Multi-Branch CNN-Based Prediction of MR RF-Induced Heating in Human Models with Pedicle Screw Implants

A multi-branch convolutional neural network (CNN) architecture is proposed for predicting the peak 1g averaged specific absorption rate (pSAR<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1g</inf>) in human models with pedicle screw systems exposed to the radiofrequency (RF) field from a 1.5T magnetic resonance (MR) scanner. This method leverages simulation data from realistic human models, which exhibit complex dielectric properties and heterogeneities, enhancing applicability to real-world scenarios. Using voxelized data from Sim4Life, key electrical parameters are extracted as network inputs. With a relatively small training dataset, the model achieves an R2 of 0.9562 on the training set and 0.8944 on the test set, demonstrating its effectiveness.