Deep Image-Pass Filter for Accelerating Dynamic MRI

The unsupervised deep image prior (DIP) method has shown promising results in various applications such as image denoising and restoration. The core concept of DIP revolves around the inherent ability of convolutional neural networks (CNNs) to comprehend and preserve images. However, for accelerating dynamic MRI, the gradient flow in DIP only moves from sampled points in k-space to the network parameters, lacking a direct pathway from the image domain to the network parameters. To address this issue, we propose a novel approach, termed Deep Image-Pass Filter (DIPF). Specifically, DIPF sets the input of the network as a learnable image rather than random noise as in DIP. The consistency between input and output is constrained within the loss function, establishing a direct gradient flow pathway from the image domain to the network parameters. This design essentially transforms the network into a filter, selectively allowing the lossless passage of undistorted images. Experimental results demonstrate that DIPF outperforms DIP and other state-of-the-art unsupervised methods in the reconstruction of both long-axis and short-axis cardiac cine MRI.