List of Papers Browse by Subject Areas Author List
Abstract
Mueller matrix imaging provides rich, physically meaningful
contrast for biomedical tissue analysis, but supervised learning is hindered by scarce dense annotations and strong domain shifts across specimens and acquisition settings. We introduce MuellerPT, a physics guided pre-training approach that learns transferable dense representations by predicting Lu-Chipman decomposition maps from per-pixel 4x4 Mueller matrices. To scale pre-training, we collected a new large Multispectral Animal Polarimetric Organ dataset (MAP-Org). The pre-trained encoder is adapted with a segmentation head for grey vs. white matter segmentation in lamb brain. A classification head is used for colorectal cancer vs. non-cancer classification. Both segmentation and classification are evaluated across few-shot learning scenarios. In segmentation, MuellerPT improves label efficiency and cross specimen transfer compared to models without pre-training, achieving an absolute DICE gain of over 20% compared to the baseline trained from scratch when using 5% of the training data. In classification, MuellerPT also enhances label efficiency, improving overall accuracy by 8% compared to the baseline when using 1% of the training data. We demonstrate MuellerPT’s robustness to domain shift with a qualitative evaluation of its predicted Lu-Chipman maps on an ex vivo human oesophagus sample. These results suggest that predicting Lu-Chipman decomposition is an effective and practical pretext task for robust biomedical inference from Mueller polarimetry and can pave the way for future work on label efficient Mueller imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2771_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/adamt222/MuellerPT
Link to the Dataset(s)
https://zenodo.org/records/20274683
BibTex
@InProceedings{TleAda_MuellerPT_MICCAI2026,
author = { Tlemsani, Adam AND Li, Yingdian AND Giot, Maxime AND Slim, Naim AND Peters, Christopher J. AND Ghosh, Abhijeet AND Elson, Daniel S.},
title = { { MuellerPT: Decomposition Driven Pre-training for Dense Learning in Mueller Polarimetry } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
MullerPT leverages self-supervised learning using a new large Multispectral Animal Polarimetric Organ Datasets (MAP-Org) to pretrain an encoder to predict Lu-Chipman decomposition maps for mueller matrices. The encoder is then applied to segmentation and classification task of grey and white matter segmentations on PoLambRimetry, and classification of colorectal cancer on ColoPola, respectively. Performance was evaluated over 30 runs with different seeds and showed that on segmentation using few-shot learning MuellerPT outperformed baseline model (HRNet-Scratch) when only 1-50% of data was provided.
- 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 core idea of using Lu-Chipman decomposition as a physics-grounded pretext task is well-motivated and elegant. Unlike generic self-supervised objectives (contrastive loss, masked autoencoders), this approach grounds representation learning directly in the optical physics of Mueller polarimetry. The dual-stream architecture separating polarimetric content (fθ) from intensity (mψ) seems reasonable and allows for situations where downstream datasets lacking the un-normalized m(0,0) channel. The label-efficiency gains are compelling. The 20% absolute DICE improvement at 5% training data and the 8% accuracy gain at 1% for classification are substantial, and the benefit to the minority class (white matter) is particularly encouraging for clinical applicability, where rare tissue classes are often the most diagnostically important.
- 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 comparison baseline is narrow. The paper only compares MuellerPT against HRNet-Scratch (random initialization). There is no comparison against: (1) ImageNet pre-trained initialization, (2) generic SSL methods (e. g. , SimCLR, MAE, DINO) applied to Mueller data, or (3) using the Lu-Chipman maps directly as input features without pre-training. This makes it difficult to attribute the gains specifically to the physics-grounded pretext task versus simply the benefit of any form of pre-training. A frozen encoder baseline would also help disentangle the benefit of pre-training from fine-tuning dynamics. The MAP-Org dataset (41 images) is quite small for a pre-training corpus, and it is composed exclusively of animal food tissue (lamb, chicken, steak, bacon). The paper does not analyze how dataset size or tissue diversity affect downstream performance. An ablation on pre-training data scale, or a comparison pre-training on a subset of MAP-Org vs. the full set, would substantially strengthen the empirical claims. Lack of ablation experiments make design choices and their relative contribution toward improvement in performance difficult to interpret. For example, two-stream architecture vs single-stream architecture is proposed but the relative contribution of single stream is not ablated. The authors acknowledge in the introduction that Lu-Chipman decompositions are sensitive to measurement noise and depolarization behaviour, yet no sensitivity analysis is conducted to examine how decomposition quality affects pre-training. Inclusion of such results would strengthen the quality of the paper and ensure that the results are generalization and due to contribution of Lu-Chipman decomposition.
