High frequency full-waveform inversion with deep learning for seismic and medical ultrasound imaging

Sound speed inversion is a key problem in seismic imaging for the purpose of high-resolution imaging and ground structure analysis. In the medical ultrasound imaging domain, longitudinal sound speed recovery, as opposed to shear wave velocity methods, has seen little use, mostly due to the resources required. However both research and new hardware are showing interesting applications and implications (Nebojsa Duric and Littrup 2018) Classic full waveform inversion (FWI) and Travel time tomography methods require large time and computational resources, and often, human in the loop interaction. This makes them inapplicable for most medical imaging applications. FWI also generally requires a good initial condition and low-frequency content in the signal for stable convergence. In this work we analyze the applicability of a deep-learning based approach for high frequency FWI. We present results on single-shot recovery using simulated data employing characteristic parameters for both medical ultrasound and seismic datasets. Results show great potential for interactive frame rate approximate solution.