Predicting MRI RF Exposure for Passive Implantable Medical Devices Using a Mesh-based Convolutional Neural Network

In this study, a mesh-based fast prediction method, using a convolutional neural network (CNN) was proposed to estimate the radio-frequency (RF) exposure for passive implantable medical devices (PIMDs) during magnetic resonance imaging (MRI). The mesh file from the FDTD grid was used as input for the CNN. The RF exposure, in terms of peak-spatial 10 gram (g) averaged specific absorption rate
$\boldsymbol{(_{\text{ps}}\text{SAR}_{10\mathrm{g}})}$
, was used as the output of the CNN. This method is implemented and validated with 576 generic orthopedic PIMDs. Among the 576 PIMDs, 403 were randomly selected and used as the training set for the CNN, while 173 were used to examine the validity of the
$\boldsymbol{_{\text{ps}}\text{SAR}_{10\mathrm{g}}}$
predicted by the CNN. The results show that the correlation between the estimation and the target
$\boldsymbol{_{\text{ps}}\text{SAR}_{10\mathrm{g}}}$
was
$\boldsymbol{\sim 0.9}$
and that the mean absolute percentage error was ~4% at 1.5 T and 3 T.