Classification of lower limb arterial stenoses from doppler blood flow signal analysis with time-frequency representation and pattern recognition techniques

Abstract

A pattern recognition system was used to classify Doppler blood flow signals for the determination of lower limb arterial stenoses. The diagnostic features were extracted from time-frequency representations of Doppler signals. Three techniques were tested to estimate time-frequency representations: the short-time Fourier transform, the autoregressive (AR) modeling, and the Bessel distribution. A boundary tracking algorithm was proposed to extract the frequency contour of the Doppler time-frequency representations. Based on the characteristics of the Doppler frequency contour, shape descriptors from an autoregressive analysis were proposed as diagnostic features. Simple algorithms were proposed to normalize these autoregressive shape descriptors. Amplitude distribution of the Doppler time-frequency representation was also found useful for stenosis classification. A total of 379 arterial segments from the aorta to the popliteal artery were classified by the pattern recognition system into three categories of diameter reduction (0–19%, 20–49%, and 50–99%). The short-time Fourier transform provided an overall accuracy of 80% (kappa = 0.38); AR modeling, 81% (kappa = 0.42); and the Bessel distribution, 82% (kappa = 0.43). All these results are better than those based on visual interpretation (accuracy = 76%, kappa = 0.29) performed by a trained technologist. The AR modeling and the Bessel distribution improved the performance of the pattern recognition system in comparison with the short-time Fourier transform. It is likely that with further improvement, the pattern recognition approach will be a useful clinical tool to quantify stenoses and to follow the disease progression with more reliability and less bias than visual interpretation as done currently in clinical practice.