Physics inspired neural network for optical property retrieval from diffuse reflectance

Introduction Optical property retrieval in diffuse reflectance imaging, like diffuse reflectance spectroscopy (DRS) and hyperspectral imaging (HSI), often involves fitting measured spectra to analytical solutions using approximations such as Diffusion Theory (DT). This method, while accurate, is not always generalizable due to the assumptions inherent in DT and results in non-unique solutions for optical properties and physiological parameters. In addition, it is computationally intensive. Physics-inspired deep learning offers generalizable data descriptions guided by physical principles but requires extensive labelled data, which is hard to obtain, especially in medical contexts. Methods We propose a deep learning approach to retrieve physiological parameters from DRS and HSI spectra using DT-simulated training data. The DT-simulated data is synthesised using a range for the optical properties: Blood Volume Fraction (BVF), Saturation, water-fat ratio (WFR), average blood vessel radius (R), scattering amplitude (SA), and scattering slope (SL). The range for these parameters we have extracted from literature. Results Our feed-forward neural network achieved median relative errors of 4% and 2% for DRS and HSI, respectively. Discussion Results suggest that the proposed method is robust and that retrieval of optical properties is possible with similar results to DT but also reducing operation time.