Generative Augmentation of Raman Spectra for Glioma Classification

Andrei Iu\c{s}an; Andrei P\u{a}un; Bogdan Oancea; Daria Voiculescu; Ion Petre; Iulian Vasile; Mihaela P\u{a}un

REVIEW 3 major objections 5 minor 17 references

Generative Augmentation of Raman Spectra for Glioma Classification

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Synthetic Raman spectra from a small-cohort generative model improve glioma classifiers when mixed with real data, even though pure synthetic training falls short.

desk verdict Careful small-cohort Raman study that honestly shows a large TS/TR domain gap and only tiny, mostly non-significant TSR/TR gains; useful protocol, overstated "consistent" claim. read the letter →

arxiv 2607.10196 v1 pith:LZH4FSMB submitted 2026-07-11 cs.LG

Andrei Iuşan , Iulian Vasile , Daria Voiculescu , Ion Petre , Andrei Păun , Bogdan Oancea , Mihaela Păun This is my paper

classification cs.LG

keywords RamanspectroscopygliomaclassificationconditionalVAEsyntheticdataaugmentationIDHstatusmethylationsubtypesdomaingappatient-isolatedcross-validation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Biomedical Raman datasets for glioma are small, imbalanced, and noisy, so standard classifiers overfit. This paper trains a class-conditioned variational autoencoder on spectra from 58 patient biopsies and shows that the synthetic spectra it produces cannot fully replace real measurements: models trained only on synthetics lag the real-data baseline. When the same synthetics are mixed into the real training set, however, accuracy rises consistently across several classifiers under strict patient-isolated cross-validation. The authors argue that the generator has captured enough spectral structure to act as a regularizer rather than a full substitute, and they also introduce a reconstruction-error classifier that works without a separate discriminative model. The practical message is that deep generative augmentation remains useful even when independent patient samples are scarce.

What carries the argument

The beta-CVAE: a 1-D convolutional encoder-decoder whose decoder is conditioned on class labels, trained with a multi-task loss (reconstruction + weighted KL + auxiliary classification) so that sampling from the latent prior yields class-balanced synthetic Raman spectra used for TSR/TR augmentation and for Classification-by-Reconstruction.

What would settle it

Re-run the identical patient-isolated protocol after replacing the manual tumor-spectrum selection with a fully automated, histopathology-validated mask; if the TSR/TR accuracy gains disappear while TR/TR stays the same, the claimed regularization benefit is an artifact of selection bias.

Watch

Extended reading notes

Core claim

Even with only 58 independent glioma biopsies, a beta-regularized conditional variational autoencoder can learn enough class-conditioned spectral structure that adding its synthetic outputs to real training data improves downstream IDH-status and methylation-subtype classification under patient-isolated evaluation, while training on synthetic data alone remains inferior because of a clear domain gap.

Load-bearing premise

The heuristic, visually guided pipeline that picks which spectra count as tumor (intensity threshold plus per-slide clustering and manual selection) produces an unbiased set of spectra whose class distributions the generator can safely learn.

Editorial extensions

If this is right

Synthetic Raman spectra can be used as a regularizer rather than a full replacement for scarce real biopsies.
Strict patient-isolated evaluation remains mandatory; spectrum-level splits would inflate the apparent gains.
Reconstruction-error classification offers a generative alternative when discriminative boundaries are unstable.
The same pipeline can be tried on other small-cohort Raman tasks (e.g., other solid tumors) where multi-class biochemical differences are subtle.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The observed domain gap (synthetic spectra hug the class mean and under-represent real variance) suggests that future generators may need explicit diversity or noise-matching terms before pure TS/TR becomes viable.
If the upper bound on accuracy is set by class overlap in the original spectral space, generative augmentation may hit a ceiling that only larger multi-site cohorts can raise.
CbR’s comparable performance to TS/TR hints that latent reconstruction error could serve as a built-in uncertainty or out-of-distribution detector for clinical Raman systems.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The punchline is simple: they train a β-CVAE on the Lita 58-slide glioma Raman set, document a clear domain gap when training only on synthetic spectra, and then report modest accuracy lifts when the synthetics are mixed back with real data under strict slide-isolated CV. That is the whole result.

