List of Papers Browse by Subject Areas Author List
Abstract
Dense prediction tasks in surgical computer vision, such as segmentation and surgical zone prediction, can provide valuable guidance for laparoscopic and robotic surgery. However, these models often suffer from distribution shifts, as training datasets rarely cover the variability encountered during deployment, leading to poor generalization. We propose Dense TRF, a self-supervised representation adaptation framework based on texture-centric attention. Our method leverages slot attention to learn texture-aware representations that capture invariant visual structures. By adapting these representations to the target distribution without supervision, DenseTRF significantly improves robustness to domain shifts. The framework is implemented through conditioning dense prediction on slot attention and model merging strategies. Experiments across multiple surgical procedures demonstrate improved cross-distribution generalization in comparison to state of art segmentation models and test distribution adaptation method for dense prediction tasks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6100_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/PCASOlab/Dense-TRF
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiaGui_DenseTRF_MICCAI2026,
author = { Liao, Guiqiu AND Jogan, Matjaž AND Hashimoto, Daniel A.},
title = { { DenseTRF: Texture-Aware Unsupervised Representation Adaptation for Surgical Scene Dense Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16892},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work presents a pixel-wise prediction framework for surgical image segmentation based on transformer-derived visual features combined with slot-based object-centric representations to capture invariant visual structures. Additionally, it incorporates a domain adaptation mechanism that combines representations learned from multi-domain data with target-specific adaptation, enabling improved robustness to domain variability across different surgical settings.
- 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.
This work addresses the important challenge of domain-specific adaptation in surgical imaging, proposing a framework that adapts representations to center- or domain-specific variability.
Additionally, the experimental evaluation is exhaustive. The method is compared against a wide range of baselines, including both conventional segmentation architectures and foundation models, and is evaluated across multiple datasets, varying data regimes, and a abblation study.
- 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.
Although the proposed method demonstrates improved performance over the evaluated baselines, the analysis is limited to frame-wise metrics and does not consider temporal consistency. In surgical video applications, temporal stability is an important factor, as predictions should be coherent across consecutive frames. The paper would benefit from an evaluation of temporal consistency and potentially from incorporating mechanisms to enforce temporal coherence. But it could be for a next version of the method.
Additionally, the performance (IoU, DICE) remains relatively modest. The paper lacks a discussion explaining the underlying reasons for these limitations and does not provide insights into potential improvements for both segmentation accuracy and temporal stability in future work.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work addresses an important challenge in surgical image analysis, namely domain-specific adaptation, and proposes a well-motivated framework based on slot representations. The experimental evaluation is comprehensive, including comparisons with both conventional methods and foundation models across multiple datasets and data regimes.
However, the paper would benefit from a clearer discussion of the moderate performance (e.g., IoU and DICE) and from outlining future directions, particularly regarding improvements in segmentation accuracy and temporal consistency.
- 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
The paper describes a general framework based on Slot Attention to adapt the dense features from a pre-trained encoder such that downstream dense predictions improve robustness to domain shifts.
- 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 describes an interesting adaptation framework, that clearly shows significant improvement on robustness to domain-shifts. Conceptually the method is sound and make sense.
The results certainly are convincing. At low data regime, this adaptation framework seems to significantly bump performance on target domains.
- 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 framework remains complex and the figure and text lack some clarity. Some learning curriculum details remain unclear.
-
During the continual learning phase, it is quite unclear from the text and figure (Sec. 2.2) which modules are frozen (encoder? dense head?), how the branches are initialised (from the base branch? both branches training at same speed?)
-
The other big question is: after each adaptation round, parameters are averaged between branch. The resulting parameters are applied to what? the target branch? both branches? I would assume only the target branch. But then why not applying a exponential moving average? these abrupt parameter changes will throw the stochastic gradient descent all over the place I would assume
-
The term texture centric is only justified by the patch permutation during continual training phase.
-
- 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
improvements on figures would benefit the camera-ready: make it clear what is learnt at what stage etc
Another unclear point: at comparison with foundation model adaptations, how are these trained? on base distribution only? at low data regime? In other words: you are comparing (1) a foundation model that has never seen the target domain images to (2) your adapted framework that was trained on (un-labelled) target domain?
