Machine-Learning-Enabled Raman Spectroscopy Refines Indocyanine Green Fluorescence Boundaries for Precise Glioblastoma Margin Delineation.

Intraoperative cancer diagnosis using Raman spectroscopy is a growing area of research, but clinical adoption is held back by relatively slow single point data acquisition. Its application to glioblastoma is particularly attractive because of the distinct spectra of malignant and healthy brain tissues, which allow detection of even diffusely invasive glioblastoma with high accuracy. Current clinical practice for intraoperative guidance uses indocyanine green (ICG) for wide-field localization, but its nonspecific accumulation beyond tumor margins leads to false-positive guidance. Critically, the strong fluorescence background from ICG under 785-nm Raman excitation overwhelms the inherently weak Raman signal, thereby preventing concurrent use of the 2 modalities. Here, we overcome this barrier by introducing a signal-processing algorithm that suppresses ICG-derived interference, enabling coregistered acquisition. We present an integrated "search-and-confirm" workflow, where ICG first identifies regions of interest and subsequent signal processing suppresses fluorescence interference, thereby enabling machine-learning-based Raman diagnosis within 3 s and achieving 85% to 90% identification accuracy in vivo. By enabling spatial correlation, we demonstrate that ICG systematically overestimates tumor extent, while Raman signatures match histopathological margins, thereby transforming Raman spectroscopy into a real-time corrective filter for clinical fluorescence guidance and offering a practical path toward precision glioblastoma resection.