Comparing the performance of two feedforward neural network training algorithms in MRI: reconstruction
Liam Chen, Michael R. Smith, Yin Pui Jason Hui · 2002
In dynamic magnetic resonance imaging, there is a trade-off between spatial and temporal resolutions. A number of techniques have been suggested to solve this problem. Yan and Mao (1993) have proposed a real valued feedforward neural network (FFNN) based nonlinear prediction algorithm to extrapolate truncated magnetic resonance data. Hui and Smith (1995) extended the method to the complex domain and obtained better results. This paper presents a preliminary results using the Levenberge-Marquardt algorithm instead of backpropagation algorithm to train the real valued FFNN. The results show that the real valued FFNN trained using Levenberge-Marquardt algorithm not only requires a smaller network size but also provides better prediction accuracy compared with the networks trained with the backpropagation algorithm used by other authors.