Related Experiment Video
Updated: Jun 30, 2026
Optical Coherence Tomography Based Biomechanical Fluid-Structure Interaction Analysis of Coronary Atherosclerosis Progression
Published on: January 15, 2022
Histological validation of artificial intelligence-driven automatic plaque characterization in coronary OCT: a
Miao Chu1,2, Francesca Razzi3, Giovanni Luigi De Maria2
1Biomedical Instrument Institute, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200030, China.
Background And Purpose:
Artificial intelligence (AI) is increasingly being applied to automate image interpretation, including in coronary optical coherence tomography (OCT), the gold standard for in vivo assessment of atherosclerotic plaques. Most AI models are trained using expert annotations; however, human interpretation is inherently subjective and may limit model performance. Therefore, validation against the true reference standard-histology-remains essential. The study aims to evaluate and compare the performance of an AI-powered plaque characterization model with that of clinicians, using co-registered histology as the reference standard.
Methods:
Matched OCT pullbacks and serial histological sections from 25 plaques in 11 swine atherosclerotic arteries were analyzed. Precise OCT-histology co-registration was achieved using a hierarchical coarse-to-fine approach, with stent edges and anatomical landmarks as references. Plaque components (fibrous, lipidic, and calcific) were manually labeled on histology and automatically segmented on OCT by the AI model. Meanwhile, three blinded readers with different levels of OCT expertise independently annotated the corresponding frames. The relative percentages of plaque components derived from OCT were compared with histology.
Results:
Across all histological sections, the median percentages of fibrous, lipidic, and calcific components were 65.0% (interquartile range [IQR]: 53.9%-92.5%), 34.3% (IQR: 6.5%-44.7%), and 0.2% (IQR: 0%-1.9%), respectively. The AI model demonstrated excellent correlation with histology, with Spearman's ρ = 0.907 (P < 0.001) for fibrous and ρ = 0.900 (P < 0.001) for lipidic components. The mean absolute discrepancy relative to histology was comparable between the AI model and the senior reader and smaller than that of the intermediate and junior readers. Agreement with histology improved with reader's experience (fibrous: intraclass correlation coefficients [ICC] = 0.666, 0.720, and 0.821; lipidic: ICC = 0.593, 0.684, and 0.803), yet remained lower than that of the AI model (fibrous: ICC = 0.938; lipidic: ICC = 0.939).
Conclusions:
Despite being trained on human-annotated data, the AI model demonstrated superior agreement with histology compared with clinicians. AI-driven plaque characterization may reduce interpretative subjectivity and enhance the clinical utility of coronary OCT in the management of coronary artery disease.
