Automatic Delineation of Human Carotid Plaque Components by Supervised Learning in ARFI Variance of Acceleration Imaging, in Vivo

Acoustic radiation force impulse (ARFI) imaging is an ultrasound-based elastography technique that enables noninvasive characterization of tissue properties. Applied to carotid atherosclerotic plaque, ARFI imaging provides insight into plaque structure and composition, with prior studies demonstrating that peak displacement (PD) differentiates soft from stiff components, and that the decadic log of the variance of acceleration, log(VoA), distinguishes calcium (CAL), collagen (COL), lipid-rich necrotic core (LRNC), and intraplaque hemorrhage (IPH). However, clinical interpretation of ARFI images is subjective and time-consuming. In this study, supervised learning classifiers were developed to automatically delineate plaque components in vivo. ARFI imaging was performed in 17 patients undergoing clinically indicated carotid endarterectomy (CEA). One-dimensional normalized cross-correlation was applied to radiofrequency (RF) echo data to extract ARFI displacement profiles, cross-correlation coefficient (CC), PD, log(VoA), and RF signal-to-noise ratio (SNR). Support Vector Machine (SVM) and Random Forest (RF) classifiers were trained on either the extracted parameters or on displacement profile segments 3–6 ms after ARFI excitation. Ground truth was established from spatially aligned histology with pathologist-annotated regions of CAL, COL, LRNC, and IPH. Classifier performance was assessed by receiver operating characteristic (ROC) analysis. Parameter-based SVM classifiers achieved high AUC (>0.85) for CAL, COL, and LRNC, and moderate AUC (>0.75) for IPH, outperforming the profile-based model. These results demonstrate that supervised learning methods enable automatic delineation of human carotid plaque components in ARFI log(VoA) imaging, in vivo.