What is new is the combination, not the idea of generative Raman augmentation (Zhao already did cGANs for skin). They add a multi-task encoder, a 6-class methylation task, Classification-by-Reconstruction, and unusually careful patient isolation plus slide-level bootstrap CIs. The experimental hygiene is better than most ML-for-spectroscopy papers: multiple classifiers, loss/latent/ratio ablations, and they themselves note that most TSR/TR deltas are small and that no multiple-testing correction was applied. The PCA/t-SNE and mean±σ plots make the domain gap visible rather than hidden. That honesty is real credit.

The soft spots are proportional, not fatal. The adverb "consistently" in the abstract and conclusion is stronger than Tables 2 and A.1–A.11 support: RF/XGBoost gains are typically +0.00 to +0.02 with CIs that include zero; only the 1D-CNN sometimes moves more. The domain gap is an order of magnitude larger than the claimed regularization benefit, so residual bias from the heuristic intensity-threshold + NMF-GMM + visual cluster selection pipeline could easily erase the tiny positive deltas. No code or data release. Hyper-parameters are hand-chosen. None of this makes the work circular or incoherent; it just means the central claim is still provisional.

This is for people who actually run Raman classifiers on tens of patients and need a realistic sense of what generative augmentation can and cannot do. It is not a methods breakthrough and not a diagnostic advance. I would send it to peer review: the protocol is solid enough and the reporting is candid enough that referees can force the language to match the statistics. I would not cite it myself unless I were writing a methods comparison on the same cohort, but I would bring the figures to a reading group as a clean example of how to measure domain gap honestly.

Referee Report

3 major / 5 minor

Summary. The manuscript studies generative data augmentation for Raman-based glioma classification on a small cohort of 58 biopsy slides. A 1D-convolutional β-CVAE with an auxiliary classification head is trained under slide-isolated 5-fold CV to produce class-conditioned synthetic spectra. These are evaluated in TS/TR and TSR/TR regimes for binary IDH-status and 6-class methylation subtype tasks, using RF+SVM, RF, XGBoost and a 1D-CNN, plus a reconstruction-error inference method (CbR). Pure synthetic training underperforms the real baseline (domain gap), while mixing synthetic with real data yields small positive accuracy deltas for most models. The authors conclude that generative models can still capture enough structure for useful regularization even with limited independent patient samples.

Significance. Data scarcity and acquisition variability are genuine bottlenecks for Raman spectroscopy ML in oncology. The work is carefully designed: strict slide-isolated folds, bootstrap CIs at the slide level, multiple classifiers, and ablations of reconstruction loss, latent dimension and real:synthetic ratio. The explicit documentation of the domain gap and the modest size of the TSR/TR gains is scientifically honest and useful for the community. If the regularization claim holds under tighter statistical control and a less heuristic tumor-spectrum selection pipeline, the paper would provide a practical, reproducible template for generative augmentation in other small-cohort biomedical spectroscopy settings. The CbR idea is a secondary but interesting generative-inference baseline.

major comments (3)

The central claim that synthetic spectra “consistently improve” classification (Abstract, §5.1, §5.3, Conclusion) is not secured by the reported statistics. Table 2 and Tables A.1–A.11 show that most TSR/TR Δs for RF/XGBoost are +0.00–+0.02 with 95 % CIs that include zero; only occasional larger gains appear for 1D-CNN. The authors themselves note that improvements are “mostly small and not statistically significant” and that no multiple-testing correction was applied (§5.3). With dozens of model×loss×latent×ratio comparisons, uncorrected bootstrap p-values cannot support the adverb “consistently.” Either apply a correction (or a pre-specified primary endpoint) or rephrase the claim to match the evidence (e.g., “small positive trends that rarely reach significance”).
The load-bearing assumption that the heuristic spatio-spectral pipeline (§3.2, intensity threshold 0.6 + 3-component GMM on NMF coefficients + manual cluster selection per slide, Figs. 2–3) yields an unbiased set of “tumor” spectra is not validated. Retention rates vary widely across slides (Fig. 3). Any class-dependent selection bias would affect both the real baseline and the CVAE training distribution and could reverse the tiny positive TSR/TR deltas. At minimum, report inter-annotator agreement or a sensitivity analysis (different thresholds / automatic cluster selection) showing that the claimed regularization benefit is stable.
The domain gap (TS/TR Δ ≈ −0.06 to −0.26) is an order of magnitude larger than the claimed regularization benefit. PCA/t-SNE and mean±σ plots (Figs. 5–6) correctly show that synthetic spectra have far lower variance and form tighter class clusters than real data. Without a quantitative fidelity/diversity metric (e.g., MMD, coverage of real spectral modes) it remains unclear whether the small TSR/TR gains arise from genuine regularization or from mild label-smoothing / class-balancing effects that could be obtained by simpler means. A short comparison against classical augmentation (noise, SMOTE, mixup alone) would clarify the added value of the β-CVAE.

