Dynamic Cardiac MRI Reconstruction via Separate Optimization of K-Space and Hybrid-Domian Spatial-Temporal Feature Fusion

Xinrui Guo, Liwen Li, Fumin Guo

Abstract: Cardiac magnetic resonance imaging (CMRI), with excellent soft-tissue contrast and non-invasive nature, is a pivotal clinical imaging technique for the diagnosis and treatment of heart diseases. Deep learning reconstruction methods developed to accelerate CMRI are typically trained and deployed for specific scenarios, which struggles to adapt to the diverse range of clinical situations encountered in practice. These challenges arise due to the variability in multiple contrasts, imaging views, vendors, undersampling patterns, and acceleration rates. To address these issues, we developed an unrolled framework that applies separated optimization on center and periphery of k-space, i.e., the low-frequency and high-frequency components of image, to achieve robust reconstruction of diverse undersampled cardiac data. The proposed approach leverages the information of adjacent frames to complement the k-space center, and uses the designed spatio-temporal reconstruction blocks to mine the prior information about the image details contained in the periphery k-space. Overall image optimization is also incorporated to further improve the reconstruction quality.

External IDs:dblp:conf/miccai/GuoLG24