Artificial neural network for myelin water imaging

Bias-reduced neural networks for parameter estimation in quantitative MRI

PURPOSE

To develop neural network (NN)-based quantitative MRI parameter estimators with minimal bias and a variance close to the Cramér-Rao bound.

THEORY AND METHODS

We generalize the mean squared error loss to control the bias and variance of the NN's estimates, which involves averaging over multiple noise realizations of the same measurements during training. Bias and variance properties of the resulting NNs are studied for two neuroimaging applications.

RESULTS

In simulations, the proposed strategy reduces the estimates' bias throughout parameter space and achieves a variance close to the Cramér-Rao bound. In vivo, we observe good concordance between parameter maps estimated with the proposed NNs and traditional estimators, such as nonlinear least-squares fitting, while state-of-the-art NNs show larger deviations.

CONCLUSION

The proposed NNs have greatly reduced bias compared to those trained using the mean squared error and offer significantly improved computational efficiency over traditional estimators with comparable or better accuracy.

Reproducible Research Practices in Magnetic Resonance Neuroimaging: A Review Informed by Advanced Language Models

MRI has progressed significantly with the introduction of advanced computational methods and novel imaging techniques, but their wider adoption hinges on their reproducibility. This concise review synthesizes reproducible research insights from recent MRI articles to examine the current state of reproducibility in neuroimaging, highlighting key trends and challenges. It also provides a custom generative pretrained transformer (GPT) model, designed specifically for aiding in an automated analysis and synthesis of information pertaining to the reproducibility insights associated with the articles at the core of this review.

Improved myelin water fraction mapping with deep neural networks using synthetically generated 3D data

We introduce a generative model for synthesis of large scale 3D datasets for quantitative parameter mapping of myelin water fraction (MWF). Our model combines a MR physics signal decay model with an accurate probabilistic multi-component parametric T2 model. We synthetically generate a wide variety of high quality signals and corresponding parameters from a wide range of naturally occurring prior parameter values. To capture spatial variation, the generative signal decay model is combined with a generative spatial model conditioned on generic tissue segmentations. Synthesized 3D datasets can be used to train any convolutional neural network (CNN) based architecture for MWF estimation. Our source code is available at: https://github.com/quin-med-harvard-edu/synthmap Reduction of acquisition time at the expense of lower SNR, as well as accuracy and repeatability of MWF estimation techniques, are key factors that affect the adoption of MWF mapping in clinical practice. We demonstrate that the synthetically trained CNN provides superior accuracy over the competing methods under the constraints of naturally occurring noise levels as well as on the synthetically generated images at low SNR levels. Normalized root mean squared error (nRMSE) is less than 7% on synthetic data, which is significantly lower than competing methods. Additionally, the proposed method yields a coefficient of variation (CoV) that is at least 4x better than the competing method on intra-session test-retest reference dataset.

REUSED: A deep neural network method for rapid whole-brain high-resolution myelin water fraction mapping from extremely under-sampled MRI

Changes in myelination are a cardinal feature of brain development and the pathophysiology of several central nervous system diseases, including multiple sclerosis and dementias. Advanced magnetic resonance imaging (MRI) methods have been developed to probe myelin content through the measurement of myelin water fraction (MWF). However, the prolonged data acquisition and post-processing times of current MWF mapping methods pose substantial hurdles to their clinical implementation. Recently, fast steady-state MRI sequences have been implemented to produce high-spatial resolution whole-brain MWF mapping within ∼20 min. Despite the subsequent significant advances in the inversion algorithm to derive MWF maps from steady-state MRI, the high-dimensional nature of such inversion does not permit further reduction of the acquisition time by data under-sampling. In this work, we present an unprecedented reduction in the computation (∼30 s) and the acquisition time (∼7 min) required for whole-brain high-resolution MWF mapping through a new Neural Network (NN)-based approach, named NN-Relaxometry of Extremely Under-SamplEd Data (NN-REUSED). Our analyses demonstrate virtually similar accuracy and precision in derived MWF values using NN-REUSED compared to results derived from the fully sampled reference method. The reduction in the acquisition and computation times represents a breakthrough toward clinically practical MWF mapping.

