Dynamic Convolution Unrolling Networks for Cardiac Cine MR Reconstruction

Motivation: Effective reconstruction of complex non-rigid cardiac motion in CINE MRI remains a challenge for traditional CNN models. Goal(s): To propose the Dynamic Convolution Unrolling Networks (DCUN) approach for improved accuracy and efficiency in cardiac CINE MRI reconstruction. Approach: Our method utilizes dynamically adaptive convolutional kernels and the KernelWarehouse method to enhance parameter utilization and address motion challenges. Results: DCUN shows significant performance improvements over DL-ESPIRiT, achieving higher PSNR and SSIM metrics in reconstructions at 20x acceleration. Moreover, our transfer testing results demonstrate that DCUN maintains effective performance across diverse external datasets. Impact: This study introduces an efficient tool for cardiac CINE MRI reconstruction, enhancing image quality for clinical use and supporting accurate diagnoses. It also provides valuable insights for advancing future dynamic reconstruction methodologies in medical imaging.