Region-Guided Channel-Wise Attention Network for Accelerated MRI Reconstruction

Magnetic resonance imaging (MRI) has been widely used in clinical practice for medical diagnosis of diseases. However, the long acquisition time hinders its development in time-critical applications. In recent years, deep learning-based methods leverage the powerful representations of neural networks to recover high-quality MR images from undersampled measurements, which shortens the acquisition process and enables accelerated MRI scanning. Despite the achieved inspiring success, it is still challenging to provide high-fidelity reconstructions under high acceleration factors. As an important mechanism in deep neural networks, attention modules have been used to improve the reconstruction quality. Due to the computational costs, many attention modules are not suitable for applying to high-resolution features or to capture spatial information, which potentially limits the capacity of neural networks. To address this issue, we propose a novel channel-wise attention which is implemented under the guidance of implicitly learned spatial semantics. We incorporate the proposed attention module in a deep network cascade for fast MRI reconstruction. In experiments, we demonstrate that the proposed framework produces superior reconstructions with appealing local visual details, compared to other deep learning-based models, validated qualitatively and quantitatively on the FastMRI knee dataset.