Self-labelled encoder-decoder (SLED) for multi-echo gradient echo-based myelin water imaging

PURPOSE

Reconstruction of high quality myelin water imaging (MWI) maps is challenging, particularly for data acquired using multi-echo gradient echo (mGRE) sequences. A non-linear least squares fitting (NLLS) approach has often been applied for MWI. However, this approach may produce maps with limited detail and, in some cases, sub-optimal signal to noise ratio (SNR), due to the nature of the voxel-wise fitting. In this study, we developed a novel, unsupervised learning method called self-labelled encoder-decoder (SLED) to improve gradient echo-based MWI data fitting.

METHODS

Ultra-high resolution, MWI data was collected from five mouse brains with variable levels of myelination, using a mGRE sequence. Imaging data was acquired using a 7T preclinical MRI system. A self-labelled, encoder-decoder network was implemented in TensorFlow for calculation of myelin water fraction (MWF) based on the mGRE signal decay. A simulated MWI phantom was also created to evaluate the performance of MWF estimation.

RESULTS

Compared to NLLS, SLED demonstrated improved MWF estimation, in terms of both stability and accuracy in phantom tests. In addition, SLED produced less noisy MWF maps from high resolution MR microscopy images of mouse brain tissue. It specifically resulted in lower noise amplification for all mouse genotypes that were imaged and yielded mean MWF values in white matter ROIs that were highly correlated with those derived from standard NLLS fitting. Lastly, SLED also exhibited higher tolerance to low SNR data.

CONCLUSION

Due to its unsupervised and self-labeling nature, SLED offers a unique alternative to analyze gradient echo-based MWI data, providing accurate and stable MWF estimations.

Method to Minimize the Errors of AI: Quantifying and Exploiting Uncertainty of Deep Learning in Brain Tumor Segmentation

Despite the unprecedented success of deep learning in various fields, it has been recognized that clinical diagnosis requires extra caution when applying recent deep learning techniques because false prediction can result in severe consequences. In this study, we proposed a reliable deep learning framework that could minimize incorrect segmentation by quantifying and exploiting uncertainty measures. The proposed framework demonstrated the effectiveness of a public dataset: Multimodal Brain Tumor Segmentation Challenge 2018. By using this framework, segmentation performances, particularly for small lesions, were improved. Since the segmentation of small lesions is difficult but also clinically significant, this framework could be effectively applied to the medical imaging field.

DIFFnet: Diffusion Parameter Mapping Network Generalized for Input Diffusion Gradient Schemes and b-Value

In MRI, deep neural networks have been proposed to reconstruct diffusion model parameters. However, the inputs of the networks were designed for a specific diffusion gradient scheme (i.e., diffusion gradient directions and numbers) and a specific b-value that are the same as the training data. In this study, a new deep neural network, referred to as DIFFnet, is developed to function as a generalized reconstruction tool of the diffusion-weighted signals for various gradient schemes and b-values. For generalization, diffusion signals are normalized in a q-space and then projected and quantized, producing a matrix (Qmatrix) as an input for the network. To demonstrate the validity of this approach, DIFFnet is evaluated for diffusion tensor imaging (DIFFnet_DTI) and for neurite orientation dispersion and density imaging (DIFFnet_NODDI). In each model, two datasets with different gradient schemes and b-values are tested. The results demonstrate accurate reconstruction of the diffusion parameters at substantially reduced processing time (approximately 8.7 times and 2240 times faster processing time than conventional methods in DTI and NODDI, respectively; less than 4% mean normalized root-mean-square errors (NRMSE) in DTI and less than 8% in NODDI). The generalization capability of the networks was further validated using reduced numbers of diffusion signals from the datasets and a public dataset from Human Connection Project. Different from previously proposed deep neural networks, DIFFnet does not require any specific gradient scheme and b-value for its input. As a result, it can be adopted as an online reconstruction tool for various complex diffusion imaging.

