Automated spectral decomposition and reconstruction of optical properties using a mixed autoencoder approach

Area of Science:

• Biophotonics
• Optical properties
• Light-tissue interactions

Background:

• Investigating optical properties (OPs) is vital in biophotonics for understanding light-tissue interactions.
• Current methods like inverse Monte Carlo simulations (IMCS) struggle to detail spectral behavior of microscopic absorbers and scatterers.

Purpose of the Study:

• Develop a customized autoencoder neural network (ANN) to automatically identify the spectral behavior of individual microscopic absorbers and scatterers.
• Enable precise computation of OPs from measurements.

Main Methods:

• Designed an asymmetric autoencoder neural network (ANN) where the bottleneck represents the number of absorbers and scatterers.
• The ANN computes OPs using a linear combination of absorbers and scatterers.
• Validated the model using intralipid as a scatterer and ink as an absorber.

Main Results:

• Successfully extracted the spectral shape of each constituent using decoder weights.
• Demonstrated the ANN's effectiveness in detailing spectral behavior of absorbers and scatterers.
• Achieved high-precision prediction of optical properties.

Conclusions:

• The developed ANN is a robust tool for extracting spectral behavior of absorbers and scatterers without requiring prior knowledge.
• Potential applications include determining relative concentrations of tissue constituents.