A Denoising UNet Model with ConvNeXt Block for MRI Reconstruction

Magnetic Resonance Imaging (MRI) is an important auxiliary tool in clinical medicine, but its long imaging time remains a significant issue, making the reconstruction of high-quality images from fewer k-space data a key research direction. Recently, deep learning has achieved promising results in the field of MRI reconstruction. However, most current deep learning methods are based on traditional ConvNet, which are limited by their receptive field size and unable to capture global features of the image. Thus, this paper proposes a novel model named MRI ConvNeXt Blocks Network (MICXN). The model consists of two main components: (1) a V-Net-like architecture with ConvNeXt blocks (UCA), serving as the main part of the denoising network. This uses ConvNeXt blocks to provide larger convolutional kernels and receptive fields, better-preserving image details, and incorporates skip connections between the encoder and decoder to merge specific hierarchical features; (2) a Data Consistency (DC) layer to ensure the integrity of the original k-space data. We conducted both quantitative evaluation and qualitative analysis of our method on the Calgary-Campinas (CC) dataset, comparing it with other deep learning approaches. The experimental results indicate that our method achieves superior reconstruction performance even with low-sampled k-space data.