Subsecond accurate myelin water fraction reconstruction from FAST-T2 data with 3D UNET

PURPOSE

To develop a 3D UNET convolutional neural network for rapid extraction of myelin water fraction (MWF) maps from six-echo fast acquisition with spiral trajectory and T₂ -prep data and to evaluate its accuracy in comparison with multilayer perceptron (MLP) network.

METHODS

The MWF maps were extracted from 138 patients with multiple sclerosis using an iterative three-pool nonlinear least-squares algorithm (NLLS) without and with spatial regularization (srNLLS), which were used as ground-truth labels to train, validate, and test UNET and MLP networks as a means to accelerate data fitting. Network testing was performed in 63 patients with multiple sclerosis and a numerically simulated brain phantom at SNR of 200, 100 and 50.

RESULTS

Simulations showed that UNET reduced the MWF mean absolute error by 30.1% to 56.4% and 16.8% to 53.6% over the whole brain and by 41.2% to 54.4% and 21.4% to 49.4% over the lesions for predicting srNLLS and NLLS MWF, respectively, compared to MLP, with better performance at lower SNRs. UNET also outperformed MLP for predicting srNLLS MWF in the in vivo multiple-sclerosis brain data, reducing mean absolute error over the whole brain by 61.9% and over the lesions by 67.5%. However, MLP yielded 41.1% and 51.7% lower mean absolute error for predicting in vivo NLLS MWF over the whole brain and the lesions, respectively, compared with UNET. The whole-brain MWF processing time using a GPU was 0.64 seconds for UNET and 0.74 seconds for MLP.

CONCLUSION

Subsecond whole-brain MWF extraction from fast acquisition with spiral trajectory and T₂ -prep data using UNET is feasible and provides better accuracy than MLP for predicting MWF output of srNLLS algorithm.

Revisiting the T2 spectrum imaging inverse problem: Bayesian regularized non-negative least squares

Multi-echo T₂ magnetic resonance images contain information about the distribution of T₂ relaxation times of compartmentalized water, from which we can estimate relevant brain tissue properties such as the myelin water fraction (MWF). Regularized non-negative least squares (NNLS) is the tool of choice for estimating non-parametric T₂ spectra. However, the estimation is ill-conditioned, sensitive to noise, and highly affected by the employed regularization weight. The purpose of this study is threefold: first, we want to underline that the apparently innocuous use of two alternative parameterizations for solving the inverse problem, which we called the standard and alternative regularization forms, leads to different solutions; second, to assess the performance of both parameterizations; and third, to propose a new Bayesian regularized NNLS method (BayesReg). The performance of BayesReg was compared with that of two conventional approaches (L-curve and Chi-square (X²) fitting) using both regularization forms. We generated a large dataset of synthetic data, acquired in vivo human brain data in healthy participants for conducting a scan-rescan analysis, and correlated the myelin content derived from histology with the MWF estimated from ex vivo data. Results from synthetic data indicate that BayesReg provides accurate MWF estimates, comparable to those from L-curve and X², and with better overall stability across a wider signal-to-noise range. Notably, we obtained superior results by using the alternative regularization form. The correlations reported in this study are higher than those reported in previous studies employing the same ex vivo and histological data. In human brain data, the estimated maps from L-curve and BayesReg were more reproducible. However, the T₂ spectra produced by BayesReg were less affected by over-smoothing than those from L-curve. These findings suggest that BayesReg is a good alternative for estimating T₂ distributions and MWF maps.

