Coil-Agnostic Attention-Based Network for Parallel MRI Reconstruction.

Magnetic resonance imaging (MRI) is widely used in clinical diagnosis. However, as a slow imaging modality, the long scan time hinders its development in time-critical applications. The acquisition process can be accelerated by types of under-sampling strategies in k-space and reconstructing images from a few measurements. To reconstruct the image, many parallel imaging methods use the coil sensitivity maps to fold multiple coil images with model-based or deep learning-based estimation methods. However, they can potentially suffer from the inaccuracy of sensitivity estimation. In this work, we propose a novel coil-agnostic attention-based framework for multi-coil MRI reconstruction which completely avoids the sensitivity estimation and performs data consistency (DC) via a sensitivity-agnostic data aggregation consistency block (DACB). Experiments were performed on the FastMRI knee dataset and show that the proposed DACB and attention module-integrated framework outperforms other deep learning-based algorithms in terms of image quality and reconstruction accuracy. Ablation studies also indicate the superiority of DACB over conventional DC methods.