Observer efficiency in discrimination tasks simulating malignant and benign breast lesions with ultrasound

We investigate an ideal observer approach to signal processing in ultrasonic imaging. In two-class discrimination tasks of the sort explored in this work, the ideal observer approach rests on the use of the likelihood ratio as a test statistic. We derive this test statistic in the domain of the radio frequency (RF) signal under multivariate Gaussian assumptions, and we describe a power series approach for inverting the large covariance matrices that result. We also show how a Wiener-filter for deconvolution emerges from a first-order truncation of the power series. We then use the ideal observer approach to investigate performance in a number of tasks idealized from the use of ultrasonic imaging for the discrimination of malignant and benign breast tissue. We consider both standard B-mode processing, and the effect of Weiner filtering the RF data. We report the statistical efficiency of human observers in these tasks - as evaluated by psychophysical studies - with respect to the ideal observer. The ideal observer allows us to compute the statistical efficiency with which suboptimal observers - such as humans - perform these tasks, and how they are influenced by signal processing parameters.