Predicting RF Heating of Conductive Leads During Magnetic Resonance Imaging at 1.5 T: A Machine Learning Approach

Abstract

The number of patients with active implantable medical devices continues to rise in the United States and around the world. It is estimated that 50-75% of patients with conductive implants will need magnetic resonance imaging (MRI) in their lifetime. A major risk of performing MRI in patients with elongated conductive implants is the radiofrequency (RF) heating of the tissue surrounding the implant's tip due to the antenna effect. Currently, applying full-wave electromagnetic simulation is the standard way to predict the interaction of MRI RF fields with the human body in the presence of conductive implants; however, these simulations are notoriously extensive in terms of memory requirement and computational time. Here we present a proof-of-concept simulation study to demonstrate the feasibility of applying machine learning to predict MRI-induced power deposition in the tissue surrounding conductive wires. We generated 600 clinically relevant trajectories of leads as observed in patients with cardiac conductive implants and trained a feedforward neural network to predict the 1g-averaged SAR at the lead tips knowing only the background field of MRI RF coil and coordinates of points along the lead trajectory. Training of the network was completed in 11.54 seconds and predictions were made within a second with R² = 0.87 and Root Mean Squared Error (RMSE) = 14.5 W/kg. Our results suggest that machine learning could provide a promising approach for safety assessment of MRI in patients with conductive leads.Clinical Relevance- Machine learning can potentially allow real-time assessment of MRI RF safety in patients with conductive leads when only the knowledge of lead's trajectory (image-based) and MRI RF coil features (vendor-specific) are in hand.

Figure 1.

Example of X-ray photograph of a patient with CIED overlaid on ANSYS human body model and manual trajectories with IPGs on right (A) as well as left (B).

Figure 2.

(A) Simulation setup in ANSYS HFSS showing homogeneous body model and MRI RF coil. The heart is shown to visualize the position of distal parts of leads and was not included in FEM simulations (B) Overlay of 600 trajectories in the body model (C) 1g-averaged SAR on a central axial plane within the 20 × 20× 20 mm³ cube surrounding the exposed lead’s tip.

Figure 3.

(A) Concatenation of 3D coordinate. (B) Structure of feedforward neural network; W_i and b_i represent weight and bias matrices for each layer; @ is followed by the number of neurons of every layer.

Figure 4.

Distribution of Simulated 1g-averaged SAR with IPGs in right as well as in left pectoral regions. Circles indicate the mean values.

Figure 5.

Training loss and validation loss with increasing epochs

Figure 6.

Performance of the feedforward neural network with predicted 1g-averaged SAR vs simulated 1g-averaged SAR. The coefficient of determination (R²) was relatively high (equals to 0.87).