Enhanced skin tone and blood estimation through AI-enabled upscaling of discrete reflectance spectra

Objective quantification of epidermal melanin content and dermal blood volume is crucial for equitable applications in dermatology and in the evaluation and engineering of optical devices, such as pulse oximeters, where variable amounts of melanin may introduce inaccuracies. Here, we demonstrate a method leveraging diffuse reflectance spectroscopy and a machine-learning framework using fully connected neural networks (FCNNs) to non-invasively estimate melanin and blood indices from sparse data through spectral upscaling. We validated the algorithm on custom phantoms, porcine skin, and in vivo human participants. In phantoms with variable blood and melanin (0–20%), the model achieved high accuracy ( R 2 > 0.90) to absolute concentrations. In porcine skin, estimated melanin strongly correlated with histology-derived density scores (Pearson r = 0.95). Across six regions of the human hand, the FCNN accurately estimated pigmentation from sparse wavelengths, even during hand-raise challenges that induce confounding blood volume changes. After upscaling, sparse-wavelength estimates strongly correlated with a model using over 450 wavelengths (Pearson r = 0.96), reducing data requirements ~ 150-fold without compromising accuracy. These findings highlight the importance of high-spectral information for disentangling melanin and blood contributions in reflectance spectra and the capacity of machine-learning to restore predictive performance from sparse inputs. These findings demonstrate that artificial-intelligence -driven spectral upscaling can enhance objective skin tone quantification in wearable optical devices.