CineJENSE: Simultaneous Cine MRI Image Reconstruction and Sensitivity Map Estimation Using Neural Representations

Ziad Al-Haj Hemidi, Nora Vogt, Lucile Quillien, Christian Weihsbach, Mattias P. Heinrich, Julien Oster

Abstract: Parallel imaging (PI) techniques have enabled accelerated magnetic resonance imaging (MRI). However, real-time imaging is still hindered by the fact that PI methods typically show insufficient reconstruction performance for high acceleration factors and limited auto-calibration signals. In this study, we introduce CineJENSE, an unsupervised implicit neural representation (INR) network designed for simultaneous image reconstruction and sensitivity map estimation in cardiac 2D+t cine MRI. Expanding upon the recently proposed IMJENSE network for 2D data, our model simultaneously processes cine frames of the entire cardiac cycle. It does not only surpass the 2D IMJENSE model in terms of computational efficiency but also enhances reconstruction quality for the cardiac MRI reconstruction challenge (CMRxRecon) dataset. We hypothesize that this enhancement can be explained by the effective learning of robust data priors from spatiotemporal redundancies in undersampled raw data, which generalize well to unseen k-space regions. To prevent the propagation of errors from pre-computed coil sensitivities, the proposed network learns sensitivity maps in an end-to-end manner, utilizing low-resolution hash grid encodings to ensure the generation of smooth estimates, while maintaining low computation times.