基于深度神经网络的时空编码磁共振成像超分辨率重建方法

单扫描时空编码磁共振成像是一种新型超快速磁共振成像技术,它对磁场不均匀和化学位移伪影有较强的抵抗性,但是其固有的空间分辨率较低,因此通常需要进行超分辨率重建,以在不增加采样点数的情况下提高时空编码磁共振图像的空间分辨率.然而,现有的重建方法存在迭代求解时间长、重建结果有混叠伪影残留等问题.为此,本文提出了一种基于深度神经网络的单扫描时空编码磁共振成像超分辨率重建方法.该方法采用模拟样本训练深度神经网络,再利用训练好的网络模型对实际采样信号进行重建.数值模拟、水模和活体鼠脑的实验结果表明,该方法能快速重建出无残留混叠伪影、纹理信息清楚的超分辨率时空编码磁共振图像.适当增加训练样本数量以及在训练样本中加入适当的随机噪声水平,有助于改善重建效果.     

Feasibility of MR fingerprinting using a high-performance 0.55 T MRI system

BackgroundMR fingerprinting (MRF) is a versatile method for rapid multi-parametric quantification. The application of MRF for lower MRI field could enable multi-contrast imaging and improve exam efficiency on these systems. The purpose of this work is to demonstrate the feasibility of 3D whole-brain T1 and T2 mapping using MR fingerprinting on a contemporary 0.55 T MRI system.Materials and methodsA 3D whole brain stack-of-spirals FISP MRF sequence was implemented for 0.55 T. Quantification was validated using the NIST/ISMRM Quantitative MRI phantom, and T1 and T2 values of white matter, gray matter, and cerebrospinal fluid were measured in 19 healthy subjects. To assess MRF performance in the lower SNR regime of 0.55 T, measurement precision was calculated from 100 simulated pseudo-replicas of in vivo data and within-session measurement repeatability was evaluated.ResultsT1 and T2 values calculated by MRF were strongly correlated to standard measurements in the ISMRM/NIST MRI system phantom (R² > 0.99), with a small constant bias of approximately 5 ms in T2 values. 3D stack-of-spirals MRF was successfully applied for whole brain quantitative T1 and T2 at 0.55 T, with spatial resolution of 1.2 mm × 1.2 mm × 5 mm, and acquisition time of 8.5 min. Moreover, the T1 and T2 quantifications had precision <5%, despite the lower SNR of 0.55 T.ConclusionA 3D whole-brain stack-of-spirals FISP MRF sequence is feasible for T1 and T2 mapping at 0.55 T.     

A multi-inversion-recovery magnetic resonance fingerprinting for multi-compartment water mapping

PurposeThis study aimed at introducing short-T1/T2 compartment to MR fingerprinting (MRF) at 3 T. Water that is bound to myelin macromolecules have significantly shorter T1 and T2 than free water and can be distinguished from free water by multi-compartment analysis.MethodsWe developed a new multi-inversion-recovery (mIR) water mapping-MRF based on an unbalanced steady-state coherent sequence (FISP). mIR pulses with an interval of 400 or 500 repetition times (TRs) were inserted into the conventional FISP MRF sequence. Data from our proposed mIR MRF was used to quantify different compartments, including myelin water, gray matter free water, and white matter free water, of brain water by virtue of the iterative non-negative least square (NNLS) with reweighting. Three healthy volunteers were scanned with mIR MRF on a clinical 3 T MRI.ResultsUsing an extended phase graph simulation, we found that our proposed mIR scheme with four IR pulses allowed differentiation between short and long T1/T2 components. For in vivo experiments, we achieved the quantification of myelin water, gray matter water, and white matter water at an image resolution of 1.17 × 1.17 × 5 mm³/pixel. As compared to the conventional MRF technique with single IR, our proposed mIR improved the detection of myelin water content. In addition, mIR MRF using spiral-in/out trajectory provided a higher signal level compared with that with spiral-out trajectory. Myelin water quantification using mIR MRF with 4 IR and 5 IR pulses were qualitatively similar. Meanwhile, 5 IR MRF showed fewer artifacts in myelin water detection.ConclusionWe developed a new mIR MRF sequence for the rapid quantification of brain water compartments.     

