List of Papers Browse by Subject Areas Author List
Abstract
Optical coherence tomography (OCT) is essential in ophthalmology, but inconsistent image quality especially in low-cost devices hinders automated analysis. To address this, we introduce a flow-matching-based test-time adaptation method that generates high-quality surrogate images from noisy inputs. Typically, domain gaps between test and training data cause pixel distribution mismatches during the denoising process. We overcome this by matching the test image’s histogram to synthetic reference trajectories, successfully aligning the input with expected distributions. Additionally, we remove the network’s time conditioning to account for slight deviations in real-world noise distributions. Our approach achieves state-of-the-art performance in segmenting critical biomarkers for two stages of Age-related Macular Degeneration (AMD). Code is available: https://github.com/Veit21/tta-flow.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1314_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Veit21/tta-flow
Link to the Dataset(s)
RETOUCH dataset: https://retouch.grand-challenge.org/
BibTex
@InProceedings{HucVei_TestTime_MICCAI2026,
author = { Hucke, Veit AND Pinetz, Thomas AND Reiter, Gregor AND Schmidt-Erfurth, Ursula AND Bogunović, Hrvoje},
title = { { Test-Time Adaptation in Optical Coherence Tomography Using Trajectory-Aligned Time-Independent Flow } },
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 proposes a test-time adaptation method based on flow matching to address the issue of inconsistent image quality in low-cost optical coherence tomography (OCT) devices, which hinders automated analysis. It introduces Trajectory Alignment, which aligns the distribution of the input test image with synthetic reference trajectories via histogram matching, thereby bridging the gap between theoretical and actual noise distributions. Furthermore, it removes time conditioning in the flow network (Time-Independent Flow), allowing the network to implicitly manage slight deviations in unknown real-world noise.
- 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 introduces Flow Matching, a cutting-edge generative model, to the test-time adaptation (TTA) task in medical imaging. It provides an inspiring design choice by removing time conditioning to address the discrepancy between real physical noise and theoretical Gaussian noise. The required reference histograms can be pre-extracted and averaged from synthetic trajectories, incurring virtually no additional computational cost during inference.
- 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 authors explicitly state in the manuscript, “Notably, our experiments demonstrate that this approach performs robustly even without the inclusion of explicit data-fidelity terms”. In medical image reconstruction, this complete abandonment of fidelity constraints on the input anatomy is dangerous. During the subsequent dozens of ODE solving steps, the generative model can easily mistake early, subtle pathological structures for noise and erase them, or “hallucinate” healthy retinal tissues that fit the high-quality distribution but do not actually exist. (2) The core histogram matching and ODE reconstruction heavily rely on a globally fixed target time point hyperparameter, s_target. Although the authors conducted an ablation study in Figure 3, claiming relatively steady performance across different reference time points, this does not mask this flaw. In real clinical scenarios, the base physical noise levels of OCT images acquired from different patients fluctuate drastically. Using a globally uniform constant to align the distribution of all images contradicts the image-specific nature of TTA and fails to dynamically adapt to the true physical state of each scan. (3) The proposed method achieves results comparable to fully supervised methods. However, for the geographic atrophy (GA) segmentation task (Topcon to Spectralis), the supervised upper bound is 77.3 , while this paper’s best performance is only 57.1.A gap of 20 percentage points indicates that the method fails to genuinely resolve the severe domain mismatch. Furthermore, the non-generative TTA baseline methods compared are rather outdated; for instance, TENT is from 2021, and CoTTA is from 2022.The lack of comparison with the latest TTA technologies from the past two years weakens the persuasiveness of their state-of-the-art claims.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper explores the use of flow matching models for TTA of OCT images, which is a direction with potential. However, as a medical image analysis paper, it presents risks regarding methodological rigor and clinical safety. While the authors explicitly claim that the method performs robustly even without explicit data-fidelity constraints. This strategy, which relies on prolonged generative evolution, poses a risk of erasing early, subtle lesions in low-quality images or “hallucinating” anatomical structures. Secondly, although the ablation study in Figure 3 attempts to demonstrate low sensitivity to the fixed reference time point s_target, this remains a rigid, “one-size-fits-all” mechanism that fundamentally fails to dynamically adapt to the drastically fluctuating physical noise levels of individual images in real clinical settings. Finally, the experimental evidence is insufficiently convincing. On the Topcon to Spectralis task (GA segmentation), the best result still falls a massive 20 percentage points short of the supervised upper bound. Furthermore, the baselines used for comparison (such as TENT from 2021 and CoTTA from 2022 ) are outdated, and the lack of comparison with the latest TTA technologies significantly weakens the authors’ state-of-the-art claims.
