Regularized autoregressive analysis of intravascular ultrasound backscatter: improvement in spatial accuracy of...

Autoregressive (AR) models are qualified for analysis of stochastic, short-time data, such as intravascular ultrasound (IVUS) backscatter. Regularization is required for AR analysis of short data lengths with an aim to increase spatial accuracy of predicted plaque composition and was achieved by determining suitable AR orders for short data records. Conventional methods of determining order were compared to the use of trend in the mean square error for determining order. Radio-frequency data from 101 fibrous, 56 fibro-lipidic, 50 calcified, and 70 lipid-core regions of interest (ROIs) were collected ex vivo from 51 human coronary arteries with 30 MHz unfocused IVUS transducers. Spectra were computed for AR model orders between 3-20 for data representing ROIs of two sizes (32 and 16 samples at 100 MHz sampling frequency) and were analyzed in the 17-42 MHz bandwidth. These spectra were characterized based on eight previously identified parameters. Statistical classification schemes were computed from 75% of the data and cross-validated with the remaining 25% using matched histology. The results determined the suitable AR order numbers for the two ROI sizes. Conventional methods of determining order did not perform well. Trend in the mean square error was identified as the most suitable factor for regularization of short record lengths.