APIR4EMC: Autocalibrated parallel imaging reconstruction for extended multi-contrast imaging

PurposeTo improve image quality of multi-contrast imaging with the proposed Autocalibrated Parallel Imaging Reconstruction for Extended Multi-Contrast Imaging (APIR4EMC).MethodsAPIR4EMC reconstructs multi-contrast images in an autocalibrated parallel imaging reconstruction framework by adding contrasts as virtual coils. Compensation of signal evolution along the echo train of different contrasts is performed to improve signal prediction for missing samples. As a proof of concept, we performed prospectively accelerated phantom and in-vivo brain acquisitions with T1, T1-fat saturated (Fatsat), T2, PD, and FLAIR contrasts. The k-space sampling patterns of these acquisitions were jointly optimized. Images were jointly reconstructed with the proposed APIR4EMC method as well as individually with GRAPPA. Root mean square error (RMSE) to fully sampled reference images and g-factor maps were computed for both methods in the phantom experiment. Visual evaluation was performed in the in-vivo experiment.ResultsCompared to GRAPPA, APIR4EMC reduced artifacts and improved SNR of the reconstructed images in the phantom acquisitions. Quantitatively, APIR4EMC substantially reduced noise amplification (g-factor) as well as RMSE compared to GRAPPA. Signal evolution compensation reduced artifacts. In the in-vivo experiments, 1 mm³ isotropic 3D images with contrasts of T1, T1-Fatsat, T2, PD, and FLAIR were acquired in as little as 7.5 min with the acceleration factor of 9. Reconstruction quality was consistent with the phantom results.ConclusionCompared to single contrast reconstruction with GRAPPA, APIR4EMC reduces artifacts and noise amplification in accelerated multi-contrast imaging.     

Black-blood T2* mapping with delay alternating with nutation for tailored excitation

PurposeTo develop a black-blood T2* mapping method using a Delay Alternating with Nutation for Tailored Excitation (DANTE) preparation combined with a multi-echo gradient echo (GRE) readout (DANTE-GRE).Materials and methodsSimulations of the Bloch equation for DANTE-GRE were performed to optimize sequence parameters. After optimization, the sequence was applied to a phantom scan and to neck and lower extremity scans conducted on 12 volunteers at 3 T using DANTE-GRE, Motion-Sensitized Driven Equilibrium (MSDE)-GRE, and multi-echo GRE. T2* values were measured using an offset model. Statistical analyses were conducted to compare the T2* values between the three sequences.ResultsSimulation results showed that blood suppression can be achieved with various DANTE parameter adjustments. T2* maps acquired by DANTE-GRE were consistent and comparable to those acquired with multi-echo GRE in phantom experiments. In the in vivo experiments, DANTE-GRE was more comparable to multi-echo GRE than MSDE-GRE regarding the measurement of muscle T2* values.ConclusionDue to its high signal intensity retention and effective blood signal suppression, DANTE-GRE allows for robust and accurate T2* quantification, superior to that of MSDE-GRE, while overcoming blood flow artifacts associated with traditional multi-echo GRE.     

Magnetic Resonance Fingerprinting with short relaxation intervals

PurposeThe aim of this study was to investigate a technique for improving the performance of Magnetic Resonance Fingerprinting (MRF) in repetitive sampling schemes, in particular for 3D MRF acquisition, by shortening relaxation intervals between MRF pulse train repetitions.Material and methodsA calculation method for MRF dictionaries adapted to short relaxation intervals and non-relaxed initial spin states is presented, based on the concept of stationary fingerprints. The method is applicable to many different k-space sampling schemes in 2D and 3D. For accuracy analysis, T₁ and T₂ values of a phantom are determined by single-slice Cartesian MRF for different relaxation intervals and are compared with quantitative reference measurements. The relevance of slice profile effects is also investigated in this case. To further illustrate the capabilities of the method, an application to in-vivo spiral 3D MRF measurements is demonstrated.ResultsThe proposed computation method enables accurate parameter estimation even for the shortest relaxation intervals, as investigated for different sampling patterns in 2D and 3D. In 2D Cartesian measurements, we achieved a scan acceleration of more than a factor of two, while maintaining acceptable accuracy: The largest T₁ values of a sample set deviated from their reference values by 0.3% (longest relaxation interval) and 2.4% (shortest relaxation interval). The largest T₂ values showed systematic deviations of up to 10% for all relaxation intervals, which is discussed. The influence of slice profile effects for multislice acquisition is shown to become increasingly relevant for short relaxation intervals. In 3D spiral measurements, a scan time reduction of 36% was achieved, maintaining the quality of in-vivo T1 and T2 maps.ConclusionsReducing the relaxation interval between MRF sequence repetitions using stationary fingerprint dictionaries is a feasible method to improve the scan efficiency of MRF sequences. The method enables fast implementations of 3D spatially resolved MRF.     