Quad-Contrast Imaging: Simultaneous Acquisition of Four Contrast-Weighted Images (PD-Weighted, T₂-Weighted, PD-FLAIR and T₂-FLAIR Images) With Synthetic T₁-Weighted Image, T₁- and T₂-Maps

Magnetic resonance imaging (MRI) can provide multiple contrast-weighted images using different pulse sequences and protocols. However, a long acquisition time of the images is a major challenge. To address this limitation, a new pulse sequence referred to as quad-contrast imaging is presented. The quad-contrast sequence enables the simultaneous acquisition of four contrast-weighted images (proton density (PD)-weighted, T₂-weighted, PD-fluid attenuated inversion recovery (FLAIR), and T₂-FLAIR), and the synthesis of T₁-weighted images and T₁- and T₂-maps in a single scan. The scan time is less than 6 min and is further reduced to 2 min 50 s using a deep learning-based parallel imaging reconstruction. The natively acquired quad contrasts demonstrate high quality images, comparable to those from the conventional scans. The deep learning-based reconstruction successfully reconstructed highly accelerated data (acceleration factor 6), reporting smaller normalized root mean squared errors (NRMSEs) and higher structural similarities (SSIMs) than those from conventional generalized autocalibrating partially parallel acquisitions (GRAPPA)-reconstruction (mean NRMSE of 4.36% vs. 10.54% and mean SSIM of 0.990 vs. 0.953). In particular, the FLAIR contrast is natively acquired and does not suffer from lesion-like artifacts at the boundary of tissue and cerebrospinal fluid, differentiating the proposed method from synthetic imaging methods. The quad-contrast imaging method may have the potentials to be used in a clinical routine as a rapid diagnostic tool.

A data-driven T2 relaxation analysis approach for myelin water imaging: Spectrum analysis for multiple exponentials via experimental condition oriented simulation (SAME-ECOS)

PURPOSE

The decomposition of multi-exponential decay data into a T₂ spectrum poses substantial challenges for conventional fitting algorithms, including non-negative least squares (NNLS). Based on a combination of the resolution limit constraint and machine learning neural network algorithm, a data-driven and highly tailorable analysis method named spectrum analysis for multiple exponentials via experimental condition oriented simulation (SAME-ECOS) was proposed.

THEORY AND METHODS

The theory of SAME-ECOS was derived. Then, a paradigm was presented to demonstrate the SAME-ECOS workflow, consisting of a series of calculation, simulation, and model training operations. The performance of the trained SAME-ECOS model was evaluated using simulations and six in vivo brain datasets. The code is available at https://github.com/hanwencat/SAME-ECOS.

RESULTS

Using NNLS as the baseline, SAME-ECOS achieved over 15% higher overall cosine similarity scores in producing the T₂ spectrum, and more than 10% lower mean absolute error in calculating the myelin water fraction (MWF), as well as demonstrated better robustness to noise in the simulation tests. Applying to in vivo data, MWF from SAME-ECOS and NNLS was highly correlated among all study participants. However, a distinct separation of the myelin water peak and the intra/extra-cellular water peak was only observed in the mean T₂ spectra determined using SAME-ECOS. In terms of data processing speed, SAME-ECOS is approximately 30 times faster than NNLS, achieving a whole-brain analysis in 3 min.

CONCLUSION

Compared with NNLS, the SAME-ECOS method yields much more reliable T₂ spectra in a dramatically shorter time, increasing the feasibility of multi-component T₂ decay analysis in clinical settings.

Ensemble learning accurately predicts the potential benefits of thrombolytic therapy in acute ischemic stroke

Background

Finding methods to accurately predict the final infarct volumes for acute ischemic stroke patients with full or no recanalization would significantly help to evaluate the potential benefits of thrombolytic therapy. We proposed such a method by constructing a model of ensemble deep learning and machine learning using diffusion-weighted imaging (DWI) only.

