Abstract
Reliable identification of lipid distribution is critical for assessing coronary vulnerability, yet conventional optical coherence tomography (OCT) lacks compositional specificity. Expanding OCT into the spectral information through spectroscopic OCT (S-OCT), in combination with deep learning, enables automated, composition-aware tissue characterization without requiring hardware modification. This study aims to develop a weakly supervised deep learning framework for lipid detection and localization from S-OCT data, minimizing the need for dense manual annotation. A ResNet-34 network incorporating convolutional block attention modules (CBAMs) was trained using frame-level binary labels of lipid presence. Gradient-weighted class activation mapping (Grad-CAM) was applied to generate interpretable activation maps highlighting lipid-associated regions. Model predictions were validated against Oil Red O-stained histology of rabbit aortas. The proposed model accurately localized lipid regions with strong spatial correspondence to histology, achieving an arc-level overlap agreement of 83.9%. Comparative analyses confirmed that incorporating spectroscopic information significantly improved lipid detection over conventional OCT. This framework demonstrates the feasibility of spectroscopically enhanced, weakly supervised deep learning for automated lipid detection in intravascular imaging. By enabling efficient lipid screening and spatial interpretation, it establishes a scalable foundation for downstream assessment of lipid burden and clinically relevant plaque characterization, with potential utility for automated risk stratification.