MRI RF-induced Heating Prediction for Complex-shaped Passive Implantable Devices using Neural Network Methods
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Magnetic Resonance Imaging (MRI) Radiofrequency (RF) -induced heating are one of the major safety concerns for patients with Passive Implantable Medical Devices (PIMDs) implanted inside the body. Evaluation of RF-induced heating includes experimental measurements and full-wave Electromagnetic (EM) simulations which will cost a significant amount of time and computational resources. Neural Network (NN) methods are introduced as a data-driven regression model which trains the parameters using device features to implement the predictions of RF-induced heating. While the previous NN models cannot predict the RF-induced heating of complex-shaped PIMDs as the device structure cannot be characterized by several parameters. Also, no discussions have been made on the strategy of training dataset selection. In this study, mesh-based Convolutional Neural Network (CNN) models are introduced to implement the heating prediction for complex-shaped PIMDs. Tibia Plating System and Spinal Fixation System device models are developed with variations on geometrical features. In-vitro and in-vivo EM simulations are performed with device mesh information and peak SAR values extracted. Incident Electric field information and mesh information are combined to form the input of the CNN models. After training and testing, CNN model convergence and data correlations are observed as a metric of CNN general prediction efficacy. The distribution of absolute errors and absolute percentage errors are shown to further investigate the prediction performance for different data. For the selection of training dataset, naïve strategy is initially introduced which uses different sizes of training dataset to find the training dataset with good prediction performance and least data possible. Principal Component Analysis (PCA) is performed on the input matrices of CNN model, which provides a threshold as the best prediction performance a training dataset can achieve with the dataset size same with the number of top-ranking Principle Components (PCs). Overall, mesh-based CNN models can predict the RF-induced heating of some complex-shaped PIMDs with acceptable performance. With the guidance of PCA analysis, the optimal size of training dataset can be determined before the simulations are performed which can save a lot of time used by obtaining excessive simulation results.