minor comments (5)

Hyper-parameter choices (β=2, α=0.2, γ, d=32, max 500 spectra/slide, 40k/120k synthetic samples) are stated but not systematically justified beyond small ablations; a short sensitivity table or fixed primary configuration would improve reproducibility.
Appendix B notes that four slides appear in two test folds; the bootstrap procedure should explicitly account for this repeated use so that uncertainty is not understated.
Notation for the reconstruction loss mixes γ and an undefined “delta” in the appendix table captions; unify with Eq. (2).
CbR is introduced as model-independent yet its accuracy (≈62 % binary) is only compared narratively to TS/TR; a direct head-to-head table would help.
Minor typos and formatting: “Ius,an”, “P˘aun”, occasional missing spaces around citations, and the arXiv date line.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: purely empirical ML evaluation; performance deltas are measured on held-out real slides, not forced by construction or self-citation.

full rationale

The paper reports an empirical study of β-CVAE synthetic augmentation for Raman glioma spectra under patient-isolated CV. All load-bearing claims (TS/TR domain gap, modest TSR/TR gains, CbR accuracy) are obtained by training generators/classifiers on training folds and scoring on held-out real test slides; none reduce by definition to a fitted parameter or to an input equation. Hyper-parameters (β=2, α=0.2, γ, d=32, max-spectra=500, mixup/shift) are chosen by hand or small ablations and then frozen; the reported accuracies and bootstrap Δs are independent measurements, not tautological. The sole external protocol citation (Lita et al. for the dataset and slide-isolated splits) is non-overlapping authorship and supplies data/protocol, not a uniqueness theorem or ansatz that forces the present results. No self-definitional loops, fitted-input-as-prediction, or renaming of known results appear. The work is therefore self-contained against its own held-out benchmarks; any weakness lies in statistical power/overclaiming of “consistent” gains, not circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The central claim rests on standard VAE mathematics, the domain premise that Raman spectra carry IDH/methylation signal, a suite of hand-chosen hyper-parameters, and one invented inference procedure (CbR). No new physical entities are postulated; the free parameters are the usual ML knobs that control generation quality and therefore the size of the reported gains.

free parameters (7)

β (KL weight) = 2
Set to 2 following Higgins et al.; controls latent regularization and therefore synthetic diversity.
α (auxiliary classification weight) = 0.2
Empirically set to 0.2 after observing overfitting at 0.5; directly affects how class-discriminative the encoder becomes.
γ (MSE vs cosine reconstruction mix) = varied in ablations
Ablated over {0, 0.5, 1}; changes the geometry of the reconstruction loss and the resulting domain gap.
latent dimension d = 32 (default)
Tested in {16, 32, 64}; most runs use 32; capacity of the generative model.
max spectra per slide = 500
Hard cap of 500 randomly drawn spectra per slide to equalize slide contribution; arbitrary but load-bearing for training balance.
intensity threshold for signal filtering = 0.6
Fixed at 0.6 on max intensity; determines which pixels enter the GMM stage.
synthetic sample counts (40k / 120k) = 40 000 (binary), 120 000 (6-class)
Number of generated spectra used in TS/TR and TSR/TR; ratio ablations also vary the real:synth mix.

assumptions (4)

domain assumption Raman spectral differences between IDH-mutant and IDH-wildtype (and among the six methylation subtypes) are large enough to be learnable above acquisition noise and fluorescence background.
Implicit throughout; without it the classification tasks and the utility of any generative model are undefined.
standard math The β-VAE ELBO with the stated multi-task loss yields a latent space from which class-conditional samples can be drawn by z ~ N(0,I).
Standard conditional VAE theory (Higgins et al., Kingma & Welling) invoked in §3.3.
ad hoc to paper Slide-level stratified 5-fold splits with manual balancing (each class ≥15 % in every test fold) prevent patient-level leakage while remaining statistically valid.
Appendix B; four slides appear in two test folds, so the design is not a pure partition.
ad hoc to paper Visual inspection of GMM clusters after NMF correctly identifies tumor versus background/low-signal spectra without introducing class-dependent selection bias.
§3.2; the only quality-control step for the ~200 k retained spectra.