Motion artifacts reduction in brain MRI by means of a deep residual network with densely connected multi-resolution blocks (DRN-DCMB)

Objective: Magnetic resonance imaging (MRI) acquisition is inherently sensitive to motion, and motion artifact reduction is essential for improving image quality in MRI. Methods: We developed a deep residual network with densely connected multi-resolution blocks (DRN-DCMB) model to reduce the motion artifacts in T1 weighted (T1W) spin echo images acquired on different imaging planes before and after contrast injection. The DRN-DCMB network consisted of multiple multi-resolution blocks connected with dense connections in a feedforward manner. A single residual unit was used to connect the input and output of the entire network with one shortcut connection to predict a residual image (i.e. artifact image). The model was trained with five motion-free T1W image stacks (pre-contrast axial and sagittal, and post-contrast axial, coronal, and sagittal images) with simulated motion artifacts. Results: In other 86 testing image stacks with simulated artifacts, our DRN-DCMB model outperformed other state-of-the-art deep learning models with significantly higher structural similarity index (SSIM) and improvement in signal-to-noise ratio (ISNR). The DRN-DCMB model was also applied to 121 testing image stacks appeared with various degrees of real motion artifacts. The acquired images and processed images by the DRN-DCMB model were randomly mixed, and image quality was blindly evaluated by a neuroradiologist. The DRN-DCMB model significantly improved the overall image quality, reduced the severity of the motion artifacts, and improved the image sharpness, while kept the image contrast. Conclusion: Our DRN-DCMB model provided an effective method for reducing motion artifacts and improving the overall clinical image quality of brain MRI.     

基于级联卷积神经网络的前列腺磁共振图像分类

针对深度学习训练成本高,以及基于磁共振图像的前列腺癌临床诊断需要大量医学常识且极为耗时的问题,本文提出了一种基于级联卷积神经网络（Convolutional Neural Network,CNN）和磁共振图像的前列腺癌（Prostate Cancer,PCa）自动分类诊断方法,该网络以Faster-RCNN作为前网络,对前列腺区域进行提取分割,用于排除前列腺附近组织器官的干扰;以基于ResNet改进的网络结构CNN40_bottleneck作为后网络,用于对前列腺区域病变进行分类.后网络由瓶颈结构串联组成,其中使用批量标准化（Batch Normalization,BN）、全局平均池化（Global Average Pooling,GAP）进行优化.实验结果证明,本文方法对前列腺癌诊断结果较好,而且缩减了训练时间和参数量,有效降低了训练成本.     

Correction of out-of-FOV motion artifacts using convolutional neural network

PurposeSubject motion during MRI scan can result in severe degradation of image quality. Existing motion correction algorithms rely on the assumption that no information is missing during motions. However, this assumption does not hold when out-of-FOV motion happens. Currently available algorithms are not able to correct for image artifacts introduced by out-of-FOV motion. The purpose of this study is to demonstrate the feasibility of incorporating convolutional neural network (CNN) derived prior image into solving the out-of-FOV motion problem.Methods and materialsA modified U-net network was proposed to correct out-of-FOV motion artifacts by incorporating motion parameters into the loss function. A motion model based data fidelity term was applied in combination with the CNN prediction to further improve the motion correction performance. We trained the CNN on 1113 MPRAGE images with simulated oscillating and sudden motion trajectories, and compared our algorithm to a gradient-based autofocusing (AF) algorithm in both 2D and 3D images. Additional experiment was performed to demonstrate the feasibility of transferring the networks to different dataset. We also evaluated the robustness of this algorithm by adding Gaussian noise to the motion parameters. The motion correction performance was evaluated using mean square error (NMSE), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).ResultsThe proposed algorithm outperformed AF-based algorithm for both 2D (NMSE: 0.0066 ± 0.0009 vs 0.0141 ± 0.008, P < .01; PSNR: 29.60 ± 0.74 vs 21.71 ± 0.27, P < .01; SSIM: 0.89 ± 0.014 vs 0.73 ± 0.004, P < .01) and 3D imaging (NMSE: 0.0067 ± 0.0008 vs 0.070 ± 0.021, P < .01; PSNR: 32.40 ± 1.63 vs 22.32 ± 2.378, P < .01; SSIM: 0.89 ± 0.01 vs 0.62 ± 0.03, P < .01). Robust reconstruction was achieved with 20% data missed due to the out-of-FOV motion.ConclusionIn conclusion, the proposed CNN-based motion correction algorithm can significantly reduce out-of-FOV motion artifacts and achieve better image quality compared to AF-based algorithm.     

