Physics-informed neural networks for estimation of volume and surface optical parameters in skin hyperspectral imaging using one- and two-layer tissue models

Hyperspectral imaging (HSI) is a non-invasive optical modality capable of characterizing biological tissues, making it suitable for medical applications such as skin lesion analysis. However, recovering physiological and optical tissue parameters from hyperspectral skin data remains a challenging inverse problem due to the high dimensionality and complexity of light–tissue interactions. We hypothesize that physics-informed neural networks (PINNs) can separate surface and volume reflections while recovering optical properties directly from hyperspectral skin images. To test this hypothesis, four PINN architectures were developed by combining diffusion-based analytical solutions of the radiative transfer equation and a surface reflectance model with parallel multilayer perceptron (PMLP) and transformer (T) backbones for one- and two-compartment skin models (OTCM and TTCM), resulting in the PMLP-OTCM, T-OTCM, PMLP-TTCM, and T-TTCM models. The PINNs were compared with conventional least-squares fitting using both synthetic and in vivo data. On synthetic data, the PINNs achieved significantly lower optical parameter estimation errors ( p < 0.001), reducing the mean absolute error (MAE) for melanin volume fraction ( f mel ), blood volume fraction ( f bl ), oxygen saturation ( S O2 ), vessel radius ( R v ), and surface reflectance coefficients ( ϵ ϕ i ). For the in vivo experiments, PMLP-TTCM achieved substantially lower mean relative absolute errors in spectral reconstruction, with reductions of 35.8--53.6% compared to least-squares fitting. In addition, the PMLP-TTCM produced parameter maps that were more consistent with skin anatomy and physiological expectations and required only 1.1 × 10 −5 s per voxel, compared to 1.4 × 10 −1 s for conventional fitting. In conclusion, we developed a PINN framework that enables fast, accurate, and interpretable optical property estimation from hyperspectral skin images and provides a scalable alternative to traditional inverse fitting methods with potential for future clinical HSI applications.