Deep learning reconstruction for cardiac magnetic resonance fingerprinting T₁ and T₂ mapping

Abstract

To develop a deep learning method for rapidly reconstructing T 1 and T 2 maps from undersampled electrocardiogram (ECG) triggered cardiac magnetic resonance fingerprinting (cMRF) images. Methods: A neural network was developed that outputs T 1 and T 2 values when given a measured cMRF signal time course and cardiac RR interval times recorded by an ECG. Over 8 million cMRF signals, corresponding to 4000 random cardiac rhythms, were simulated for training. The training signals were corrupted by simulated k-space undersampling artifacts and random phase shifts to promote robust learning. The deep learning reconstruction was evaluated in Monte Carlo simulations for a variety of cardiac rhythms and compared with dictionary-based pattern matching in 58 healthy subjects at 1.5T. Results: In simulations, the normalized root-mean-square error (nRMSE) for T 1 was below 1% in myocardium, blood, and liver for all tested heart rates. For T 2 , the nRMSE was below 4% for myocardium and liver and below 6% for blood for all heart rates. The difference in the mean myocardial T 1 or T 2 observed in vivo between dictionary matching and deep learning was 3.6 ms for T 1 and −0.2 ms for T 2. Whereas dictionary generation and pattern matching required more than 4 min per slice, the deep learning reconstruction only required 336 ms. Conclusion: A neural network is introduced for reconstructing cMRF T 1 and T 2 maps directly from undersampled spiral images in under 400 ms and is robust to arbitrary cardiac rhythms, which paves the way for rapid online display of cMRF maps. K E Y W O R D S deep learning, magnetic resonance fingerprinting, neural network, T 1 mapping, T 2 mapping, tissue characterization 2128 | HAMILTON eT AL.

Deep Learning and Its Application to Function Approximation for MR in Medicine: An Overview

“…In the case of MR fingerprinting, 117 ANNs approximated a function that received a series of images as the inputs and produced tissue H. Takeshima parameters as the outputs; 118 and a function that received a series of images and cardiac R-R intervals as the inputs, and produced tissue parameters as the outputs. 119 For estimating T1 and T2 values using a MOLLI sequence, 120 a DNN was used to approximate a function that received a time series of 1-dimensional signals and time stamps as the inputs, and produced T1 and T2 values as the outputs. 121 DNNs were used for estimating electrical properties tomography (EPT) by approximating functions, which produce EPT from B1+ amplitudes, transceiver phase, and existence of tissue.…”

Technical overview of magnetic resonance fingerprinting and its applications in radiation therapy

“…Therefore, it is critical to calculate the MRF dictionary efficiently for cardiac imaging to accelerate postprocessing. 27 The original 2D MRF method has recently been extended for volumetric quantitative imaging with improved spatial coverage. With a stack-of -spirals trajectory, Ma et al developed a 3D MRF method for quantitative T 1 , T 2, and M 0 mapping.…”

“…Depending on the computation resources and the size of the MRF dictionary, the time for generating one whole dictionary varies from a few seconds up to tens of minutes. Therefore, it is critical to calculate the MRF dictionary efficiently for cardiac imaging to accelerate postprocessing 27 …”

Accelerated cardiac T1 mapping in four heartbeats with inline MyoMapNet: a deep learning-based T1 estimation approach

“…They extended this algorithm using deep learning (DL) for rapid T 1 map reconstruction [30]. Similarly, Zhang et al and Hamilton et al used DL to rapidly reconstruct T 1 and T 2 maps from images collected using MR fingerprinting [29,31]. To reduce motion artifacts, an interleaved T 1 mapping sequence with radial sampling used a convolutional neural network model to reconstruct highly accelerated T 1 -weighted image to minimize the acquisition window of the single-shot image [32].…”