- 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
There is significant concern regarding reproducibility. The authors defer implementations details to code repository not yet available at review time, which could have been solved through the use of anonym. 4open. science. Examples of missing parameters include hyperparamters, such as learning rate, batch size, pre-training epochs, and optimizers. Open-source datasets are utilized for evaluation but no information regarding MAP-Org data availability is provided. Statistical reporting over the 30 runs improves confidence in the presented results, however with no code or training details, independent verification of the results is impossible.
- 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?
While the core idea is physically well-motivated and the few-shot results are encouraging, the paper falls short of the reproducibility standards expected for acceptance. The methodology is not self-contained — loss functions, optimizer, learning rate, batch size, number of training epochs, and architectural details of the refinement block and decoders are all absent from the manuscript, with implementation details deferred entirely to a code repository not available at review time. A reader cannot reimplement this work from the paper alone.
- Reviewer confidence
Somewhat confident (2)
- [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 MuellerPT, a physics-guided pre-training strategy for dense learning on Mueller matrix polarimetry data. The central idea is to pre-train encoders by predicting Lu-Chipman decomposition maps—depolarization, retardance, and diattenuation—from multispectral per-pixel Mueller matrices, using these decomposition-derived quantities as a self-/weakly-supervised pretext task. The paper also introduces a new pre-training dataset, MAP-Org, consisting of multispectral animal organ polarimetric images, and evaluates the learned representations on two downstream biomedical tasks: grey/white matter segmentation on PoLambRimetry and colorectal cancer classification on ColoPola, both under few-shot settings. The paper argues that this pretext task improves label efficiency and transferability under domain shift compared with training the same architecture from scratch.
- 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.An interesting and relatively under-researched problem setting.
This paper explores Mueller matrix polarization measurements, an area far less saturated than mainstream medical imaging. The proposed approach—designing a pre-trained target based on polarization physics principles, rather than simply importing a generic SSL target—is well-motivated and potentially valuable to the field. Using Lu-Chipman decomposition as the structured target is particularly appealing because it is physically meaningful and requires minimal manual annotation.
2.A physics-based pre-training task is a significant contribution.
The mapping obtained from the predicted decomposition is a plausible pre-training signal: it is task-relevant, associated with known optical structures, and more consistent with downstream domains than generic contrast or reconstruction targets. This paper strongly demonstrates that this supervised approach can encourage the generation of representations that reflect tissue microstructure rather than merely surface image statistics.
3.This paper contributes a new dataset resource.
The MAP-Org dataset itself appears to be a valuable contribution. Despite its relatively small absolute size, collecting 41 multispectral Mueller matrix images covering various animal tissues for pre-training is no easy feat in this field, and the public availability of this dataset enhances the potential impact of this work.
4.Evaluation on two downstream tasks is a wise design choice.
This paper does not validate on only one downstream benchmark. It tests the pre-trained representations on two tasks: dense segmentation (PoLambRimetry) and image-level classification (ColoPola), making the conclusions of transfer learning more convincing than single-task evaluations.
5.The focus on few-shot training is appropriate.
Performance improvements are primarily concentrated in low-label regions, consistent with the research motivation of this paper. In the segmentation task, significant performance improvements are achieved compared to the initial baseline at 1% and 5% label availability; in the classification task, improvements at 1% and 5% label availability are also practically meaningful. This is precisely the area where domain-specific pre-training methods are most effective.
