1. Introduction

2. Materials and Methods

2.2. Multi-Model Light Sheet Microscope Design

The Multi-modal Raman Light Sheet Microscope is based on the OpenSPIM platform[3], with significant enhancements to incorporate multi-modal imaging capabilities. The optical system consists of three primary components: the beam-shaping and illumination optics, the spectral selection and imaging optics, and the precision sample positioning system(ref. Figure 2).

• Excitation Lasers: Two continuous-wave lasers, with emission wavelengths of 660 nm and 785 nm, are used to excite Rayleigh and Raman scattering, respectively. The 660 nm laser is optimized for fluorescence imaging while minimizing autofluorescence, and the 785 nm laser enhances Raman scattering signals. Adjustable power outputs (0.5–130 mW for 660 nm and 0.5–200 mW for 785 nm) allow fine control of illumination intensity. Both laser beams are collimated and aligned coaxially using a series of broadband mirrors and dichroic splitters.
• Beam Shaping: The laser beams are expanded using a Keplerian telescope system formed by achromatic doublets, which increase the illuminated field of view without compromising beam focus. A cylindrical lens focuses the expanded beam into a static light sheet, projected into the sample chamber through a 10x water immersion objective.
• Imaging Optics: Photons scattered and emitted by the sample are collected by a 20x water immersion detection objective, positioned orthogonally to the light sheet for optimal detection. An Acousto-Optic Tunable Filter (AOTF) enables precise spectral selection with a 2 nm bandwidth, allowing for fine control over the wavelengths collected. Additional long-pass and short-pass filters further refine the detected signal. A high-sensitivity sCMOS camera (Hamamatsu ORCA Flash 4.0 LT+) is used for image acquisition, operating in a 1024x1024 pixel mode optimized for low-light conditions.

Table 1. Optical and Mechanical Components Used in the Raman Light Sheet Microscope.

Table 1. Optical and Mechanical Components Used in the Raman Light Sheet Microscope.

No.	Component Specification	Manufaturer
1	LuxX Laser 785 nm, adjustable laser power 0.5–200 mW	Omicron GmbH
2	LuxX Laser 660 nm, adjustable laser power 0.5–130 mW	Omicron GmbH
3,5,8	Broadband mirror, Ø25.4 mm, EO2 coated, mounted in Polaris K1 Kinematic Mirror Mount	Thorlabs GmbH
4,6	Broadband mirror, Ø25.4 mm, EO3 coated, mounted in Polaris K1 Kinematic Mirror Mount	Thorlabs GmbH
7	BrightLine laser dichroic beamsplitter, 25 mm x 36 mm, reflection band 350–671 nm, transmission band 702–1200 nm	Semrock
9	Mounted achromatic doublet lens, Ø12.7 mm, focal length 25 mm, anti-reflex coating 400–1100 nm	Thorlabs GmbH
10	Mounted achromatic doublet lens, Ø12.7 mm, focal length 50 mm, anti-reflex coating 400–1100 nm	Thorlabs GmbH
11	Mounted cylindrical achromatic doublet lens, Ø25.4 mm, focal length 50 mm, anti-reflex coating 650–1050 nm	Thorlabs GmbH
12	UMPLFLN10XW water dipping objective, magnification 10x, numerical aperture 0.3, working distance 3.5 mm	Evident
13	UMPLFLN20XW water dipping objective, magnification 20x, numerical aperture 0.5, working distance 3.5 mm	Evident
14	Acousto-Optic Tunable Filter (AOTF), spectral range 550–1000 nm	Brimrose
15	Polarization filter	Thorlabs GmbH
16	6-position motorized filter wheel	Thorlabs GmbH
17	Longpass filter, 660 nm	Semrock
18	Notch filter, 660 nm	Semrock
19	Shortpass filter, 660 nm	Semrock
20	Longpass filter, 785 nm	Semrock
21a	Tube lens U-TLU and C-mount (U-TV0.5XC-3)	Evident
21b	Aspheric condenser lens, Ø25 mm, focal length 20 mm, anti-reflex coating 650–1050 nm	Thorlabs GmbH
22a	sCMOS camera ORCA Flash 4.0 LT+	Hamamatsu
22b	CXY1 two-axis translating lens mount, Ø550 µm optic fiber	Thorlabs GmbH
23	USB-4D stage (X, Y, Z, R)	Picard-Industries
24	Sample chamber, aluminum mounting frame, acrylic water chamber	CeMOS
25	MultiSpec®Raman spectrometer	tec5 GmbH

