Multi-parametric imaging using multiple-echoes with optimized simultaneous acquisition for highly-efficient quantitative magnetic resonance imaging

The multi-echo gradient echo sequence and neural network-based reconstruction in qMRI enable simultaneous estimation of multiple parameters, addressing efficiency and alignment issues, resulting in improved tissue characterization and reduced artifacts.

WO2026236752A1PCT designated stage Publication Date: 2026-11-12THE GENERAL HOSPITAL CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027653
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2026-05-11
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing quantitative magnetic resonance imaging (qMRI) techniques face challenges in efficiently acquiring and aligning multiple parameters such as Ti, T2, T2*, and susceptibility mapping, leading to long acquisition times and misalignment issues that hinder clinical and research applications.

Method used

A method involving a multi-echo gradient echo sequence with variable flip angles and a neural network-based reconstruction approach is employed to simultaneously estimate multiple quantitative parameters from a single acquisition, using a magnetization preparation RF pulse followed by readout acquisitions and k-space sampling, enabling co-registered parametric maps without separate acquisitions.

Benefits of technology

This approach reduces acquisition time, minimizes motion-related artifacts, and improves tissue characterization by providing comprehensive, accurately aligned multi-parametric maps within clinically feasible times.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026027653_12112026_PF_FP_ABST
    Figure US2026027653_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A method for multi-parametric quantitative magnetic resonance imaging (MRI) is provided herein. The method includes applying a magnetization preparation radio frequency (RF) pulse to a subject using an MRI system. Following application of the magnetization preparation RF pulse, the method includes performing a plurality of readout acquisitions as longitudinal magnetization recovers back to equilibrium. The plurality of readout acquisitions includes one or more single-echo readouts and a multi-echo gradient echo (MGRE) readout that acquires data at a plurality of echo times. The method further includes sampling k-space data during the plurality of readout acquisitions using a variable-density k-space sampling strategy. The method also includes reconstructing multi-contrast images from the sampled k-space data.
Need to check novelty before this filing date? Find Prior Art

Description

MGH 2025-095-02125141.05025MULTI-PARAMETRIC IMAGING USING MULTIPLE-ECHOES WITH OPTIMIZED SIMULTANEOUS ACQUISITION FOR HIGHLY-EFFICIENT QUANTITATIVE MAGNETIC RESONANCE IMAGINGCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Patent Application Serial No. 63 / 803,492, filed on May 9, 2025, and entitled “MULTI-PARAMETRIC IMAGING USING MULTIPLE-ECHOES WITH OPTIMIZED SIMULTANEOUS ACQUISITION FOR HIGHLY-EFFICIENT QUANTITATIVE MRI,” which is herein incorporated by reference in its entirety.STATEMENT OF FEDERALLY SPONSORED RESEARCH

[0002] This invention was made with government support under EB032378 awarded by the National Institutes of Health. The government has certain rights in the invention.BACKGROUND

[0003] Quantitative magnetic resonance imaging (qMRI) measures physical parameters, such as relaxation times and magnetic susceptibility, to provide information related to the local structural environment of tissues. For example, Ti relaxation time may be associated with molecular mobility, T2 relaxation time may reflect water molecular environment, and T2* relaxation time may provide insights into tissue microenvironment and iron deposition. Quantitative susceptibility mapping (QSM) measures the spatial distribution of magnetic susceptibility, and paramagnetic and diamagnetic susceptibility sources in brain tissues may help in characterizing tissue iron and myelin concentration. Such quantitative parameters have been used for the study of aging, brain development, and the assessment of disease and treatment.

