Fast and accurate calculation of myocardial T1 and T2 values using deep learning Bloch equation simulations (DeepBLESS)

MyoFold: Joint myocardial tissue composition and wall motion quantification via a highly folded sequence

PURPOSE

This study aims to develop and evaluate a novel cardiovascular MR sequence, MyoFold, designed for the simultaneous quantifications of myocardial tissue composition and wall motion.

METHODS

MyoFold is designed as a 2D single breathing-holding sequence, integrating joint T1/T2 mapping and cine imaging. The sequence uses a 2-fold accelerated balanced SSFP (bSSFP) for data readout and incorporates electrocardiogram synchronization to align with the cardiac cycle. MyoFold initially acquires six single-shot inversion-recovery images, completed during the diastole of six successive heartbeats. T2 preparation (T2-prep) is applied to introduce T2 weightings for the last three images. Subsequently, over the following six heartbeats, segmented bSSFP is performed for the movie of the entire cardiac cycle, synchronized with an electrocardiogram. A neural network trained using numerical simulations of MyoFold is used for T1 and T2 calculations. MyoFold was validated through phantom and in vivo experiments, with comparisons made against MOLLI, SASHA, T2-prep bSSFP, and the conventional cine.

RESULTS

In phantom studies, MyoFold exhibited a 10% overestimation in T1 measurements, whereas T2 measurements demonstrated high accuracy. In vivo experiments revealed that MyoFold T1 had comparable accuracy to SASHA and precision similar to MOLLI. MyoFold demonstrated good agreement with T2-prep bSSFP in myocardial T2 measurements. No significant differences were observed in the quantification of left-ventricle wall thickness and function between MyoFold and the conventional cine.

CONCLUSION

MyoFold presents as a rapid, simple, and multitasking approach for quantitative cardiovascular MR examinations, offering simultaneous assessment of tissue composition and wall motion. The sequence's multitasking capabilities make it a promising tool for comprehensive cardiac evaluations in clinical settings.

The role of artificial intelligence in cardiovascular magnetic resonance imaging

Cardiovascular magnetic resonance (CMR) imaging is the gold standard test for myocardial tissue characterization and chamber volumetric and functional evaluation. However, manual CMR analysis can be time-consuming and is subject to intra- and inter-observer variability. Artificial intelligence (AI) is a field that permits automated task performance through the identification of high-level and complex data relationships. In this review, we review the rapidly growing role of AI in CMR, including image acquisition, sequence prescription, artifact detection, reconstruction, segmentation, and data reporting and analysis including quantification of volumes, function, myocardial infarction (MI) and scar detection, and prediction of outcomes. We conclude with a discussion of the emerging challenges to widespread adoption and solutions that will allow for successful, broader uptake of this powerful technology.

Recurrent neural network-based simultaneous cardiac T1, T2, and T1ρ mapping

The purpose of the current study was to explore the feasibility of training a deep neural network to accelerate the process of generating T1, T2, and T1ρ maps for a recently proposed free-breathing cardiac multiparametric mapping technique, where a recurrent neural network (RNN) was utilized to exploit the temporal correlation among the multicontrast images. The RNN-based model was developed for rapid and accurate T1, T2, and T1ρ estimation. Bloch simulation was performed to simulate a dataset of more than 10 million signals and time correspondences with different noise levels for network training. The proposed RNN-based method was compared with a dictionary-matching method and a conventional mapping method to evaluate the model's effectiveness in phantom and in vivo studies at 3 T, respectively. In phantom studies, the RNN-based method and the dictionary-matching method achieved similar accuracy and precision in T1, T2, and T1ρ estimations. In in vivo studies, the estimated T1, T2, and T1ρ values obtained by the two methods achieved similar accuracy and precision for 10 healthy volunteers (T1: 1228.70 ± 53.80 vs. 1228.34 ± 52.91 ms, p > 0.1; T2: 40.70 ± 2.89 vs. 41.19 ± 2.91 ms, p > 0.1; T1ρ: 45.09 ± 4.47 vs. 45.23 ± 4.65 ms, p > 0.1). The RNN-based method can generate cardiac multiparameter quantitative maps simultaneously in just 2 s, achieving 60-fold acceleration compared with the dictionary-matching method. The RNN-accelerated method offers an almost instantaneous approach for reconstructing accurate T1, T2, and T1ρ maps, being much more efficient than the dictionary-matching method for the free-breathing multiparametric cardiac mapping technique, which may pave the way for inline mapping in clinical applications.

