Space-Time Deformable Attention Parallel Imaging Reconstruction for Highly Accelerated Cardiac MRI

Cardiac magnetic resonance imaging (MRI) provides excellent soft tissue contrast resolution and stands as a pivotal noninvasive modality for assessing cardiac structure and function. However, owing to the intricate balance between spatial and temporal resolution, the reconstruction of cardiac cine MRI sequences dedicated to the heart presents a more complex challenge compared to the swift reconstruction of general magnetic resonance images. While numerous deep learning techniques have emerged to MRI reconstruction, a majority of these endeavors have tended to overlook the dynamic nuances of cardiac motion. In response to this gap, we propose a Space-Time Deformable Attention Parallel Imaging Reconstruction (STDAPIR) framework. This approach is initially refined through the utilization of the Variational Network (VarNet), where the subsequently reconstructed high-frequency data serves as a means to attain enhanced precision in coil sensitivity map estimation. Then, we extend this framework through the integration of Nonlinear Activation Free Network (NAFNet), incorporating the Space-Time Deformable Attention (STDA) module to accommodate spatiotemporal considerations. By introducing these advancements, our methodology aims to elevate the quality of reconstructed images within the cardiac domain. Empirical findings gleaned from our experiments underscore the efficacy of our proposed method, revealing a notable enhancement in both precision and perceptual fidelity of the resulting reconstructed images.