Deep-learning-assisted Fourier transform imaging spectroscopy for hyperspectral fluorescence imaging.

Area of Science:

• Optics and Photonics
• Biomedical Imaging
• Computational Science

Background:

• Hyperspectral fluorescence imaging is crucial for distinguishing probes with similar emission spectra.
• Fourier transform imaging spectroscopy (FTIS) offers high spectral resolution but suffers from low imaging throughput due to extensive interferogram sampling.
• Existing FTIS systems require He-Ne correction, increasing cost and complexity.

Purpose of the Study:

• To investigate the application of deep learning to reduce interferogram sampling in FTIS.
• To assess the impact of deep learning on image quality and system robustness.
• To explore the potential for simplifying FTIS systems by eliminating the need for He-Ne correction.

Main Methods:

• Deep learning models were developed and trained to reconstruct hyperspectral images from undersampled interferograms.
• The approach was validated using biological samples (endothelial cells) and diverse fluorescent beads.
• Neural network models were optimized using the Hyperband algorithm and compared to manual optimization.

Main Results:

• Deep learning reduced interferogram sampling by an order of magnitude with no noticeable degradation in image quality.
• The deep learning approach demonstrated increased robustness against translation stage errors and environmental vibrations.
• The necessity for He-Ne correction in FTIS was successfully bypassed, simplifying the system.

Conclusions:

• Deep learning offers a powerful method to accelerate hyperspectral fluorescence imaging with FTIS.
• This technique improves imaging efficiency and system practicality by reducing data requirements and enhancing robustness.
• The optimized deep learning models pave the way for more cost-effective and compact FTIS systems.