SwinUNet: a multiscale feature learning approach to cardiovascular magnetic resonance parametric mapping for myocardial tissue characterization

Objective.Cardiovascular magnetic resonance (CMR) can measure T1 and T2 relaxation times for myocardial tissue characterization. However, the CMR procedure for T1/T2 parametric mapping is time-consuming, making it challenging to scan heart patients routinely in clinical practice. This study aims to accelerate CMR parametric mapping with deep learning.Approach. A deep-learning model, SwinUNet, was developed to accelerate T1/T2 mapping. SwinUNet used a convolutional UNet and a Swin transformer to form a hierarchical 3D computation structure, allowing for analyzing CMR images spatially and temporally with multiscale feature learning. A comparative study was conducted between SwinUNet and an existing deep-learning model, MyoMapNet, which only used temporal analysis for parametric mapping. The T1/T2 mapping performance was evaluated globally using mean absolute error (MAE) and structural similarity index measure (SSIM). The clinical T1/T2 indices for characterizing the left-ventricle myocardial walls were also calculated and evaluated using correlation and Bland-Altman analysis.Main results. We performed accelerated T1 mapping with ≤4 heartbeats and T2 mapping with 2 heartbeats in reference to the clinical standard, which required 11 heartbeats for T1 mapping and 3 heartbeats for T2 mapping. SwinUNet performed well in all the experiments (MAE < 50 ms, SSIM > 0.8, correlation > 0.75, and Bland-Altman agreement limits < 100 ms for T1 mapping; MAE < 1 ms, SSIM > 0.9, correlation > 0.95, and Bland-Altman agreement limits < 1.5 ms for T2 mapping). When the maximal acceleration was used (2 heartbeats), SwinUNet outperformed MyoMapNet and gave measurement accuracy similar to the clinical standard.Significance. SwinUNet offers an optimal solution to CMR parametric mapping for assessing myocardial diseases quantitatively in clinical cardiology.

The future of cardiovascular magnetic resonance: All-in-one vs. real-time (Part 1)

Cardiovascular magnetic resonance (CMR) protocols can be lengthy and complex, which has driven the research community to develop new technologies to make these protocols more efficient and patient-friendly. Two different approaches to improving CMR have been proposed, specifically "all-in-one" CMR, where several contrasts and/or motion states are acquired simultaneously, and "real-time" CMR, in which the examination is accelerated to avoid the need for breathholding and/or cardiac gating. The goal of this two-part manuscript is to describe these two different types of emerging rapid CMR. To this end, the vision of each is described, along with techniques which have been devised and tested along the pathway of clinical implementation. The pros and cons of the different methods are presented, and the remaining open needs of each are detailed. Part 1 will tackle the "all-in-one" approaches, and Part 2 the "real-time" approaches along with an overall summary of these emerging methods.

Scanner-Independent MyoMapNet for Accelerated Cardiac MRI T1 Mapping Across Vendors and Field Strengths

BACKGROUND

In cardiac T1 mapping, a series of T1 -weighted (T1 w) images are collected and numerically fitted to a two or three-parameter model of the signal recovery to estimate voxel-wise T1 values. To reduce the scan time, one can collect fewer T1 w images, albeit at the cost of precision or/and accuracy. Recently, the feasibility of using a neural network instead of conventional two- or three-parameter fit modeling has been demonstrated. However, prior studies used data from a single vendor and field strength; therefore, the generalizability of the models has not been established.

PURPOSE

To develop and evaluate an accelerated cardiac T1 mapping approach based on MyoMapNet, a convolution neural network T1 estimator that can be used across different vendors and field strengths by incorporating the relevant scanner information as additional inputs to the model.

STUDY TYPE

Retrospective, multicenter.

POPULATION

A total of 1423 patients with known or suspected cardiac disease (808 male, 57 ± 16 years), from three centers, two vendors (Siemens, Philips), and two field strengths (1.5 T, 3 T). The data were randomly split into 60% training, 20% validation, and 20% testing.

FIELD STRENGTH/SEQUENCE

A 1.5 T and 3 T, Modified Look-Locker inversion recovery (MOLLI) for native and postcontrast T1 .

ASSESSMENT

Scanner-independent MyoMapNet (SI-MyoMapNet) was developed by altering the deep learning (DL) architecture of MyoMapNet to incorporate scanner vendor and field strength as inputs. Epicardial and endocardial contours and blood pool (by manually drawing a large region of interest in the blood pool) of the left ventricle were manually delineated by three readers, with 2, 8, and 9 years of experience, and SI-MyoMapNet myocardial and blood pool T1 values (calculated from four T1 w images) were compared with conventional MOLLI T1 values (calculated from 8 to 11 T1 w images).

STATISTICAL TESTS

Equivalency test with 95% confidence interval (CI), linear regression slope, Pearson correlation coefficient (r), Bland-Altman analysis.

