(TUSNet) A Deep-Learning Model for One-Shot Transcranial Ultrasound Simulation and Phase Aberration Correction | arXiv 

A 2D U-Net-based network where every block scans the input image (2D slice of the skull Ct slice, waveguide lines, and target) with LSTMs in four directions before convolving with the intuition that LSTM “memory” along the propagation direction should capture wave structure that local convs miss. The transducer and target are fed in as drawn into the image. Trained entirely on synthetic SkullGAN-generated skull slices but tested on real patient CTs.

98.3% focal pressure recovery and 0.18mm position error in 21ms (vs k-Wave’s 25.8 seconds) is good, but notice that targets were sampled from only 56 fixed grid points and its unclear how well it generalizes?