Towards machine learning for heterogeneous inverse scattering in 3D microscopy.

Area of Science:

• Biophotonics
• Computational Imaging
• Medical Physics

Background:

• Light scattering in biological tissues is complex due to varying refractive properties of particles like cells.
• Coherent speckle effects are crucial in microscopy imaging and understanding them is key for accurate reconstruction.
• Current methods for volumetric reconstruction of scattering parameters can be limited.

Purpose of the Study:

• To develop a learning-based framework for volumetric reconstruction of scattering parameters, specifically particle densities.
• To leverage physically accurate speckle simulations for improved training data in microscopy applications.
• To analyze network design, training data, and input features for learning-based imaging systems.

Main Methods:

• Utilizing physically accurate speckle simulators to model light propagation and scattering.
• Incorporating speckle statistics, such as the memory effect, into the learning framework.
• Analyzing various aspects of network design, including architecture, training data, and input features.

Main Results:

• Demonstrated the importance of realistic speckle modeling for successful learning-based reconstruction.
• Explored the integration of speckle statistics to enhance the learning framework.
• Provided an analysis of key components for designing effective learning-based imaging systems.

Conclusions:

• Physically accurate speckle simulation is essential for training learning-based imaging systems in scattering media.
• The study provides insights into network design and data requirements for volumetric reconstruction.
• This work lays the foundation for advanced learning-based imaging systems in biophotonics and medical imaging.