Algorithms for Intraoperative Neurovascular Inclusion Detection, Diameter and Depth Prediction Based on Frequency Domain Near Infrared Spectroscopy.

This study proposes an improved method for subsurface detection of neurovascular structures and their diameter and depth prediction as crucial feedback to neurosurgeons to prevent critical damage. The method relies on frequency-domain near infrared spectroscopy and machine learning algorithms based on numerical modeling data. The tasks solved include: analyzing the impact of the technical implementation of the spectrometer, forming effective feature vectors for classification and regression, selecting algorithms, developing training methods, and experimentally testing the results. Variational autoencoder-based algorithms demonstrate superior performance in classification and strong results in regression. A key advantage of these algorithms is their ability to train on unlabeled data while preserving the physical meaning of the latent space due to the applied custom constraint. It is essential that the light detectors of the spectrometers have a high internal gain. Experimental tests confirm the feasibility of partial training on simulated data.