Methods

The proposed prediction model (named AUNet) combines an adaptive linear ensemble model (ALEM) of machine learning and a deep U-Net network with an accelerated non-local module (U-NL-Net) to learn voxel-wise and spatial features, respectively. Of 40 patients with acute ischemic stroke who received thrombolytic therapy, 17 were fully recanalized, 14 were not recanalized, and nine were partially recanalized. The AUNet was separately trained for full recanalization conditions (AUNetR) and no recanalization (AUNetN) as the best and worst outcomes of thrombolysis, respectively.

Results

AUNet performed significantly better in predicting the final infarct volumes in both the recanalization and non-recanalization conditions [area under the receiver operating characteristic curve (AUC) =0.898±0.022, recanalization; AUC =0.875±0.036, non-recanalization: Matthew's correlation coefficient (MCC) =0.863±0.033, recanalization; MCC =0.851±0.025, non-recanalization] than the fixed-thresholding method (AUC =0.776±0.021, P<0.0001, recanalization; AUC =0.692±0.023, P<0.0001, non-recanalization: MCC =0.742±0.035, recanalization; MCC =0.671±0.024, non-recanalization), the logistic regression method (AUC =0.797±0.023, P<0.003, recanalization; AUC =0.751±0.030, P<0.003, non-recanalization: MCC =0.762±0.035, recanalization; MCC =0.730±0.031, non-recanalization), and a recently developed convolutional neural network (AUC =0.814±0.013, P<0.003, recanalization; AUC =0.781±0.027, P<0.003, non-recanalization: MCC =792±0.022, recanalization; MCC =0.758±0.016, non-recanalization). The potential benefit of thrombolysis calculated from AUNetR and AUNetN showed large individual differences (from 12.81% to 239.73%).

Conclusions

AUNet improved predictive accuracy over current state-of-the-art methods. More importantly, the accurate prediction of infarct volumes under different recanalization conditions may provide benefitial information for physicians in selecting suitable patients for thrombolytic therapy.

Improved estimation of myelin water fractions with learned parameter distributions

PURPOSE

To improve estimation of myelin water fraction (MWF) in the brain from multi-echo gradient-echo imaging data.

METHODS

A systematic sensitivity analysis was first conducted to characterize the conventional exponential models used for MWF estimation. A new estimation method was then proposed for improved estimation of MWF from practical gradient-echo imaging data. The proposed method uses an extended signal model that includes a finite impulse response filter to compensate for practical signal variations. This new model also enables the use of prelearned parameter distributions as well as low-rank signal structures to improve parameter estimation. The resulting parameter estimation problem was solved optimally in the Bayesian sense.

RESULTS

Our sensitivity analysis results showed that the conventional exponential models were very sensitive to measurement noise and modeling errors. Our simulation and experimental results showed that our proposed method provided a substantial improvement in reliability, reproducibility, and robustness of MWF estimates over the conventional methods. Clinical results obtained from stroke patients indicated that the proposed method, with its improved capability, could reveal the loss of myelin in lesions, demonstrating its translational potentials.

CONCLUSION

This paper addressed the problem of robust MWF estimation from gradient-echo imaging data. A new method was proposed to provide improved MWF estimation in the presence of significant noise and modeling errors. The performance of the proposed method has been evaluated using both simulated and experimental data, showing significantly improved robustness over the existing methods. The proposed method may prove useful for quantitative myelin imaging in clinical applications.

