TCT-684 Evaluation of the Performance of Deep Learning-Based Methods in Assessing Plaque Pathology in OCT and NIRS IVUS: A Head-to-Head Comparison With Histology

Background
Accurate classification of plaque composition is essential for risk stratification and treatment planning. The use of intravascular imaging is limited because image interpretation can be challenging and time-consuming. To overcome this, deep learning (DL) methods have been developed allowing fast characterization of plaque components and burden quantification. The aim of the study is to examine performance of optical coherence tomography (OCT) and near-infrared spectroscopy-intravascular

Methods
Matched histology and OCT frames from 20 cadaveric human hearts (n = 165 cross-sections) and corresponding histology and NIRS-IVUS images from 15 hearts (n = 116) were included. The OCT-DL method was trained using estimations of experts in an established core laboratory; the NIRS-IVUS-DL method has been trained from histological sections and relied on the output of NIRS and pixel intensity in plaque to characterize phenotypes. The predictions of these methods were superimposed on the

Results
Coregistration of NIRS-IVUS and OCT images with histology was feasible in all frames. NIRS-IVUS had higher efficacy in assessing external elastic lamina than OCT, which exhibited a larger bias in heavily diseased frames with increased plaque burden (accuracy 0.95 and 0.87, respectively). Both modalities accurately detected fibrotic tissue, but NIRS-IVUS was superior to OCT for detecting calcific (recallNIRS-IVUS: 0.81, precisionNIRS-IVUS: 0.80, and F-scoreNIRS-IVUS: 0.80 vs 0.36, 0.98, and 0.52

Conclusion
NIRS-IVUS-DL enabled more accurate assessment of vessel wall and plaque composition. This may result from superior discriminatory capacity and greater tissue penetration of combined intravascular imaging modalities of NIRS-IVUS alongside the more robust histological training used in development of NIRS-IVUS-DL