RESULTS

The proposed SI-MyoMapNet successfully created T1 maps. Native and postcontrast T1 values measured from SI-MyoMapNet were strongly correlated with MOLLI, despite using only four T1 w images, at both field-strengths and vendors (all r > 0.86). For native T1 , SI-MyoMapNet and MOLLI were in good agreement for myocardial and blood T1 values in institution 1 (myocardium: 5 msec, 95% CI [3, 8]; blood: -10 msec, 95%CI [-16, -4]), in institution 2 (myocardium: 6 msec, 95% CI [0, 11]; blood: 0 msec, [-18, 17]), and in institution 3 (myocardium: 7 msec, 95% CI [-8, 22]; blood: 8 msec, [-14, 30]). Similar results were observed for postcontrast T1 .

DATA CONCLUSION

Inclusion of field strength and vendor as additional inputs to the DL architecture allows generalizability of MyoMapNet across different vendors or field strength.

EVIDENCE LEVEL

2.

TECHNICAL EFFICACY

Stage 2.

DeepFittingNet: A deep neural network-based approach for simplifying cardiac T1 and T2 estimation with improved robustness

PURPOSE

To develop and evaluate a deep neural network (DeepFittingNet) for T1 /T2 estimation of the most commonly used cardiovascular MR mapping sequences to simplify data processing and improve robustness.

THEORY AND METHODS

DeepFittingNet is a 1D neural network composed of a recurrent neural network (RNN) and a fully connected (FCNN) neural network, in which RNN adapts to the different number of input signals from various sequences and FCNN subsequently predicts A, B, and Tx of a three-parameter model. DeepFittingNet was trained using Bloch-equation simulations of MOLLI and saturation-recovery single-shot acquisition (SASHA) T1 mapping sequences, and T2 -prepared balanced SSFP (T2 -prep bSSFP) T2 mapping sequence, with reference values from the curve-fitting method. Several imaging confounders were simulated to improve robustness. The trained DeepFittingNet was tested using phantom and in-vivo signals, and compared to the curve-fitting algorithm.

RESULTS

In testing, DeepFittingNet performed T1 /T2 estimation of four sequences with improved robustness in inversion-recovery T1 estimation. The mean bias in phantom T1 and T2 between the curve-fitting and DeepFittingNet was smaller than 30 and 1 ms, respectively. Excellent agreements between both methods was found in the left ventricle and septum T1 /T2 with a mean bias <6 ms. There was no significant difference in the SD of both the left ventricle and septum T1 /T2 between the two methods.

CONCLUSION

DeepFittingNet trained with simulations of MOLLI, SASHA, and T2 -prep bSSFP performed T1 /T2 estimation tasks for all these most used sequences. Compared with the curve-fitting algorithm, DeepFittingNet improved the robustness for inversion-recovery T1 estimation and had comparable performance in terms of accuracy and precision.

Uncertainty-aware physics-driven deep learning network for free-breathing liver fat and R2 * quantification using self-gated stack-of-radial MRI

PURPOSE

To develop a deep learning-based method for rapid liver proton-density fat fraction (PDFF) and R2 * quantification with built-in uncertainty estimation using self-gated free-breathing stack-of-radial MRI.

METHODS

This work developed an uncertainty-aware physics-driven deep learning network (UP-Net) to (1) suppress radial streaking artifacts because of undersampling after self-gating, (2) calculate accurate quantitative maps, and (3) provide pixel-wise uncertainty maps. UP-Net incorporated a phase augmentation strategy, generative adversarial network architecture, and an MRI physics loss term based on a fat-water and R2 * signal model. UP-Net was trained and tested using free-breathing multi-echo stack-of-radial MRI data from 105 subjects. UP-Net uncertainty scores were calibrated in a validation dataset and used to predict quantification errors for liver PDFF and R2 * in a testing dataset.

RESULTS

Compared with images reconstructed using compressed sensing (CS), UP-Net achieved structural similarity index >0.87 and normalized root mean squared error <0.18. Compared with reference quantitative maps generated using CS and graph-cut (GC) algorithms, UP-Net achieved low mean differences (MD) for liver PDFF (-0.36%) and R2 * (-0.37 s-1 ). Compared with breath-holding Cartesian MRI results, UP-Net achieved low MD for liver PDFF (0.53%) and R2 * (6.75 s-1 ). UP-Net uncertainty scores predicted absolute liver PDFF and R2 * errors with low MD of 0.27% and 0.12 s-1 compared to CS + GC results. The computational time for UP-Net was 79 ms/slice, whereas CS + GC required 3.2 min/slice.

CONCLUSION

UP-Net rapidly calculates accurate liver PDFF and R2 * maps from self-gated free-breathing stack-of-radial MRI. The pixel-wise uncertainty maps from UP-Net predict quantification errors in the liver.

