Refine
Document Type
- Preprint (5)
- Conference Paper (3)
- Article (2)
Language
- English (10)
Is part of the Bibliography
- no (10)
Review
- nein (5)
- peer-review (4)
- ja (1)
Version
- published (2)
We study a deep learning approach to address the heavy storage and computation re- quirements of the baseline dictionary-matching (DM) for Magnetic Resonance Fingerprint- ing (MRF) reconstruction. The MRF-Net provides a piece-wise affine approximation to the (temporal) Bloch response manifold projection. Fed with non-iterated back-projected images, the network alone is unable to fully resolve spatially-correlated artefacts which ap- pear in highly undersampling regimes. We propose an accelerated iterative reconstruction to minimize these artefacts before feeding into the network. This is done through a convex regularization that jointly promotes spatio-temporal regularities of the MRF time-series.
Compressive MRI quantification using convex spatiotemporal priors and deep encoder-decoder networks
(2020)
This work proposes an end-to-end deep fully convolutional neural network for MRF reconstruction (MRF-FCNN), which firstly employs linear dimensionality reduction and then uses a neural network to project the data into the tissue parameters. The MRF dictionary is only used for training the network and not during image reconstruction. We show that MRF-FCNN is capable of achieving accuracy comparable to the ground-truth maps thanks to capturing spatio-temporal data structures without a need for the non-scalable dictionary matching step used in the baseline reconstructions.