Fig. 1.

Illustration of our proposed DNN goals (bottom) in comparison to the traditional approach (top). Traditionally, raw channel data undergo DAS beamforming followed by envelope detection, log compression, and filtering to produce an interpretable DAS beamformed image, which is then passed to a segmentation algorithm to isolate a desired segment of the image. We propose to replace this sequential process with an FCNN architecture, consisting of a single encoder and two decoders, which simultaneously outputs both a DNN image and a DNN segmentation directly from raw ultrasound channel data received after a single plane wave insonification. The input is in-phase/quadrature (IQ) ultrasound data, presented as a 3-D tensor.