Rapid anatomical imaging of the neonatal brain using T2 -prepared 3D balanced steady-state free precession

PURPOSE

To develop a new approach to 3D gradient echo-based anatomical imaging of the neonatal brain with a substantially shorter scan time than standard 3D fast spin echo (FSE) methods, while maintaining a high SNR.

METHODS

T2 -prepration was employed immediately prior to image acquisition of 3D balanced steady-state free precession (bSSFP) with a single trajectory of center-out k-space view ordering, which requires no magnetization recovery time between imaging segments during the scan. This approach was compared with 3D FSE, 2D single-shot FSE, and product 3D bSSFP imaging in numerical simulations, plus phantom and in vivo experiments.

RESULTS

T2 -prepared 3D bSSFP generated image contrast of gray matter, white matter, and CSF very similar to that of reference T2 -weighted imaging methods, without major image artifacts. Scan time of T2 -prepared 3D bSSFP was remarkably shorter compared to 3D FSE, whereas SNR was comparable to that of 3D FSE and higher than that of 2D single-shot FSE. Specific absorption rate of T2 -prepared 3D bSSFP remained within the safety limit. Determining an optimal imaging flip angle of T2 -prepared 3D bSSFP was critical to minimizing blurring of images.

CONCLUSION

T2 -prepared 3D bSSFP offers an alternative method for anatomical imaging of the neonatal brain with dramatically reduced scan time compared to standard 3D FSE and higher SNR than 2D single-shot FSE.

Cascade of Denoising and Mapping Neural Networks for MRI R2* Relaxometry of Iron-Loaded Liver

MRI of effective transverse relaxation rate (R2*) measurement is a reliable method for liver iron concentration quantification. However, R2* mapping can be degraded by noise, especially in the case of iron overload. This study aimed to develop a deep learning method for MRI R2* relaxometry of an iron-loaded liver using a two-stage cascaded neural network. The proposed method, named CadamNet, combines two convolutional neural networks separately designed for image denoising and parameter mapping into a cascade framework, and the physics-based R2* decay model was incorporated in training the mapping network to enforce data consistency further. CadamNet was trained using simulated liver data with Rician noise, which was constructed from clinical liver data. The performance of CadamNet was quantitatively evaluated on simulated data with varying noise levels as well as clinical liver data and compared with the single-stage parameter mapping network (MappingNet) and two conventional model-based R2* mapping methods. CadamNet consistently achieved high-quality R2* maps and outperformed MappingNet at varying noise levels. Compared with conventional R2* mapping methods, CadamNet yielded R2* maps with lower errors, higher quality, and substantially increased efficiency. In conclusion, the proposed CadamNet enables accurate and efficient iron-loaded liver R2* mapping, especially in the presence of severe noise.

Impact of deep learning architectures on accelerated cardiac T1 mapping using MyoMapNet

The objective of the current study was to investigate the performance of various deep learning (DL) architectures for MyoMapNet, a DL model for T1 estimation using accelerated cardiac T1 mapping from four T1 -weighted images collected after a single inversion pulse (Look-Locker 4 [LL4]). We implemented and tested three DL architectures for MyoMapNet: (a) a fully connected neural network (FC), (b) convolutional neural networks (VGG19, ResNet50), and (c) encoder-decoder networks with skip connections (ResUNet, U-Net). Modified Look-Locker inversion recovery (MOLLI) images from 749 patients at 3 T were used for training, validation, and testing. The first four T1 -weighted images from MOLLI5(3)3 and/or MOLLI4(1)3(1)2 protocols were extracted to create accelerated cardiac T1 mapping data. We also prospectively collected data from 28 subjects using MOLLI and LL4 to further evaluate model performance. Despite rigorous training, conventional VGG19 and ResNet50 models failed to produce anatomically correct T1 maps, and T1 values had significant errors. While ResUNet yielded good quality maps, it significantly underestimated T1 . Both FC and U-Net, however, yielded excellent image quality with good T1 accuracy for both native (FC/U-Net/MOLLI = 1217 ± 64/1208 ± 61/1199 ± 61 ms, all p < 0.05) and postcontrast myocardial T1 (FC/U-Net/MOLLI = 578 ± 57/567 ± 54/574 ± 55 ms, all p < 0.05). In terms of precision, the U-Net model yielded better T1 precision compared with the FC architecture (standard deviation of 61 vs. 67 ms for the myocardium for native [p < 0.05], and 31 vs. 38 ms [p < 0.05], for postcontrast). Similar findings were observed in prospectively collected LL4 data. It was concluded that U-Net and FC DL models in MyoMapNet enable fast myocardial T1 mapping using only four T1 -weighted images collected from a single LL sequence with comparable accuracy. U-Net also provides a slight improvement in precision.

