Non-resonant background removal in broadband CARS microscopy using deep-learning algorithms

Area of Science:

• Biomedical Optics
• Spectroscopy
• Machine Learning

Background:

• Broadband Coherent anti-Stokes Raman (BCARS) microscopy offers rapid, full Raman spectra acquisition of biological samples.
• Non-resonant background (NRB) in CARS signals distorts spectral data and reduces chemical contrast.
• Traditional NRB removal requires manual algorithms and expert knowledge.

Purpose of the Study:

• To review existing deep learning models for NRB removal in BCARS.
• To introduce and evaluate two novel deep learning architectures (CNN+GRU and GAN) for NRB removal.
• To assess the performance of various models using an improved, generalized training dataset.

Main Methods:

• Review of existing deep learning models (SpecNet, VECTOR, LSTM, Bi-LSTM).
• Development and implementation of CNN+GRU and GAN architectures.
• Training and testing models on a generalized dataset and experimental BCARS data.
• Application of models in spectral unmixing pipelines for BCARS images.

Main Results:

• CNN+GRU and VECTOR models demonstrate the highest accuracy in NRB removal.
• GAN models identify the highest number of true positive peaks in experimental data.
• GAN and VECTOR models are most suitable for real-time BCARS image processing.

Conclusions:

• Deep learning significantly enhances NRB removal in BCARS microscopy.
• Novel CNN+GRU and GAN architectures offer competitive or superior performance.
• Optimized models facilitate more accurate and efficient chemical contrast imaging in biological samples.