2.4. Zero-Shot Deconvolution Network (ZS-DeconvNet)

ZS-DeconvNet is an advanced machine learning-based denoising algorithm designed to enhance image quality by reducing noise and preserving fine structural details. Unlike traditional supervised deep learning methods, ZS-DeconvNet operates in an unsupervised manner, requiring no ground-truth data or pre-trained models [2,16].

Mathematical Model of ZS-DeconvNet

The network minimizes noise in the acquired images using a self-supervised learning framework. The objective function for ZS-DeconvNet is formulated as:

$$L\left(\theta \right)={\left|\left|I_{raw}-I_{denoised}\left(\theta \right)\right|\right|}^2+\lambda \left|\left|\nabla I_{denoised}\left(\theta \right)\right|\right|$$

Where:

$I_{raw}$ represents the noisy input image.

$I_{denoised}\left(\theta \right)$ represents the output of the neural network after applying the network’s parameters $\left(\theta \right)$

$\lambda$ is a regularization term controlling the smoothness of the denoised image.

$\nabla I_{denoised}\left(\theta \right)$represents the image gradients, ensuring that edges are preserved during denoising.

The Zero-Shot Deconvolution Network (ZS-DeconvNet) was implemented to improve the quality of the Raman Light Sheet Microscopy images by reducing noise and enhancing resolution.

Network Architecture: ZS-DeconvNet

The ZS-DeconvNet is built on a CNN-based encoder-decoder architecture. The encoder compresses the input image into a lower-dimensional latent space through a sequence of Conv2D layers, each followed by batch normalization and max pooling. These layers are designed to extract critical structural features from the image while progressively reducing its dimensionality, enabling the model to focus on the most important information. The decoder mirrors this process, gradually reconstructing the image by applying upsampling and Conv2D layers. Additionally, skip connections between the encoder and decoder allow the model to retain high-resolution details by concatenating features from earlier layers. Finally, a sigmoid-activated Conv2D layer produces the denoised output.

During training, ZS-DeconvNet follows a zero-shot learning approach. Two corrupted versions of the same image are generated—Denoised Image A and Denoised Image B(ref. Figure 4) —by adding and subtracting noise, respectively. The model learns to map Denoised Image A (input) to Denoised Image B (target) without requiring a clean reference image. This is achieved using a Mean Squared Error (MSE) loss function, which minimizes the pixel-wise difference between the predicted and target images. By optimizing the MSE, the network progressively improves its ability to remove noise and restore details from noisy input data. Once the ZS-DeconvNet model is trained, it can be applied to new, unseen noisy images, resulting a denoised version of the input image by leveraging the learned features from the training process.

The ZS-DeconvNet is designed to generalize well to unseen data. The lack of a need for pre-trained datasets enhances its flexibility, allowing it to be applied in image enhancement tasks. This feature is particularly advantageous in applications such as cell imaging and video microscopy, where pre-trained datasets may not be available or applicable. (ref. Figure 4)

Pre-processing: Before being fed into the ZS-DeconvNet, the raw images(Original Image) undergo a series of crucial preprocessing steps aimed at preparing the data for optimal denoising and model training. In the initial step I, two corrupted image pair(Image A, Image B) were generated by introducing Gaussian noise to the original image, where Image A was with added noise and Image B was with inverted noise. After creating corrupted image pair from Original Image, median filtering was applied to these images to remove high-intensity “salt-and-pepper” noise, which is commonly seen in images captured through noisy channels, such as biomedical imaging. A median filter with a kernel size of 3 is applied, effectively suppressing noise while preserving edges and fine details within the image. This enhances the network’s ability to focus on meaningful structural elements during training and inference, contributing to more precise noise removal. (ref. Figure 4)

These pairs serve as input(Denoised Image A) and target (Denoised Image B) images during the training phase of ZS-DeconvNet. The model learns to predict Denoised Image B from Denoised Image A, simulating a noise-to-noise learning framework.

