pith. sign in

arxiv: 2605.23840 · v1 · pith:R6W7BDSYnew · submitted 2026-05-22 · 💻 cs.CV

MuellerPT: Decomposition Driven Pretraining for Dense Learning in Mueller Polarimetry

Adam Tlemsani , Yingdian Li , Maxime Giot , Naim Slim , Christopher J. Peters , Abhijeet Ghosh , Daniel S. Elson This is my paper

Pith reviewed 2026-05-25 04:48 UTC · model grok-4.3

classification 💻 cs.CV
keywords Mueller polarimetrypretrainingLu-Chipman decompositionlabel efficiencytissue segmentationcancer classificationbiomedical imagingdomain shift
0
0 comments X

The pith

Pretraining by predicting Lu-Chipman decomposition maps from Mueller matrices enables more label-efficient segmentation and classification in biomedical polarimetry.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MuellerPT, a pretraining method that teaches an encoder to predict Lu-Chipman decomposition maps directly from per-pixel 4x4 Mueller matrices. This physics-guided pretext task is meant to produce dense representations that transfer to downstream tasks despite scarce labels and shifts across specimens. When the pretrained encoder is adapted with a segmentation head, it improves grey-versus-white-matter separation on lamb brain data; with a classification head, it improves colorectal cancer detection. Both tasks show clear gains in few-shot regimes compared with training from scratch, and the approach is tested for robustness on an ex-vivo human esophagus sample.

Core claim

MuellerPT learns transferable representations by pretraining an encoder to predict Lu-Chipman decomposition maps from 4x4 Mueller matrices collected in a new Multispectral Animal Polarimetric Organ dataset. The same encoder is then fine-tuned for grey-versus-white-matter segmentation and for colorectal cancer classification. In the segmentation task the method yields an absolute DICE gain of over 20 percent relative to a from-scratch baseline at 5 percent labeled data; in classification it raises overall accuracy by 8 percent at 1 percent labeled data. Qualitative inspection of the predicted decomposition maps on an unseen human esophagus sample further indicates robustness to domain shift.

What carries the argument

Prediction of Lu-Chipman decomposition maps as the pretext task whose targets are derived from the input Mueller matrices.

If this is right

  • Fewer labeled examples suffice to reach high DICE scores on grey-versus-white-matter segmentation.
  • Overall accuracy in colorectal cancer classification rises even when only 1 percent of the training labels are available.
  • Cross-specimen transfer improves without task-specific adaptation beyond the initial pretraining.
  • The learned representations remain stable enough to produce plausible decomposition maps on tissue types absent from the pretraining set.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same decomposition-prediction objective could be applied to other Mueller-matrix tasks such as birefringence mapping or collagen orientation estimation.
  • Large unlabeled polarimetric datasets collected under varying acquisition conditions may become more valuable once this pretraining route is used.
  • If decomposition targets prove effective across modalities, analogous physics-derived pretext tasks could be designed for other imaging techniques that admit matrix or tensor decompositions.

Load-bearing premise

Predicting Lu-Chipman decomposition maps produces representations that transfer effectively to the segmentation and classification tasks.

What would settle it

A controlled run in which a model pretrained to predict random or shuffled decomposition targets shows no performance advantage over a from-scratch baseline on the same low-data segmentation and classification splits.

Figures

Figures reproduced from arXiv: 2605.23840 by Abhijeet Ghosh, Adam Tlemsani, Christopher J. Peters, Daniel S. Elson, Maxime Giot, Naim Slim, Yingdian Li.