- 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.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes TTA-Flow, a flow-matching-based test-time adaptation method for OCT that adds two simple but effective ideas: histogram-based trajectory alignment and removal of explicit time conditioning. The method is evaluated on cross-device OCT adaptation for fluid and GA segmentation, and it consistently outperforms classical TTA baselines, diffusion-based baselines, and plain flow matching, with especially strong gains on Cirrus→Spectralis and clear FID improvement on the paired in-house dataset.
- 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 method is clearly motivated, technically coherent, and the design is easy to follow. The experiments are solid: comparisons include TENT, CoTTA, CPDM, DDA, and standard flow matching; results are reported on two tasks and multiple devices; and the ablation supports both proposed components.
- 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 main limitation is novelty: the contribution is more of a strong adaptation of existing generative TTA ideas than a fundamentally new framework. As mentioned in the paper, authors utilize methods from Lipman et al. [13, 14] ] and Sun et al. [21].
-
The method uses histogram matching to place a real test scan onto a learned flow trajectory, but histogram matching aligns only the global intensity distribution, not the local OCT texture, speckle statistics, or lesion structure that actually define the domain gap. In addition, the start point s_target is a fixed global hyperparameter chosen by grid search, which further weakens the claim that the alignment is truly adaptive to each test scan.
-
The evaluation is also still limited in scale, with 70 RETOUCH volumes and 40 paired GA patients, and only annotated B-scans used for quantitative analysis. I suggest authors should investigate bigger dataset such as AI-READI to make the evaluation more robust and reliable.
-
In addition, the approach is computationally heavy for a TTA method, requiring a large 103M flow model, about 80 hours of training, and about 2.2 seconds per B-scan at inference. So although the method works, it is not lightweight TTA in the usual sense. It is closer to test-time generative restoration with a heavy pre-trained model.
-
The paper writing needs to be updated. Paragraph lengths are not balanced.
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Despite that the novelty remains the main concern, adapting known techniques for practical applications in medical field seems constructive approach for me.
I lean towards Weak Accept but a rejection is also fine if authors did not clarify the weaknesses.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The rebuttal partially addresses my concerns, mainly on hyperparameter robustness.
However, key technical issues remain: the trajectory alignment is still heuristic because histogram matching only aligns global intensity statistics; the method still lacks an explicit data-fidelity constraint for preserving subtle pathology (explanation with Dice is not totally correct); and the cross-device generalizability is harder to verify since part of the evaluation relies on a small in-house paired GA dataset without clear release information.
Review #3
- Please describe the contribution of the paper
The paper proposes a method that tries to turn noisy optical coherence tomography scans from lower-quality scanners into cleaner scans that look more like high-quality Spectralis scans before segmentation. It does this by matching the intensity histogram of the test scan to a reference path from a flow model, and by removing explicit time input from the model so it can better handle real scanner noise.
- 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 tackles an important real problem. Segmentation models often drop in performance when scan quality changes across scanners, especially for lower-cost devices. The most interesting part is the use of histogram alignment to match the test scan to the model’s own reference path before reconstruction. That is easy to understand and easy to implement.
- 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 novelty is limited. Flow matching itself is standard and comes from references [13] and [14], and the choice to remove time input is already motivated by reference [21]. The added histogram step is practical, but it does not feel like a major new method on its own. Some important study details are unclear. The paper says the target reference point was picked by trying different settings, but it does not clearly explain how this was done for the target scanners. It is also not clear how much of the improvement comes from histogram matching alone versus the full method. Since the method changes medical images without a direct constraint that keeps the output close to the input, the paper should show more clearly that important structure is preserved.
Also, some setup is not clear, which is understandable due to the page limit, but it should be explained more in the camera-ready version if accepted.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper actually resonates with me. I like the paper as it is addressing an important problem in clinics, but the novelty is very limited in machine learning. To sum up, I believe this paper, after addressing some of the mentioned issues, can be accepted.
- 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 for their constructive feedback and are encouraged that all reviewers found our general framework to be useful and practical. We address the primary concerns below.
Exclusion of data-fidelity term (R1, R3): While we agree that generative models can theoretically hallucinate structures, we did not observe this in practice. We found the bias of the starting point to be strong enough to be sufficient and data-fidelity terms very sensitive to the hyperparameters. Notably, our method outperforms multiple baselines that explicitly use data-fidelity terms (e.g., DDA and CPDM), including CPDM which utilizes a fidelity term strictly designed for OCT. Furthermore, our reported Dice scores inherently penalize such hallucinations.
