Classification of atherosclerotic plaque composition by spectral analysis of intravascular ultrasound data

Knowledge of atherosclerotic plaque composition in vivo would be a tremendous aid towards selection of appropriate clinical intervention and timely detection of unstable atheroma. Spectral analysis of backscattered intravascular ultrasound (IVUS) data has demonstrated ability to characterize plaque. Spectral parameters such as slope, midband fit and y-intercept are indicative of scatterer size, concentration and acoustic impedance. In this study, we compared the ability of spectral parameters computed via the Welch power spectrum (PSD) and autoregressive (AR) models to classify plaque composition. Data was collected in vitro from 32 left anterior descending coronary arteries, with 30 MHz non-focused IVUS transducers in saline at physiologic pressure. Regions of interest (ROI), selected from histology, comprised 64 collagen-rich, 24 fibro-lipidic, 23 calcified and 37 calcified-necrotic regions. A novel quantitative method was used for IVUS data correlation with corresponding histology sections. Periodograms of IVUS samples, identified for each ROI, were used to calculate various spectral parameters (such as the midband fit, slope, y-intercept, and maximum power). Statistical classification trees (CT) were computed with 75% of the data for plaque characterization. The remaining data were utilized to assess the accuracy of the CTs. The overall accuracies with the Welch PSD and AR model (order 14) were 85.6% and 81.1% on the training data, 64.9% and 37.8% on the test data, respectively. These numbers were improved to 91.9% and 89.2% on training data, 73% and 59.5% on test data, when the calcified and calcified-necrotic regions were combined for analysis. Further analysis with AR model order number and level of zero padding increased the overall accuracy, with individual sensitivities to collagen of 91.7%, calcium of 88.9%, and fibro-lipid of 83.3%. Most of the CTs mis-classified few fibro-lipidic regions as collagen, which is histologically acceptable.