while smaller p adapts more freely but raises forgetting risk. For citation, we will replace the mismatched reference with Ahmad et al., Quant. Imaging Med. Surg. 16(4):321 (2026), a review on real-world shift/robustness in CXR foundation models.
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.
This paper presents a Test-Time Adaptation (TTA) approach for medical foundation models. It employs the Mean Teacher paradigm with Entropy-Consistency Quadrants to segregate high-risk inputs and adapt. The paper has merit; however, there are multiple concerns raised by the reviewers that must be addressed. Particularly, ambiguity in the interpretation of low-risk and high-risk quadrants, lack of clarity, and sensitivity analysis. Authors are encouraged to address all the reviewers’ concerns.
- 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 paper address an important problem of test-time adaptation of foundation models in medical domain. The approach is simple, well-formulated, and effective. Rebuttal has addressed reviewers’ concerns and all reviewers are positive post-rebuttal.
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.
Following the rebuttal phase, the authors successfully addressed the reviewers’ core concerns. Given the consistent empirical improvements and the thoroughness of the post-rebuttal clarifications, the paper is recommended for 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.
The authors have addressed the reviewers’ concern sufficiently well, and all agree that the paper should be accepted.
