Abstract:
Diffuse reflectance is a rapid, low-cost and non-invasive optical diagnostic technique that detects compositions or abnormalities in tissues or turbid materials. The inverse problem needs to be solved to estimate the optical properties from measured diffuse reflectance spectra, and further to analyze compositions of the medium. However, challenges in solving the inverse problems come up from the failure of diffusion approximation near the light source, ill-posedness of numerical PDE approaches and the high computational cost of computational methods such as iterative Monte Carlo (MC) and the lookup table inverse model. With the advent of machine learning (ML) as a powerful tool o learn mappings or patterns from datasets, studies have been carried out to resolve the challenges, leading to better accuracy and higher efficiency. Our study aims at developing a comprehensive and robust ML-based inverse model to solve the optical properties from measured diffuse reflectance spectra accurately. First of all, the calibration problem was taken care of with the transfer learning technique utilizing simulation data and a small amount of experimental data. Subsequently, a ML-based surrogate model was developed to fit the diffuse reflectance spectra over given geometry from optical properties. Finally, the trained ML-based surrogate forward model was reloaded to generate lookup tables to compare with the traditional MC-based lookup table method. The trained ML forward model is then employed iteratively to find the optimal optical solution of the optical properties that most closely match the measured diffuse reflectance spectrum. Our ML-based inverse model exhibited comparable accuracy to the MC-based lookup table model while performing with superior versatility, reduced computational costs and faster implementation.
Join Teams Meeting