High-Throughput Molecular Imaging via Deep-Learning-Enabled Raman Spectroscopy.

Area of Science:

• Biophotonics
• Spectroscopy
• Machine Learning

Background:

• Raman spectroscopy offers label-free molecular contrast but suffers from slow acquisition speeds, limiting high-throughput applications.
• Existing methods for improving Raman signal quality and resolution are often insufficient for rapid imaging.

Purpose of the Study:

• To develop a deep learning framework (DeepeR) for significantly increasing the throughput of Raman spectroscopy.
• To enhance both the signal quality and spatial resolution of hyperspectral Raman images.

Main Methods:

• Trained a deep learning model on over 1.5 million hyperspectral Raman spectra (400 hours of data).
• Implemented deep learning for denoising and reconstruction of low signal-to-noise Raman spectra, achieving a 10x improvement in mean-squared error.
• Developed a neural network for 2-4x spatial super-resolution of hyperspectral Raman images while preserving molecular information.
• Applied transfer learning to extend the framework from cellular to tissue-scale imaging.

Main Results:

• Achieved Raman imaging speed-ups of 40-90x for high-resolution cellular imaging in under 1 minute.
• Demonstrated a 160x speed-up for lower-resolution applications like rapid screening and spectral pathology.
• Successfully extended the deep learning approach to tissue-scale imaging.

Conclusions:

• The DeepeR framework significantly accelerates Raman spectroscopy, overcoming previous throughput limitations.
• This advancement enables high-quality, high-resolution molecular imaging at unprecedented speeds.
• DeepeR provides a foundation for diverse high-throughput Raman spectroscopy and molecular imaging applications in biomedicine.