MRF-ZOOM for the unbalanced steady-state free precession (ubSSFP) magnetic resonance fingerprinting

In magnetic resonance fingerprinting (MRF), tissue parameters are determined by finding the best-match to the acquired MR signal from a predefined signal dictionary. This dictionary searching (DS) process is generally performed in an exhaustive manner, which requires a large predefined dictionary and long searching time. A fast MRF DS algorithm, MRF-ZOOM, was recently proposed based on DS objective function optimization. As a proof-of-concept study, MRF-ZOOM was only tested with one of the earliest MRF sequences but not with the recently more popular unbalanced steady state free precession MRF sequence (MRF-ubSSFP, or MRF-FISP). Meanwhile noise effects on MRF and MRF-ZOOM have not been examined. The purpose of this study was to address these open questions and to verify whether MRF-ZOOM can be combined with a dictionary-compression based method to gain further speed. Numerical simulations were performed to evaluate the DS objective function properties, noise effects on MRF, and to compare MRF-ZOOM with other methods in terms of speed and accuracy. In-vivo experiments were performed as well. Evaluation results showed that premises of MRF-ZOOM held for MRF-FISP; noise did not affect MRF-ZOOM more than the conventional MRF method; when SNR ≥ 1, MRF quantification yielded accurate results. Dictionary compression introduced quantification errors more to T2 quantification. MRF-ZOOM was thousands of times faster than the conventional MRF method. Combining MRF-ZOOM with dictionary compression showed no benefit in terms of fitting speed. In conclusion, MRF-ZOOM is valid for MRF- FISP, and can remarkably save MRF dictionary generation and searching time without sacrificing matching accuracy.     

Towards predicting the encoding capability of MR fingerprinting sequences

Sequence optimization and appropriate sequence selection is still an unmet need in magnetic resonance fingerprinting (MRF). The main challenge in MRF sequence design is the lack of an appropriate measure of the sequence's encoding capability. To find such a measure, three different candidates for judging the encoding capability have been investigated: local and global dot-product-based measures judging dictionary entry similarity as well as a Monte Carlo method that evaluates the noise propagation properties of an MRF sequence. Consistency of these measures for different sequence lengths as well as the capability to predict actual sequence performance in both phantom and in vivo measurements was analyzed. While the dot-product-based measures yielded inconsistent results for different sequence lengths, the Monte Carlo method was in a good agreement with phantom experiments. In particular, the Monte Carlo method could accurately predict the performance of different flip angle patterns in actual measurements. The proposed Monte Carlo method provides an appropriate measure of MRF sequence encoding capability and may be used for sequence optimization.     

一种基于双分支改良编解码器的图像去噪算法

针对传统图像去噪算法多噪声去除难,深层卷积神经网络去噪模型网络复杂、训练时间长等问题,提出一种基于自编码器结构的双分支改良编解码网络,实现高效图像去噪。双分支结构之一采用降-升采样实现点噪声消除,另一分支专注于宏观的图像修复和伪像去除,后端利用残差结构进行整合,实现数字图像混合噪声去噪。实验结果显示:对于含有标准差为15,均值为0的高斯噪声、噪声密度为5%的椒盐噪声和散粒噪声的混合噪声图像测试集,实验去噪效果相较于输入混合噪声图像峰值信噪比,平均提升了5.3%。与12层全卷积神经网络相比,去噪效果相当,训练速度提升了25.4%,体现了其“轻量级”的优点。实验表明:该方法相较于深层卷积神经网络,训练速度快,网络简单;相较于传统图像去噪算法,噪声去除效果也较为明显。该算法可应用于轻量级视觉平台后端去噪。     

Deep-learning-based prediction of living cells mitosis via quantitative phase microscopy

We present a deep learning approach for living cells mitosis classification based on label-free quantitative phase imaging with transport of intensity equation methods. In the approach, we applied a pretrained deep convolutional neural network using transfer learning for binary classification of mitosis and non-mitosis. As a validation, we demonstrated the performances of the network trained by phase images and intensity images, respectively. The convolutional neural network trained by phase images achieved an average accuracy of 98.9% on the validation data, which outperforms the average accuracy 89.6% obtained by the network trained by intensity images. We believe that the quantitative phase microscopy in combination with deep learning enables researchers to predict the mitotic status of living cells noninvasively and efficiently.     

