CoChat

The DOI matches its Crossref record; 118 of the 118 references it cites were found in a registry or an index; 11 of the 21 sources its claims cite were checked for support.

Hyperspectral unmixing for Raman spectroscopy via physics-constrained autoencoders

Open access
Georgiev, Dimitar, Fernández-Galiana, Álvaro, Vilms Pedersen, Simon, Papadopoulos, Georgios, Xie, Ruoxiao, Stevens, Molly M., Barahona, Mauricio · 2024 · Proceedings of the National Academy of Sciences

Abstract

Raman spectroscopy is widely used across scientific domains to characterize the chemical composition of samples in a nondestructive, label-free manner. Many applications entail the unmixing of signals from mixtures of molecular species to identify the individual components present and their proportions, yet conventional methods for chemometrics often struggle with complex mixture scenarios encountered in practice. Here, we develop hyperspectral unmixing algorithms based on autoencoder neural networks, and we systematically validate them using both synthetic and experimental benchmark datasets created in-house. Our results demonstrate that unmixing autoencoders provide improved accuracy, robustness, and efficiency compared to standard unmixing methods. We also showcase the applicability of autoencoders to complex biological settings by showing improved biochemical characterization of volumetric Raman imaging data from a monocytic cell.

Overview

CoChat Summary

TL;DR

The paper introduces physics-constrained autoencoder neural networks for blind hyperspectral unmixing of Raman spectroscopy data. Across synthetic mixtures, experimental sugar mixtures, and volumetric Raman imaging of a THP-1 cell, the autoencoders generally outperform conventional N-FINDR/VCA + NNLS/FCLS methods, especially under noise, artifacts, and no pure-pixel conditions.

Key findings

  • Autoencoders outperformed conventional methods on synthetic Raman mixtures across 11 dataset variants, with improvement increasing for noisier and more artifact-contaminated scenarios; they matched conventional methods on the simplest ideal Chessboard datasets. Experiments used 6 models × 11 dataset variants × 25 replicates = 1,650 experiments.
  • On experimental sugar mixtures, autoencoders performed comparably to conventional methods on clean high-SNR data when pure-solution reference spectra were included, but substantially outperformed them in low-SNR settings and when pure pixels were absent. The sugar data comprised 240 mixtures and high-SNR (1,920 spectra, 5 s integration) and low-SNR (7,680 spectra, 0.5 s integration) measurements.
  • Autoencoders were computationally faster than N-FINDR+FCLS and VCA+FCLS on synthetic datasets up to 250,000 spectra, even using CPU-only computation and including training time.
  • In a low-SNR volumetric Raman scan of a THP-1 cell, autoencoder unmixing produced more distinct endmember signatures and well-defined abundance features than VCA+NNLS, and detected cholesterol, which the original VCA+NNLS approach did not.

Methods

  • The framework treats the autoencoder latent representation as fractional abundances and the decoder as a mixing function; physics-inspired constraints include nonnegative endmembers, abundance nonnegativity, and sum-to-one abundances. Four encoders were tested (Dense, Convolutional, Transformer, Convolutional Transformer) with linear or bilinear Fan-model decoders, trained self-supervised using spectral angle distance loss.
  • Synthetic data were generated with known ground truth: 5 endmembers, 1,000 spectral bands, and 100×100 abundance scenes (10,000 spectra per dataset), spanning Chessboard, Gaussian, and Dirichlet scenes plus linear/bilinear mixtures and artifacts.
  • Experimental validation used 240 in-house sugar mixtures of glucose, sucrose, fructose, and maltose, plus a volumetric THP-1 cell Raman scan; performance was measured by MSE for abundances and SAD for endmembers, with Hungarian matching to ground truth when available.

Limitations

  • The paper states limitations including potentially higher data requirements for stable performance, challenging interpretability of learned internal representations, and sensitivity to hyperparameters requiring careful tuning.
  • The text also notes that synthetic data are inherently idealized and that findings from synthetic data alone should be extrapolated with caution.
  • For the THP-1 cell application, the paper states that direct comparisons are challenging due to the lack of ground truth.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1073/pnas.2407439121