[0004] Many diseases exhibit a wide range of pathological changes that affect different aspects of tissues, making it difficult to relate changes of a single parameter to a specific pathological process. Additionally, single parameter mapping techniques often suffer from long acquisition times, low efficiency, and misalignment with other complementary acquisitions, hampering their application in clinical and research settings. Simultaneous multi-parametric mapping techniques use tailored pulse sequences to estimate a set of parameters with complementary information while eliminating the need for co-registration, offering the potential1QB\125141.05025\102311859.2MGH 2025-095-02125141.05025to enable comprehensive tissue characterization and improve clinical diagnosis. However, there exist few techniques that can provide Ti, T2, T2*, and susceptibility mapping simultaneously. Therefore, improved systems and methods for multi-parametric quantitative MRI are desired.SUMMARY OF THE DISCLOSURE

[0005] According to an aspect of the present disclosure, a method for multi-parametric quantitative magnetic resonance imaging (MRI) is provided. The method includes applying a magnetization preparation radio frequency (RF) pulse to a subject using an MRI system. Following application of the magnetization preparation RF pulse, the method includes performing a plurality of readout acquisitions as longitudinal magnetization recovers back to equilibrium. The plurality of readout acquisitions includes one or more single-echo readouts and a multi-echo gradient echo (MGRE) readout that acquires data at a plurality of echo times. The method further includes sampling k-space data during the plurality of readout acquisitions using a variable-density k-space sampling strategy. The method also includes reconstructing multi-contrast images from the sampled k-space data.

[0006] According to another aspect of the present disclosure, a method for reconstructing multi-contrast images using a neural network is provided. The method includes accessing, with a computer system, undersampled multi-contrast k-space data acquired using a multi-parametric MRI pulse sequence. The method includes accessing, with the computer system, a trained neural network, where the trained neural network implements an unrolled network architecture including a plurality of iterative blocks, each iterative block including one or more convolutional layers followed by a data consistency layer. The method further includes inputting the undersampled multi-contrast k-space data to the trained neural network. The method also includes generating, as output from the trained neural network, reconstructed multi -contrast images.

[0007] According to another aspect of the present disclosure, a method for training a neural network for multi-contrast image reconstruction is provided. The method includes accessing, with a computer system, undersampled multi -contrast k-space data. The method includes partitioning an undersampling mask into a plurality of disjoint subsets including a training mask, a validation mask, and a loss mask. The method includes training a neural network on the undersampled multicontrast k-space data using a zero-shot self-supervised learning approach. Training includes2QB\125141.05025\102311859.2MGH 2025-095-02125141.05025applying the training mask to enforce data consistency constraints within the neural network, applying the loss mask to compute a training loss and update network parameters, and applying the validation mask to compute a validation loss as a stopping criterion. The method further includes storing the trained neural network for later use.BRIEF DESCRIPTION OF THE DRAWINGS

[0008] FIG. 1 illustrates an example pulse sequence for multi-parametric quantitative MRI.

[0009] FIG. 2A illustrates a first pulse sequence configuration with readouts followed by a multi-echo gradient echo module, according to aspects of the present disclosure.

[0010] FIG. 2B illustrates a second pulse sequence configuration with a multi-echo gradient echo module inserted before a final readout, according to aspects of the present disclosure.

[0011] FIG. 2C illustrates a third pulse sequence configuration with readouts followed by a multi-echo gradient echo module, according to aspects of the present disclosure.

[0012] FIG. 2D illustrates a fourth pulse sequence configuration without delays between readouts, according to aspects of the present disclosure.

[0013] FIG. 2E illustrates a fifth pulse sequence configuration without delays between readouts, according to aspects of the present disclosure.

[0014] FIG. 2F illustrates a sixth pulse sequence configuration without delays between readouts, according to aspects of the present disclosure.

[0015] FIG. 3 is a flowchart of an example method for multi-parametric quantitative MRI.

[0016] FIG. 4 is a flowchart of an example method for reconstructing multi-contrast images using a neural network.

[0017] FIG. 5 illustrates an overview of a multi-contrast multi-slice zero-shot selfsupervised learning reconstruction.

[0018] FIG. 6 is a flowchart of an example method for training a neural network using zero-shot self-supervised learning.

[0019] FIG. 7 illustrates an example system for multi-parametric quantitative MRI.

[0020] FIG. 8 illustrates example hardware for implementing the system of FIG. 7.