- 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 factors leading to that score are:
- the method and goals are clear
- some learning curriculum details are unclear
- the framework remains complex
- 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
This paper proposes DenseTRF, a framework for unsupervised test-distribution adaptation in surgical dense prediction tasks. The core idea is to use slot attention to learn texture-aware, object-centric representations via reconstruction, then adapt these representations to target deployment distributions using only unlabeled data. Experiments on three surgical datasets under extreme low-data regimes show improvements over segmentation architectures and foundation model 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.Well-identified problem. The paper targets a good point in surgical AI: models trained on limited annotated data from one set of procedures degrade under distribution shift at deployment. The insight that texture-based representations may be more stable than geometry-based ones in surgery is intuitive and grounded in domain knowledge. 2.Principled integration of object-centric learning with adaptation. The combination of slot attention with a periodic model-merging strategy is a coherent design. The merging mechanism addresses the tension between target specialization and catastrophic forgetting, and the paper provides evidence (the “w SA w/o ada” ablation) that naive slot attention without adaptation can actually hurt performance, validating the need for the proposed strategy. 3.Thorough experimental design. The evaluation includes five independent runs with different seeds, statistical significance testing, three complementary metrics, performance curves across varying annotation ratios, and comparisons against both task-specific architectures and foundation model baselines. The ablation study is well-structured with clearly defined variants that isolate each component’s contribution. 4.Clinically relevant evaluation regime. Testing under 1–5% annotation availability is a realistic scenario for surgical applications, where dense annotation is prohibitively ex pensive. The demonstration that DenseTRF maintains stable performance even at 1% is practically significant.
- 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 “texture-aware” claim lacks direct evidence. This is the paper’s central conceptual claim, yet it is not directly validated. The paper asserts that patch-order permutation (borrowed from SPOT [10]) encourages texture-centric slot grouping, but provides no visualization or analysis showing that the learned slots actually capture texture properties rather than other visual cues. Without more visual and quantitative evidence, “texture-aware” remains an aspirational label rather than a demonstrated property. 2.Limited novelty of individual components. Each component of DenseTRF is drawn from existing work: slot attention from Locatello et al. [13], the DINOSAUR reconstruction framework from Seitzer et al. [20], patch-order permutation from SPOT [10], and model merging from continual learning [27]. The paper’s contribution lies in their combination and application to surgical adaptation, which is reasonable, but the paper could more explicitly discuss what is genuinely new versus assembled. 3.Incomplete ablation of the merging strategy. The periodic merging is presented with a fixed 50/50 weight (Eq. 4), but several design choices are left unexplored: How many merging rounds are performed? How sensitive is performance to the merging frequency? What happens with asymmetric weights ? The current ablation tests whether adaptation helps at all (“w SA w/o ada”) but does not dissect the merging mechanism itself. 4.Efficiency discussion. Table 1 shows DenseTRF has 388 MB parameters — larger than most baselines except Mask2Former. For a method motivated by deployment in real settings, there is no discussion of inference speed, latency, or computational cost. A comparison of FLOPs or inference FPS would contextualize the deployment claim. 5.Missing qualitative results. For a dense-prediction paper, there are no segmentation visualizations showing predicted masks against ground truth. Such visualizations would help readers understand where DenseTRF improves. The slot attention masks are shown in Fig. 2(b), but not analyzed in detail.
- 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?
The paper presents a well-motivated approach to unsupervised adaptation for surgical dense prediction, with a rigorous experiments featuring multiple runs, statistical testing, and consistent improvements across three surgical domains. However, several factors above prevent a stronger recommendation. I lean toward acceptance, but would encourage the authors to strengthen the work.
- 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
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper has received 3 positive reviews with all reviewers acknowledging its methodological novelty, well-thought method design, problem importance and extensive validation. However all reviewers agree on the need for more details on the method development (all) and make key remarks on the need for additional results discussions to highlight the model’s contribution (ablations - Rev3), explain limitations and modest gains ( Rev #1), perform a computational complexity analysis (Rev #3). The authors should take care of these concerns.
My recommendation is for the paper to be accepted