Raman hyperspectral microscopy is a valuable tool in biological and biomedical imaging. Because Raman scattering is often weak in comparison to other phenomena, prevalent spectral fluctuations and contaminations have brought advancements in analytical and chemometric methods for Raman spectra. These chemometric advances have been key contributors to the applicability of Raman imaging to biological systems. As studies increase in scale, spectral contamination from extrinsic background, intensity from sources such as the optical components that are extrinsic to the sample of interest, has become an emerging issue. Although existing baseline correction schemes often reduce intrinsic background such as autofluorescence originating from the sample of interest, extrinsic background is not explicitly considered, and these methods often fail to reduce its effects. Here we show that extrinsic background can significantly affect a classification model using Raman images, yielding misleadingly high accuracies in the distinction of benign and malignant samples of follicular thyroid cell lines. To mitigate its effects, we develop extrinsic background correction (EBC) and demonstrate its use in combination with existing methods on Raman hyperspectral images. EBC isolates regions containing the smallest amounts of sample materials that retain extrinsic contributions that are specific to the device or environment. We perform classification both with and without the use of EBC, and we find that EBC retains biological characteristics in the spectra while significantly reducing extrinsic background. We also address its possible generalization for inhomogeneous illumination profile.

Spectral decomposition, a pivotal process in hyperspectral imaging, involves separating mixed signals into their constituent parts, known as endmembers, to extract meaningful information. The Bayesian Information Criterion, a statistical metric derived from Bayesian probability theory, serves as a valuable tool for model selection in spectral decomposition reducing the risk of overfitting and enhancing the robustness of the unmixing analysis.

In this work we utilise BIC in spectral decomposition through fitting models with varying numbers of endmembers and assessing the trade-off between model complexity and data fidelity, allowing the selection of the most parsimonious representation that best captures the underlying structure of the spectral data. This methodology results is a more refined and interpretable spectral decomposition, aiding in molecular interpretation of data science models in chemical imaging.

Sample size planning (SSP) is crucial for experimental planning but is not well-established for spectroscopic and image data, especially in combination with deep learning. The existing approaches are typically quite complex for routine use in experimental planning. To make the existing approaches more accessible, we developed web-based tools for the existing approaches. Besides, we extended the approach to imaging data and deep learning by introducing transfer learning in the SSP pipeline.

ACKNOWLEDGMENT: Financial support from the EU, the TMWWDG, the TAB, the BMBF, the DFG, the Carl-Zeiss Foundation, and the Leibniz Association is greatly acknowledged. This work is supported by the BMBF, funding program Photonics Research Germany (LPI-BT3-IPHT, FKZ: 13N15708) and is integrated into the Leibniz Center for Photonics in Infection Research (LPI). The LPI initiated by Leibniz-IPHT, Leibniz-HKI, Friedrich Schiller University Jena, and Jena University Hospital is part of the BMBF national roadmap for research infrastructures.

Transfer learning is an important technique to improve the model generalizability, which is crucial to apply Raman spectroscopy in biological applications considering the substantial spectral variations between batches. We systematically investigated the limit of existent model transfer methods and explored the possibilities of deep learning-based approaches in cases of limited sample size.

The gold standard in histology is to use chemical stains or genetic modified tissue, where some internal structures emit a fluorescent signal. These methods require trained staff and several hours or days of preparation. Virtual staining employs trained neural networks to take over the staining process. Based on an unlabeled microscopic images the network can predict the corresponding fluorescent image for DAPI and Phalloidin488 staining, enabling studies on cell nuclei and the cytoskeleton.

Neural networks usually need a huge amount of training data, so the possibilities of transfer learning for a reduction of the dataset size were investigated. In addition, we also present first studies the interpretability of the trained network to find ideal image acquisition techniques and optimize the training.

Raman imaging has become a vital tool for studying biological processes thanks to its label-free and non-invasive nature. However, photodamage has been a long-term concern in Raman imaging which requires large number of photons to produce enough contrast due to its inherent weak scattering efficiency. In this paper, we proposed to optimize the instrument slit-width and leverage deep learning approach to accelerate the Raman imaging, and thereby reduce the photodamage. Experiment results have shown that the collaborative effort yields Raman image with high SNR, SR and SpR, whose quality is comparable to the image obtained using narrow slits, small scan steps, and long integration times, while an 80-fold improved imaging speed allows the photodamage to be reduced greatly. Subsequently, we have been successfully employed it to observe the dynamic changes of cytochrome c in a single cell during the apoptotic before reaching the photodamage limit. It is a new endeavor in the study of cell dynamics and provides a reliable tool for more observation of other biochemical processes.

Optical microscopy is widely applied to the investigation of biomedical samples, and a variety of image processing approaches have been established to reduce artifacts generated by the measurement process. However, a standardized and reliable method for assessing image quality is still lacking. Our study contributes to the investigation of image evaluation methods for fluorescence microscopy. We present a set of no-reference metrics that can be used for the characterization of experimental artifacts. In addition, our method is incorporated into a machine learning approach for automatic classification of single artifacts. The metrics identify reliable markers for single artifacts in fluorescence microscopy measurements, can be easily interpreted, and allow the selection of the best image based on specific quality requirements. Our study provides a simple evaluation tool for optical microscopy that can also be extended to the different stages of the processing pipeline.

In this study, we propose a novel approach for enhancing the resolution of quantitative phase images using Generative Adversarial Networks (GANs).

Photonic techniques are optimal tools to characterise samples in various research disciplines such as remote sensing, materials characterisation, life sciences and medicine. To exploit the full potential of these techniques, the entire data life cycle of photonic data needs to be investigated and optimised. The photonic data lifecycle starts with data generation and planning of the corresponding study/experiment, followed by data modelling using artificial intelligence (AI) techniques such as chemometrics, machine learning (ML) and deep learning (DL), and it ends by data storage and archiving.

In this contribution, we will present our studies aimed at the generation of correction procedures and inverse modelling tools for photonic data and heir measurement processes using data science methods. We will also present our research activities towards a repository for sharing vibrational spectroscopic data (VibSpecDB), which is embedded in the National Research Data Infrastructure Initiative in Germany (NFDI) and its chemistry consortium (NFDI4Chem).

Acknowledgements

This work is supported by the BMBF, funding program Photonics Research Germany (13N15466 (LPI-BT1-FSU), 13N15710 (LPI-BT3-FSU), 13N15708 (LPI-BT3-IPHT)) and is integrated into the Leibniz Center for Photonics in Infection Research (LPI). The LPI initiated by Leibniz-IPHT, Leibniz-HKI, Friedrich Schiller University Jena and Jena University Hospital is part of the BMBF national roadmap for research infrastructures. Parts are funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 441958208 (NFDI4Chem).

Deep learning models are widely used because of their high accuracy in solving classification problems in spectroscopy, but they lack interpretability. The challenge lies in the balance between interpretability and accuracy. Current interpretive methods, such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), can sometimes provide mathematical meaning but physically implausible interpretations by perturbing individual feature values. To address this gap, our research proposes a group-focused methodology that targets 'spectral zones' to estimate the impact of collective spectral features directly. This approach enhances the interpretability of deep learning models, diminishes noisy data, and provides a more comprehensive understanding of model behaviors. By applying group perturbations, the resultant interpretations are not only more intuitive but also offer results that are easily comparable with domain expertise, thus leading to an enriched analysis of the model's decision-making processes.