Evaluating and Predicting Non-linear Acoustics Using Multi-Axis Gradient Variations and Convolutional Neural Networks

Motivation: MRI scanners produce substantial acoustic noise that may affect sensitive groups. Traditional acoustic models, assuming linear sound pressure level (SPL) scale and combine linearly along three separate axes, may oversimplify system dynamics. Goal(s): To evaluate SPL linearity and develop a predictive model that improves accuracy over traditional methods. Approach: We implemented a custom pulse sequence design to test gradient amplitudes and frequencies and used a convolutional neural network(CNN) to predict SPL changes on an ultra-high performance head-only system. Results: Non-linear SPL responses were observed; our CNN more accurately predicted SPL compared to linear models, demonstrating potential for improved patient comfort and image quality. Impact: This study challenges existing linear acoustic models in MRI, offering a more complex and accurate understanding of SPL dynamics. With improved acoustic noise estimation, more informed choices can be made when developing pulse sequences with reduced SPL.