Cites

118 of 118 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1. Hyperspectral unmixing for Raman spectroscopy using autoencoder neural networks. (A) Illustration of hyperspectral unmixing. A set of Raman measurements X={xi}, where each xi is a spectrum, are collected from a sample of interest. The spectra xi are viewed as mixtures of a set of Raman spectral signatures M (endmembers), related to the chemical species present, and their respective fractional abundances ai. Hyperspectral unmixing aims to infer the set of unknown underlying endmembers M and fractional abundances ai from the set of measurements X, thereby obtaining qualitative and quantitative compositional information about the sample. (B) Hyperspectral unmixing as a self-supervised autoencoder learning problem: the decoder D learns endmember signatures during training and the encoder E learns to derive the corresponding fractional abundances. (C) Encoders can accommodate different concepts from representation learning, such as convolutional layers and attention, to improve feature extraction and provide more accurate and robust unmixing. (D) Decoders can be structured to model different linear and nonlinear mixing models. Physics-inspired constraints are indicated in red in C and D.
Fig. 1. Hyperspectral unmixing for Raman spectroscopy using autoencoder neural networks. (A) Illustration of hyperspectral unmixing. A set of Raman measurements X={xi}, where each xi is a spectrum, are collected from a sample of interest. The spectra xi are viewed as mixtures of a set of Raman spectral signatures M (endmembers), related to the chemical species present, and their respective fractional abundances ai. Hyperspectral unmixing aims to infer the set of unknown underlying endmembers M and fractional abundances ai from the set of measurements X, thereby obtaining qualitative and quantitative compositional information about the sample. (B) Hyperspectral unmixing as a self-supervised autoencoder learning problem: the decoder D learns endmember signatures during training and the encoder E learns to derive the corresponding fractional abundances. (C) Encoders can accommodate different concepts from representation learning, such as convolutional layers and attention, to improve feature extraction and provide more accurate and robust unmixing. (D) Decoders can be structured to model different linear and nonlinear mixing models. Physics-inspired constraints are indicated in red in C and D.CC BY · PubMed Central
Fig. 2. Benchmarking autoencoders on synthetic Raman mixtures. (A) Schematic of our synthetic data generation workflow, which allows us to create synthetic mixtures with known ground truth. (B) Representative unmixing results (endmembers (Left), fractional abundances (Right)) obtained by each of the six algorithms considered here when applied to one of our synthetic datasets (Chessboard scene, +artifacts scenario). (C and D) Summary of the unmixing performance of the six algorithms on 11 types of synthetic datasets, which span variable mixing levels and complexity. Linear mixtures are shown in C, and bilinear mixtures in D. CI are given by one SD around the sample mean (each bar represents n=25 evaluations: 5 datasets with 5 model repetitions each).
Fig. 2. Benchmarking autoencoders on synthetic Raman mixtures. (A) Schematic of our synthetic data generation workflow, which allows us to create synthetic mixtures with known ground truth. (B) Representative unmixing results (endmembers (Left), fractional abundances (Right)) obtained by each of the six algorithms considered here when applied to one of our synthetic datasets (Chessboard scene, +artifacts scenario). (C and D) Summary of the unmixing performance of the six algorithms on 11 types of synthetic datasets, which span variable mixing levels and complexity. Linear mixtures are shown in C, and bilinear mixtures in D. CI are given by one SD around the sample mean (each bar represents n=25 evaluations: 5 datasets with 5 model repetitions each).CC BY · PubMed Central
Fig. 3. Computational efficiency analysis. The computational cost (measured as wall time) of autoencoders and conventional methods on synthetic datasets (Chessboard scene, +artifacts scenario) as the number of spectra is increased. Each dot represents the average across three evaluations (CI based on one SD are small and not visible to the eye). AE models are equipped with decoders for linear unmixing.
Fig. 3. Computational efficiency analysis. The computational cost (measured as wall time) of autoencoders and conventional methods on synthetic datasets (Chessboard scene, +artifacts scenario) as the number of spectra is increased. Each dot represents the average across three evaluations (CI based on one SD are small and not visible to the eye). AE models are equipped with decoders for linear unmixing.CC BY · PubMed Central
Fig. 4. Experimental validation on Raman spectroscopy data from sugar mixtures. (A) Schematic diagram of sugar mixture preparation. Two sets of Raman measurements are acquired from a library of 240 sugar mixtures—high and low signal-to-noise ratio (SNR) data collected using integration times of 5s and 0.5s, respectively. (B) Ground-truth endmember signatures were estimated by taking the median of high SNR reference spectra measured from pure solutions containing a single sugar species. (C and D) Summary of unmixing performance for: (C) idealized scenario where spectra collected from pure solutions are included; and (D) original data without the spectra from pure solutions (i.e., no pure pixels). CI are given as one SD around the sample mean (n=5).
Fig. 4. Experimental validation on Raman spectroscopy data from sugar mixtures. (A) Schematic diagram of sugar mixture preparation. Two sets of Raman measurements are acquired from a library of 240 sugar mixtures—high and low signal-to-noise ratio (SNR) data collected using integration times of 5s and 0.5s, respectively. (B) Ground-truth endmember signatures were estimated by taking the median of high SNR reference spectra measured from pure solutions containing a single sugar species. (C and D) Summary of unmixing performance for: (C) idealized scenario where spectra collected from pure solutions are included; and (D) original data without the spectra from pure solutions (i.e., no pure pixels). CI are given as one SD around the sample mean (n=5).CC BY · PubMed Central
Fig. 5. Volumetric Raman imaging of a THP-1 cell. (A) Brightfield image of the studied THP-1 cell. (B) Reconstruction of a cross-section of the cell obtained by overlaying fractional abundances derived by: VCA+NNLS, our Dense AE model, and our Deep Dense AE model. (C) Detailed overview of the unmixing results obtained with our Deep Dense AE model, displaying the spatial distribution (layer and volume) of the individual fractional abundances (Left) and the associated endmember signatures (Right). Fractional abundance maps are normalized for consistent visualization. Brightfield image and raw Raman spectroscopy data are from Kallepitis et al. (86).
Fig. 5. Volumetric Raman imaging of a THP-1 cell. (A) Brightfield image of the studied THP-1 cell. (B) Reconstruction of a cross-section of the cell obtained by overlaying fractional abundances derived by: VCA+NNLS, our Dense AE model, and our Deep Dense AE model. (C) Detailed overview of the unmixing results obtained with our Deep Dense AE model, displaying the spatial distribution (layer and volume) of the individual fractional abundances (Left) and the associated endmember signatures (Right). Fractional abundance maps are normalized for consistent visualization. Brightfield image and raw Raman spectroscopy data are from Kallepitis et al. (86).CC BY · PubMed Central

