Coronary lipid detection by fusing OCT-derived near-infrared spectra and image morphology.

Intravascular assessment of lipid burden and its spatial distribution is important for characterizing high-risk coronary plaques. Conventional near-infrared spectroscopy has shown clinical utility in identifying lipid-rich plaques, but it provides limited depth information for localizing features related to thin-cap fibroatheroma. Although spectroscopic OCT aims to simultaneously provide characteristic spectra and depth information, severe scattering interference renders weak lipid-related spectral cues highly challenging to isolate. To address this, we propose NIR-Spectral-Morphology Network (NIRSMorNet), a physics-informed framework that combines OCT-derived spectral sub-band contrast with image morphology for coronary lipid detection. Specifically, it constructs a multi-dimensional near-infrared spectral sub-band input from raw interferometric signals, learns spectral-morphological features through implicit fusion, utilizes a cross-scale state-space module to stabilize local responses via circumferential continuity, and employs a fixed physics-guided inference scheme to generate model-derived depth-resolved lipid-evidence maps. Evaluation using OCT-derived expert consensus labels on internal and external cohorts yielded AUCs of 0.971 and 0.934, respectively, for A-line-level lipid detection. NIRSMorNet further provides model-derived depth-resolved lipid-evidence maps, enabling traceable 2-mm longitudinal summaries of local lipid-evidence burden and an exploratory summary of superficial lumen-adjacent lipid evidence.