PredictUS: A Method to Extend the Resolution-Precision Trade-Off in Quantitative Ultrasound Image Reconstruction

Abstract

We present PredictUS, a novel Quantitative Ultrasound (QUS) parameter estimation technique with improved resolution and precision using augmented ultrasound data. The ultrasound data is generated using a sequence-to-sequence convolutional neural network based on WaveNet. The spectral-based QUS techniques are limited by the well-studied trade-off between the precision of the estimated QUS parameters and the window size used in estimation, limiting the practical utility of the QUS techniques. In this paper, we present a method to increase the window size by predicting the next data points of a given window. The method provides better estimates of local tissue properties with high resolution by virtually extending the property to a larger region. Our proof-of-concept study based on attenuation coefficient estimate (ACE), an important QUS parameter, attains a resolution reduction up to 50% while maintaining comparable estimation precision. This result shows the promise to extend the precision-resolution trade-off, which, in turn, would have implications in small lesion detection or heterogeneous tissue characterization.