TMET-55. Glioma Grade Prediction Using Hyperspectral Imaging and Machine Learning

Abstract Hyperspectral imaging (HSI) is an advanced optical technique that captures high-resolution spectral information across a wide wavelength range, enabling rapid and label-free visualization of biological tissues. The integration with machine learning has recently expanded its diagnostic application, particularly in disease classification and real-time intraoperative support. Gliomas represent a biologically and histologically heterogeneous group of primary brain tumors, for which timely and accurate grading is critical to optimizing surgical and therapeutic strategies. This study aimed to develop a machine learning–based classification model for glioma grade prediction using autofluorescence spectra acquired by HSI. We utilized a hyperspectral microscope capable of resolving the 450–850 nm range into 40 discrete spectral bands. A total of 192 regions of interest (ROIs) were analyzed from nine glioma specimens resected between February and September 2024 (6 glioblastomas, 2 oligodendrogliomas grade 3, and 1 oligodendroglioma grade 2). Spectral data were acquired under UV (330–385 nm) and blue light (460–490 nm) excitation. Principal component analysis (PCA) was conducted to extract PC1–PC10 as candidate features. Using XGBoost and Gain-based feature selection, the top five PCs were selected to train the final classification model (parameters: max_depth = 6, η = 0.3). Model performance was evaluated using accuracy and macro-F1 score. Consistent Fluorescence peaks were observed at 475 and 635 nm under UV excitation and at 545 and 635 nm under blue excitation. The nested 5-fold cross-validation yielded a mean accuracy of 0.88 and macro-F1 score of 0.87. Key spectral features corresponded to NADH, FAD, PpIX, and porphyrin aggregates, reflecting underlying metabolic diversity. These findings suggest that HSI-based autofluorescence profiling, combined with machine learning, shows strong potential for accurate glioma grade classification. With further validation in larger and more diverse cohorts, this method may enhance intraoperative decision-making and advance tumor metabolic characterization of brain tumors.