Location-specific prediction of vulnerable plaque using IVUS, virtual histology, and spatial context

Early detection of the high-risk lesions such as thin-cap fibroatheroma (TCFA) is highly desired in the clinic. Our group recently addressed the task of prediction of future TCFAs based on baseline virtual histology intravascular ultrasound (VH-IVUS) data with prediction performance not sufficient for routine clinical use. To achieve clinical relevance of our TCFA prediction, an improved strategy is presented here that introduces a spatial context between adjacent IVUS-frame locations and uses a 3-frame TCFA definition. We compared performance of four types of feature set (VH-based, IVUS-based, biomarkers, and combined features), two feature selection approaches (support vector machine recursive feature elimination [SVM RFE] and mutual information [MI]), and two classifiers (SVM and random forests [RF]) when analyzing 24 baseline-follow-up patient datasets. The experimental results indicated that the best prediction performance achieved nearly 10% improvement compared to our previous context-free method — AUC = 0.86, sensitivity=82.6%, specificity=82.1%.