List of Papers Browse by Subject Areas Author List
Abstract
Medical imaging models often degrade when deployed at new clinical sites due to differences in imaging equipment, protocols, and patient populations. Test-time adaptation (TTA) addresses this by updating a pretrained model using only unlabeled target data, without access to source data. However, existing TTA methods were designed for single-label classification on natural image benchmarks, minimizing entropy uniformly across all samples without considering label dependencies. This overlooks a key property of multi-label medical imaging: pathologies do not occur independently but exhibit structured co-occurrence patterns. In this work, we propose Co-occurrence Weighted Adaptation (CoWA), which leverages disease co-occurrence patterns as a reliability signal for adaptation. CoWA estimates label co-occurrence structure from model predictions and downweights samples that deviate from expected patterns, enabling adaptation to rely more on consistent predictions while reducing the impact of noisy ones. We evaluate CoWA on chest X-ray benchmarks under domain shifts and demonstrate consistent improvements over established baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4786_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/woojin716/CoWA
Link to the Dataset(s)
CheXpert: https://stanfordmlgroup.github.io/competitions/chexpert/
MIMIC-CXR: https://physionet.org/content/mimic-cxr/
NIH ChestX-ray14: https://nihcc.app.box.com/v/ChestXray-NIHCC
VinDr-CXR: https://physionet.org/content/vindr-cxr/
BibTex
@InProceedings{JeoWoo_Leveraging_MICCAI2026,
author = { Jeong, Woojin AND Choi, Yujin AND Kim, Dongbin AND Park, Soyeon AND Lee, Jaewook},
title = { { Leveraging Pathology Co-occurrence for Test-Time Adaptation in Chest X-Ray Diagnosis } },
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
This paper studies test-time adaptation for multi-label chest X-ray diagnosis by exploiting pathology co-occurrence patterns. The problem is important, and the proposed CoWA framework is clinically motivated: it uses an estimated co-occurrence matrix from target-domain predictions to weight samples during entropy minimization, rather than treating all labels and all samples independently. The idea is intuitive and differs from standard TTA pipelines. The empirical results are generally positive, but the performance gains are overall modest, and the current validation remains limited to chest X-ray classification.
- 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 addresses an important and practical problem. Test-time adaptation for multi-label medical diagnosis is highly relevant, and incorporating disease dependency structure is a meaningful direction for this setting. 2.The proposed method is simple and well motivated. Using pathology co-occurrence as a reliability signal for sample weighting is intuitive, easy to implement, and compatible with standard entropy-based TTA. 3.The experimental section is reasonably complete. The paper compares against several established TTA baselines and includes additional analyses on co-occurrence matrix evolution, sample-weight reliability, and norm ablation.
- 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 empirical gains are relatively limited. Although CoWA is often competitive and frequently ranks first or second, the absolute improvement over strong baselines is generally modest, which weakens the overall impact of the method. Table 1 shows that many of the gains are small. 2.The validation scope is narrow. All experiments are conducted on chest X-ray classification, so it remains unclear whether the proposed idea would generalize to other medical imaging modalities or other multi-label diagnostic tasks. 3.The method relies on dataset-level co-occurrence structure, which may introduce or amplify bias. The paper itself notes that co-occurrence patterns vary across datasets. In practice, such variation may reflect differences in label definitions, patient populations, or institutional case mix rather than stable disease mechanisms. Since CoWA directly estimates and uses these correlations during adaptation, it may reinforce dataset-specific associations rather than clinically robust relationships. This limitation is important and currently under-discussed.
- 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 should expand the discussion of limitations, especially the risk that co-occurrence-based adaptation may amplify dataset-specific bias caused by annotation differences, population shift, or non-causal correlations. The authors should frame the method more cautiously as exploiting target-domain statistical dependence, rather than implying that it captures a stable clinical structure across domains. The paper would be strengthened by additional analysis showing when co-occurrence-based weighting helps and when it may fail, particularly under stronger shifts in label prevalence, population composition, or labeling criteria.
- 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 #2
- Please describe the contribution of the paper
This paper introduces a novel idea for unsupervised TTA. The authors argue that most TTA approaches aren’t designed for multi-class tasks. In biomedical scenarios, diseases often co-occur. Maybe that could be used to help with the adaptation part. They calculate an average co-occurrence matrix over the entire target dataset using the pretrained source model. Then they calculate the distance of each data point from this average co-occurrence matrix. The lower this distance, the higher the certainty and thus the higher the loss weight. This helps give more stable gradients for 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.Good visualizations and well presented results. 2.Interesting analysis in Table2 of the performance of the different L-norms.
