CDF-Net: Cross-Domain Fusion Network for Accelerated MRI Reconstruction

Abstract

Accurate reconstruction of accelerated Magnetic Resonance Imaging (MRI) would produce myriad clinical benefits including higher patient throughput and lower examination cost. Traditional approaches utilize statistical methods in the frequency domain combined with Inverse Discrete Fourier Transform (IDFT) to interpolate the under-sampled frequency domain (referred as k-space) and often result in large artifacts in spatial domain. Recent advances in deep learning-based methods for MRI reconstruction, albeit outperforming traditional methods, fail to incorporate raw coil data and spatial domain data in an end-to-end manner. In this paper, we introduce a cross-domain fusion network (CDF-Net), a neural network architecture that recreates high resolution MRI reconstructions from an under-sampled single-coil k-space by taking advantage of relationships in both the frequency and spatial domains while also having an awareness of which frequencies have been omitted. CDF-Net consists of three main components, a U-Net variant operating on the spatial domain, another U-Net performing inpainting in k-space, and a ‘frequency informed’ U-Net variant merging the two reconstructions as well as a skip connected zero-filled reconstruction. The proposed CDF-Net represents one of the first end-to-end MRI reconstruction network that leverages relationships in both k-space and the spatial domain with a novel ‘frequency information pathway’ that allows information about missing frequencies to flow into the spatial domain. Trained on the largest public fastMRI dataset, CDF-Net outperforms both traditional statistical interpolation and deep learning-based methods by a large margin.