Simultaneous quantification of SPIO and gadolinium contrast agents using MR fingerprinting

PurposeDevelop a magnetic resonance fingerprinting (MRF) methodology with R₂¹ quantification, intended for use with simultaneous contrast agent concentration mapping, particularly gadolinium (Gd) and iron labelled CD8+ T cells.MethodsVariable-density spiral SSFP MRF was used, modified to allow variable TE, and with an exp.(−TE·R₂¹) dictionary modulation. In vitro phantoms containing SPIO labelled cells and/or gadolinium were used to validate parameter maps, probe undersampling capacity, and verify dual quantification capabilities. A C57BL/6 mouse was imaged using MRF to demonstrate acceptable in vivo resolution and signal at 8× undersampling necessary for a 25-min scan.ResultsStrong agreement was found between conventional and MRF-derived values for R₁, R₂, and R₂¹. Expanded MRF allowed quantification of iron-loaded CD8+ T cells. Results were robust to 8× undersampling and enabled recreation of relaxation profiles for both a Gd agent and iron labelled cells simultaneously. In vivo data demonstrated sufficient SNR in undersampled data for parameter mapping to visualise key features.ConclusionMRF can be expanded to include R₁, R₂, and R₂¹ mapping required for simultaneous quantification of gadolinium and SPIO in vitro, allowing for potential implementation of a variety of future in vivo studies using dual MR contrast agents, including molecular imaging of labelled cells.     

RF slice profile effects in magnetic resonance fingerprinting

The radio frequency (RF) slice profile effects on T1 and T2 estimation in magnetic resonance fingerprinting (MRF) are investigated with respect to time-bandwidth product (TBW), flip angle (FA) level and field inhomogeneities. Signal evolutions are generated incorporating the non-ideal slice selective excitation process using Bloch simulation and matched to the original dictionary with and without the non-ideal slice profile taken into account. For validation, phantom and in vivo experiments are performed at 3T. Both simulations and experiments results show that T1 and T2 error from non-ideal slice profile increases with increasing FA level, off-resonance, and low TBW values. Therefore, RF slice profile effects should be compensated for accurate determination of the MR parameters.     

深度学习在核物理中的应用

深度学习是目前最好的模式识别工具,预期会在核物理领域帮助科学家从大量复杂数据中寻找与某些物理最相关的特征。本文综述了深度学习技术的分类,不同数据结构对应的最优神经网络架构,黑盒模型的可解释性与预测结果的不确定性。介绍了深度学习在核物质状态方程、核结构、原子核质量、衰变与裂变方面的应用,并展示如何训练神经网络预测原子核质量。结果发现使用实验数据训练的神经网络模型对未参与训练的实验数据拥有良好的预测能力。基于已有的实验数据外推,神经网络对丰中子的轻原子核质量预测结果与宏观微观液滴模型有较大偏离。此区域可能存在未被宏观微观液滴模型包含的新物理,需要进一步的实验数据验证。     

Sensor and Actuator Fault Diagnosis for Robot Joint Based on Deep CNN

This paper proposes a data-driven method-based fault diagnosis method using the deep convolutional neural network (DCNN). The DCNN is used to deal with sensor and actuator faults of robot joints, such as gain error, offset error, and malfunction for both sensors and actuators, and different fault types are diagnosed using the trained neural network. In order to achieve the above goal, the fused data of sensors and actuators are used, where both types of fault are described in one formulation. Then, the deep convolutional neural network is applied to learn characteristic features from the merged data to try to find discriminative information for each kind of fault. After that, the fully connected layer does prediction work based on learned features. In order to verify the effectiveness of the proposed deep convolutional neural network model, different fault diagnosis methods including support vector machine (SVM), artificial neural network (ANN), conventional neural network (CNN) using the LeNet-5 method, and long-term memory network (LTMN) are investigated and compared with DCNN method. The results show that the DCNN fault diagnosis method can realize high fault recognition accuracy while needing less model training time.     

