Exploratory study demonstrates accurate classification of glioma subtypes in patient tissue, indicating potential for rapid intraoperative characterization.
Key Points
To evaluate whether spectral and dielectric parameters obtained via terahertz time-domain spectroscopy can differentiate major adult diffuse glioma subtypes without chemical labeling.
Extracted 492 spectral and dielectric features across 82 frequency points (0.2 to 1.4 THz) from 523 tissue slices across 63 glioma patients.
Implemented a patient-wise split with 50 training patients and 13 held-out validation patients to prevent slice-level data leakage.
Built a hierarchical classification workflow combining LASSO-based feature selection, principal component analysis, and random forest models.
The hierarchical model achieved a slice-level classification accuracy of 0.818 on the held-out validation cohort of 13 patients.
Model performance yielded a macro-F1 score of 0.760 across glioblastoma, astrocytoma, and oligodendroglioma tissue slices.