Simultaneous multi-parametric acquisition and reconstruction techniques in cardiac magnetic resonance imaging: Basic concepts and status of clinical development

Simultaneous multi-parametric acquisition and reconstruction techniques (SMART) are gaining attention for their potential to overcome some of cardiovascular magnetic resonance imaging's (CMR) clinical limitations. The major advantages of SMART lie within their ability to simultaneously capture multiple "features" such as cardiac motion, respiratory motion, T1/T2 relaxation. This review aims to summarize the overarching theory of SMART, describing key concepts that many of these techniques share to produce co-registered, high quality CMR images in less time and with less requirements for specialized personnel. Further, this review provides an overview of the recent developments in the field of SMART by describing how they work, the parameters they can acquire, their status of clinical testing and validation, and by providing examples for how their use can improve the current state of clinical CMR workflows. Many of the SMART are in early phases of development and testing, thus larger scale, controlled trials are needed to evaluate their use in clinical setting and with different cardiac pathologies.

Motion correction for native myocardial T1 mapping using self-supervised deep learning registration with contrast separation

In myocardial T₁ mapping, undesirable motion poses significant challenges because uncorrected motion can affect T₁ estimation accuracy and cause incorrect diagnosis. In this study, we propose and evaluate a motion correction method for myocardial T₁ mapping using self-supervised deep learning based registration with contrast separation (SDRAP). A sparse coding based method was first proposed to separate the contrast component from T₁ -weighted (T1w) images. Then, a self-supervised deep neural network with cross-correlation (SDRAP-CC) or mutual information as the registration similarity measurement was developed to register contrast separated images, after which signal fitting was performed on the motion corrected T1w images to generate motion corrected T₁ maps. The registration network was trained and tested in 80 healthy volunteers with images acquired using the modified Look-Locker inversion recovery (MOLLI) sequence. The proposed SDRAP was compared with the free form deformation (FFD) registration method regarding (1) Dice similarity coefficient (DSC) and mean boundary error (MBE) of myocardium contours, (2) T₁ value and standard deviation (SD) of T₁ fitting, (3) subjective evaluation score for overall image quality and motion artifact level, and (4) computation time. Results showed that SDRAP-CC achieved the highest DSC of 85.0 ± 3.9% and the lowest MBE of 0.92 ± 0.25 mm among the methods compared. Additionally, SDRAP-CC performed the best by resulting in lower SD value (28.1 ± 17.6 ms) and higher subjective image quality scores (3.30 ± 0.79 for overall quality and 3.53 ± 0.68 for motion artifact) evaluated by a cardiologist. The proposed SDRAP took only 0.52 s to register one slice of MOLLI images, achieving about sevenfold acceleration over FFD (3.7 s/slice).

Clinical evaluation of the Multimapping technique for simultaneous myocardial T1 and T2 mapping

The Multimapping technique was recently proposed for simultaneous myocardial T₁ and T₂ mapping. In this study, we evaluate its correlation with clinical reference mapping techniques in patients with a range of cardiovascular diseases (CVDs) and compare image quality and inter- and intra-observer repeatability. Multimapping consists of an ECG-triggered, 2D single-shot bSSFP readout with inversion recovery and T₂ preparation modules, acquired across 10 cardiac cycles. The sequence was implemented at 1.5T and compared to clinical reference mapping techniques, modified Look-Locker inversion recovery (MOLLI) and T₂ prepared bSSFP with four echo times (T₂bSSFP), and compared in 47 patients with CVD (of which 44 were analyzed). In diseased myocardial segments (defined as the presence of late gadolinium enhancement), there was a high correlation between Multimapping and MOLLI for native myocardium T₁ (r² = 0.73), ECV (r² = 0.91), and blood T₁ (r² = 0.88), and Multimapping and T₂bSSFP for native myocardial T₂ (r² = 0.80). In healthy myocardial segments, a bias for native T₁ (Multimapping = 1,116 ± 21 ms, MOLLI = 1,002 ± 21, P < 0.001), post-contrast T₁ (Multimapping = 479 ± 31 ms, MOLLI = 426 ± 27 ms, 0.001), ECV (Multimapping = 21.5 ± 1.9%, MOLLI = 23.7 ± 2.3%, P = 0.001), and native T₂ (Multimapping = 48.0 ± 3.0 ms, T₂bSSFP = 53.9 ± 3.5 ms, P < 0.001) was observed. The image quality for Multimapping was scored as higher for all mapping techniques (native T₁, post-contrast T₁, ECV, and T₂bSSFP) compared to the clinical reference techniques. The inter- and intra-observer agreements were excellent (intraclass correlation coefficient, ICC > 0.9) for most measurements, except for inter-observer repeatability of Multimapping native T₁ (ICC = 0.87), post-contrast T₁ (ICC = 0.73), and T₂bSSFP native T₂ (ICC = 0.88). Multimapping shows high correlations with clinical reference mapping techniques for T₁, T₂, and ECV in a diverse cohort of patients with different cardiovascular diseases. Multimapping enables simultaneous T₁ and T₂ mapping and can be performed in a short breath-hold, with image quality superior to that of the clinical reference techniques.

