Evolutionary Exploration of a Neural Network for Predicting Ultrasound Propagation

The search for the optimal treatment plan in a focused ultrasound-based procedure is a complex multi-modal problem. It aims to deliver a solution within a clinically relevant time frame while maintaining precision above a critical threshold. We must balance clinical speed with precision. Machine learning offers a promising solution, as a recent neural predictor for acoustic skull propagation speeds up simulations significantly. To delve deeper into the design, we attempted to improve the solver using an evolutionary algorithm, questioning the significance of different building blocks. By utilizing Genetic Programming, we significantly enhanced the solution, resulting in a solver with approximately an order of magnitude better Root Mean Square Error (RMSE) for the predictor, all while delivering solutions within a reasonable time frame. Additionally, a second study explored the impact of multi-resolution encoding on network precision, offering insights for further research on memory blocks and convolution kernel sizes in Partial Differential Equation (PDE) Recurrent Convolutional Neural Network (RCNN) solvers.