MuellerPT: Decomposition Driven Pre-training for Dense Learning in Mueller Polarimetry

Adam Tlemsani, Yingdian Li, Maxime Giot, Naim Slim, Christopher J. Peters, Abhijeet Ghosh, Daniel S. Elson · 2026

Read MuellerPT: Decomposition Driven Pre-training for Dense Learning in Mueller Polarimetry by Adam Tlemsani, Yingdian Li, Maxime Giot, Naim Slim, Christopher J. Peters, Abhijeet Ghosh, Daniel S. Elson (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-23840.txt.

Introduction

Adam Tlemsani and colleagues attack a data scarcity problem in Mueller polarimetry. This imaging reads rich tissue contrast from polarized light. Their MuellerPT pretraining teaches an encoder to predict Lu-Chipman decomposition maps from per pixel Mueller matrices. A new multispectral animal organ dataset scales the training up. With only five percent of training data, lamb brain grey versus white matter segmentation jumps by more than twenty percent above training from scratch. Cancer classification gains eight percent at one percent of data.

Catalog record

CREATORAdam Tlemsani, Yingdian Li, Maxime Giot, Naim Slim, Christopher J. Peters, Abhijeet Ghosh, Daniel S. Elson
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21085
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.23840

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MuellerPT: Decomposition Driven Pre-training for Dense Learning in Mueller Polarimetry Adam Tlemsani, Yingdian Li, Maxime Giot, Naim Slim, Christopher J. Peters, Abhijeet Ghosh, Daniel S. Elson 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. 1 Introduction Polarization imaging encodes information that conventional intensity imaging is often insensitive to, particularly in turbid biological tissue where scattering, absorption anisotropy and tissue structural organisation affect the measured polarimetric response [ 10 ] . In the Stokes-Mueller formalism, a \textbf{M}\in\mathbb{R}^{4\times{4}} Mueller matrix describes the transformation of the incident Stokes vector into an emergent Stokes vector [ 17 ] . Mueller Matrix Polarimetry (MMP) is particularly beneficial when microstructural changes in the tissue are diagnostic, such as the anisotropy changes evident in cancers including cervix, skin, colon, pancreas and brain [ 10 , 18 , 19 , 1 , 8 ] . Despite the promise of MMP for a variety of medical diagnostic tasks, severe barriers impede the adoption of MMP systems including the time-sequential acquisition [ 7 ] , large bulky systems [ 3 ] and how to efficiently utilise the Mueller matrix for clinical tasks [ 10 ] . These constraints motivate the need for learning pipelines which are label efficient, robust to acquisition artefacts, resilient to domain shift and capable of producing representations which align with physical tissue contrast mechanisms. The Lu-Chipman decomposition [ 12 ] factorises measured Mueller matrices into a sequence of matrices interpreted as a diattenuator, retarder and depolarizer, providing a clear route from 16 element Mueller matrices to a small set of physically interpretable parameters. Whilst including fewer parameters, Lu-Chipman decomposition can retain almost as much of the clinically relevant information as full 16 element MMP [ 14 ] . This motivates working with these decompositions rather than the 16 element Mueller matrix as the parameters can be interpreted whilst maintaining much of the diagnostic information. These Lu-Chipman parameter maps are attractive for pre-training because they can be computed, without manual annotation, from unlabelled Mueller image acquisitions, providing structured supervision grounded in optics and a natural target which encourage encoders to learn features aligned with tissue microstructure. At the same time, these decomposition-derived maps are not error free but can be sensitive to measurement noise and depolarization behaviour, underscoring the need for robust learning objectives and evaluation [ 11 ] . In polarimetric imaging, the data-driven polarimetry literature is expanding rapidly, but much of it remains supervised or focused on task-specific restoration rather than general purpose encoder pre-training [ 21 ] . Explicit reusable pre-training schemes tailored to Mueller physics are scarce in the literature. Adjacent polarisation domains have demonstrated the utility of Self-Supervised Learning (SSL), for example, self-supervised contrastive learning has been proposed for 3D polarized light imaging of nerve fibre architecture [ 15 ] , yet to the best of our knowledge there exists no general Mueller pre-trained encoders. The core problem this paper addresses is that Mueller imaging models must learn informative, clinically transferable representations from high-dimensional polarimetric measurements under limited labels and diverse acquisition conditions (noise, calibration drift, device differences and tissue/domain shift) [ 11 ] . Existing medical SSL pipelines, while powerful, are typically agnostic to Mueller matrix information and many Mueller imaging studies either (1) train from scratch on small datasets (2) rely on generic model initialisations, or (3) use Lu-Chipman maps primarily as supervised features and not a pre-training signal. We contribute the following. 1. MuellerPT, a novel pre-training pipeline for Mueller imaging encoders that uses Lu-Chipman parameter prediction as the pretext task 2. We demonstrate improved downstream segmentation and classification performance, particularly in the limited data setting, on representative Mueller imaging tasks compared to training from scratch with random weight initialisation 3. We demonstrate robustness to domain shift as MuellerPT is resilient to various datasets and simulated systems 4. A large open source dataset of reflected Mueller matrix images on a diverse range of animal tissue types set for release with this manuscript, named the Multispectral Animal Polarimetric Organ dataset (MAP-Org). 2 Methodology 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, f_{\theta} and m_{\psi} 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 liver/breast/wings, steak, bacon) sourced from local butchers. The images were collected with a reflection configuration Mueller matrix imaging system. The system is broken down into two parts, a polarization state generator (PSG) and polarization state analyser (PSA).The PSG contains a collimated white light emitting diode (LED, MCWHLP3, Thorlabs, Newton, New Jersey, United States), a filter wheel with six different wavelengths ranging from 450 to 680 nm (FW102C, Thorlabs), a linear polarizer and a rotating quarter wave plate. This PSG generates four incident illumination polarization states. After the light reflects of the sample, a polarization sensitive camera (Blackfly S, model BFS-U3-51S5P-C, Teledyne FLIR) captures the reflected polarization state after passing through a slider containing either air or another quarter wave plate. Using a collection of known reference samples and the eigenvalue calibration method [ 4 ] , the calibration error of our system was below 4.1% for each wavelength. For a given measured Mueller image \mathbf{M}\in\mathbb{R}^{H\times{W}\times{6}\times{4}\times{4}} , where H and W denote the spatial dimensions, 6 the number of wavelengths and 4\times{4} the elements of the Mueller matrix which will be denoted by m(0,0) for the first element and m(3,3) for the last. We first normalise our matrices by dividing out by m(0,0) to get intensity independent representations. Measured Mueller matrices often lead to physically unrealisable matrices due to noise, saturation and poor alignment. We filter out these physically unrealisable matrices by testing for the condition that the eigenvalues of the coherency matrix are greater than or equal to 0 [ 6 ] . Matrices which fail this test

Read the complete text