Decoding the microstructural properties of white matter using realistic models

Multi-echo gradient echo (ME-GRE) magnetic resonance signal evolution in white matter has a strong dependence on the orientation of myelinated axons with respect to the main static field. Although analytical solutions have been able to predict some of the white matter (WM) signal behaviour of the hollow cylinder model, it has been shown that realistic models of WM offer a better description of the signal behaviour observed. In this work, we present a pipeline to (i) generate realistic 2D WM models with their microstructure based on real axon morphology with adjustable fiber volume fraction (FVF) and g-ratio. We (ii) simulate their interaction with the static magnetic field to be able to simulate their MR signal. For the first time, we (iii) demonstrate that realistic 2D WM models can be used to simulate a MR signal that provides a good approximation of the signal obtained from a real 3D WM model derived from electron microscopy. We then (iv) demonstrate in silico that 2D WM models can be used to predict microstructural parameters in a robust way if ME-GRE multi-orientation data is available and the main fiber orientation in each pixel is known using DTI. A deep learning network was trained and characterized in its ability to recover the desired microstructural parameters such as FVF, g-ratio, free and bound water transverse relaxation and magnetic susceptibility. Finally, the network was trained to recover these micro-structural parameters from an ex vivo dataset acquired in 9 orientations with respect to the magnetic field and 12 echo times. We demonstrate that this is an overdetermined problem and that as few as 3 orientations can already provide comparable results for some of the decoded metrics.

Quantitative Multicomponent T2 Relaxation Showed Greater Sensitivity Than Flair Imaging to Detect Subtle Alterations at the Periphery of Lower Grade Gliomas

Purpose

To demonstrate that quantitative multicomponent T2 relaxation can be more sensitive than conventional FLAIR imaging for detecting cerebral tissue abnormalities.

Methods

Six patients affected by lower-grade non-enhancing gliomas underwent T2 relaxation and FLAIR imaging before a radiation treatment by proton therapy (PT) and were examined at follow-up. The T2 decay signal obtained by a thirty-two-echo sequence was decomposed into three main components, attributing to each component a different T2 range: water trapped in the lipid bilayer membrane of myelin, intra/extracellular water and cerebrospinal fluid. The T2 quantitative map of the intra/extracellular water was compared with FLAIR images.

Results

Before PT, in five patients a mismatch was observed between the intra/extracellular water T2 map and FLAIR images, with peri-tumoral areas of high T2 that typically extended outside the area of abnormal FLAIR hyper-intensity. Such mismatch regions evolved into two different types of patterns. The first type, observed in three patients, was a reduced extension of the abnormal regions on T2 map with respect to FLAIR images (T2 decrease pattern). The second type, observed in two patients, was the appearance of new areas of abnormal hyper-intensity on FLAIR images matching the anomalous T2 map extension (FLAIR increase pattern), that was considered as asymptomatic radiation induced damage.

Conclusion

Our preliminarily results suggest that quantitative T2 mapping of the intra/extracellular water component was more sensitive than conventional FLAIR imaging to subtle cerebral tissue abnormalities, deserving to be further investigated in future clinical studies.

Model-informed machine learning for multi-component T2 relaxometry

Recovering the T₂ distribution from multi-echo T₂ magnetic resonance (MR) signals is challenging but has high potential as it provides biomarkers characterizing the tissue micro-structure, such as the myelin water fraction (MWF). In this work, we propose to combine machine learning and aspects of parametric (fitting from the MRI signal using biophysical models) and non-parametric (model-free fitting of the T₂ distribution from the signal) approaches to T₂ relaxometry in brain tissue by using a multi-layer perceptron (MLP) for the distribution reconstruction. For training our network, we construct an extensive synthetic dataset derived from biophysical models in order to constrain the outputs with a priori knowledge of in vivo distributions. The proposed approach, called Model-Informed Machine Learning (MIML), takes as input the MR signal and directly outputs the associated T₂ distribution. We evaluate MIML in comparison to a Gaussian Mixture Fitting (parametric) and Regularized Non-Negative Least Squares algorithms (non-parametric) on synthetic data, an ex vivo scan, and high-resolution scans of healthy subjects and a subject with Multiple Sclerosis. In synthetic data, MIML provides more accurate and noise-robust distributions. In real data, MWF maps derived from MIML exhibit the greatest conformity to anatomical scans, have the highest correlation to a histological map of myelin volume, and the best unambiguous lesion visualization and localization, with superior contrast between lesions and normal appearing tissue. In whole-brain analysis, MIML is 22 to 4980 times faster than the non-parametric and parametric methods, respectively.

