Deep learning for denoising in a Mueller matrix microscope.

Area of Science:

• Biomedical optics
• Microscopy
• Image processing

Background:

• Mueller matrix microscopy is vital for analyzing complex biological microstructures.
• Key performance metrics, measurement accuracy and acquisition time, often present a trade-off, increasing instrument complexity and cost.

Purpose of the Study:

• To develop a learning-based approach to simultaneously enhance measurement accuracy and reduce acquisition time in Mueller matrix microscopy.
• To overcome the inherent limitations of conventional Mueller matrix microscopes.

Main Methods:

• A learning-based method was implemented using a rotating polarizer and waveplate polarization state generator.
• A modified U-Net network with channel attention was employed to denoise low-quality images.
• High-quality Mueller matrix data acquired over long durations served as the ground truth for training.

Main Results:

• The modified U-Net effectively reduced noise in images acquired with significantly shorter acquisition times.
• The method achieved high measurement accuracy comparable to traditional, longer acquisition methods.
• Simultaneous improvement in both accuracy and speed was demonstrated.

Conclusions:

• The proposed learning-based method offers a viable solution for improving Mueller matrix microscopy performance.
• This approach can lead to more efficient and accurate characterization of biological microstructures.
• It has the potential to reduce the complexity and cost associated with advanced polarization imaging techniques.