MRI reconstruction from truncated data using a complex domain backpropagation neural network

We propose a new data (extrapolation) modeling approach for reconstructing truncated magnetic resonance (MR) data. In our method, available low-frequency MR data are used to train a complex domain backpropagation neural network. This network is used to extrapolate the MR data and recover the missing high-frequency components. The performance of the proposed approach is demonstrated with a comparison to an existing real-valued neural network based method. Better results are obtained with the new approach because the complex-valued network makes use of the correlated information in the complex data instead of treating the data as separate real and imaginary parts