Improving accuracy of myocardial T1 estimation in MyoMapNet

PURPOSE

To improve the accuracy and robustness of T₁ estimation by MyoMapNet, a deep learning-based approach using 4 inversion-recovery T₁ -weighted images for cardiac T₁ mapping.

METHODS

MyoMapNet is a fully connected neural network for T₁ estimation of an accelerated cardiac T₁ mapping sequence, which collects 4 T₁ -weighted images by a single Look-Locker inversion-recovery experiment (LL4). MyoMapNet was originally trained using in vivo data from the modified Look-Locker inversion recovery sequence, which resulted in significant bias and sensitivity to various confounders. This study sought to train MyoMapNet using signals generated from numerical simulations and phantom MR data under multiple simulated confounders. The trained model was then evaluated by phantom data scanned using new phantom vials that differed from those used for training. The performance of the new model was compared with modified Look-Locker inversion recovery sequence and saturation-recovery single-shot acquisition for measuring native and postcontrast T₁ in 25 subjects.

RESULTS

In the phantom study, T₁ values measured by LL4 with MyoMapNet were highly correlated with reference values from the spin-echo sequence. Furthermore, the estimated T₁ had excellent robustness to changes in flip angle and off-resonance. Native and postcontrast myocardium T₁ at 3 Tesla measured by saturation-recovery single-shot acquisition, modified Look-Locker inversion recovery sequence, and MyoMapNet were 1483 ± 46.6 ms and 791 ± 45.8 ms, 1169 ± 49.0 ms and 612 ± 36.0 ms, and 1443 ± 57.5 ms and 700 ± 57.5 ms, respectively. The corresponding extracellular volumes were 22.90% ± 3.20%, 28.88% ± 3.48%, and 30.65% ± 3.60%, respectively.

CONCLUSION

Training MyoMapNet with numerical simulations and phantom data will improve the estimation of myocardial T₁ values and increase its robustness to confounders while also reducing the overall T₁ mapping estimation time to only 4 heartbeats.

Fast, Accurate, and Robust T2 Mapping of Articular Cartilage by Neural Networks

For T2 mapping, the underlying mono-exponential signal decay is traditionally quantified by non-linear Least-Squares Estimation (LSE) curve fitting, which is prone to outliers and computationally expensive. This study aimed to validate a fully connected neural network (NN) to estimate T2 relaxation times and to assess its performance versus LSE fitting methods. To this end, the NN was trained and tested in silico on a synthetic dataset of 75 million signal decays. Its quantification error was comparatively evaluated against three LSE methods, i.e., traditional methods without any modification, with an offset, and one with noise correction. Following in-situ acquisition of T2 maps in seven human cadaveric knee joint specimens at high and low signal-to-noise ratios, the NN and LSE methods were used to estimate the T2 relaxation times of the manually segmented patellofemoral cartilage. In-silico modeling at low signal-to-noise ratio indicated significantly lower quantification error for the NN (by medians of 6−33%) than for the LSE methods (p < 0.001). These results were confirmed by the in-situ measurements (medians of 10−35%). T2 quantification by the NN took only 4 s, which was faster than the LSE methods (28−43 s). In conclusion, NNs provide fast, accurate, and robust quantification of T2 relaxation times.

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

PURPOSE

To develop and evaluate MyoMapNet, a rapid myocardial T1 mapping approach that uses fully connected neural networks (FCNN) to estimate T1 values from four T1-weighted images collected after a single inversion pulse in four heartbeats (Look-Locker, LL4).

METHOD

We implemented an FCNN for MyoMapNet to estimate T1 values from a reduced number of T1-weighted images and corresponding inversion-recovery times. We studied MyoMapNet performance when trained using native, post-contrast T1, or a combination of both. We also explored the effects of number of T1-weighted images (four and five) for native T1. After rigorous training using in-vivo modified Look-Locker inversion recovery (MOLLI) T1 mapping data of 607 patients, MyoMapNet performance was evaluated using MOLLI T1 data from 61 patients by discarding the additional T1-weighted images. Subsequently, we implemented a prototype MyoMapNet and LL4 on a 3 T scanner. LL4 was used to collect T1 mapping data in 27 subjects with inline T1 map reconstruction by MyoMapNet. The resulting T1 values were compared to MOLLI.

