Plaque burden estimated from optical coherence tomography with deep learning: In vivo validation using co-registered intravascular ultrasound

Objectives: The objective of the present study was to compare plaque burden (PB) calculated from optical coherence tomography (OCT) using deep learning (DL) with PB derived from co-registered intravascular ultrasound (IVUS).

Background: A DL algorithm was developed for automated plaque characterization and PB quantification from OCT images. However, the performance of this algorithm for PB quantification has not been validated.

Methods: Five-year follow-up OCT and IVUS images from 15 patients implanted with bioresorbable vascular scaffold (BVS) at baseline were analyzed. Precise co-registration for 72 anatomical slices was achieved utilizing unique BVS radiopaque markers. PB derived from OCT DL and IVUS were compared. OCT cross-sections were divided into four subgroups with different media visibility level. The impact of media visibility on the numerical difference between OCT-derived and IVUS-derived PB was investigated. The stent sizes selected by OCT DL and IVUS were compared.

Results: Sixty-four paired OCT and IVUS cross-sections were compared. OCT DL showed good concordance with IVUS for PB assessment (ICC = 0.81, difference = -3.53 ± 6.17%, p < 0.001). The numerical difference between OCT DL-derived PB and IVUS-derived PB was not substantially impacted by missing segments of media visualization (p = 0.21). OCT DL showed a diagnostic accuracy of 92% in identifying PB > 65%. The stent sizes selected by OCT DL were smaller compared to the ones selected by IVUS (difference = 0.30 ± 0.34 mm, p < 0.001).

Conclusions: The DL algorithm provides a feasible and reliable method for automated PB estimation from OCT, irrespective of media visibility. OCT DL showed good diagnostic accuracy in identifying PB > 65%, revealing its potential to complement conventional OCT imaging.