Deep Learning-Assisted Terahertz Metasurface Biosensing for Rapid Ex Vivo Molecular Assessment of Diffuse Glioma Frozen Sections

Abstract

Abstract Molecularly informed maximal safe resection of diffuse glioma requires rapid assessment of surgical specimens, but conventional molecular assays are not suited to repeated ex vivo testing. We developed a deep learning-assisted terahertz (THz) metasurface workflow in which a K/I-patterned aluminum metasurface both serves as the cryosection-receiving substrate and converts tissue-induced dielectric perturbations into multi-resonance THz responses. After preliminary thickness optimization, 50 μm sections were prepared directly on the active sensing area. The study included 53 patients, 59 tissue specimens, and 590 nested block-level spectra. A SEResNet1D model analyzed the complete 0.50–1.12 THz transmittance curve using a patient/specimen-grouped hold-out design. Distinct from prior conventional THz molecular studies and metasurface studies focused mainly on tumor–normal contrast, cultured cells, or animal imaging, the workflow integrates direct on-metasurface cryosectioning, resonant multi-mode THz transduction, full-spectrum deep learning, and leakage-aware grouped validation for five-class molecularly relevant assessment of human diffuse glioma frozen sections. Across five independently initialized runs, SEResNet1D achieved a mean five-class accuracy of 97.48 ± 0.87%. One THz spectrum required approximately 8 s, and a ready-to-scan tissue–sensor unit could be prepared in approximately 3 min. These findings support metasurface-enhanced THz sensing with full-spectrum deep learning as a rapid ex vivo adjunct for molecularly relevant assessment of diffuse glioma specimens.