Autocorrelation Kernel Support Vector Machines for Doppler Ultrasound M:Mode Images Denoising

Abstract

Doppler ultrasound M-mode images are routinely used in clinical echocardiography, and they have been proposed for non-invasive estimation of the intracardiac pressure gradients in the heart, a process that has been shown to be sensitive to spline interpolation. In this work, we scrutinized the effect of interpolation with a new approach using support vector machines (SVM) for estimation using autocorrelation Mercer kernels in ultrasound images. The SVM algorithm was modified to provide the estimation of a whole image in terms of a reduced set of pixels used as training data set. The autocorrelation of the color Doppler M-mode (CDMM) image was estimated with conventional cross-correlation by considering the complete image, and it was used as the Mercer kernel required by SVM. Several subsampling strategies were scrutinized, namely, a heuristic approach, a criterion based on the edges, and a criterion based on the amplitudes. In order to evaluate the proposed methods, we analyzed a previously proposed Doppler image synthetic model, as well as case study with a real image. Results in terms of mean absolute error showed that the minimum error is obtained when information from the edges is considered, yielding 7.37 for single width radial basis function, 2.80 for double with radial basis function, and 0.80 for autocorrelation kernel. The autocorrelation kernel provides with an accurate estimation of the spatiotemporal distribution of flow velocity within the heart using CDMM images. This methodology can be further exploited for enhancing the sensitivity of the pressure gradients estimation to polynomial interpolation and then for improving noninvasive cardiovascular diagnosis.