Rapid Glioma Subtype Classification Using Label-Free Terahertz Time-Domain Spectroscopy and Hierarchical Machine Learning

Abstract

Rapid, label-free characterization of glioma tissue remains challenging in time-sensitive clinical workflows. This exploratory study evaluated whether terahertz time-domain spectroscopy (THz-TDS)-derived spectral and dielectric features could support classification of major adult-type diffuse glioma subtypes, including glioblastoma (GBM), astrocytoma, and oligodendroglioma. A total of 523 glioma tissue slices from 63 patients were analyzed using a unified THz-TDS workflow. For each slice, 492 features were extracted from six dielectric-parameter families across 82 frequency points from 0.2 to 1.4 THz. To reduce slice-level information leakage, model development and validation used a strict patient-wise split, with 50 training patients and 13 held-out validation patients. A hierarchical framework combining LASSO-based feature selection, principal component analysis, and random forest classification was constructed. In validation, the final framework achieved a slice-level accuracy of 0.818 and a macro-F1 score of 0.760, supporting further investigation of THz-TDS for rapid ex vivo glioma tissue characterization.