ARD-Unet: An Attention-based Redisual Dense U-net for Accelerated Multi-modal MRI Reconstruction

In recent times, there has been a noteworthy focus on deep learning techniques for accelerating magnetic resonance imaging (MRI). It has been shown in recent research that by taking into account the redundancy between different modalities, the reconstruction of under-sampled MRI modality images can be more efficiently improved using a fully-sampled reference MRI modality. However, the long acquisition time of MRI can have negative effects on multi-modal reconstruction, such as patients' discomfort and motion artifacts, leading to spatial misalignment between modalities. To address these issues, this paper proposes a residual dense U-net for enhancing the quality of multi-modal reconstruction. Additionally, a dual-branch Squeeze-Excitation attention module is developed to effectively extract features, capturing valuable information in both spatial and channel-wise upsampling processes. Experimental results on testing datasets demonstrate that our proposed network achieves promising performance in reconstruction tasks, preserving more detailed brain tissue information and higher contrast compared to alternative methods.