A Hybrid Vision Transformer-BiRNN Architecture for Direct k-Space to Image Reconstruction in Accelerated MRI

Long scan times remain a fundamental challenge in Magnetic Resonance Imaging (MRI). Accelerated MRI, which undersamples k-space, requires robust reconstruction methods to solve the ill-posed inverse problem. Recent methods have shown promise by processing image-domain features to capture global spatial context. However, these approaches are often limited, as they fail to fully leverage the unique, sequential characteristics of the k-space data themselves, which are critical for disentangling aliasing artifacts. This study introduces a novel, hybrid, dual-domain deep learning architecture that combines a ViT-based autoencoder with Bidirectional Recurrent Neural Networks (BiRNNs). The proposed architecture is designed to synergistically process information from both domains: it uses the ViT to learn features from image patches and the BiRNNs to model sequential dependencies directly from k-space data. We conducted a comprehensive comparative analysis against a standard ViT with only an MLP head (Model 1), a ViT autoencoder operating solely in the image domain (Model 2), and a competitive UNet baseline. Evaluations were performed on retrospectively undersampled neuro-MRI data using R = 4 and R = 8 acceleration factors with both regular and random sampling patterns. The proposed architecture demonstrated superior performance and robustness, significantly outperforming all other models in challenging high-acceleration and random-sampling scenarios. The results confirm that integrating sequential k-space processing via BiRNNs is critical for superior artifact suppression, offering a robust solution for accelerated MRI.