Current browse context:
eess
Change to browse by:
References & Citations
Electrical Engineering and Systems Science > Image and Video Processing
Title: Compressive MR Fingerprinting reconstruction with Neural Proximal Gradient iterations
(Submitted on 27 Jun 2020 (v1), last revised 6 Jul 2020 (this version, v3))
Abstract: Consistency of the predictions with respect to the physical forward model is pivotal for reliably solving inverse problems. This consistency is mostly un-controlled in the current end-to-end deep learning methodologies proposed for the Magnetic Resonance Fingerprinting (MRF) problem. To address this, we propose ProxNet, a learned proximal gradient descent framework that directly incorporates the forward acquisition and Bloch dynamic models within a recurrent learning mechanism. The ProxNet adopts a compact neural proximal model for de-aliasing and quantitative inference, that can be flexibly trained on scarce MRF training datasets. Our numerical experiments show that the ProxNet can achieve a superior quantitative inference accuracy, much smaller storage requirement, and a comparable runtime to the recent deep learning MRF baselines, while being much faster than the dictionary matching schemes. Code has been released at this https URL
Submission history
From: Dongdong Chen [view email][v1] Sat, 27 Jun 2020 03:52:22 GMT (8188kb,D)
[v2] Tue, 30 Jun 2020 19:05:51 GMT (16338kb,D)
[v3] Mon, 6 Jul 2020 11:51:40 GMT (8169kb,D)
Link back to: arXiv, form interface, contact.