Figure 1
Figure 1. Figure 1: MuellerPT outline - we first pre-trained on our new MAP-Org dataset and used Lu-Chipman prediction as our pretext task. The encoders for the Mueller and M00 images, fθ and mψ respectively, were then used and fine-tuned for the downstream tasks including segmentation and classification. A dataset of 41 multispectral Mueller matrix images was acquired of fresh animal tissue (lamb heart/kidney/chops, chicken … view at source ↗
Figure 2
Figure 2. Figure 2: Representative example Mueller matrices from our MAP-Org dataset. The m00 element displayed is un-normalized We report both macro and per class DICE scores [5] for PoLambRimetry. We report overall accuracy, sensitivity and specificity for ColoPola. To assess performance in data limited scenarios, we report few shot performance for (1, 5, 25, 50 and 100 percent of the training data). The mean DICE scores fo… view at source ↗
Figure 3
Figure 3. Figure 3: Visualisations of the PoLambRimetry segmentation results for the HRNet￾Scratch (top row) and our MuellerPT (bottom row) for different few shot learning scenarios. Grey matter is blue and white matter is yellow in this figure. The ground truth result is presented in the last column. To understand how pre-training impacts both classes we report the per class DICE score in [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 4
Figure 4. Figure 4: Mueller matrix and MuellerPT Lu-Chipman decomposition Retardance (R), Depolarization (∆) and Diattenuation (D) results for a sample of oesophagus Mucinous adenocarcinoma (MAC) from a patient. The decompositions are computed for a set of reduced Mueller matrix configurations. We present the results for the ColoPola dataset in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces MuellerPT, a physics-guided pretraining approach for Mueller polarimetry that learns dense representations by predicting Lu-Chipman decomposition maps from per-pixel 4x4 Mueller matrices. It collects a new large Multispectral Animal Polarimetric Organ (MAP-Org) dataset to scale pretraining. The pretrained encoder is adapted for few-shot grey/white matter segmentation in lamb brain and colorectal cancer classification, reporting absolute DICE gains >20% vs. scratch baseline at 5% labels and accuracy gains of 8% at 1% labels, plus qualitative robustness on human oesophagus data. The central claim is that Lu-Chipman prediction is an effective and practical pretext task for label-efficient, domain-robust inference.

Significance. If the central claim holds after controls, the work would establish a concrete route for embedding physical priors into self-supervised pretraining for Mueller matrix imaging, addressing annotation scarcity and domain shift in biomedical applications. The MAP-Org dataset would also be a reusable resource. The approach is distinctive in tying the pretext task directly to a standard polarimetric decomposition rather than generic reconstruction or contrastive objectives.

major comments (1)
  1. [Experiments / Results] Experiments (results on segmentation and classification): The manuscript reports large gains from MuellerPT but does not include an ablation that applies an alternative self-supervised objective (e.g., pixel-wise reconstruction or contrastive loss) to the identical MAP-Org 4x4 Mueller matrices using the same encoder backbone. Without this control, the headline improvements (DICE +20% at 5% labels, accuracy +8% at 1% labels) cannot be attributed specifically to the Lu-Chipman decomposition pretext rather than any large-scale pretraining on the target domain. This directly undermines the claim that the decomposition constitutes an effective pretext task.
minor comments (2)
  1. [Abstract / §3] Abstract and §3: Quantitative claims are stated without error bars, dataset statistics (e.g., number of specimens, pixels, or acquisition conditions in MAP-Org), or ablation tables; these details are required to assess the reliability of the reported gains.
  2. [Methods] Methods: The precise formulation of the pretext loss (how Lu-Chipman parameters are regressed from the 4x4 matrix input) and the encoder architecture are not described at a level that permits reproduction; add equations and pseudocode.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. The major comment correctly identifies a missing control that would help isolate the contribution of the Lu-Chipman pretext task. We address this point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Experiments / Results] Experiments (results on segmentation and classification): The manuscript reports large gains from MuellerPT but does not include an ablation that applies an alternative self-supervised objective (e.g., pixel-wise reconstruction or contrastive loss) to the identical MAP-Org 4x4 Mueller matrices using the same encoder backbone. Without this control, the headline improvements (DICE +20% at 5% labels, accuracy +8% at 1% labels) cannot be attributed specifically to the Lu-Chipman decomposition pretext rather than any large-scale pretraining on the target domain. This directly undermines the claim that the decomposition constitutes an effective pretext task.