invented entities (1)

Classification by Reconstruction (CbR)
purpose: Inference procedure that assigns a test spectrum the class label minimizing reconstruction MSE under that class condition.
Introduced in §3.4 as an alternative to discriminative classifiers; performance is reported but no external validation of the procedure exists outside this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Generative Augmentation of Raman Spectra for Glioma Classification." pith.science (2026). https://pith.science/paper/LZH4FSMB

@misc{pith2026260710196,
  author       = {Pith},
  title        = {Pith review of: Generative Augmentation of Raman Spectra for Glioma Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LZH4FSMB}},
  note         = {Machine review of arXiv:2607.10196}
}

abstract

Access to sufficiently large biomedical datasets remains a major obstacle for machine learning in Raman spectroscopy-based diagnostics. In particular, for glioma analysis, datasets are typically small and heterogeneous, affected by acquisition-specific variability. This work investigates the utility of deep generative augmentation in such a small-cohort setting. We analyze glioma biopsy spectra acquired from 58 tumor samples and consider both binary IDH-status classification and 6-class methylation subtype classification problems. To address the limited size and imbalance of the dataset, we develop a conditional variational autoencoder ($\beta$-CVAE) capable of generating class-conditioned synthetic Raman spectra. The generated data are evaluated in Train-on-Synthetic, Test-on-Real (TS/TR) and Train-on-Synthetic+Real, Test-on-Real (TSR/TR) settings under a strict patient-isolated cross-validation protocol. Models trained exclusively on synthetic data underperform models trained on real spectra, indicating a substantial domain gap between synthetic and real distributions. However, augmenting the real training data with synthetic spectra consistently improves classification performance across multiple models. These findings indicate that, even with a limited number of independent patient samples, generative models can capture sufficient structure to provide useful regularization for downstream classifiers. We also investigate a reconstruction-based inference strategy, termed Classification by Reconstruction (CbR), in which class prediction is based on reconstruction error under different class conditions. Overall, the results support the use of deep generative augmentation as a practical strategy for improving machine learning robustness in Raman spectroscopy applications characterized by limited biomedical datasets.

Figures

Figures reproduced from arXiv: 2607.10196 by the authors.

**Figure 1.** (a) False coloring of a slide, with Red = Phenylalanine marker (990–1020 cm−1 ), Green = Lipids (2800–3000 cm−1 ), Blue = CH2 deformations (1430–1460 cm−1 ). (b) Zoomed region for 3 × 3 median kernel visualization. (c) Raw vs. Median filtered spectra. 1. Median filtering. A 3 × 3 median kernel was applied on the spatial dimensions (H, W) of each slide for denoising purposes, filtering out cosmic radiation and Gaussi… view at source ↗

**Figure 2.** Spectral clustering of two slides (top, middle). (a) Slide 1 cluster assignment; (b) Slide 1 Mean spectra per cluster: blue = weak signal cluster, discarded, red = medium intensity signal, discarded, green = strong signal, selected as tumor; (c) Slide 2 cluster assignment; (d) Slide 2 Mean spectra per cluster: red, green = strong tissue signal, selected as tumor, blue = weak signal, discarded; (e, f) Histograms of m… view at source ↗

**Figure 3.** Histogram of retention rates. This approach ensured that each slide was processed independently and visually inspected to verify cluster consistency. The resulting binary tumor labeling yields approximately 200,000 spectra retained for downstream analysis from the original ∼270,000. We draw attention to the fact that the percentage of retained spectra from each slide varied significantly (see [PITH_FULL_IMAGE:figu… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Architecture of the 1D-Convolutional Conditional Variational Autoencoder (β-CVAE). The encoder is composed of 1-d convolutional layers with 3 blocks: filters [32, 64, 128], kernels [11, 5, 3], and strides [2, 2, 2], compressing the input spectrum to the ddimensional l…

**Figure 5.** Figure 5: PCA (top) and t−SNE (bottom) embeddings of real and synthetic spectra, coloured by source (left) and class (right). Taken together, the PCA and t-SNE plots offer insight into the spatial structure of the spectral data. The PCA plots reveal a variance difference betwee…

**Figure 6.** Figure 6: Mean ±σ Raman spectra, real vs. synthetic, per class. An important observation is that useful augmentation effects were obtained despite the relatively small number of independent patient samples available for training the generative model. In contrast to conventional…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 9 canonical work pages