- 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.This work seems like an importance sampling (IS) based approach for TTA applications. However, no theoretical connections or formulations were drawn against that. PRIORITIZED EXPERIENCE REPLAY https://arxiv.org/pdf/1511.05952.There are some other things as well like being careful with the distribution of the IS-weights o.w. There are some stability concerns etc, which our paper fixes in a very adhoc way via max(w_min,w_i). Maybe there are some better annealing strategies for this?
2.That leads me to another question: IS helps accelerate the convergence speed, so maybe if the authors showed us the validation curves of the classification performance on the val/test datasets over time, we might see something along the lines that other methods might be coming close to the performance of COWA but taking slightly longer. It will also be interesting to validate that an increase in the convergence speed is indeed in fact happening.
3.No ablation studies: What if we replace max(w_min,w_i) with just w_i? What happens then? a. Some numerical metrics or visualizations of the stability of the gradients (measured perhaps by the var(grad) over time ) vs simple L = H, L = wi*H, L = max(w_min,wi) * H would be helpful. b. The H(.) is the shannon entropy right?
4.Otherwise, marginal improvements over AdaBN and also sometimes AdaBN beats this work. So the results aren’t even showing consistent improvements either.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The idea of the paper is indeed very novel. However, it needs a lot more theoretical and experimental validation and exploration to be good enough for MICCAI standards.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
After reading the reviews and the rebuttal, I feel that this work is promising but not yet ready for publication at this venue. The authors need to incorporate the suggested improvements into the submission for a reevaluation.
Review #3
- Please describe the contribution of the paper
This paper introduces Co-occurrence Weighted Adaptation (CoWA), a novel source-free test-time adaptation method that dynamically estimates a pathology co-occurrence matrix from target domain predictions to serve as a per-sample reliability signal for multi-label medical image classification. Unlike standard approaches that indiscriminately minimize entropy for all predictions regardless of label dependencies, CoWA downweights samples with clinically inconsistent disease combinations and assigns higher weights to structurally coherent ones. By incorporating this underlying clinical dependency structure into the adaptation process, the method improves the diagnostic robustness of medical imaging models against cross-institutional domain shifts without requiring any source data or target annotations.
- 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 proposed question is a important yet hardly studied one, which being solved will greatly imporve the generalization ability of existing multi-label prediction models. 2.The method prevents severe performance degradation on vulnerable or low-prevalence disease classes by explicitly suppressing unreliable gradient updates that standard methods would otherwise reinforce. The overall demonstration of methods is clear and easy to follow. 3.Comprehensive evaluations across multiple chest X-ray domain shift scenarios show that CoWA consistently achieves superior or highly competitive diagnostic accuracy compared to established TTA baselines.
- 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.It seems when adapt to the target domain the threshold hyperparameter is fixed. Although the author did ablation studies to show the effectiveness of the chosen threshold, it could be better if the author could make the threshold also adapt to the domain field. 2.When adapt to the domain field, CoWA heavily relies on the model’s initial predictions on the target domain to dynamically estimate the pathology co-occurrence matrix. If an extreme domain shift causes the pre-trained model to produce highly inaccurate initial predictions, the resulting empirical matrix will be flawed. It would be better if author can discuss the possibility of this kind of situation.
- 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?
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 author has identified an important problem and proposed an effective method to solve it, but there are still some concerns, as noted in the Weakness section.
- 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.
Thanks authors for their detailed responses which clearly address my concern and I believe it should be accepted given its methodology novelty and good experiment results.
Author Feedback
We thank the reviewers and AC for the valuable feedback.
[R1w1, R2w4: Performance Gains] Avoiding severe degradation on vulnerable classes is as clinically important as maximizing mean AUROC, where CoWA’s gains concentrate. As Tab.1 shows, on Pt. under NIH→MIMIC, No-Adapt achieves 57.5 yet TENT drops to 49.1 and AdaBN to 48.3, while CoWA preserves 55.7.As R1 noted, CoWA achieves the best or second-best mean AUROC across all shifts and is the only method without a severe drop, whereas baselines exhibit large regressions relative to No-Adapt on multiple shifts (e.g., AdaBN on NIH→CheX, TENT on CheX→NIH, EATA on NIH→VinDr; Tab.1). We attribute this robustness to CoWA’s modeling of dependency-based inconsistencies during adaptation, which AdaBN cannot capture as it only updates BN statistics.
[R1w3, R3w2: Robustness and Bias] CoWA’s reliance on target-derived signals raises two concerns: that poor initial predictions could corrupt M (R3w2), and that dataset-specific co-occurrence may reinforce biases rather than reflect stable clinical structure (R1w3). For the first: (i) threshold binarization (Eq. 1) blocks low-confidence activations from entering M; (ii) M accumulates across batches, dampening early-stage noise (Fig. 3). Empirically, NIH→MIMIC already captures this regime (Tab.1): source predictions on rare classes (e.g., Pt.) are unreliable yet CoWA achieves the highest mean AUROC. We agree with R1, and note this is intrinsic to unsupervised TTA: any target-derived signal can reinforce target-domain biases. We will reframe the contribution as exploiting target-domain dependence rather than stable clinical structure.
