On the influence of training data quality in k‐t BLAST reconstruction

This work investigated how the quality of prior information (i.e., data acquired during the training stage) influences k-t BLAST reconstruction. The impact of several factors, such as the amount of training data, the presence of spatial misregistration in the training data, and the effects of filtering, was investigated with simulations and in vivo data. It is shown that k-t BLAST outperforms sliding window reconstruction, even with very limited training data. By increasing the amount of training data, reconstruction error continues to decrease, albeit by a diminishing amount. However, an increased amount of training data also increases susceptibility to misregistration of the training data. Filtering of the training data with the goal of reducing truncation artifacts had only minor impact on reconstruction errors. Considering the balance among obtaining the most benefit from the training data, minimizing susceptibility to misregistration, and keeping data acquisition to a minimum, it is concluded that in cardiac imaging the training datasets should be limited to 10-20 profiles in k-space for a typical field of view. The training data may be acquired in a separate breathhold without much penalty, if care is taken to minimize misregistration, such as with a navigator.