Carotid plaque stroke risk assessment using multiscale AM-FM analysis based on DoG filterbanks

Abstract

The objective of this work was the investigation of multiscale Amplitude Modulation - Frequency Modulation (AM-FM) analysis based on Difference of Gaussians (DoG) filterbanks representations in order to predict the risk of stroke by analysing carotid plaques ultrasound images of individuals with asymptomatic carotid stenosis. We computed the instantaneous amplitude, instantaneous phase and the magnitude of instantaneous frequency to extract histogram features on each plaque region. The Support Vectors Machine classifier was implemented to classify asymptomatic versus symptomatic plaques. A dataset of 100 carotid plaque images (50 asymptomatic and 50 symptomatic) were tested, and showed that the AM-FM features based on DoG filterbanks and simple histograms performed better than the traditional AM-FM features. Best results were obtained when an eight scale filterbank with a combination of scales was used reaching the accuracy of 75%.