Effectively Training MRI Reconstruction Network via Sequentially Using Undersampled k-Space Data with Very Low Frequency Gaps

Convolutional Neural Networks (CNNs) have achieved great advances on Magnetic Resonance Imaging (MRI) reconstruction. However, CNNs are still suffering from significant aliasing artifacts for undersampled data with high acceleration rates. This is mainly due to the huge gap between the highly undersampled k-space data and its fully-sampled counterpart. To mitigate this problem, we constructed a series of well-organized undersampled k-space data, each of which has very small frequency gap with its neighbors. By sequentially using these undersampled data and their fully-sampled ones to train a given CNN model N, the model N can gradually know how to fill the progressively increased frequency gaps and thus reduce the aliasing artifacts. Experiments on the MSSEG dataset demonstrated the effectiveness of the proposed training method.