[0021] FIG. 9 illustrates an example MRI system.DETAILED DESCRIPTION3QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0022] Described here are systems and methods for multi-parametric quantitative MRI using multi-echo acquisition. The systems and methods described herein enable simultaneous estimation of multiple quantitative parameters, including Ti, T2, T2*, and magnetic susceptibility, from a single acquisition sequence. In some embodiments, a multi-echo gradient echo sequence with variable flip angles is employed to acquire data that encodes information for all target parameters. The acquired multi-echo data can be processed using model-based fitting approaches that jointly estimate the relaxation times and susceptibility values, thereby providing inherently co-registered parametric maps without the need for separate acquisitions or post-hoc registration. The techniques described herein offer improved acquisition efficiency compared to conventional single-parameter mapping methods, enabling comprehensive tissue characterization within clinically feasible scan times. Additionally, the simultaneous acquisition approach reduces motion-related artifacts and misalignment errors that can arise when combining separately acquired parametric maps.

[0023] Quantitative MRI measures physical parameters (e.g., relaxation times and magnetic susceptibility) to provide information related to the local structural environment of tissues. For example, Ti relaxation time may be associated with molecular mobility, T2 relaxation time may reflect water molecular environment, and T2* relaxation time may provide insights into tissue microenvironment and iron deposition. Quantitative susceptibility mapping (QSM) measures the spatial distribution of magnetic susceptibility, and paramagnetic and diamagnetic susceptibility sources in brain tissues may help in characterizing tissue iron and myelin concentration. Such quantitative parameters may be used for the study of aging, brain development, and the assessment of disease and treatment.

[0024] In some aspects, the systems and methods described herein provide multiparametric quantitative MRI using multi-echo acquisition. The techniques may enable simultaneous estimation of multiple quantitative parameters, including Ti, T2, T2*, proton density , and magnetic susceptibility, from a single acquisition sequence. In some cases, the multiparametric acquisition approach may eliminate the need for co-regi strati on between separately acquired parametric maps, as the multiple parameters may be derived from a common dataset. The simultaneous acquisition of multiple parameters may provide complementary information about tissue composition and microstructure.4QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0025] In some embodiments, the techniques described herein may enable susceptibility source separation from a single scan. Paramagnetic and diamagnetic susceptibility sources may be separated to provide information on tissue iron and myelin concentration. The separation of susceptibility sources may offer insights into tissue composition that may not be available from total susceptibility measurements alone.

[0026] As described in more detail below, the multi-parametric mapping approach described herein may use a tailored pulse sequence that encodes information for multiple target parameters within a single acquisition. In some cases, the acquired data may be processed using model-based fitting approaches that jointly estimate the relaxation times and susceptibility values. The joint estimation approach may provide inherently co-registered parametric maps without the need for separate acquisitions or post-hoc registration.

[0027] In some aspects, the techniques described herein may offer acquisition efficiency compared to conventional single-parameter mapping methods. The simultaneous acquisition of multiple parameters from a single scan may reduce total scan time and may enable comprehensive tissue characterization within clinically feasible acquisition times. Additionally, the simultaneous acquisition approach may reduce motion-related artifacts and misalignment errors that can arise when combining separately acquired parametric maps.

[0028] Referring now to FIG. 1, an example pulse sequence that encodes information for multiple target parameters within a single acquisition is illustrated. In some embodiments, the pulse sequence and related imaging method may be referred to as Multi-parametric Imaging using Multiple-echoes with Optimized Simultaneous Acquisition (MIMOSA). The pulse sequence may include a magnetization preparation radio frequency (RF) pulse 102 followed by a plurality of multi-echo acquisitions that are performed as longitudinal magnetization recovers back to equilibrium following application of the magnetization preparation RF pulse 102.

[0029] In some embodiments, the magnetization preparation RF pulse 102 may include an inversion recovery RF pulse. The inversion recovery RF pulse may invert the longitudinal magnetization, creating magnetization-prepared longitudinal magnetization that subsequently recovers toward equilibrium overtime. In some cases, the magnetization preparation RF pulse 102 may include an adiabatic inversion recovery pulse, which may provide uniform inversion across a range of RF field strengths.5QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0030] The pulse sequence may also include a T2 preparation module 104. In some embodiments, the T2 preparation module 104 may include a 90-degree tip-down RF pulse followed by a series of adiabatic RF refocusing pulses and a 90-degree tip-up RF pulse. The T2 preparation module 104 may encode T2 contrast information into the longitudinal magnetization prior to subsequent readout acquisitions. T2 relaxation may occur during the T2-sensitizing period of the T2 preparation module 104. In some aspects, the T2 relaxation that occurs during the T2 preparation module may be described by the relationship M2= MxexpT£r^prep^, where Mi and M2 are thelongitudinal magnetizations before and after the T2 preparation module respectively, and TEi2prep is the duration of the T2 preparation block.

