Spectral-based quantitative ultrasound imaging: A model free approach

Abstract

Quantitative ultrasound (QUS) imaging can improve the diagnostic capabilities of ultrasound. Spectral-based QUS approaches utilize backscattered ultrasound signals to extract additional information about tissue state. These technique have traditionally relied on scattering models to differentiate tissue state. However, the use of machine learning approaches may obviate the need for models. A model-based and model-free (principle component analysis (PCA)) approach to spectral-based QUS were compared for their ability to differentiate fatty from non-fatty liver in a rabbit model. PCA was observed to provide better differentiation of fatty from non-fatty liver, i.e., PCA predicted fatty liver 86% of the time versus 36% for model-based. In addition, three calibration approaches for spectral-based QUS were compared: the traditional reference phantom, reference free, and an in situ calibration target. To test the calibration procedures, a phantom was scanned with and without a lossy layer placed on top and integrated backscatter coefficients (IBSCs) were calculated from the scattered data. Utilizing an in situ calibration target provided the ability to account for transmission losses and attenuation. The root mean square error between IBSCs estimated from the phantom with and without the lossy layer were 8.28 for the reference phantom versus 1.24 for the in situ calibration approach.