Evolutionary Optimization of a Focused Ultrasound Propagation Predictor Neural Network

The search for the optimal treatment plan of a focused ultrasound based
procedure is a complex multi-modal problem, trying to deliver the
solution in clinically relevant time while not sacrificing the precision
bellow a critical threshold. To test a solution, a multitude of
computationally expensive simulations need to be evaluated, often
thousands of times. Recent renaissance of machine learning could provide
a solution to this. Indeed, a state-of-the-art neural predictor of the
Acoustic Propagation through a human skull was published recently,
speeding up the simulation significantly. The utilized architecture,
however, could use some improvements in precision. To explore their
design more deeply, we made an attempt to improve the solver by use of
an evolutionary algorithm, challenging the importance of different
building blocks. Utilizing Genetic Programming, we managed to improve
their solution significantly, resulting in a solver with approximately
an order of magnitude better RMSE of the predictor, while still
delivering solutions in reasonable time frame. Furthermore, a second
study was conducted to gauge the effects of the multi-resolution
encoding on precision of the network, providing interesting topics for
further research on the effects of the memory blocks and convolution
kernel sizes for PDE RCNN solvers.