RESULTS

MyoMapNet trained using a combination of native and post-contrast T1-weighted images had excellent native and post-contrast T1 accuracy compared to MOLLI. The FCNN model using four T1-weighted images yields similar performance compared to five T1-weighted images, suggesting that four T1 weighted images may be sufficient. The inline implementation of LL4 and MyoMapNet enables successful acquisition and reconstruction of T1 maps on the scanner. Native and post-contrast myocardium T1 by MOLLI and MyoMapNet was 1170 ± 55 ms vs. 1183 ± 57 ms (P = 0.03), and 645 ± 26 ms vs. 630 ± 30 ms (P = 0.60), and native and post-contrast blood T1 was 1820 ± 29 ms vs. 1854 ± 34 ms (P = 0.14), and 508 ± 9 ms vs. 514 ± 15 ms (P = 0.02), respectively.

CONCLUSION

A FCNN, trained using MOLLI data, can estimate T1 values from only four T1-weighted images. MyoMapNet enables myocardial T1 mapping in four heartbeats with similar accuracy as MOLLI with inline map reconstruction.

Recurrent inference machines as inverse problem solvers for MR relaxometry

In this paper, we propose the use of Recurrent Inference Machines (RIMs) to perform T₁ and T₂ mapping. The RIM is a neural network framework that learns an iterative inference process based on the signal model, similar to conventional statistical methods for quantitative MRI (QMRI), such as the Maximum Likelihood Estimator (MLE). This framework combines the advantages of both data-driven and model-based methods, and, we hypothesize, is a promising tool for QMRI. Previously, RIMs were used to solve linear inverse reconstruction problems. Here, we show that they can also be used to optimize non-linear problems and estimate relaxometry maps with high precision and accuracy. The developed RIM framework is evaluated in terms of accuracy and precision and compared to an MLE method and an implementation of the Residual Neural Network (ResNet). The results show that the RIM improves the quality of estimates compared to the other techniques in Monte Carlo experiments with simulated data, test-retest analysis of a system phantom, and in-vivo scans. Additionally, inference with the RIM is 150 times faster than the MLE, and robustness to (slight) variations of scanning parameters is demonstrated. Hence, the RIM is a promising and flexible method for QMRI. Coupled with an open-source training data generation tool, it presents a compelling alternative to previous methods.

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

This article presents an overview of deep learning (DL) and its applications to function approximation for MR in medicine. The aim of this article is to help readers develop various applications of DL. DL has made a large impact on the literature of many medical sciences, including MR. However, its technical details are not easily understandable for non-experts of machine learning (ML).The first part of this article presents an overview of DL and its related technologies, such as artificial intelligence (AI) and ML. AI is explained as a function that can receive many inputs and produce many outputs. ML is a process of fitting the function to training data. DL is a kind of ML, which uses a composite of many functions to approximate the function of interest. This composite function is called a deep neural network (DNN), and the functions composited into a DNN are called layers. This first part also covers the underlying technologies required for DL, such as loss functions, optimization, initialization, linear layers, non-linearities, normalization, recurrent neural networks, regularization, data augmentation, residual connections, autoencoders, generative adversarial networks, model and data sizes, and complex-valued neural networks.The second part of this article presents an overview of the applications of DL in MR and explains how functions represented as DNNs are applied to various applications, such as RF pulse, pulse sequence, reconstruction, motion correction, spectroscopy, parameter mapping, image synthesis, and segmentation.

Deep learning-enhanced T1 mapping with spatial-temporal and physical constraint

PURPOSE

To propose a reconstruction framework to generate accurate T₁ maps for a fast MR T₁ mapping sequence.

METHODS

A deep learning-enhanced T₁ mapping method with spatial-temporal and physical constraint (DAINTY) was proposed. This method explicitly imposed low-rank and sparsity constraints on the multiframe T₁ -weighted images to exploit the spatial-temporal correlation. A deep neural network was used to efficiently perform T₁ mapping as well as denoise and reduce undersampling artifacts. Additionally, the physical constraint was used to build a bridge between low-rank and sparsity constraint and deep learning prior, so the benefits of constrained reconstruction and deep learning can be both available. The DAINTY method was trained on simulated brain data sets, but tested on real acquired phantom, 6 healthy volunteers, and 7 atherosclerosis patients, compared with the narrow-band k-space-weighted image contrast filter conjugate-gradient SENSE (NK-CS) method, kt-sparse-SENSE (kt-SS) method, and low-rank plus sparsity (L+S) method with least-squares T₁ fitting and direct deep learning mapping.

