Rapid hyperspectral Raman imaging of cells with depthwise separable 3D MultiResU-Net

Area of Science:

• Biomedical Optics
• Computational Imaging
• Spectroscopy

Background:

• Hyperspectral Raman imaging (HRI) offers label-free molecular mapping of cells.
• Weak Raman scattering signals necessitate long integration times, limiting HRI applications.
• Existing denoising methods struggle to preserve spectral integrity and cellular morphology.

Purpose of the Study:

• To develop an efficient computational framework for accelerating HRI.
• To enhance signal quality and reduce integration times in HRI without sacrificing spectral integrity.
• To enable rapid, label-free cellular imaging for high-throughput and dynamic biomedical analysis.

Main Methods:

• A depthwise separable three-dimensional MultiResU-Net was developed for joint spatial-spectral denoising of HRI data cubes.
• The network integrates multi-scale feature fusion and depthwise separable convolutions to capture spatial-spectral correlations.
• The method was validated on both synthetic and experimentally acquired HRI datasets.

Main Results:

• The proposed deep learning method substantially enhances signal quality in HRI data.
• It accurately reconstructs spectral features and preserves cellular morphology.
• The approach enables rapid cellular Raman imaging at short integration times, outperforming traditional filters and 1D neural networks.

Conclusions:

• The developed MultiResU-Net provides an efficient and generalizable computational framework for accelerating HRI.
• This advancement supports broader applications in high-throughput and dynamic biomedical analysis.
• The method facilitates label-free molecular mapping of cells with improved speed and fidelity.