Real-time MR cardiac image registration during respiration: A neural network approach

Real-time (RT), 2D MR technology is developing to guide cardiac interventional procedures. RT 2D MR, however is suffering from low resolution and image quality. To enhance RT 2D images, we propose to register RT MR images into MR images captured in breath-hold mode for only one cardiac cycle. But registration however is not that sharp to compensate respiratory motion in RT situation. Thus, neural network time series predictor is used to predict the heart displacement caused by respiratory motion. The entire framework was tested via three complex respiration simulations. The results show that the framework is reliable and can perform matching in limited RT circumstance. Mean misalignment for the proposed framework is less than 2 mm which is absolutely acceptable in clinical situation.