6.This paper includes cross-sample segmentation evaluation.
For the PoLambRimetry, sample-level splitting using nested cross-validation is appropriate and helps reduce data leakage. This is an advantage, especially in small biomedical datasets where leakage between slices or samples is a common problem.
7.This paper includes domain offset analysis.
Qualitative experiments on isolated human esophageal tissue, particularly with a simplified Mueller matrix configuration, are an interesting attempt to show that the learned representations can generalize beyond the precise pre-training domain, even tolerating hardware simplifications. While this is not a complete quantitative validation, it is a useful and thoughtful addition.
- 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 comparison scope is too narrow.
The primary comparison is with the same HRNet architecture trained entirely from scratch. This is a fair and important baseline, but insufficient to prove the correctness of the proposed pre-training task. The paper demonstrates the value of Lu-Chipman prediction as a pre-training task, but does not compare it with other stronger pre-training schemes.
2.The pre-training dataset is small, and the paper may have exaggerated its scalability.
MAP-Org contains 41 multispectral Mueller matrix images from fresh animal tissues. This is useful in this subdomain, but the paper also describes the method as a way to “scale up pre-training.” Considering the size of the real-world dataset, this statement seems somewhat idealistic.
3.Downstream results are positive, but the advantage gradually disappears as the label ratio increases.
The results show that the method is mainly effective with few samples, and when the label ratio reaches 25%, 50%, and 100%, the difference between MuellerPT and the Scratch baseline in both segmentation and classification becomes very small. This contribution is best viewed as an improvement in low-label efficiency rather than a general performance improvement.
4.Ablation experiments need enhancement.
This method combines several elements: Lu-Chipman prediction, a two-stream architecture with separated normalized Mueller data and m(0,0) channels, hardware-heuristic dropout, and Mueller-specific data augmentation. Since only the overall method is compared to the initial training, it is unclear which components are most important. This is a major technical shortcoming of this paper.
5.Limited comparison with task-specific non-pre-trained alternatives.
This paper proposes a pre-trained task partly for the interpretability of Lu-Chipman parameters, but does not compare it with simpler alternatives. Therefore, it is impossible to determine whether the performance improvement stems from pre-training itself or simply from injecting decomposition-derived prior information in some way.
- 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?
I tend to be positive about this paper because it proposes a domain-specific and physically meaningful pre-training task applied to a niche but important field of biomedical imaging. Using Lu-Chipman decomposition as the pre-training signal is reasonable and makes more sense than simply applying a general representation learning method to the Mueller dataset. The evaluation also covers two different downstream tasks—segmentation and classification—and the improvements achieved under few-shot conditions are meaningful and consistent with the paper’s motivation. Sample-level cross-validation for the segmentation task is another highlight.
My reservations mainly lie in the depth of the empirical validation. The paper only compares to training from scratch, making it difficult to determine whether the proposed pre-training task outperforms other reasonable pre-training or representation learning methods. Furthermore, the method incorporates multiple design choices without ablation, making it difficult to pinpoint exactly what factors are at play. The evidence for domain shift is interesting but only qualitative, and these advantages largely disappear under high-label settings.
Overall, I think this paper has sufficient innovation and encouraging evidence to slightly exceed the standard, especially considering its domain of expertise and the practical difficulty of collecting such data. However, since comparative validation is still limited, it is not yet fully acceptable.
- 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
The paper presetns MuellerPT, a physics-guided pretraining framework for Mueller matrix polarimetry. The core idea is to pre-train dense image encoders by predicting Lu–Chipman decomposition parameters, depolarization, retardance, and diattenuation, from unlabeled Mueller matrices. These parameters are physically meaningful and can be computed without manual labels. This makes them suitable as structured self-supervised or weakly supervised petraining targets.