Suitable Baselines (R1, R2): While the classical Test-Time Adaptation (TTA) methods we compare against are older, our primary focus is on recent generative-based TTA baselines such as DDA and CPDM. Because these modern methods explicitly leverage powerful generative priors, they are theoretically expected to inherently outperform classical TTA approaches. Consequently, they serve as the most rigorous and relevant baselines to benchmark our approach against, whereas the older methods were included strictly for completeness. While our model uses 103M parameters, the diffusion baselines use the same model and same number of parameters. The runtime of the diffusion based baselines is longer as there are more neural function evaluations (100 vs 1000). Speeding up our approach is the focus of future work.
Reliance on hyperparameters (R1, R2, R3): Although our system requires selecting a single hyperparameter for each target scanner, performance is highly robust to this choice (see Fig. 3). In fact, oracle selection (choosing the absolute best hyperparameter per individual patient) improves overall results by only 1 percentage point. This minimal difference represents the theoretical upper bound of what an automated tuning system could achieve, demonstrating that the manual choice is not overly sensitive to begin with. We believe that this is because the unconditional version of our approach is intrinsically robust to a variety of noise levels as can be seen in Fig. 3.
Supervised performance (R1): We acknowledge the performance gap between the Topcon and Spectralis devices, however, the signal-to-noise ratio (SNR) in the Topcon images is inherently much lower and the annotation was performed on each device separately. Consequently, it is unclear if the theoretical upper bound of the Spectralis device is even achievable, as certain fine structures may simply not be captured by the Topcon sensor and the annotation might be less certain. Despite this physical limitation, our approach provides a substantial performance boost compared to standard predictions directly on Topcon images.
Limited novelty (R2, R3): We strongly believe that the simplicity of our approach is its greatest strength. It is highly accessible, straightforward to implement, and significantly enhances the output of generative models currently prevalent in medical imaging. Furthermore, while this approach is particularly well-suited for OCT, techniques like histogram matching have a long, proven track record of broad utility in other modalities, such as MRI.
Confusing writing (R2, R3): We will utilize the additional space provided for the camera-ready version to expand and clarify the methodology section to ensure reproducibility and ease of reading. Furthermore, we have prepared a code repository to be released upon acceptance to further avoid any misconceptions.
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.
All three reviewers agree on acceptance, although only slightly above the bar. My own assessment goes in line with this method potentially being a meaningful contribution to the literature. The trajectory alignment is interesting to the community and the application is challenging. However, despite the overall assessment above the acceptance bar, in total there are a large number of arguments by reviewers that need clarification, especially regarding claimed novelty, lack of explicit data fidelity together with danger of hallucinations, as well as shortcomings in the evaluation regarding competing, more recent TTA methods. I invite authors to address these and other concerns in an organized manner in a rebuttal (focusing on the major issues sorted by perceived importance from the author’s point of view), to clarify the multiple mentioned issues in the reviews.
- 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.
Taking into account the rebuttal in addition to the reviews, I tend to agree that this work demonstrate a good contribution which might spark some discussion at the conference. I think in case of acceptance the authors should clearly state limitations, e.g. potential hallucinations (although not appearing in their tests), and carefully clarify all points addressed in the rebuttal, e.g. the details on hyperparameters. Concerns regarding generalizability of results are noted, but I think these could be addressed in a follow up work.
In case of acceptance, authors should tidy up the references, as these are not formatted in a proof-ready version yet (pages missing, inconsistent use of hyperlinks, arxiv versions of published works). Moreover, I suggest a slightly more thorough inclusion of recent works using flow matching in medical image synthetization or reconstruction.
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.
Reviewers summarized strengths of your work, novel aspects and noted that your solution to the specific application area may be of potential interest. The rebuttal helped to clarify open issues and misunderstandings. Would the paper be accepted, authors are strongly encouraged to improve the final write-up by including recommendations by reviewers, in particular improving clarity of notations and descriptions.
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.
All three reviewers said the paper above the bar in the initial review and agreed the paper tackles an important OCT cross-device problem with an easy-to-implement histogram-alignment idea and a clean time-conditioning removal (R1, R3).
I believe the rebuttal was most convincing on hyperparameter robustness, showing that per-patient oracle tuning buys only about a point over the single global s_target, and on why the generative-TTA baselines (DDA, CPDM) are the relevant comparisons. The residual worry (R2), is that histogram matching aligns only global intensity and the method lacks an explicit data-fidelity term, so subtle pathology preservation is argued rather than proven; I agree this is a real limitation, but note the method still beats fidelity-using baselines and that R2’s concerns were largely known at the original weak-accept stage. With R1 confirming accept, R3 unchanged and positive, and only R2 moving down on pre-existing grounds, I recommend acceptance, with the authors encouraged to revise the data-fidelity claims and expand the methodology in the camera-ready.
