Single-fiber co-registered Raman-OCT technique with deep learning enhances in vivo oral tissue assessment

Conventional oral cancer diagnosis relies on white light imaging and tissue biopsy that fail to capture critical microstructural and biomolecular information, thereby limiting early detection of precancerous and malignant lesions. To address these challenges, we present a unique coaxial and co-registered morpho-chemical imaging platform that integrates Raman spectroscopy (RS) and optical coherence tomography (OCT) within a compact, single-fiber probe. This integration uniquely enables video-rate OCT imaging together with sub-second Raman acquisition, providing spatially aligned microstructural and biochemical tissue information in vivo. Specifically, the single-fiber RS-OCT probe is constructed using a double-clad fiber (DCF) and gradient index (GRIN) lens architecture. The DCF's single-mode core delivers and collects OCT light for high-resolution structural imaging, whereas its high-numerical-aperture multimode inner cladding coupled to the GRIN fiber lens facilitates efficient Raman excitation and collection along an identical optical axis, ensuring truly coaxial and co-registered tissue measurements. A free-space coupling scheme incorporating spatial filtering effectively suppresses both fiber-originated Raman background and DCF-induced OCT multipath artifacts, achieving artifact-free biochemical-microstructural integration within a compact intraoral probe. The resulting system achieves $\sim 100\mathrm{dB}$ OCT imaging sensitivity and acquires both fingerprint (880 to $1800{\mathrm{cm}}^{-1}$ ) and high-wavenumber (2800 to $3600{\mathrm{cm}}^{-1}$ ) tissue Raman spectra within sub-seconds. We validated the hybrid RS-OCT system through in vivo oral measurements across nine anatomically distinct intraoral sites in healthy volunteers, including challenging posterior tongue regions. To synergistically fuse the co-registered morpho-chemical datasets, we implemented a cross-modality deep learning framework. This model, integrating OCT patch-voting with a dual-region Raman branch (fingerprint and high wavenumber), achieved a superior overall classification accuracy of 92.38% across all nine sites. This represents a significant diagnostic improvement of 16.55% and 21.19% over Raman-only (75.83%) and OCT-only (71.19%) modalities, respectively. Notably, the RS-OCT cross-modality deep learning model attained 93.4% accuracy within high-risk tongue sub-regions. This work establishes the single-fiber-enabled hybrid RS-OCT technique as a promising tool for enhanced in vivo oral tissue diagnosis with strong potential for broader clinical translations.