Pixel-wise and real-time estimation of optical mean path length using deep learning: application for intraoperative functional brain mapping

Optical imaging is a non-invasive technique that is able to monitor hemodynamic and metabolic brain response following neuronal activation during neurosurgery. However, it still lacks robustness to be used as a clinical standard. In particular, the quantification of the biomarkers of brain functionality needs to be improved. The quantification relies on the modified Be er Lambert law, which needs a correct estimation of the optical mean path length of traveled photons. Monte Carlo simulations are used for estimating the optical path length, but it is time-consuming, especially when modeling a patient's brain cortex. In this study, we developed a neural network based on the UNET architecture for a pixel-wise and real-time estimation of optical mean path length. The neural network was trained with segmentation of brain cortex as input and mean path length data as target. This deep learning approach allows a real time estimation of the optical mean path length. The results can be beneficial and useful within the framework of our EU-funded HyperProbe project, which a ims at transforming neuronavigation during glioma resection using novel hyperspectral imaging technology.