[1]

James J. P. Alix et al. “Non-negative matrix factorisation of Raman spectra finds common patterns relating to neuromuscular disease across differing equipment configurations, preclinical models and human tis- sue”. In:Journal of Raman Spectroscopy54.3 (Dec. 2022), pp. 258–268. issn: 1097-4555.doi:10.1002/jrs.6480

work page doi:10.1002/jrs.6480 2022
[2]

Balancing reconstruction error and Kullback-Leibler divergence in Variational Autoencoders

Andrea Asperti and Matteo Trentin.Balancing reconstruction error and Kullback-Leibler divergence in Variational Autoencoders. 2020.doi: 10.48550/ARXIV.2002.07514

work page Pith review arXiv doi:10.48550/arxiv.2002.07514 2020
[3]

Digital medicine and the curse of dimensionality

Visar Berisha et al. “Digital medicine and the curse of dimensionality”. In:npj Digital Medicine4.1 (2021), p. 153.doi:10.1038/s41746-021- 00521-5

doi:10.1038/s41746-021- 2021
[4]

Raman based molecular imaging and analytics: a magic bullet for biomedical applications?

Thomas Bocklitz et al. “Raman based molecular imaging and analytics: a magic bullet for biomedical applications?” In:Analyst141.2 (2016), pp. 505–515.doi:10.1039/C5AN01806G. 22

work page doi:10.1039/c5an01806g 2016
[5]

Maximum Likelihood from Incomplete Data Via the ¡i¿EM¡/i¿ Algorithm

A. P. Dempster, N. M. Laird, and D. B. Rubin. “Maximum Likelihood from Incomplete Data Via the ¡i¿EM¡/i¿ Algorithm”. In:Journal of the Royal Statistical Society Series B: Statistical Methodology39.1 (Sept. 1977), pp. 1–22.issn: 1467-9868.doi:10.1111/j.2517-6161.1977. tb01600.x

doi:10.1111/j.2517-6161.1977 1977
[6]

Real- valued (medical) time series generation with recurrent conditional gans

Crist´ obal Esteban, Stephanie L Hyland, and Gunnar R¨ atsch. “Real- valued (medical) time series generation with recurrent conditional gans”. In:arXiv preprint arXiv:1706.02633(2017)

arXiv 2017
[7]

Raman Spectroscopy Spectral Fingerprints of Biomarkers of Traumatic Brain Injury

Georgia Harris et al. “Raman Spectroscopy Spectral Fingerprints of Biomarkers of Traumatic Brain Injury”. In:Cells12.22 (Nov. 2023), p. 2589.issn: 2073-4409.doi:10.3390/cells12222589.url:http: //dx.doi.org/10.3390/cells12222589

work page doi:10.3390/cells12222589.url:http: 2023
[8]

beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework

Irina Higgins et al. “beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework”. In:International Conference on Learning Representations. 2017

2017

Show all 17 references

[9]

Synthetic Data – what, why and how?

James Jordon et al. “Synthetic Data – what, why and how?” In: (2022). doi:10.48550/ARXIV.2205.03257

2022 doi arXiv
[10]

Application of non-negative matrix factorization in oncology

H. Kato et al. “Application of non-negative matrix factorization in oncology”. In:Precision Cancer Medicine5 (2022)

2022
[11]

Biomedical ap- plications of Raman and infrared spectroscopy to diagnose tissues

Christoph Krafft, Gerwin Steiner, and Reiner Salzer. “Biomedical ap- plications of Raman and infrared spectroscopy to diagnose tissues”. In: Journal of Physics D: Applied Physics42.18 (2009), p. 183001.doi: 10.1088/0022-3727/42/18/183001

2009 doi
[12]

Glioblastoma – A Contemporary Overview of Epidemiology, Classification, Pathogenesis, Diagnosis, and Treatment: A Review Article

K. Kr´ olikowska et al. “Glioblastoma – A Contemporary Overview of Epidemiology, Classification, Pathogenesis, Diagnosis, and Treatment: A Review Article”. In:International Journal of Molecular Sciences 26.24 (2025).doi:10.3390/ijms262412162

2025 doi
[13]

Learning the parts of objects by non-negative matrix factorization

Daniel D. Lee and H. Sebastian Seung. “Learning the parts of objects by non-negative matrix factorization”. In:Nature401.6755 (Oct. 1999), pp. 788–791.issn: 1476-4687.doi:10.1038/44565