    Authors: We agree that the current experiments, which compare MuellerPT only against training from scratch, do not fully isolate whether the gains arise from the specific Lu-Chipman decomposition objective or from large-scale pretraining on the MAP-Org domain in general. An ablation with alternative self-supervised objectives (pixel-wise reconstruction and contrastive loss) on the identical 4x4 Mueller matrices and encoder would strengthen attribution to the physics-guided pretext. In the revised manuscript we will add these controls using the same backbone and dataset, allowing direct comparison of the resulting downstream performance in the few-shot segmentation and classification settings. This revision will clarify the contribution of the decomposition task. revision: yes

Circularity Check

0 steps flagged

No significant circularity; pretraining objective defined independently of downstream tasks

full rationale

The paper defines the pretext task as predicting Lu-Chipman decomposition maps (a standard, externally defined physical procedure applied to input 4x4 Mueller matrices) and evaluates transfer on separate downstream segmentation and classification tasks using few-shot data. No step reduces a claimed result to a fitted parameter, self-citation chain, or definitional equivalence; the decomposition target is computed deterministically from the data rather than learned or renamed from the target metrics. The absence of an ablation comparing to generic pretraining is a methodological gap but does not create circularity in the derivation chain itself.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no information on free parameters, axioms, or invented entities.

pith-pipeline@v0.9.0 · 5813 in / 1169 out tokens · 25077 ms · 2026-05-25T04:48:17.175541+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Photodiagnosis and photodynamic therapy30, 101708 (2020)

    Ahmad, I., Khaliq, A., Iqbal, M., Khan, S.: Mueller matrix polarimetry for char- acterization of skin tissue samples: A review. Photodiagnosis and photodynamic therapy30, 101708 (2020)

    work page 2020

  2. [2]

    Journal of Biomed- ical Optics29(9), 096002 (2024)

    Alonso, V.M., Anichini, G., Qi, J., O’Neill, K., Conde, O.M., Elson, D.S.: PoLam- bRimetry: a multispectral polarimetric atlas of lamb brain. Journal of Biomed- ical Optics29(9), 096002 (2024). https://doi.org/10.1117/1.JBO.29.9.096002, https://doi.org/10.1117/1.JBO.29.9.096002

    work page doi:10.1117/1.jbo.29.9.096002 2024

  3. [3]

    IEEE Transactions on Medical Imaging (2025)

    Chae, S., Huang, T., Rodríguez-Núñez, O., Lucas, T., Vanel, J.C., Vizet, J., Pierangelo, A., Piavchenko, G., Genova, T., Ajmal, A., et al.: Machine learning approach to 3×4 mueller polarimetry for complete reconstruction of diagnostic polarimetric images of biological tissues. IEEE Transactions on Medical Imaging (2025)

    work page 2025

  4. [4]

    Applied optics38(16), 3490–3502 (1999)

    Compain, E., Poirier, S., Drevillon, B.: General and self-consistent method for the calibration of polarization modulators, polarimeters, and mueller-matrix ellipsome- ters. Applied optics38(16), 3490–3502 (1999)

    work page 1999

  5. [5]

    Dice,L.R.:Measuresoftheamountofecologicassociationbetweenspecies.Ecology 26(3), 297–302 (1945)

    work page 1945

  6. [6]

    Applied Optics55(20), 5449–5455 (2016)

    Gil, J.J.: On optimal filtering of measured mueller matrices. Applied Optics55(20), 5449–5455 (2016)

    work page 2016

  7. [7]

    Optics express 15(12), 7393–7400 (2007)

    Guyot, S., Anastasiadou, M., Deléchelle, E., De Martino, A.: Registration scheme suitable to mueller matrix imaging for biomedical applications. Optics express 15(12), 7393–7400 (2007)

    work page 2007

  8. [8]

    Optics express33(20), 43379–43392 (2025)

    Hahne, C., Diaz, I., Rodríguez-Núñez, O., Gros, É., Blatter, M., Lucas, T., Hasler, D., Novikova, T., Maragkou, T., Schucht, P., et al.: Polarimetric feature analysis of mueller matrices for brain tumor image segmentation. Optics express33(20), 43379–43392 (2025)

    work page 2025

  9. [9]

    IEEE Transactions on Image Processing (2025) 10 A

    Hahne, C., Rodríguez-Núñez, O., Gros, É., Lucas, T., Hewer, E., Novikova, T., Maragkou, T., Schucht, P., McKinley, R.: Physically consistent image augmenta- tion for deep learning in mueller matrix polarimetry. IEEE Transactions on Image Processing (2025) 10 A. Tlemsani et al

    work page 2025

  10. [10]

