AI-empowered OCT is non-inferior to IVUS for media and plaque burden quantification: head-to-head histopathology validation using human coronary arteries
J Huang, X Yu, R Dai, R He, Z Fan, Y Wang, W Wijns, S TuAbstract
Background
Optical coherence tomography (OCT) provides excellent resolution at the cost of shallow penetration depth, limiting its ability for complete assessment of the vessel wall and, consequently, plaque burden (PB). A deep learning (DL) algorithm was therefore developed for fully automated media delineation and PB estimation from OCT.
Purpose
To perform the first head-to-head comparison of AI-enabled OCT versus intravascular ultrasound (IVUS) for quantifying media area and PB, using histopathology as the reference standard.
Methods
Ten post-mortem human hearts were harvested, yielding 29 major epicardial coronary arteries. The arteries were ligated at the side branches and connected to a coronary perfusion system during OCT and IVUS pullbacks. After imaging, the arteries were subdivided into different coronary segments and serially sectioned at 1-mm intervals and stained with haematoxylin-eosin, Masson trichrome, and Alizarin Red when applicable (Figure 1). The matching between histopathologic and OCT and IVUS cross-sections relied primarily on distance from the barbed connector, supplemented by corresponding luminal configurations and anatomical landmarks including side branches or plaque morphology. The Lumen area, media area and PB from OCT and IVUS was fully automatically derived using DL, and semiautomatically derived using a developed prototype in a blinded fashion and compared to histopathology (Figure 2). Statistical analyses accounted for within-heart clustering.
Results
A total of 834 paired OCT-IVUS cross-sections were obtained and compared. Using histopathology as the reference standard, the lumen area concordance was high for both modalities (r=0.89 [95% CI: 0.88, 0.91] for OCT versus r=0.90 [95% CI: 0.89, 0.91] for IVUS, p=0.30). For the assessment of IEL area, AI-powered OCT was numerically lower compared to IVUS (r=0.85 [95% CI: 0.83-0.87] for OCT versus r=0.86 [95% CI: 0.84-0.88] for IVUS), but statistically non-significant (p=0.45). AI-powered OCT were non-inferior to IVUS for assessing PB (r= 0.76 [95% CI: 0.73-0.79] for OCT versus r=0.79 [95% CI: 0.76, 0.81] for IVUS, p=0.12) (Figure 1). In identifying histologically significant cross-sections (PB>65%), the AI-powered OCT achieved a non-inferior discrimination compared with IVUS (area under the receiver operating characteristic curve compared [AUC]=0.93 [95% CI: 0.91-0.94] for OCT and 0.93 [95% CI: 0.91-0.94] for IVUS, p= 0.75).
Conclusions
Validated against histopathology, the DL algorithm enables reliable, fully automated IEL delineation and plaque burden estimation from OCT, with performance non-inferior to co-localised IVUS. AI-empowered OCT demonstrates high diagnostic accuracy and may broaden the quantitative capabilities of OCT in clinical practice.Figure 1For image description, please refer to the figure legend and surrounding text.Figure 2For image description, please refer to the figure legend and surrounding text.