Neural and statistical modeling of ultrasound backscatter

The work in this dissertation concerns neural and statistical analysis of ultrasound backscatter. Backscatter echo modeling in ultrasound (US) is an important task that can facilitate interpreting US B-scans (brightness scans). Backscatter echo can be modeled by different statistical distributions, such as Rayleigh, Nakagami, generalized Nakagami, Rician, K, generalized K, and homodyned K. Each of these distributions has specific parameters that can be clinically important in tissue characterization.; The K, generalized K, and homodyned K distributions are the most general models whose parameters are related to scatterer density, spacing, and amplitude. However, their probability density functions (pdfs) involve transcendental functions, and estimating their parameters cannot be performed analytically. Artificial neural networks (ANNs), especially kernel-function and recurrent networks, can be useful in parameter estimation. ANNs can uncover highly non-linear relationships between parameters and function values that are computed from the data (e.g. functions of moments). Furthermore, ANNs are robust with respect to noisy data, and therefore can be used as complementary devices to assist in backscatter characterization.; In this dissertation, hybrid methods using entropy, ANNs, and other mathematical constructs are presented for the parameter estimation of the general backscatter distributions. Visualization of ultrasound envelopes from simulated RF data based on parameter maps of Nakagami and K distribution parameters is also described. This approach may be used to complement B-scans in obtaining additional information from ultrasonographic data.; Finally, a new approach to characterization of ultrasound backscatter echo based on generalized entropies is introduced. This approach makes no assumptions about the specific scattering distribution. Low order Renyi and Tsallis entropies have a higher dynamic range than Shannon entropy with respect to a wide range of scattering conditions, and are therefore potentially useful in estimating scatterer density, regularity, and amplitude. A neural network estimator is constructed to illustrate the validity of this approach.