[R2w1: Relation to IS] We thank R2 for the IS perspective, which sharpened our framing. CoWA shares IS’s intuition of upweighting reliable samples, but is better framed as confidence-based instance reweighting [1,2]: IS weights are likelihood ratios that make the objective an unbiased estimator under the target distribution, whereas CoWA weights are not likelihood ratios and Eq. 4 is not constructed for unbiased estimation. Their role is instead to suppress unreliable training signals, with weight magnitude reflecting per-sample reliability (Fig. 4). The functional form in Eq. 3 follows a Gibbs (maximum entropy) solution under an expected cost constraint [3], providing principled grounding rather than an arbitrary choice.
[R2w1, R2w3a: Stability] CoWA addresses R2’s stability concern both structurally and empirically. By construction, CoWA weights are bounded in [w_min, 1] (Eq. 3, 4), precluding the heavy-tail and weight-collapse behavior of unbounded IS weights. Empirically, we measured BN-affine gradient L2-norm variance over adaptation on CheX→MIMIC under three losses, where H(·) is the (Shannon) binary entropy per class, summed over pathologies (R2w3b):
- (A) L=H: 2.38e-5
- (B) L=w_i·H: 1.05e-5
- (C) L=max(w_min,w_i)·H: 1.02e-5 (A)’s 2.3× higher variance is consistent with TENT’s instability on rare classes (Tab.1). The near-identical variance of (B) and (C) addresses R2w3a: consistency-based reweighting accounts for most of the stabilization, while w_min serves as an early-stage safeguard before M stabilizes (Fig. 3); Fig. 4 further shows the weight distribution lies well above w_min after adaptation. More elaborate annealing strategies are left for future work.
[R2w2: Convergence] To test R2’s hypothesis that baselines may approach CoWA with more iterations, we measured AUROC validation curves across adaptation steps. The gap widens rather than narrows: CoWA exceeds TENT at every iteration, and after both methods peak at iter 30, TENT’s post-peak AUROC declines 1.42× faster.
[Future Work] (R1w2) As CoWA is modality-agnostic, extending it to other modalities would be a valuable next step. (R3w1) We also agree that a principled adaptive threshold for M estimation is a worthwhile direction.
Ref. [1] Kumar et al. NIPS 2010, Self-Paced [2] Ren et al. ICML 2018, Reweight [3] Jaynes, 1957, Info Theory
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 introduces CoWA, a test-time adaptation framework for multi-label chest X-ray diagnosis that uses pathology co-occurrence patterns to weight samples during entropy minimization. While clinically intuitive, reviewers hold divided opinions due to limited performance gains and theoretical gaps.
-R2’s concern regarding the marginal empirical improvements, particularly instances where baselines like AdaBN outperform CoWA.
-Discuss the risk of the method reinforcing dataset-specific biases rather than capturing stable clinical structures, especially across different populations (R1).
-Connect the formulation to Importance Sampling and provide ablation studies for the weight truncation strategy to prove gradient stability (R2).
-Explain how the framework handles severe initial prediction errors where the estimated co-occurrence matrix might be fundamentally flawed (R3).
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This submission proposes an intuitive test-time adaptation framework that incorporates pathology co-occurrence for multi-label medical classification, initially drawing divided reviews regarding empirical gains and theoretical grounding. In the rebuttal, the authors successfully addressed major concerns by providing concrete evidence of gradient stability, clarifying the instance-reweighting formulation, and pledging to reframe the scope to account for dataset-specific bias. Given that the majority of reviewers (2/3) found the clarifications compelling and favor acceptance, the Area Chair recommends a Weak Accept, contingent upon the authors fully integrating these post-rebuttal revisions into the final manuscript.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
I recommend reject: R2 (Faculty) maintained Reject independent of rebuttal on marginal empirical gains (AdaBN sometimes outperforming CoWA) and unresolved theoretical gaps (importance-sampling connection, weight-truncation stability), and the two soft “4 — would not mind if rejected” votes from R1 (no post-rebuttal update) and R3 do not outweigh a Faculty Reject to a documented performance concern that the rebuttal could not resolve.
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.
I agree with most reviewers that the paper highlights an important yet currently overlooked problem in test-time adaptation: diseases are co-occurring. Most reviewers find the method novel and the experiments convincing. Although the paper lacks some in-depth performance analysis and ablation studies, it would still spark meaningful discussion at MICCAI.
