Enhanced stimulated Raman and fluorescence imaging by single-frame trained BDN

Area of Science:

• Biomedical imaging
• Materials science
• Spectroscopic imaging

Background:

• Hyperspectral and multispectral imaging offer rich data but are often degraded by noise.
• Noise limits the extraction of crucial image features, especially in sensitive specimens.
• Existing machine learning methods require large datasets, posing a practical challenge.

Purpose of the Study:

• To introduce an efficient deep learning network for denoising spectroscopic images.
• To improve the extraction of local and global image features obscured by noise.
• To develop a versatile denoising solution applicable across different imaging modalities.

Main Methods:

• Development of the BiFormer denoising network (BDN) utilizing local and global connections.
• Implementation of sparse architectures and fine-tuning strategies within BDN.
• Adaptation and testing of BDN on stimulated Raman scattering (SRS) and fluorescence imaging data.

Main Results:

• BDN significantly enhanced signal-to-noise ratio (SNR) in SRS images by up to 16-fold.
• Subtle features at higher spatial frequencies were notably improved in denoised images.
• BDN achieved substantial SNR gains and reduced exposure times in fluorescence imaging.

Conclusions:

• BDN effectively denoises spectroscopic images, improving feature visibility and data quality.
• The network's versatility is demonstrated across SRS and fluorescence imaging modalities.
• BDN holds significant potential for advancing biomedical and materials science research through enhanced imaging.