Data-Driven Speed-of-Sound Reconstruction for Medical Ultrasound: Impacts of Training Data Format and Imperfections on Convergence

Abstract

AbstractB-mode imaging is a qualitative method and its interpretation depends on users’ experience. Quantitative tissue information can increase precision and decrease user ambiguity. For example, Speed-of-Sound (SoS) in tissue is known to carry diagnostic information. Studies showed the possibility of SoS reconstruction from ultrasound raw data (a.k.a., RF data) using deep neural networks (DNNs). However, many ultrasound systems are designed to process demodulated data (i.e., IQ data) and often decimate data in early stages of acquisition. In this study we investigated the impacts of input data format and decimation on convergence of the DNNs for SoS reconstruction. Our results show that fully data-driven SoS reconstruction is possible using demodulated ultrasound data presented in Cartesian or Polar format using an encoder-decoder network. We performed a study using only amplitude and only phase information of ultrasound data for SoS reconstruction. Our results showed that distortion of the phase information results in inconsistent SoS predictions, indicating sensitivity of the investigated approach to phase information. We demonstrated that without losing significant accuracy, decimated IQ data can be used for SoS reconstruction.KeywordsSpeed-of-SoundDeep Neural NetworkUltrasound