MoSAIK (Motion Search with AI in K-space) is a method that retrospectively estimates in‑plane rigid‑body motion directly from standard multi‑coil TSE k‑space data. It requires no motion‑free reference, external navigators, or extra acquisitions; instead it exploits violations of k‑space self‑consistency to infer motion parameters and feeds the predicted rotation and translation into a model‑based, motion‑aware reconstruction pipeline rather than performing post‑hoc image correction.
Implementation-wise, the complex coil data are split into log‑magnitude and phase channels. A decoupled convolutional embedding extracts local features, which are then processed by a Transformer encoder that models the sequence of phase‑encoding line indices. The network outputs a rotation angle and a two‑dimensional translation vector for each phase‑encoding line. Training uses motion‑free clinical scans with synthetically added motion; evaluation is performed on four held‑out subjects.
Performance is measured by component‑wise error, structural similarity index (SSIM), peak signal‑to‑noise ratio (PSNR), and in‑vivo entropy focus criterion (EFC). Results show very low rotation error (MSE $<10^{-4}$) and sub‑pixel translation error. Reconstruction gains include an SSIM increase of ~0.03, a PSNR boost of ~1.5 dB, and a reduced EFC, indicating clearer images.
Review