Computer Vision News - May 2022

51 Elisabeth Preuhs in the fingerprints (Fig. 1). The results show that RNNs reconstruct high detail maps with clearly reduced artifacts compared to previous DL methods (Fig. 1). In addition, the reconstruction time can be reduced up to a factor 30 with RNNs compared to the SOTA method. In a second project, DL- based cardiac navigation for MRI of a beating heart was developed. A motion- resolved reconstruction (i.e., a movie of a beating heart) requires the knowledge of the cardiac motion in a scan. The ECG-sensor is the clinical SOTA to derive this motion, which is not always reliable within a MR scanner. My framework [ 2 ] directly uses parts of the acquiredMRI data of a scan as input to a fully convolutional neural network (FCNN) which estimates the R-wave timepoints, i.e., the cardiac phases (Fig. 2). This simplifies the workflow, as no external ECG-sensor needs to be attached to the subject. The results also showed, that by using the features from the MRI data instead of the electrical ECG-signal, limitations of the ECG-signal (like missed R-waves because of its interference with the magnetic fields) could have been overcome (Fig. 2).