AI Extracts Hidden Spectral Data from OCT to Flag High-Risk Coronary Plaques

Coronary artery disease kills primarily through one mechanism: a lipid-rich plaque ruptures, triggers clot formation, and blocks blood flow to the heart muscle. The deadliest plaques are not necessarily the largest ones - thin-cap fibroatheromas with large lipid cores are disproportionately likely to rupture even when they cause only modest arterial narrowing. Identifying those high-risk plaques before they rupture is a major goal of interventional cardiology.

Optical coherence tomography, the near-infrared imaging technology used during catheter-based coronary procedures, provides extremely detailed cross-sectional images of vessel walls. But standard OCT images reveal structure, not composition. A plaque that looks stable might harbor a vulnerable lipid core that the image alone cannot reliably detect. A new approach from Korea Advanced Institute of Science and Technology (KAIST), published in Biomedical Optics Express, exploits wavelength-dependent information already embedded in the OCT signal - information that standard processing discards - to fill that compositional gap.

The spectral information hidden in the OCT signal

Standard OCT processing collapses the wavelength-dependent backscattering information into a single intensity value at each point. But different tissue types - lipid, fibrous tissue, calcium - interact with light at different wavelengths in subtly different ways. Those spectral signatures are present in the raw OCT data; they are just not extracted by conventional processing pipelines.

The KAIST team, led by Hyeong Soo Nam, developed a method to extract that spectral information and feed it into a deep learning model. The model learns to associate specific wavelength-dependent signal patterns with lipid-rich tissue, then automatically highlights suspicious regions throughout the vessel cross-section.

"Plaques with more lipid and certain patterns of lipid distribution are strongly associated with the risk of major cardiac events," Nam said. "By analyzing wavelength-dependent information hidden in the OCT signal and combining it with AI, we were able to identify the presence and distribution of lipid within the vessel wall."

Weakly supervised training - a practical advantage

A major obstacle to training AI models on medical images is annotation effort. Pixel-level labeling - manually outlining lipid regions in hundreds of OCT cross-sections - is extremely time-consuming and requires expert cardiologists or pathologists. The subjectivity involved also introduces inter-annotator variability.

The KAIST model uses a weakly supervised approach: instead of requiring pixel-level labels, it trains on frame-level annotations that indicate only whether lipid is present or absent in a given image slice. The model then learns to localize likely lipid regions through the weaker supervision signal.

"Unlike many conventional AI systems that require experts to painstakingly label lipid regions at the pixel level - an extremely time-consuming and subjective process - our approach learns from much simpler frame-level annotations that indicate only whether lipid is present or absent," Nam said. "This substantially lowers the annotation burden and makes the method far more practical for real-world clinical use."

Validation in a rabbit atherosclerosis model

The researchers validated the method using intravascular OCT data acquired from a rabbit model of atherosclerosis - a standard preclinical model for coronary plaque studies. AI-generated lipid predictions were compared against histopathology results from lipid-specific tissue staining, which serves as the reference standard for ground-truth plaque composition.

The model achieved strong classification performance - correctly identifying image frames containing lipid-rich plaques - and the spatial regions it highlighted showed good agreement with pathological findings. The validation is a meaningful step forward, but it is limited by the rabbit model: rabbit atherosclerotic plaques differ from human lesions in composition, lipid type, and distribution, and the performance observed in this model may not transfer directly to human coronary arteries.

No hardware modifications required

A notable practical feature is that the method works with existing clinical OCT hardware without any modifications. The spectral information it exploits is already present in the raw signal from standard systems. Deploying the analysis requires only software integration into the existing image processing pipeline, which lowers the barrier to eventual clinical use compared to approaches that require new hardware or additional imaging agents.

What the team is working on next

The current validation is preclinical. The researchers are working on improving processing speed to support real-time analysis during catheterization procedures - currently a necessary step before clinical deployment. They are also planning validation studies using human coronary artery data and developing strategies to integrate the method into existing clinical workflows without adding friction for physicians.

"During a coronary intervention, this method could provide clinicians with additional information to support risk assessment, procedural planning and evaluation of treatment response," Nam said.