1999 doi
[14]

Raman-based machine learning platform reveals unique metabolic differences between IDHmut and IDHwt glioma

Adrian Lita et al. “Raman-based machine learning platform reveals unique metabolic differences between IDHmut and IDHwt glioma”. In: Neuro-oncology26 (June 2024).doi:10.1093/neuonc/noae101. 23

2024 doi
[15]

Raman Spectroscopy in the Diagnosis of Brain Gliomas: A Literature Review

E. V. Stupak et al. “Raman Spectroscopy in the Diagnosis of Brain Gliomas: A Literature Review”. In:Cureus17.2 (2025).doi:10.7759/ cureus.79165

2025
[16]

Overfitting in prediction models – is it a problem only in high dimensions?

Jyothi Subramanian and Richard Simon. “Overfitting in prediction models – is it a problem only in high dimensions?” In:Contempo- rary Clinical Trials36.2 (2013), pp. 636–641.issn: 1551-7144.doi: https://doi.org/10.1016/j.cct.2013.06.011

2013 doi
[17]

Improving skin cancer detection by Raman spectroscopy using convolutional neural networks and data augmentation

J. Zhao et al. “Improving skin cancer detection by Raman spectroscopy using convolutional neural networks and data augmentation”. In:Fron- tiers in Oncology14 (2024), p. 1320220.doi:10.3389/fonc.2024. 1320220. 24 A Extended Experiment Results This appendix reports the per-expe...

2024 doi

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.

[1] [1]

James J. P. Alix et al. “Non-negative matrix factorisation of Raman spectra finds common patterns relating to neuromuscular disease across differing equipment configurations, preclinical models and human tis- sue”. In:Journal of Raman Spectroscopy54.3 (Dec. 2022), pp. 258–268. issn: 1097-4555.doi:10.1002/jrs.6480

work page doi:10.1002/jrs.6480 2022

[2] [2]

Balancing reconstruction error and Kullback-Leibler divergence in Variational Autoencoders

Andrea Asperti and Matteo Trentin.Balancing reconstruction error and Kullback-Leibler divergence in Variational Autoencoders. 2020.doi: 10.48550/ARXIV.2002.07514

work page Pith review arXiv doi:10.48550/arxiv.2002.07514 2020

[3] [3]

Digital medicine and the curse of dimensionality

Visar Berisha et al. “Digital medicine and the curse of dimensionality”. In:npj Digital Medicine4.1 (2021), p. 153.doi:10.1038/s41746-021- 00521-5

doi:10.1038/s41746-021- 2021

[4] [4]

Raman based molecular imaging and analytics: a magic bullet for biomedical applications?

Thomas Bocklitz et al. “Raman based molecular imaging and analytics: a magic bullet for biomedical applications?” In:Analyst141.2 (2016), pp. 505–515.doi:10.1039/C5AN01806G. 22

work page doi:10.1039/c5an01806g 2016

[5] [5]

Maximum Likelihood from Incomplete Data Via the ¡i¿EM¡/i¿ Algorithm

A. P. Dempster, N. M. Laird, and D. B. Rubin. “Maximum Likelihood from Incomplete Data Via the ¡i¿EM¡/i¿ Algorithm”. In:Journal of the Royal Statistical Society Series B: Statistical Methodology39.1 (Sept. 1977), pp. 1–22.issn: 1467-9868.doi:10.1111/j.2517-6161.1977. tb01600.x

doi:10.1111/j.2517-6161.1977 1977

[6] [6]

Real- valued (medical) time series generation with recurrent conditional gans

Crist´ obal Esteban, Stephanie L Hyland, and Gunnar R¨ atsch. “Real- valued (medical) time series generation with recurrent conditional gans”. In:arXiv preprint arXiv:1706.02633(2017)

arXiv 2017

[7] [7]

Raman Spectroscopy Spectral Fingerprints of Biomarkers of Traumatic Brain Injury

Georgia Harris et al. “Raman Spectroscopy Spectral Fingerprints of Biomarkers of Traumatic Brain Injury”. In:Cells12.22 (Nov. 2023), p. 2589.issn: 2073-4409.doi:10.3390/cells12222589.url:http: //dx.doi.org/10.3390/cells12222589

work page doi:10.3390/cells12222589.url:http: 2023

[8] [8]

beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework

Irina Higgins et al. “beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework”. In:International Conference on Learning Representations. 2017

2017