Abstract 19246: Histology-Validated Neural Networks Enable Accurate Plaque Tissue and Thin-Capped Fibroatheroma Characterization Through Intravascular Optical Coherence Tomography
Introduction: Although intravascular ultrasound can successfully classify plaque composition in an automated fashion through ultrasound virtual histology, this has not been accomplished with intravascular optical coherence tomography (IVOCT). The complexity of IVOCT images has limited its clinical acceptance despite its greater resolution. This is particularly true with thin-capped fibroatheroma (TCFA), which are uniquely recognized by IVOCT, but prone to be falsely identified by expert viewers.
Hypothesis: We assessed the hypothesis that histology validated supervised neural networks can reliably characterize arterial tissue and TCFA in IVOCT images.
Methods: Neural network features and nodes were optimized to best classify arterial fibrous, calcium and lipid tissue in IVOCT images. Grey level co-occurrence matrix and windowed texture analysis was used to develop network features. Lipid pixels were used to isolate TCFA. The network was trained on 81000 pixels and tested on 61000 pixels, sampled separately from 21 plaque lesions in 11 left anterior descending and 4 right coronary arteries. Imaging was conducted on 11 human hearts (3 women, 8 men) within 24 hours of death. The age at death was 65 ± 11 years. Accuracy was calculated by comparing network classification to histology assessment.
Results: Fibrous pixels were classified with 94.1% ± 0.38% sensitivity and 90.6% ± 0.24% specificity, calcium pixels with 87.5% ± 0.38% sensitivity and 91.8% ± 0.76% specificity, lipid pixels with 94.1% ± 0.39% sensitivity and 95.9% ± 0.24% specificity and TCFA pixels with 94.1% ± 0.52% sensitivity and 84.1% ± 0.32% specificity (n=61000). Accuracies reported herein exceed previously reported values for automated IVOCT plaque tissue classification. Histology validation makes the presented method more reliable than expert observer-validated approaches.
Conclusion: In conclusion, we have developed an IVOCT-based automated plaque classification algorithm that serves as a building block for high-resolution virtual histology of human coronary arteries. Gaining insight into the composition of plaque can enable precise TCFA identification and provides a valuable diagnostic methodology to assess the efficacy of cardiovascular therapeutic interventions.
Author Disclosures: V.L. Baruah: None. A. Zahedivash: None. T.B. Hoyt: None. D. Vela: None. L. Buja: None. T.E. Milner: None. M.D. Feldman: None.
- © 2016 by American Heart Association, Inc.