Partial Fourier MRI: AR models with SVD

The authors investigated several partial Fourier imaging techniques to reconstruct magnetic resonance images from the central fifty percent phase encodings. Parametric models were chosen and compared for accuracy in predicting the full time-domain data, and for final image quality. The auto-regressive model, with parameter estimation from Single Value Decomposition (AR-SVD) was compared against the iterative Levinson-Durbin approach, and against auto-regressive moving-average models suggested by Marple (1987) and Smith (1986). Suitable MR images from a phantom were obtained for testing the models. AR-SVD reconstruction most closely matched the full image intensity function, giving the lowest MSE, especially with gradient-echo and fast spin echo images, thus demonstrating superior tolerance to noise, at the cost of increased computing load