[0031] Following the magnetization preparation RF pulse 102, the pulse sequence may include a series of readouts 106, which in some instances may include applying RF excitation pulses at different flip angles. For example, the one or more single-echo readouts 106, which may include Fast Low Angle Shot (FLASH) readouts. The FLASH readout may combine low flip angle excitation with short repetition time to enable fast acquisition. In some cases, the pulse sequence may include multiple FLASH readouts performed at different time points during the magnetization recovery period.

[0032] In some embodiments, the pulse sequence may include a multi-echo gradient echo (MGRE) readout 108. The MGRE readout 108 may contain multiple echoes at different echo times to generate T2* and magnetic susceptibility contrast. As a non-limiting example, the MGRE readout 108 may include six echoes at echo times such as TE = 2.7, 7.0, 11.3, 15.6, 19.9, and 24.2 ms at 3T field strength, though other suitable echo time intervals may be selected based on the specific application requirements and field strength. In one example implementation, the repetition time of the MGRE module 108 may be approximately 27.5 ms, though other repetition times may be employed depending on the desired echo spacing and number of echoes. In some embodiments, the MGRE readout 108 may include four or more echoes spanning a range of echo times.

[0033] In some aspects, the steady-state transverse magnetization signal during the MGRE module 108 may be described by the relationship Mn(TEi) = Mo■ sin (g)(i e MGRE )e-TEi / T2 ei(i<p0+2n&.f-TEi)whereygjtfre ech0tjmeTRMGRE is the repetitiontime of the MGRE module 108, a is the flip angle, T2* is the transverse relaxation time, (po is the phase of magnetization at TE = 0, and Af is the total magnetic field which includes background 6QB\125141.05025\102311859.2MGH 2025-095-02125141.05025field induced by background susceptibility sources and tissue field induced by tissue susceptibility sources.

[0034] Each multi-echo acquisition may sample a different portion of the longitudinal magnetization recovery curve as the magnetization recovers back to equilibrium following application of the magnetization preparation RF pulse 102. By performing multiple multi-echo acquisitions during different portions of the magnetization recovery curve, data may be acquired from multiple different effective inversion times (TIs). For instance, each multi-echo acquisition may sample blocks of the longitudinal magnetization recovery curve, where each block has a different effective inversion time. As a non-limiting example, effective inversion times may include values such as TI = 1000, 1500, 2000, and 2500 ms, though other suitable time intervals may be selected based on the Ti values of the tissues of interest and the desired sampling of the magnetization recovery curve.

[0035] In some embodiments, the pulse sequence structure may include a T2 preparation module 104 followed by a first FLASH readout 106, and a magnetization preparation RF pulse 102 that includes an inversion recovery preparation pulse followed by additional FLASH readouts 106 and an MGRE readout 108. As a non-limiting example, the total repetition time (TR) of the pulse sequence may be approximately 6030 ms, though other total repetition times may be employed depending on the number of readout blocks and the desired sampling of the magnetization recovery curve. In some cases, one dummy scan may be incorporated at the beginning of the sequence to allow the signal to reach steady state. The multiple readouts performed at different time points during the magnetization recovery may provide data with different levels of Ti-weighted tissue contrast corresponding to the different effective inversion times.

