Limitations of Tissue Differentiation by Spectral Measures of Backscattered Ultrasound

A set of spectral measures and two decision algorithms have been investigated with respect to classification accuracy and robustness to measurement variables. Liver and spleen in normal subjects were selected as test tissues. Tissue classification was based on eight conventional measures derived from the digital Fourier transform of the signal. Decision algorithms based on both Euclidean and Mahalanobis distances were investigated. Classification accuracy of 80 percent was achieved with the latter. Classification accuracy is limited by inter and intra-subject variance as well as by the interdependence of the measures. Nearly all the data lie in a two dimensional plane within the eight dimensional feature space. Further improvement in classification accuracy requires identification of a discriminant measure which is more orthogonal to that data plane.