Post-processing: Following the ZS-DeconvNet denoising process, additional post-processing steps are performed, such as applying a region of interest (ROI) mask, applying Median Filter with Kernal size 3, and enhancing contrast to refine the output image. Here we also calculated PSNR, SSIM, RMSE, FRC metrics for performance evaluation[17,18]. These enhancements further improve the clarity and usability of the denoised image by focusing on key structures. (ref. Figure 4)

These operations help in emphasizing small structures within the spheroids, particularly in the cellular boundaries. If required, edge detection algorithms (such as Canny edge detection) can also be applied to highlight critical boundaries and structural details. This step is essential in the analysis of biological images, where the accurate delineation of subcellular components plays a crucial role in data interpretation.

3. Results

3.2. Quantitative Evaluation of Image Quality

The denoising capabilities of ZS-DeconvNet were quantitatively assessed using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Root Mean Square Error (RMSE), as shown in Figure 9. These metrics provide a comprehensive evaluation of how well the denoised images preserve structural integrity and reduce noise, enhancing the overall quality of the data for further analysis.

Figure 9. Denoising Performance of ZS-DeconvNet on 11B Spheroid Untreated Image using 660nm Laser for Rayleigh Spectrum: Original Image, Denoised Image after ZS-DeconvNet, FRC Analysis and denoising metrics(PSNR,SSIm and RMSE).

Figure 9. Denoising Performance of ZS-DeconvNet on 11B Spheroid Untreated Image using 660nm Laser for Rayleigh Spectrum: Original Image, Denoised Image after ZS-DeconvNet, FRC Analysis and denoising metrics(PSNR,SSIm and RMSE).

PSNR: The denoised image achieved a PSNR value of 16.74 dB, indicating a substantial reduction in noise when compared to the raw image. Although this PSNR value reflects moderate image fidelity, it represents a significant improvement in signal clarity, allowing for better visualization of features previously masked by noise. This improvement underscores the effectiveness of ZS-DeconvNet in restoring image quality in high-noise conditions such as Rayleigh scattering at 660 nm.

SSIM: The SSIM score of 0.168 indicates that some structural information was preserved post-denoising, although there is room for further optimization. Despite the relatively low score, this increase in structural similarity highlights ZS-DeconvNet’s ability to recover key features from the noisy input, enabling a more interpretable output. This is particularly relevant in imaging modalities where fine structural details are critical for accurate analysis, such as in the treated spheroids exposed to cisplatin.

RMSE: The RMSE value of 0.146 demonstrates the model’s efficiency in minimizing the error between the original noisy image and the denoised output. The reduction in RMSE confirms that ZS-DeconvNet effectively suppresses noise without introducing artifacts, thereby preserving essential structural and molecular details crucial for the accurate evaluation of treatment effects in cisplatin-treated spheroids.

In addition to these metrics, Fourier Ring Correlation (FRC) analysis (Figure 1) was conducted to further evaluate the resolution enhancement. The FRC curve shows improved frequency preservation across multiple shells, indicating that ZS-DeconvNet enhanced the image’s spatial resolution while maintaining relevant frequency details. This result highlights the model’s capability to improve image quality even in noisy and low-signal environments.

Overall, these quantitative metrics affirm the robustness of ZS-DeconvNet in effectively denoising images, particularly in the context of high-noise biomedical imaging applications, facilitating clearer feature extraction and more reliable data interpretation for both treated and untreated conditions.

