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

Significance: Investigating optical properties (OPs) is crucial in the field of biophotonics, as it has a broad impact on understanding light-tissue interactions. However, current techniques, such as inverse Monte Carlo simulations (IMCS), have limitations in extracting detailed information about the spectral behavior of microscopic absorbers and scatterers. Aim: We aim to develop a customized autoencoder neural network (ANN) that can automatically identify the spectral behavior of each microscopic absorber and scatterer responsible for generating OP. Approach: The ANN is designed to compute OP from measurements, in which the bottleneck corresponds to the number of absorbers and scatterers. The presented ANN functions asymmetrically and computes the OP using a linear combination of absorbers and scatterers. Validation was conducted using intralipid as a scatterer and ink as an absorber. Results: The employment of the decoder weights facilitated the successful extraction of the spectral shape of every constituent, demonstrating the effectiveness of the ANN in extracting detailed information about the spectral behavior of absorbers and scatterers. At the same time, the OP can be predicted with high precision. Conclusions: The presented ANN is a viable tool for extracting the spectral behavior of absorbers and scatterers without the need for prior knowledge of these components in the test and training data. Potential future applications could include the extraction of relative concentrations of constituents in tissue.