3D System Matrix Calibration by Using Coil Information and Transformer

Abstract

System Matrix-based image reconstruction approach requires a time-consuming calibration measurement. Existing methods such as compressed sensing and deep learning-based methods treat each row of the system matrix as independent data sample and lack the ability to modelling the relationships between SM rows. We firstly propose to model SM row relationships by the coil position and frequency value, which can be regarded as the additional and multimodal information. we propose a transformer-based neural network for 3D fast SM calibration, which encodes the information of coil position and frequency value into SM with self-attention mechanism in transformer.