Acceleration of Magnetic Resonance Fingerprinting Reconstruction Using Denoising and Self-Attention Pyramidal Convolutional Neural Network

Abstract

Magnetic resonance fingerprinting (MRF) based on echo-planar imaging (EPI) enables whole-brain imaging to rapidly obtain T1 and T2* relaxation time maps. Reconstructing parametric maps from the MRF scanned baselines by the inner-product method is computationally expensive. We aimed to accelerate the reconstruction of parametric maps for MRF-EPI by using a deep learning model. The proposed approach uses a two-stage model that first eliminates noise and then regresses the parametric maps. Parametric maps obtained by dictionary matching were used as a reference and compared with the prediction results of the two-stage model. MRF-EPI scans were collected from 32 subjects. The signal-to-noise ratio increased significantly after the noise removal by the denoising model. For prediction with scans in the testing dataset, the mean absolute percentage errors between the standard and the final two-stage model were 3.1%, 3.2%, and 1.9% for T1, and 2.6%, 2.3%, and 2.8% for T2* in gray matter, white matter, and lesion locations, respectively. Our proposed two-stage deep learning model can effectively remove noise and accurately reconstruct MRF-EPI parametric maps, increasing the speed of reconstruction and reducing the storage space required by dictionaries.

Systematic review of reconstruction techniques for accelerated quantitative MRI

“…Section 3.1). A neural network $$$ \mathscr{H}\left(\overset{\check }{\boldsymbol{Z}}|\boldsymbol{\theta} \right) $$$ is used where the estimated parameter maps are the output of the neural network: $$$ \hat{\boldsymbol{X}}=\mathscr{H}\left(\overset{\check }{\boldsymbol{Z}}|\boldsymbol{\theta} \right) $$$ 38,42,85,132,178,209,224,247‐259 …”

“…A neural network (Ž|𝜽) is used where the estimated parameter maps are the output of the neural network: X = (Ž|𝜽). 38,42,85,132,178,209,224,[247][248][249][250][251][252][253][254][255][256][257][258][259] The network training procedure is similar to Equation ( 20) but with Z t as the set of (aliased) training images and W t denoting the artifact-free training parameter maps.…”

Deep learning‐assisted preclinical MR fingerprinting for sub‐millimeter T1 and T2 mapping of entire macaque brain

“…Many human MRF studies use fully convolutional network (FCN) as the hundreds‐fold more lightweight and computationally efficient replacement of the dictionary‐matching approach. FCNs extract effective frame‐course features from one fingerprint for parametric quantification 14–20 . These networks are typically trained on simulated fingerprint dictionaries, and no longer require the dictionaries once trained.…”

“…FCNs extract effective frame-course features from one fingerprint for parametric quantification. [14][15][16][17][18][19][20] These networks are typically trained on simulated fingerprint dictionaries, and no longer require the dictionaries once trained. Further, to allow 1D neural networks to handle the aliasing artifacts in experimental fingerprints, several solutions have been experimented, including reducing aliasing artifacts in acquisition 15,[20][21][22] and reconstruction stage, 19,23,24 and adding simulated aliasing artifacts to training data.…”

“…[14][15][16][17][18][19][20] These networks are typically trained on simulated fingerprint dictionaries, and no longer require the dictionaries once trained. Further, to allow 1D neural networks to handle the aliasing artifacts in experimental fingerprints, several solutions have been experimented, including reducing aliasing artifacts in acquisition 15,[20][21][22] and reconstruction stage, 19,23,24 and adding simulated aliasing artifacts to training data. 25,26 However, all these studies tested the 1D neural networks within the undersampling capacity of established MRF protocols.…”