Prediction of MRI RF Exposure for Implantable Plate Devices Using Artificial Neural Network

Abstract

The purpose of this paper is to present a fast method to predict the radio frequency (RF) exposure for multi-configuration implantable devices in the magnetic resonance imaging (MRI) environment by using an artificial neural network (ANN). A synthesizing framework is developed to improve the ANN, of which the inputs are the geometric dimensions of the targeted device and the output is the peak 1-g average SAR (SAR1g) of the device. The synthesizing framework integrates feature selection and performance optimization techniques, achieved by using the mean impact value (MIV) algorithm and genetic algorithm (GA), to identify the most impactful features and improve the performance of the ANN. The framework was implemented and validated with the samples of 576 implantable plate devices with various geometric dimensions. The dimensions of the device were determined by six parameters and the peak SAR1g of the devices was numerically calculated by using a full-wave electromagnetic solver based on the finite-difference time-domain method. The efficiency and accuracy of the proposed framework were systematically evaluated. Comparing with the unimproved ANN, the mean square error (MSE) of the predicted peak SAR1g decreased by 28.06% for the ANN using MIV algorithm, while the MSE decreased by 55.29% for the presented synthesizing framework. The MSE of the predicted peak SAR1g was 8.16 W2/kg2 and the correlation between the predicted and calculated peak SAR1g was 0.994 for the improved ANN in the studied cases.