Automatic Classification of Human Carotid Plaque Features, in Vivo, Using Multiple Forms of ARFI Data

Rupture potential of atherosclerotic plaques in carotid arteries is conferred by both composition and structure of plaques. Previous studies have shown that from in vivo collected data, carotid plaque components such as collagen, calcium, necrotic core and intraplaque hemorrhage can be automatically detected by an ARFI imaging-derived machine learning classifier. Automatic classification considered normalized cross-correlation measurements of ARFI-induced displacement, signal-to-noise ratio and cross-correlation coefficients from an on-axis ARFI acquisition. We now extend our prior work by hypothesizing that using multiple ARFI data forms improves plaque feature detection and FC thickness measurement in human carotid plaques, relative to a single form of ARFI data. Carotid plaques were imaged in vivo prior to surgery in 20 patients undergoing carotid endarterectomy (CEA), and extracted plaque specimens were harvested after CEA for histological processing. ARFI data were acquired with cardiac gating to diastole with push (DP) and to systole without push (SNP) using fundamental low (FL), fundamental high (FH), and harmonic (H) tracking frequencies. Combinations of the resulting displacement profiles were used as inputs to the SVM classifier. The classifier was evaluated by 5-fold cross-validation, with the histological samples acting as gold standards. From the output SVM likelihood matrices, ROC curves were calculated for separating collagen from calcium and lipid-rich necrotic core from intraplaque hemorrhage. For all examined plaques, DP+FH+H achieved the highest average AUC of 0.926 (Sensitivity = 0.932, Specificity = 0.915) but required 2 acquisitions. The best data combination requiring only one acquisition (DP+FH) achieved an AUC of 0.917 (Sensitivity = 0.902, Specificity = 0.880). These results suggest that using multiple forms of frequency and gating of ARFI data as inputs to an automatic classifier improves discrimination of carotid plaque components, relevant for rupture vulnerability assessment.