In addition, it contributes a new multispectral animal-tissue Mueller matrix dataset, MAP-Org, which is used for pretraining. The learned encoders are then transferred to two downstream biomedical tasks: grey/white matter segmentation on PoLambRimetry and cancer/non-cancer classification on ColoPola. The main empirical finding is that MuellerPT improves label efficiency, especially in low-data regimes, with the largest reported gains at 5% training data for segmentation and 1–5% training data for classification. The paper also includes a qualitative domain-shift experiment on ex vivo human oesophagus tissue.
- 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 paper presents a physics-guided pretraining approach that uses Lu–Chipman decomposition maps as pretraining targets. This is a strong idea because these maps encode meaningful polarimetric tissue properties while avoiding the need for manual labels. Since Lu–Chipman decomposition provides a standard physical interpretation of Mueller matrices into diattenuation, retardance, and depolarization-related components, the pretext task is aligned with the imaging physics rather than relying on a generic computer vision task. The paper also addresses the important practical problem of label efficiency, as dense annotations in biomedical polarimetry are scarce, and the reported few shot experiments directly target this limitation. The strongest gains are observed in low-label settings, especially for white matter segmentation with 5% training data and colorectal classification accuracy with 1% and 5% training data. The evaluation uses relevant public biomedical polarimetry datasets, including PoLambRimetry for segmentation and ColoPola for colorectal cancer classification. The two-stream design is also reasonable, as separating the intensity like m(0,0) channel from the normalized polarimetric stream is physically motivated and consistent with earlier Mueller learning work such as MuellerNet. The reduced Mueller experiment is interesting and clinically relevant because reduced polarimetric systems can be faster and simpler.
- 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 paper has several limitations that should be addressed. The main comparison is too limited because MuellerPT is evaluated mainly against the same HRNet architecture trained from scratch. While this isolates the benefit of pre-training over random initialization, it does not show whether Lu–Chipman prediction is better than other plausible pretraining strategies. Stronger baselines should include at least one generic self-supervised learning method, such as contrastive learning or masked autoencoding, and preferably an ImageNet- or medical-pretrained baseline adapted to the same input structure. The ablation studies are also insufficient, as the paper does not clearly quantify the contribution of each design component, including the Lu–Chipman pretext target, Mueller element dropout, the two-stream m(0,0)/polarimetric design, physical realizability filtering, Mueller-specific rotations, and MAP-Org pre-training scale. Without these ablations, it is difficult to determine whether the gains come from the proposed physics-guided target or from other training details. The MAP-Org dataset is described as large, but only 41 multispectral Mueller matrix images are reported, so the manuscript should provide the effective number of samples or patches, tissue-level splits, wavelengths, spatial resolution, number of specimens per organ type, and pretraining/validation split. The classification split may also risk data leakage because random 60/20/20 splits are used for ColoPola, and the manuscript should clarify whether multiple slices could come from the same patient or tissue block. The qualitative human esophagus experiment is promising, but it is not enough to support strong robustness claims because it appears to rely on one ex vivo human sample and qualitative assessment only. The paper should soften claims about robustness to tissue or domain shift unless quantitative evaluation across more human samples and acquisition settings is added. Statistical testing is also missing, as the paper reports mean ± standard deviation over runs but does not provide confidence intervals, paired tests, or effect-size analysis, which is important because some improvements at 25–100% labels are small. Finally, the implementation details are incomplete, and the manuscript should include the loss function, optimizer, learning rates, batch size, patch size, training epochs, augmentation probabilities, dropout configurations, and exact downstream heads to support reproducibility.
- 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?
I recommend Weak Accept. The main reason is that the paper proposes a useful and physically grounded pretraining task for Mueller polarimetry, and the reported results show meaningful label-efficiency gains in limited annotated settings. The approach is relevant to biomedical imaging because manual dense annotations are scarce and domain shift is a major barrier for clinical translation. The main factors preventing a stronger recommendation are the limited baseline comparison, lack of ablation studies, incomplete reproducibility details, and an overstatement of robustness based on a mostly qualitative human-sample evaluation.
