Assessing spectral algorithms to predict atherosclerotic plaque composition with normalized and raw intravascular ultrasound data

Abstract

Spectral analysis of backscattered intravascular ultrasound (IVUS) data has demonstrated the ability to characterize plaque. We compared the ability of spectral parameters (e.g., slope, midband fit and y-intercept), computed via classic Fourier transform (CPSD), Welch power spectrum (WPSD) and autoregressive (MPSD) models, to classify plaque composition. Data were collected ex vivo from 32 human left anterior descending coronary arteries. Regions-of-interest (ROIs), selected from histology, comprised 64 collagen-rich, 24 fibrolipidic, 23 calcified and 37 calcified-necrotic regions. A novel quantitative method was used to correlate IVUS data with corresponding histologic sections. Periodograms of IVUS samples, identified for each ROI, were used to calculate spectral parameters. Statistical classification trees (CT) were computed with 75% of the data for plaque characterization. The remaining data were used to assess the accuracy of the CTs. The overall accuracies for normalized spectra with CPSD, WPSD and MPSD were, respectively, 84.7%, 85.6% and 81.1% (training data) and 54.1%, 64.9% and 37.8% (test data). These numbers were improved to 89.2%, 91.9% and 89.2% (training) and 62.2%, 73% and 59.5% (test) when the calcified and calcified-necrotic regions were combined for analysis. Most CTs misclassified a few fibrolipidic regions as collagen, which is histologically acceptable, and the unnormalized and normalized spectra results were similar.