Robust principal component analysis and clustering methods for automated classification of tissue response to ARFI excitation.

We introduce a new method for automatic classification of acoustic radiation force impulse (ARFI) displacement profiles using what have been termed "robust" methods for principal component analysis (PCA) and clustering. Unlike classical approaches, the robust methods are less sensitive to high variance outlier profiles and require no a priori information regarding expected tissue response to ARFI excitation. We first validate our methods using synthetic data with additive noise and/or outlier curves. Second, the robust techniques are applied to classifying ARFI displacement profiles acquired in an atherosclerotic familial hypercholesterolemic (FH) pig iliac artery in vivo. The in-vivo classification results are compared with parametric ARFI images showing peak induced displacement and time to 67% recovery and to spatially correlated immunohistochemistry. Our results support that robust techniques outperform conventional PCA and clustering approaches to classification when ARFI data are inclusive of low to relatively high noise levels (up to 5 dB average signal-to-noise [SNR] to amplitude) but no outliers: for example, 99.53% correct for robust techniques vs. 97.75% correct for the classical approach. The robust techniques also perform better than conventional approaches when ARFI data are inclusive of moderately high noise levels (10 dB average SNR to amplitude) in addition to a high concentration of outlier displacement profiles (10% outlier content): for example, 99.87% correct for robust techniques vs. 33.33% correct for the classical approach. This work suggests that automatic identification of tissue structures exhibiting similar displacement responses to ARFI excitation is possible, even in the context of outlier profiles. Moreover, this work represents an important first step toward automatic correlation of ARFI data to spatially matched immunohistochemistry.