SiamMAE-MRI for dynamic MR imaging

Dynamic magnetic resonance imaging (dMRI) aims to capture high-resolution temporal sequences of anatomical structures under severe k-space undersampling constraints. Inspired by masked autoencoders in computer vision, we propose SiamMAE-MRI, a novel reconstruction framework that leverages a time-average k-space data as a global spatial-temporal prior to guide the reconstruction of individual temporal frames. Unlike existing methods that require a fully-sampled reference frame, which is infeasible in dynamic MRI due to per-frame undersampling, our approach constructs a fully-sampled k-space by aggregating the time dimension via complementary sampling patterns. This time-average k-space is then used to reconstruct a guidance that serves as a reference input to a Siamese-style masked autoencoder. Each undersampled frame is treated as a masked input, and the model is trained to predict the missing pixels using cross-frame attention and spatial context fusion. Extensive experiments were conducted on 118 in-house 2D+t cardiac MR datasets using two distinct undersampling strategies: the VISTA mask and a newly introduced Sheared Grid Mask (SGM). The results demonstrate that our method consistently outperforms state-of-the-art baselines such as L+S, DcCNN, and k-GIN across various acceleration factors. For example, under R=4, our method achieves an average PSNR gain of 2.5 dB and SSIM improvement of 0.6. Even under highly challenging settings (R=8, SGM), our network preserves fine anatomical details and demonstrates strong robustness and generalization across mask types and dynamic scenarios.