4. Discussion

5. Conclusions

References

1. Krieger, M., et al. Multi-Modal Light Sheet Microscope for the Detection of Rayleigh, Raman, and Fluorescence Signals in 3D Cell Cultures. Micromachines 2023. [Google Scholar] [CrossRef]
2. Li, T., et al. Zero-shot learning enables instant denoising and super-resolution in real-time across multiple imaging modalities. Nature Communications 2024. [Google Scholar] [CrossRef]
3. OpenSPIM platform details from OpenSPIM.org.
4. Ravi, M.; Paramesh, V.; Kaviya, S.R.; Anuradha, E.; Solomon, F.D. 3D Cell Culture Systems: Advantages and Applications. J. Cell. Physiol. 2014, 230, 16–26. [Google Scholar] [CrossRef] [PubMed]
5. Park, Y.; Huh, K.; Kang, S.-W. Applications of Biomaterials in 3D Cell Culture and Contribution of 3D Cell Culture to Drug Development and Basic Biomedical Research. Int. J. Mol. Sci. 2021, 22, 2491. [Google Scholar] [CrossRef] [PubMed]
6. Koenigstein, J. Introduction to the Theory of the Raman Effect, 1st ed.; D. Reidel Publishing Company: Dordrecht-Holland, The Netherlands, 1972; pp. 79–133. [Google Scholar]
7. Eberhardt, K.; Stiebing, C.; Matthäus, C.; Schmitt, M.; Popp, J. Advantages and limitations of Raman spectroscopy for molecular diagnostics: An update. Expert Rev. Mol. Diagn. 2015, 15, 773–787. [Google Scholar] [CrossRef] [PubMed]
8. Oshima, Y.; Sato, H.; Kajiura-Kobayashi, H.; Kimura, T.; Naruse, K.; Nonaka, S. Light sheet-excited spontaneous Raman imaging of a living fish by optical sectioning in a wide field Raman microscope. Opt. Express 2012, 20, 16195–16204. [Google Scholar] [CrossRef]
9. Sage, D.; et al. DeconvolutionLab2: An open-source software for deconvolution microscopy. Methods 2017, 115, 28–41. [Google Scholar] [CrossRef] [PubMed]
10. Zhao, W.; et al. Sparse deconvolution improves the resolution of live-cell super-resolution fluorescence microscopy. Nat. Biotechnol. 2021, 40, 40–617. [Google Scholar] [CrossRef] [PubMed]
11. Wang, H.; et al. Deep learning enables cross-modality super-resolution in fluorescence microscopy. Nat.Methods 2019, 16, 103–110. [Google Scholar] [CrossRef] [PubMed]
12. Qiao, C.; et al. Evaluation and development of deep neural networksfor image super-resolution in optical microscopy. Nat. Methods 2021, 18, 194–202. [Google Scholar] [CrossRef] [PubMed]
13. Yanny, K., Monakhova, K., Shuai, R. W. &Waller, L. Deep learning for fast spatially varying deconvolution. Optica 2022, 9, 96–99. [Google Scholar] [CrossRef]
14. Shah, Z. H.; et al. Deep-learning based denoising and reconstruction of super-resolution structured illumination microscopy images. Photonics Res. 2021, 9, B168–B181. [Google Scholar] [CrossRef]
15. Lehtinen, J. et al. in Proceedings of the International Conference on Machine Learning 2965–2974 (2018).
16. Qiao, C. et al. Zero-shot learning enables instant denoising and super-resolution in optical fluorescence microscopy. ZS-Deconv- Net. 2024. [Google Scholar] [CrossRef]
17. Nieuwenhuizen, R. P.; et al. Measuring image resolution in optical nanoscopy. Nat. Methods 2013, 10, 557–562. [Google Scholar] [CrossRef] [PubMed]
18. Diekmann, R.; et al. Photon-free (s)CMOS camera characterization for artifact reduction in high- and super-resolutionmicroscopy. Nat. Commun. 2022, 13, 3362. [Google Scholar] [CrossRef] [PubMed]
19. Chen X, Zhang C, Lin P, Huang K-C, Liang J, Tian J, Cheng J-X. Volumetric che-mical imaging by stimulated Raman projection microscopy and tomography. Na-ture communications 2017, 8, 15117. [Google Scholar] [CrossRef] [PubMed]
20. Pully VV, Lenferink A, Otto C. Raman-fluorescence hybrid microspectroscopy of cell nuclei. Vibrational Spectroscopy 2010, 53, 12–8. [Google Scholar] [CrossRef]
21. Evans JW, Zawadzki RJ, Liu R, Chan JW, Lane SM, Werner JS. Optical cohe-rence tomography and Raman spectroscopy of the ex-vivo retina. Journal of bio-photonics 2009, 2, 398–406. [Google Scholar]