- 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 #4
- Please describe the contribution of the paper
The paper introduces MuellerPT, a physics-guided self-supervised learning (SSL) framework designed for Mueller matrix polarimetric imaging. Given the scarcity of densely annotated polarimetric data, the authors cleverly formulate the Lu-Chipman polar decomposition (yielding retardance, diattenuation, and depolarization) as a pretext task for pre-training a two-stream encoder. Additionally, the authors collected and released MAP-Org, a dataset comprising 41 multispectral animal polarimetric organ images to facilitate this pre-training. The pre-trained model is evaluated on downstream few-shot segmentation (PoLambRimetry) and classification (ColoPola) tasks.
- 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 most significant strength of this paper is the choice of the pretext task. Predicting physically interpretable Lu-Chipman parameters from 4x4 Mueller matrices is a highly rational, domain-specific application of self-supervised learning. It forces the network to learn representations directly tied to tissue microstructure.
In niche fields like optical polarimetry, the lack of large-scale annotated datasets is a massive hurdle. This framework offers a practical and label-efficient solution, demonstrated by absolute DICE gains of over 20% in a 5% few-shot segmentation regime.
The release of the MAP-Org dataset, even if moderately sized, is a commendable contribution to the open-science community in this specialized domain.
- 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 primary weakness of the evaluation is the choice of baselines. The authors only compare MuellerPT against an identical architecture trained from scratch (“HRNet-Scratch”). To truly demonstrate the superiority of the decomposition-driven pretext task, it should ideally be compared against standard generic SSL pretext tasks adapted for this domain (e.g., a simple autoencoder for reconstruction, or contrastive methods like SimCLR/MAE applied to the channels). Beating a from-scratch model in a 1% or 5% data regime is expected; proving that this specific pre-training is the best approach requires stronger baselines.
While the downstream tasks include a colorectal cancer dataset (ColoPola), the pre-training dataset (MAP-Org) heavily relies on ex vivo animal tissues sourced from a butcher. The generalization to human tissue is only demonstrated qualitatively on a single oesophagus sample. A more rigorous quantitative assessment on human datasets would strengthen the clinical relevance.
The proposed MAP-Org dataset contains 41 multispectral Mueller matrix images. While collecting such data is non-trivial and valuable, this is still a relatively small pre-training dataset. The paper should discuss the limitations of scale more explicitly and clarify how much diversity the dataset provides across organs, specimens, wavelengths, and acquisition conditions.
The main novelty lies in adapting physically meaningful Lu-Chipman decomposition prediction as a pretext task for Mueller polarimetry. This is valuable for the specific domain, but the overall deep learning formulation—pre-train an encoder on a dense prediction task and fine-tune it downstream—is relatively standard. The paper would benefit from clearer positioning against prior physics-guided pre-training and self-supervised learning methods in biomedical imaging.
- 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?
The paper addresses a specialized but important problem in biomedical Mueller matrix polarimetry, where labeled data are scarce and physically meaningful representations are valuable. The proposed use of Lu-Chipman decomposition prediction as a pretext task is well motivated and domain-specific. The results show meaningful improvements in low-label regimes for both segmentation and classification, and the release of the MAP-Org dataset would be useful for this niche research community.
However, the paper is borderline because the evaluation is not yet strong enough to fully support all claims. The main baseline is training from scratch, and there are no comparisons with generic self-supervised or pre-training strategies. The ablation study is limited, making it difficult to isolate the contributions of Lu-Chipman prediction, the two-stream design, Mueller element dropout, and the fusion/refinement modules. The domain-shift robustness claim is also supported mainly by a qualitative single-sample human tissue experiment. Therefore, while the paper is promising and relevant, its methodological and experimental evidence is not strong enough for a higher score.
Overall, I view this as a useful and domain-relevant contribution for a niche MICCAI area, but with moderate methodological novelty and incomplete experimental validation.
