ReconNext: A Encoder-Decoder Skip Cross Attention Based Approach to Reconstruct Cardiac MRI

Cardiac magnetic resonance imaging (MRI) is an advanced medical imaging technique widely used for the diagnosis and assessment of cardiovascular diseases. However, the acquisition time for cardiac MRI is generally longer, and cardiac motion can easily introduce artifacts that negatively impact image quality. Reconstructing cardiac MRI from under-sampled K-Space data has emerged as a viable approach to effectively reduce cardiac MRI acquisition time. In recent years, methods based on Deep Learning have been employed for image denoising and dehazing, yielding promising results. In this paper, we propose an MRI reconstruction network called ReconNext, based on MedNext and Encoder-Decoder Skip Cross Attention (EDSCA). The backbone of the proposed network is built upon MedNext, which leverages large-kernel convolutions to achieve a balance between local and non-local detail reconstruction. We introduce an Encoder-Decoder Skip Cross Attention structure, akin to self-attention, into the connection between the encoder and decoder. EDSCA incorporates cross-attention inputs from both the encoder and decoder. Compared to traditional skip connection, EDSCA better integrates information from the encoder and decoder, addressing issues of information loss. Experiment results demonstrate that ReconNext outperforms other conventional encoder-decoder architectures in cardiac MRI reconstruction, showcasing superior reconstruction effectiveness.