Deep learning-based plaque characterization in hybrid IVUS-OCT images is superior to single-modality deep learning analysis and human experts: head-to-head comparison against histology

Le résumé fourni par la source

AIMS: Hybrid intravascular ultrasound-optical coherence tomography (IVUS-OCT) can enable more accurate plaque characterization than single-modality intravascular imaging, enhancing treatment planning and vulnerable plaque detection. However, image interpretation in IVUS-OCT is challenging and time-consuming. To overcome this limitation, we introduce a novel histology-trained deep learning (DL)-classifier for plaque component classification in IVUS-OCT images and compare its performance against single-modality DL and expert analysts. METHODS AND RESULTS: IVUS-OCT frames and matched histological sections from 10 cadaveric human hearts were included in this analysis. The histological data were used to define fibrotic, calcific, and necrotic core tissue regions of interest (ROIs) in IVUS-OCT and used to train three DL-classifiers for IVUS, OCT, or hybrid IVUS-OCT image analysis (992 frames) and test their performance (264 frames). The test set was additionally annotated by experts from three different core labs, and their estimations and those of the DL-classifiers were compared with histology.The IVUS-OCT DL-classifier had a superior performance to the IVUS-DL, OCT-DL, and the expert analysts in detecting plaque phenotypes (Kappa 0.60 vs. 0.19, 0.35, and 0.53, respectively) and accurately classified 68% of histologically defined fibroatheromas. The hybrid IVUS-OCT DL-classifier also had a better performance than single-modality DL-classifiers and the experts in assessing tissue types in ROIs annotated by histology (overall accuracy 86.7% compared with 73.2% for IVUS-DL, 66.6% for OCT-DL, and 70.6% for the experts). CONCLUSION: Plaque characterization using a histology-trained hybrid IVUS-OCT DL-classifier is feasible and enables more accurate detection of plaque components and phenotype classification than single-modality DL-classifiers and expert analysts.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.