    Light: Science & Applications10(1), 194 (2021)

    He, C., He, H., Chang, J., Chen, B., Ma, H., Booth, M.J.: Polarisation optics for biomedical and clinical applications: a review. Light: Science & Applications10(1), 194 (2021)

    work page 2021

  11. [11]

    Biomedical Optics Express12(8), 5160–5178 (2021)

    Huynh, R.N., Nehmetallah, G., Raub, C.B.: Mueller matrix polarimetry and po- lar decomposition of articular cartilage imaged in reflectance. Biomedical Optics Express12(8), 5160–5178 (2021)

    work page 2021

  12. [12]

    Journal of the Optical Society of America A13(5), 1106–1113 (1996)

    Lu, S.Y., Chipman, R.A.: Interpretation of mueller matrices based on polar de- composition. Journal of the Optical Society of America A13(5), 1106–1113 (1996)

    work page 1996

  13. [13]

    Applied Optics60(22), 6682–6694 (2021)

    Ma, D., Lu, Z., Xia, L., Liao, Q., Yang, W., Ma, H., Liao, R., Ma, L., Liu, Z.: Muellernet: a hybrid 3d–2d cnn for cell classification with mueller matrix images. Applied Optics60(22), 6682–6694 (2021)

    work page 2021

  14. [14]

    Novikova, T., Ramella-Roman, J.C.: Is a complete mueller matrix necessary in biomedical imaging? Optics Letters47(21), 5549–5552 (2022)

    work page 2022

  15. [15]

    Imaging Neuroscience2, imag–2 (2024)

    Oberstrass, A., Muenzing, S.E., Niu, M., Palomero-Gallagher, N., Schiffer, C., Axer, M., Amunts, K., Dickscheid, T.: Self-supervised representation learning for nerve fiber distribution patterns in 3d-pli. Imaging Neuroscience2, imag–2 (2024)

    work page 2024

  16. [16]

    Giga- Science14, giaf120 (2025)

    Pham, T.T.H., Vo, Q.H.Q., Nguyen, T.V., Nguyen, T.H., Phan, Q.H., Le, T.H.: Colopola: A polarimetric imaging dataset for colorectal cancer detection. Giga- Science14, giaf120 (2025)

    work page 2025

  17. [17]

    Journal of biophotonics10(8), 950–982 (2017)

    Qi, J., Elson, D.S.: Mueller polarimetric imaging for surgical and diagnostic appli- cations: a review. Journal of biophotonics10(8), 950–982 (2017)

    work page 2017

  18. [18]

    Journal of biomedical optics28(10), 102904–102904 (2023)

    Robinson, D., Hoong, K., Kleijn, W.B., Doronin, A., Rehbinder, J., Vizet, J., Pierangelo, A., Novikova, T.: Polarimetric imaging for cervical pre-cancer screening aided by machine learning: ex vivo studies. Journal of biomedical optics28(10), 102904–102904 (2023)

    work page 2023

  19. [19]

    Scientific reports13(1), 16417 (2023)

    Sampaio, P., Lopez-Antuña, M., Storni, F., Wicht, J., Sökeland, G., Wartenberg, M., Márquez-Neila, P., Candinas, D., Demory, B.O., Perren, A., et al.: Müller matrix polarimetry for pancreatic tissue characterization. Scientific reports13(1), 16417 (2023)

    work page 2023

  20. [20]

    IEEE transactions on pattern analysis and machine intelligence43(10), 3349–3364 (2020)

    Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recog- nition. IEEE transactions on pattern analysis and machine intelligence43(10), 3349–3364 (2020)

    work page 2020

  21. [21]

    Opto-Electronic Science3(2), 230042–1 (2024)

    Yang, K., Liu, F., Liang, S., Xiang, M., Han, P., Liu, J., Dong, X., Wei, Y., Wang, B., Shimizu, K., et al.: Data-driven polarimetric imaging: a review. Opto-Electronic Science3(2), 230042–1 (2024)

    work page 2024