Combining IVUS + OCT Data, Biomechanical Models and Machine Learning Method for Accurate Coronary Plaque Morphology Quantification and Cap Thickness and Stress/Strain Index Predictions

Abstract

Assessment and prediction of vulnerable plaque progression and rupture risk are of utmost importance for diagnosis, management and treatment of cardiovascular diseases and possible prevention of acute cardiovascular events such as heart attack and stroke. However, accurate assessment of plaque vulnerability assessment and prediction of its future changes require accurate plaque cap thickness, tissue component and structure quantifications and mechanical stress/strain calculations. Multi-modality intravascular ultrasound (IVUS), optical coherence tomography (OCT) and angiography image data with follow-up were acquired from ten patients to obtain accurate and reliable plaque morphology for model construction. Three-dimensional thin-slice finite element models were constructed for 228 matched IVUS + OCT slices to obtain plaque stress/strain data for analysis. Quantitative plaque cap thickness and stress/strain indices were introduced as substitute quantitative plaque vulnerability indices (PVIs) and a machine learning method (random forest) was employed to predict PVI changes with actual patient IVUS + OCT follow-up data as the gold standard. Our prediction results showed that optimal prediction accuracies for changes in cap-PVI (C-PVI), mean cap stress PVI (meanS-PVI) and mean cap strain PVI (meanSn-PVI) were 90.3% (AUC = 0.877), 85.6% (AUC = 0.867) and 83.3% (AUC = 0.809), respectively. The improvements in prediction accuracy by the best combination predictor over the best single predictor were 6.6% for C-PVI, 10.0% for mean S-PVI and 8.0% for mean Sn-PVI. Our results demonstrated the potential using multi-modality IVUS + OCT image to accurately and efficiently predict plaque cap thickness and stress/strain index changes. Combining mechanical and morphological predictors may lead to better prediction accuracies.