Purpose: Radio frequency power amplifier (RFPA) is a core part in MRI system. Distortion of the RFPA still occur during MRI scan, which leads to ghost artifacts and eventually affect disease diagnosis. A pre-distortion system for linearization of RFPA is needed clinically.
Methods: In this abstract, we presented an artificial neural network structure based on indirect learning for pre-distortion of RFPA, which included two multi-layer perceptrons (artificial neural network). One is used as a distortion simulator to train the network coefficients, the other is applied as pre-distorter of the RFPA with tuned coefficients from the distortion simulator. The input of the distortion simulator is from the RFPA via power divider and the input of the pre-distorter is the forward signal from MRI signal generator. Error value is calculated as mean square error between distortion simulator and pre-distorter. Levenberg-Marquardt algorithm and Bayesian regularization were used to iteratively update the coefficients of distortion simulator based on error value and eliminate overfitting. And the coefficients of distortion simulator were provided to the pre-distorter simultaneously for real-time pre-distortion of the RFPA. The artificial neural network structure was downloaded to an FPGA chip in an experimental RFPA system with signal generator, RFPA, FPGA chip and frequency conversion module to simulate a real-time MRI RFPA system.
Results: The artificial neural network structure was tested in the experimental RFPA system. Results showed that the proposed neural network structure can improve system performance and reduce adjacent channel error power ratio about 25 dB. Bayesian regularization can ensure continuous convergence of the algorithm. Phase deviation of the RFPA was close to 0, and linearity of the RFPA was enhanced.
Conclusion: Pre-distortion for MRI RFPA linearization can be achieved in real-time using an artificial neural network structure based on indirect learning.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by National Key Research and Development Program of China (2016YFC0103400), Key Research and Development Program of Shandong Province (2017GGX201010), Jianfeng Q. was supported by the Taishan Scholars Program of Shandong Province (TS201712065).