Estimating myelin-water content from anatomical and diffusion images using spatially undersampled myelin-water imaging through machine learning

Myelin is vital for healthy neuronal development, and can therefore provide valuable information regarding neuronal maturation. Anatomical and diffusion weighted images (DWI) possess information related to the myelin content and the current study investigates whether quantitative myelin markers can be extracted from anatomical and DWI using neural networks. Thirteen volunteers (mean age 29y) are included, and for each subject, a residual neural network was trained using spatially undersampled reference myelin-water markers. The network is trained on a voxel-by-voxel basis, resulting in a large amount of training data for each volunteer. The inputs used are the anatomical contrasts (cT1w, cT2w), the standardized T1w/T2w ratio, estimates of the relaxation times (T1, T2) and their ratio (T1/T2), and common DWI metrics (FA, RD, MD, λ₁, λ₂, λ₃). Furthermore, to estimate the added value of the DWI metrics, neural networks were trained using either the combined set (DWI, T1w and T2w) or only the anatomical (T1w and T2w) images. The reconstructed myelin-water maps are in good agreement with the reference myelin-water content in terms of the coefficient of variation (CoV) and the intraclass correlation coefficient (ICC). A 6-fold undersampling using both anatomical and DWI metrics resulted in ICC = 0.68 and CoV = 5.9%. Moreover, using twice the training data (3-fold undersampling) resulted in an ICC that is comparable to the reproducibility of the myelin-water imaging itself (CoV = 5.5% vs. CoV = 6.7% and ICC = 0.74 vs ICC = 0.80). To achieve this, beside the T1w, T2w images, DWI is required. This preliminary study shows the potential of machine learning approaches to extract specific myelin-content from anatomical and diffusion-weighted scans.

Deep Reinforcement Learning Designed Shinnar-Le Roux RF Pulse Using Root-Flipping: DeepRFSLR

A novel approach of applying deep reinforcement learning to an RF pulse design is introduced. This method, which is referred to as DeepRF_SLR, is designed to minimize the peak amplitude or, equivalently, minimize the pulse duration of a multiband refocusing pulse generated by the Shinar Le-Roux (SLR) algorithm. In the method, the root pattern of SLR polynomial, which determines the RF pulse shape, is optimized by iterative applications of deep reinforcement learning and greedy tree search. When tested for the designs of the multiband pulses with three and seven slices, DeepRF_SLR demonstrated improved performance compared to conventional methods, generating shorter duration RF pulses in shorter computational time. In the experiments, the RF pulse from DeepRF_SLR produced a slice profile similar to the minimum-phase SLR RF pulse and the profiles matched to that of the computer simulation. Our approach suggests a new way of designing an RF by applying a machine learning algorithm, demonstrating a "machine-designed" MRI sequence.

DeepResp: Deep learning solution for respiration-induced B0 fluctuation artifacts in multi-slice GRE

Respiration-induced B₀ fluctuation corrupts MRI images by inducing phase errors in k-space. A few approaches such as navigator have been proposed to correct for the artifacts at the expense of sequence modification. In this study, a new deep learning method, which is referred to as DeepResp, is proposed for reducing the respiration-artifacts in multi-slice gradient echo (GRE) images. DeepResp is designed to extract the respiration-induced phase errors from a complex image using deep neural networks. Then, the network-generated phase errors are applied to the k-space data, creating an artifact-corrected image. For network training, the computer-simulated images were generated using artifact-free images and respiration data. When evaluated, both simulated images and in-vivo images of two different breathing conditions (deep breathing and natural breathing) show improvements (simulation: normalized root-mean-square error (NRMSE) from 7.8 ± 5.2% to 1.3 ± 0.6%; structural similarity (SSIM) from 0.88 ± 0.08 to 0.99 ± 0.01; ghost-to-signal-ratio (GSR) from 7.9 ± 7.2% to 0.6 ± 0.6%; deep breathing: NRMSE from 13.9 ± 4.6% to 5.8 ± 1.4%; SSIM from 0.86 ± 0.03 to 0.95 ± 0.01; GSR 20.2 ± 10.2% to 5.7 ± 2.3%; natural breathing: NRMSE from 5.2 ± 3.3% to 4.0 ± 2.5%; SSIM from 0.94 ± 0.04 to 0.97 ± 0.02; GSR 5.7 ± 5.0% to 2.8 ± 1.1%). Our approach does not require any modification of the sequence or additional hardware, and may therefore find useful applications. Furthermore, the deep neural networks extract respiration-induced phase errors, which is more interpretable and reliable than results of end-to-end trained networks.

