Does Ultrasonic Data Format Matter for Deep Neural Networks?

Abstract

Received ultrasonic data are carrier-modulated broadband signals and are converted to different formats depending on the application. Common formats extracted from the raw radio-frequency (RF) data include a complex-valued analytic signal, the envelope/magnitude, demodulated in-phase and quadrature (IQ) components, and the phase angle. Deep neural networks (DNNs) have been applied to a variety of ultrasound signal processing tasks, yet how the format of input data affects DNN results has not been well-characterized. Here, we investigate how the data format affects DNN performance and robustness for two tasks: speckle reduction and displacement estimation. Simulated data were used for training, and multiple networks were trained for each task and each input format. Network loss was compared on test data with either added white noise or a different imaging frequency. For speckle noise reduction, networks using magnitude or IQ data were more robust to changes in imaging frequency than those using the carrier-modulated RF or analytic signals. Networks using magnitude were the least robust against added white noise. For displacement estimation, networks required an input data format with phase information to perform well. Performance for all input formats were equally affected by added noise, but the RF and analytic signals were the most robust to changes in center frequency.