RESULTS

The DAINTY method can generate more accurate T₁ maps and higher-quality T₁ -weighted images compared with other methods. For atherosclerosis patients, the intraplaque hemorrhage can be successfully detected. The computation speed of DAINTY was 10 times faster than traditional methods. Meanwhile, DAINTY can reconstruct images with comparable quality using only 50% of k-space data.

CONCLUSION

The proposed method can provide accurate T₁ maps and good-quality T₁ -weighted images with high efficiency.

Deep convolution neural networks based artifact suppression in under-sampled radial acquisitions of myocardial T 1 mapping images

We developed a deep convolutional neural network (CNN) based method to remove streaking artefact from accelerated radial acquisitions of myocardial T ₁-mapping images. A deep CNN based on a modified U-Net architecture was developed and trained to remove the streaking artefacts from under-sampled T ₁ mapping images. A total of 2090 T ₁-weighted images for 33 patients (55 ± 15 years, 19 males) and five healthy subjects (30 ± 14 years, 2 males) were used for training and testing the network. The images were acquired using radial slice interleaved T ₁ mapping sequence (STONE) and retrospectively under-sampled to achieve acceleration rate of 4 (corresponding to 48 spokes). The dataset was split into training and testing subsets with 23 subjects (60%) and 15 subjects (40%), respectively. For generating voxel-wise T ₁ maps, a two-parameter fitting model was used. Network performance was evaluated using normalized mean square error (NMSE), structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) metrics. The proposed network allowed fast (<0.3 s/image) removal of the artefact from all T ₁-weighted testing images and the corresponding T ₁ maps with PSNR = 64.3 ± 1.02, NMSE = 0.2 ± 0.09 and SSIM = 0.9 ± 0.3 × 10^-4. There was no statistically significant difference between the measured T ₁ maps for both per-subject (reference: 1085 ± 37 ms, CNN: 1088 ± 37 ms, p = 0.4) and per-segment (reference: 1084 ± 48 ms, CNN: 1083 ± 58 ms, p = 0.9) analyses. In summary, deep CNN allows fast and reliable removal of streaking artefact from under-sampled radial T ₁ mapping images. Our results show that the highly non-linear operations of deep CNN processing of T ₁ mapping images do not impact accurate reconstruction of myocardial T ₁ maps.

Fast and accurate calculation of myocardial T1 and T2 values using deep learning Bloch equation simulations (DeepBLESS)

PURPOSE

To propose and evaluate a deep learning model for rapid and accurate calculation of myocardial T1 /T2 values based on a previously proposed Bloch equation simulation with slice profile correction (BLESSPC) method.

METHODS

Deep learning Bloch equation simulations (DeepBLESS) models are proposed for rapid and accurate T1 estimation for the MOLLI T1 mapping sequence with balanced SSFP readouts and T1 /T2 estimation for a radial simultaneous T1 and T2 mapping (radial T1 -T2 ) sequence. The DeepBLESS models were trained separately based on simulated radial T1 -T2 and MOLLI data, respectively. The DeepBLESS T1 -T2 estimation accuracy was evaluated based on simulated data with different noise levels. The DeepBLESS model was compared with BLESSPC in simulation, phantom, and in vivo studies for the MOLLI sequence at 1.5 T and radial T1 -T2 sequence at 3 T.

RESULTS

After DeepBLESS was trained, in phantom studies, DeepBLESS and BLESSPC achieved similar accuracy and precision in T1 -T2 estimations for both MOLLI and radial T1 -T2 (P > .05). For in vivo, DeepBLESS and BLESSPC generated similar myocardial T1 /T2 values for radial T1 -T2 at 3 T (T1 : 1366 ± 31 ms for both methods, P > .05; T2 : 37.4 ms ± 0.9 ms for both methods, P > .05), and similar myocardial T1 values for the MOLLI sequence at 1.5 T (1044 ± 20 ms for both methods, P > .05). DeepBLESS generated a T1 /T2 map in less than 1 second.

CONCLUSION

The DeepBLESS model offers an almost instantaneous approach for estimating accurate T1 /T2 values, replacing BLESSPC for both MOLLI and radial T1 -T2 sequences, and is promising for multiparametric mapping in cardiac MRI.

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

PURPOSE

To develop a deep learning method for rapidly reconstructing T₁ and T₂ maps from undersampled electrocardiogram (ECG) triggered cardiac magnetic resonance fingerprinting (cMRF) images.

METHODS

A neural network was developed that outputs T₁ and T₂ 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 T₁ was below 1% in myocardium, blood, and liver for all tested heart rates. For T₂ , the nRMSE was below 4% for myocardium and liver and below 6% for blood for all heart rates. The difference in the mean myocardial T₁ or T₂ observed in vivo between dictionary matching and deep learning was 3.6 ms for T₁ and -0.2 ms for T₂ . 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 T₁ and T₂ 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.