- Reviewer confidence
Somewhat confident (2)
- [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 would like to thank the reviewers and meta reviewer for their time and thoughtful feedback. We are encouraged that the reviewers had a generally positive view on MuellerPT. We address the concerns raised and minor adjustments will be made to the camera-ready submission. As the reviews have several similar comments, we present the response to the general concerns mentioned by the meta-reviewer first and then address the remaining individual reviewer comments. Baselines (R1, R2, R3, R4, MR): Our study was designed as a first evaluation. The architecture, splits and training schedules are fixed with only the weight initialisation changing. This isolates whether decomposition driven pretraining provides useful representations under scarce labels. Comparisons to generic SSL methods is an interesting suggestion for future work and will allow an empirical evaluation of the Lu-Chipman pretext task in comparison to generic SSL methods. Ablation study (R1, R2, R3, MR): This study evaluates the complete MuellerPT pipeline and does not provide experiments disentangling the contribution of the two-stream architecture, Mueller specific augmentations, element dropout, and Lu-Chipman target. The two stream design was found to lead to more physically plausible Lu-Chipman reconstructions, the Mueller augmentations increased dataset diversity, leading to improved generalizability and the element dropout was used to encourage learning of an ill-formed Lu-Chipman process (as opposed to simply using the numerical derivation) as seen in Figure 4.A comprehensive quantitative ablation study is an important area for future work to pinpoint which design choices lead to the best performance, although we believe this experiment would not fundamentally change the argument of our study which supports the use of Lu-Chipman as a pretraining task for biomedical Mueller polarimetry. Reproducibility (R1, R3, MR): The full code repository and MAP-Org dataset will be released at publication and included in the camera-ready submission to allow for re-implementation. Data leakage (R3, MR): The train/val/test split employed for the ColoPola dataset is similar to the split used by the original author. The ColoPola paper has the limitation that it does not provide information on whether the slices come from different patients or tissue blocks. This is therefore beyond our control, and we agree that caution should be taken when evaluating the benefits of MuellerPT on ColoPola. We nonetheless believe it to be a valuable source of preliminary evidence of image level transfer. Size of MAP-ORG (R1, R2, R3, R4): Each acquisition contains six wavelengths and dense 4x4 Mueller matrices over high resolution fields of view. However, we agree that MAP-Org is modest in terms of specimen level count and our intention is that MAP-Org is useful as a pretraining resource for biomedical Mueller polarimetry, where publicly available Mueller matrix datasets are limited. Human oesophagus experiment (R3, R4): We agree that the human oesophagus experiment should not be interpreted as quantitative evidence of robust clinical generalisation due to its limited size. The sample was chosen as a representative example human case, and although results were consistent across multiple samples, further quantitative analysis is needed. Sensitivity analysis for Lu-Chipman decomposition quality (R1): In our pipeline Lu-Chipman targets are computed after normalization and validity checks. Invalid or numerically unstable pixels are excluded to prevent any noise bias. A sensitivity analysis on how Lu-Chipman quality impacts training is an interesting extension although given the focus of this study on physically valid pixels, this would not fundamentally change the argument of our study. Methodology novelty (R4): The novelty is not the architectural or optimization framework but instead the formulation of a Mueller polarimetry specific dense pretext task based on Lu-Chipman decomposition.
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.
This manuscript introduces a physics guided pre-training approach for segmentation and classification tasks in Mueller Polarimetry, based on a newly acquired dataset. Reviewers thought that this work represents a clear, well-motivated and elegant technical contribution in an under-researched area, combined with a new dataset resource. Their main concern was narrow experimental comparisons, in particular, not accounting for any plausible alternative pre-training strategies. Moreover, given that the proposed method consists of multiple components, ablations would help identify those that are most relevant. Since no new experimental results are allowed in the rebuttal, I do not expect it would provide substantial new information for making a decision. As it stands, reviewers are on the positive side overall. In the final version, authors will need to address the potential data leakage in their classification results (suspected by reviewer 3) and issues of reproducibility, which were another major concern among reviewers.