Artificial neural network for multi-echo gradient echo-based myelin water fraction estimation

PURPOSE

To demonstrate robust myelin water fraction (MWF) mapping using an artificial neural network (ANN) with multi-echo gradient-echo (GRE) signal.

METHODS

Multi-echo gradient-echo signals simulated with a three-pool exponential model were used to generate the training data set for the ANN, which was designed to yield the MWF. We investigated the performance of our proposed ANN for various conditions using both numerical simulations and in vivo data. Simulations were conducted with various SNRs to investigate the performance of the ANN. In vivo data with high spatial resolutions were applied in the analyses, and results were compared with MWFs derived by the nonlinear least-squares algorithm using a complex three-pool exponential model.

RESULTS

The network results for the simulations show high accuracies against noise compared with nonlinear least-squares MWFs: RMS-error value of 5.46 for the nonlinear least-squares MWF and 3.56 for the ANN MWF at an SNR of 150 (relative gain = 34.80%). These effects were also found in the in vivo data, with reduced SDs in the region-of-interest analyses. These effects of the ANN demonstrate the feasibility of acquiring high-resolution myelin water images.

CONCLUSION

The simulation results and in vivo data suggest that the ANN facilitates more robust MWF mapping in multi-echo gradient-echo sequences compared with the conventional nonlinear least-squares method.

qMTNet: Accelerated quantitative magnetization transfer imaging with artificial neural networks

PURPOSE

To develop a set of artificial neural networks, collectively termed qMTNet, to accelerate data acquisition and fitting for quantitative magnetization transfer (qMT) imaging.

METHODS

Conventional and interslice qMT data were acquired with two flip angles at six offset frequencies from seven subjects for developing the networks and from four young and four older subjects for testing the generalizability. Two subnetworks, qMTNet-acq and qMTNet-fit, were developed and trained to accelerate data acquisition and fitting, respectively. qMTNet-2 is the sequential application of qMTNet-acq and qMTNet-fit to produce qMT parameters (exchange rate, pool fraction) from undersampled qMT data (two offset frequencies rather than six). qMTNet-1 is one single integrated network having the same functionality as qMTNet-2. qMTNet-fit was compared with a Gaussian kernel-based fitting. qMT parameters generated by the networks were compared with those from ground truth fitted with a dictionary-driven approach.

RESULTS

The proposed networks achieved high peak signal-to-noise ratio (>30) and structural similarity index (>97) in reference to the ground truth. qMTNet-fit produced qMT parameters in concordance with the ground truth with better performance than the Gaussian kernel-based fitting. qMTNet-2 and qMTNet-1 could accelerate data acquisition at threefold and accelerate fitting at 5800- and 4218-fold, respectively. qMTNet-1 showed slightly better performance than qMTNet-2, whereas qMTNet-2 was more flexible for applications.

CONCLUSION

The proposed single (qMTNet-1) and two joint neural networks (qMTNet-2) can accelerate qMT workflow for both data acquisition and fitting significantly. qMTNet has the potential to accelerate qMT imaging for clinical applications, which warrants further investigation.