Pattern Analysis of the Kinematics in Ultrasound Videos of the Common Carotid Artery – Application to Cardiovascular Risk Evaluation

The clinical context of this study is the prediction of cardiovascular risk by analyzing ultrasound images of the common carotid artery. The principal methodological contribution of the present work is the implementation of image processing algorithms to characterize the pattern of the artery-wall spatio-temporal trajectory during the cardiac cycle. Normalized signals corresponding to the trajectory of the biological tissues were gathered via an initial phase of motion tracking based on Kalman filtering. The originality of the present work is the introduction of three complementary statistical approaches to interrogate these signals. First, a Machine Learning strategy was carried out with the AdaBoost algorithm to automatically identify healthy and at-risk subjects. Second, the Dynamic Time Wrapping method was applied to measure the pairwise similarity between signals and identify clusters. Third, a Principal Component Analysis was performed to randomly generate unseen patterns using Point Distribution Modeling. A total of 84 subjects (42 healthy volunteers and 42 at-risk patients) were involved in this study. Two significantly different profile archetypes could be reconstructed from the two populations, showing the effect of the atherosclerosis pathology on the artery. Results demonstrate that the healthy and at-risk signals can successfully be classified with an accuracy of 73%. Quantification of the pairwise distance between all signals indicated that healthy patterns are more similar to each other, whereas there is a wider variability between at-risk patterns. Finally, new signals were generated using a statistical model, possibly hinting towards new patterns characteristics.