Jianfeng Zheng1, Qianlong Lan1, Wolfgang Kainz2, Stuart A Long1, Ji Chen1. 1. Department of Electrical and Computer Engineering, University of Houston, Houston, Texas, USA. 2. Center for Devices and Radiological Health, Food and Drug Administration, Rockville, Maryland, USA.
Abstract
PURPOSE: This paper presents a method to search for the worst-case configuration leading to the highest RF exposure for a multiconfiguration implantable fixation system under MRI. METHODS: A two-step method combining an artificial neural network and a genetic algorithm is developed to achieve this purpose. In the first step, the level of RF exposure in terms of peak 1-g and/or 10-g averaged specific absorption rate (SAR1g/10g ), related to the multiconfiguration system, is predicted using an artificial neural network. A genetic algorithm is then used to search for the worst-case configuration of this multidimensional nonlinear problem within both the enumerated discrete sample space and generalized continuous sample space. As an example, a generic plate system with a total of 576 configurations is used for both 1.5T and 3T MRI systems. RESULTS: The presented method can effectively identify the worst-case configuration and accurately predict the SAR1g/10g with no more than 20% of the samples in the studied discrete sample space, and can even predict the worst case in the generalized continuous sample space. The worst-case prediction error in the generalized continuous sample space is less than 1.6% for SAR1g and less than 1.3% for SAR10g compared with the simulation results. CONCLUSION: The combination of an artificial neural network with genetic algorithm is a robust technique to determine the worst-case RF exposure level for a multiconfiguration system, and only needs a small amount of training data from the entire system.
PURPOSE: This paper presents a method to search for the worst-case configuration leading to the highest RF exposure for a multiconfiguration implantable fixation system under MRI. METHODS: A two-step method combining an artificial neural network and a genetic algorithm is developed to achieve this purpose. In the first step, the level of RF exposure in terms of peak 1-g and/or 10-g averaged specific absorption rate (SAR1g/10g ), related to the multiconfiguration system, is predicted using an artificial neural network. A genetic algorithm is then used to search for the worst-case configuration of this multidimensional nonlinear problem within both the enumerated discrete sample space and generalized continuous sample space. As an example, a generic plate system with a total of 576 configurations is used for both 1.5T and 3T MRI systems. RESULTS: The presented method can effectively identify the worst-case configuration and accurately predict the SAR1g/10g with no more than 20% of the samples in the studied discrete sample space, and can even predict the worst case in the generalized continuous sample space. The worst-case prediction error in the generalized continuous sample space is less than 1.6% for SAR1g and less than 1.3% for SAR10g compared with the simulation results. CONCLUSION: The combination of an artificial neural network with genetic algorithm is a robust technique to determine the worst-case RF exposure level for a multiconfiguration system, and only needs a small amount of training data from the entire system.