Synthetic signal injection using inductive coupling

The limited bandwidths of volume selective RF pulses in localized in vivo MRS experiments introduce spatial artifacts that complicate spectral quantification of J-coupled metabolites. These effects are commonly referred to as a spatial interference or "four compartment" artifacts and are more pronounced at higher field strengths. The main focus of this study is to develop a generalized approach to numerical simulations that combines full density matrix calculations with 3D localization to investigate the spatial artifacts and to provide accurate prior knowledge for spectral fitting. Full density matrix calculations with 3D localization using experimental pulses were carried out for PRESS (TE=20, 70 ms), STEAM (TE=20, 70 ms) and LASER (TE=70 ms) pulse sequences and compared to non-localized simulations and to phantom solution data at 4 T. Additional simulations at 1.5 and 7 T were carried out for STEAM and PRESS (TE=20 ms). Four brain metabolites that represented a range from weak to strong J-coupling networks were included in the simulations (lactate, N-acetylaspartate, glutamate and myo-inositol). For longer TE, full 3D localization was necessary to achieve agreement between the simulations and phantom solution spectra for the majority of cases in all pulse sequence simulations. For short echo time (TE=20 ms), ideal pulses without localizing gradients gave results that were in agreement with phantom results at 4 T for STEAM, but not for PRESS (TE=20). Numerical simulations that incorporate volume localization using experimental RF pulses are shown to be a powerful tool for generation of accurate metabolic basis sets for spectral fitting and for optimization of experimental parameters.     

Quantitative MRI: Defining repeatability,reproducibility and accuracy for prostate cancer imaging biomarker development

IntroductionQuantitative MRI (qMRI) parameters have been increasingly used to develop predictive models to accurately monitor treatment response in prostate cancer after radiotherapy. To reliably detect changes in signal due to treatment response, predictive models require qMRI parameters with high repeatability and reproducibility. The purpose of this study was to measure qMRI parameter uncertainties in both commercial and in-house developed phantoms to guide the development of robust predictive models for monitoring treatment response.Materials and methodsADC, T1, and R2* values were acquired across three 3 T scanners with a prostate-specific qMRI protocol using the NIST/ISMRM system phantom, RSNA/NIST diffusion phantom, and an in-house phantom. A B1 field map was acquired to correct for flip angle inhomogeneity in T1 maps. All sequences were repeated in each scan to assess within-session repeatability. Weekly scans were acquired on one scanner for three months with the in-house phantom. Between-session repeatability was measured with test-retest scans 6-months apart on all scanners with all phantoms. Accuracy, defined as percentage deviation from reference value for ADC and T1, was evaluated using the system and diffusion phantoms. Repeatability and reproducibility coefficients of variation (%CV) were calculated for all qMRI parameters on all phantoms.ResultsOverall, repeatability CV of ADC was <2.40%, reproducibility CV was <3.98%, and accuracy ranged between −8.0% to 2.7% across all scanners. Applying B1 correction on T1 measurements significantly improved the repeatability and reproducibility (p<0.05) but increased error in accuracy (p<0.001). Repeatability and reproducibility of R2* was <4.5% and <7.3% respectively in the system phantom across all scanners.ConclusionRepeatability, reproducibility, and accuracy in qMRI parameters from a prostate-specific protocol was estimated using both commercial and in-house phantoms. Results from this work will be used to identify robust qMRI parameters for use in the development of predictive models to longitudinally monitor treatment response for prostate cancer in current and future clinical trials.     

snapMRF: GPU-accelerated magnetic resonance fingerprinting dictionary generation and matching using extended phase graphs

Purpose:Magnetic resonance fingerprinting (MRF) is a state-of-the-art quantitative MRI technique with a computationally demanding reconstruction process, the accuracy of which depends on the accuracy of the signal model employed. Having a fast, validated, open-source MRF reconstruction would improve the dependability and accuracy of clinical applications of MRF.Methods:We parallelized both dictionary generation and signal matching on the GPU by splitting the simulation and matching of dictionary atoms across threads. Signal generation was modeled using both Bloch equation simulation and the extended phase graph (EPG) formalism. Unit tests were implemented to ensure correctness. The new package, snapMRF, was tested with a calibration phantom and an in vivo brain.Results:Compared with other online open-source packages, dictionary generation was accelerated by 10–1000× and signal matching by 10–100×. On a calibration phantom, T₁ and T₂ values were measured with relative errors that were nearly identical to those from existing packages when using the same sequence and dictionary configuration, but errors were much lower when using variable sequences that snapMRF supports but that competitors do not.Conclusion:Our open-source package snapMRF was significantly faster and retrieved accurate parameters, possibly enabling real-time parameter map generation for small dictionaries. Further refinements to the acquisition scheme and dictionary setup could improve quantitative accuracy.