Deep learning Mueller matrix feature retrieval from a snapshot Stokes image.

Area of Science:

• Optical physics
• Biomedical imaging
• Computational imaging

Background:

• Mueller matrix (MM) imaging characterizes complex media and histopathological features.
• Traditional MM derivation requires multiple exposures, sensitive to system/sample instability.
• This sensitivity limits MM imaging in dynamic or unstable conditions.

Purpose of the Study:

• To develop a deep learning approach for retrieving MM-based polarimetry basis parameters (PBPs) from a single Stokes vector snapshot.
• To overcome limitations of multi-exposure MM imaging, reducing errors and hardware complexity.
• To enable accurate MM imaging for dynamic samples and unstable environments.

Main Methods:

• A deep learning translation model based on a generative adversarial network (GAN) was designed.
• Customized loss functions were implemented within the GAN architecture.
• The method retrieves MM-PBPs from a single snapshot Stokes vector, acting as a post-processing technique.

Main Results:

• The deep learning approach successfully retrieved MM-PBPs from snapshot Stokes vectors.
• The method demonstrated effectiveness on liver and breast tissue slices and blood smears.
• Quantitative similarity assessments confirmed performance at both pixel and image levels.

Conclusions:

• The proposed deep learning method enables accurate MM imaging from single snapshots.
• This approach mitigates errors from multi-exposure imaging and reduces system complexity.
• It holds significant potential for MM imaging in challenging, dynamic, or unstable environments.