Deep Learning with Domain Adaptation for Accelerated Projection Reconstruction MR
Purpose: A radial k-space trajectory is one of well-established sampling trajectory in magnetic resonance imaging. However, the radial k-space trajectory requires a large number of radial lines for high-resolution reconstruction. Increasing the number of lines causes longer sampling times, making it more difficult for routine clinical use. If we reduce the radial lines to reduce the sampling time, streaking artifact patterns are unavoidable. To solve this problem, we propose a novel deep learning approach to reconstruct high-resolution MR images from the under-sampled k-space data. Methods: The proposed deep network estimates the streaking artifacts. Once the streaking artifacts are estimated, an artifact-free image is then obtained by subtracting the estimated streaking artifacts from the distorted image. In the case of the limited number of available radial acquisition data, we apply a domain adaptation scheme, which first pre-trains the network with a large number of x-ray computed tomography (CT) data sets and then fine-tunes it with only a few MR data sets. Results: The proposed deep learning method shows better performance than the existing compressed sensing algorithms, such as total variation and PR-FOCUSS. In addition, the calculation time is several order of magnitude faster than total variation and PR-FOCUSS methods. Conclusion: The proposed deep learning method surpasses the image quality as well as the computation times against the existing compressed sensing algorithms. In addition, we demonstrate the possibilities of domain-adaptation approach when a limited number of MR data is available.
READ FULL TEXT