Verification

Claims the paper makes

8 Valid · 0 Not valid · 3 Not checked · 36 Own results and interpretations

18 more claims

Create a free account to see all 47 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 13
  • It may state two separate claims as one. 12

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 111 references

Research works printed but not linked: 5

  1. M. E. Winter, N-FINDR, An algorithm for fast autonomous spectral end-member determination in hyperspectral data in Imaging Spectrometry V. SPIE 3753, 266–275 (1999).
  2. I. Goodfellow Y. Bengio A. Courville Autoencoders in Deep learning (MIT Press 2016) pp. 499–523.
  3. A. Vaswani . “Attention is all you need” in Advances in Neural Information Processing Systems 30 I. Guyon et al. Eds. (NIPS 2017).
  4. P. H. Eilers, H. F. Boelens, Baseline correction with asymmetric least squares smoothing. Leiden Univ. Med. Centre Rep. 1, 5 (2005).
  5. M. Abadi . “A system for large-scale machine learning” in OSDI (Savannah GA 2016) vol. 16 pp. 265–283.

Cited by

See all 50

Metadata

Authors
Georgiev, D., Fernández-Galiana, Á., Vilms Pedersen, S., Papadopoulos, G., Xie, R., Stevens, M. M., & Barahona, M.
Affiliations
  • Imperial College London (GB)
  • Department of Materials, Faculty of Engineering, Imperial College London, London SW7 2AZ, United Kingdom
  • University of Oxford (GB)
Corresponding author
  • Fernández-Galiana, Álvaro
  • Stevens, Molly M.
  • Barahona, Mauricio
Venue
Proceedings of the National Academy of Sciences
Published
2024-10-29
Volume
121
Issue
45
Pages
e2407439121-e2407439121
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
OpenAlex
W4403863275
Cited by
50
References
111
Citations by year
2026: 30 · 2025: 17 · 2024: 3
FWCI
9.61
Type
journal article
Language
en
Open access
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • Spiral (Imperial College London) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • UK Research and Innovation (EP/S023283/1)
  • Schmidt Science Fellowship (N/A)
  • Rhodes Fellowship (N/A)
  • Independent Research Fund Denmark (N/A)
  • UKRI | Engineering and Physical Sciences Research Council (EP/P00114/1, EP/T020792/1, EP/N014529/1, EP/T027258/1)
  • Royal Academy of Engineering (CiET2021\94)
  • Bio Innovation Institute (N/A)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Biophysics › Spectroscopy Techniques in Biomedical and Chemical Research
  • Physical Sciences › Chemistry › Analytical Chemistry › Spectroscopy and Chemometric Analyses
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Remote-Sensing Image Classification
Keywords
hyperspectral unmixing, Raman spectroscopy, autoencoder, spectral unmixing, chemometrics, monocytic cells, biochemical characterization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Hyperspectral unmixing for Raman spectroscopy via physics-constrained autoencoders. https://app.cochat.ai/papers/10.1073/pnas.2407439121
The paper
Georgiev, D., Fernández-Galiana, Á., Vilms Pedersen, S., Papadopoulos, G., Xie, R., Stevens, M. M., & Barahona, M. (2024). Hyperspectral unmixing for Raman spectroscopy via physics-constrained autoencoders. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2407439121

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.