[0036] In some aspects, the pulse sequence may be configured according to different configurations to optimize acquisition efficiency and parameter estimation accuracy. In some embodiments, a first configuration may include four FLASH readouts followed by the MGRE module 108 with delays between the FLASH readouts 106, as illustrated in FIG. 2A. A second configuration may include four FLASH readouts 106 with the MGRE module 108 inserted before the last FLASH readout 106 with delays between the FLASH readouts 106, as illustrated in FIG.2B. A third configuration may include three FLASH readouts 106 followed by the MGRE module 108 with delays between the FLASH readouts 106, as illustrated in FIG. 2C. In some cases, the 7QB\125141.05025\102311859.2MGH 2025-095-02125141.05025same three configurations may be implemented without delays between the FLASH readouts 106, as illustrated in FIGS. 2D-2F. In some embodiments, a configuration with three FLASH readouts 106 followed by the MGRE module 108 without delays between the FLASH readouts 106 may be selected as an optimized configuration, as illustrated in FIG. 2F, as this configuration may achieve lower estimation errors while reducing the total repetition time by approximately 12% compared to configurations with delays.

[0037] The k-space sampling order may vary between different readout types. In some cases, a center-out k-space sampling order may be employed in the FLASH readout 106 for each spoke, while an opposite out-center sampling order may be applied in the MGRE readout 108. The out-center sampling order in the MGRE readout 108 may allow for longer recovery time and may provide higher signal-to-noise ratio for T2* mapping.

[0038] By acquiring data at multiple different echo times, multiple levels of T2* contrast may be obtained. By acquiring data at multiple different effective inversion times, multiple levels of Ti contrast may be obtained. The combination of multiple echo times and multiple effective inversion times may provide a set of images with varying Ti and T2* contrast from a single acquisition sequence.

[0039] In some embodiments, the magnetization preparation and multi-echo time, multiinversion time acquisition may be repeated multiple times in an outer loop to complete the full data acquisition. The data acquired across the multiple repetitions may collectively form multicontrast data that encodes information for Ti, T2, T2*, proton density, and magnetic susceptibility mapping.

[0040] In some aspects, the multi-parametric quantitative MR1 techniques described herein may be implemented at ultra-high field strengths (e.g., 7T) to enable mesoscale quantitative mapping at submillimeter resolution. Mesoscale quantitative mapping may allow for the characterization of neuroanatomical structures with high spatial resolution, providing detailed insight into the composition of brain tissue. As a non-limiting example, mesoscale multiparametric mapping may be performed at 7T with approximately 750 pm isotropic resolution, though other submillimeter resolutions may be achieved depending on the available scan time and signal-to-noise requirements. Given the shortened T2* values at higher magnetic field strength, the MGRE module may be optimized by reducing the number of echoes, for example from six to four, though other echo counts may be selected based on the specific field strength and T2*8QB\125141.05025\102311859.2MGH 2025-095-02125141.05025characteristics of the tissues of interest. In one example implementation, the MGRE module at 7T may acquire data at echo times such as TE = 3, 10, 17, and 24 ms, though other suitable echo time intervals may be employed. As a non-limiting example, the total acquisition time for whole-brain mesoscale multi -parametric mapping at 7T may be approximately 13 minutes, though other acquisition times may be achieved depending on the selected resolution, acceleration factor, and coverage requirements.

[0041] In some aspects, the T2 preparation module 104 may be modified for 7T implementation to mitigate the effects of increased Bi+inhomogeneity at ultra-high field strengths. As a non-limiting example, the length of the adiabatic refocusing pulses used in the T2 preparation module 104 may be prolonged from approximately 3 ms to approximately 8 ms to boost BE robustness, though other pulse durations may be selected based on the specific Bi+characteristics and field strength. The prolonged adiabatic refocusing pulses may provide more uniform refocusing across a range of RF field strengths encountered at 7T.

[0042] In some aspects, the multi-parametric quantitative MRI techniques described herein may employ an accelerated data acquisition scheme to reduce total scan time while maintaining image quality. The accelerated data acquisition scheme may involve undersampling of k-space data along one or more phase-encoding or partition-encoding directions.

[0043] In some embodiments, a variable-density k-space sampling strategy may be employed to enable accelerated data acquisition. As a non-limiting example, the variable-density k-space sampling strategy may include a Variable-Density sampling and Radial view ordering (VDRad) Cartesian trajectory. The VDRad trajectory may generate spiral-like spokes on a Cartesian grid, where each spiral-like spoke contains an echo train length (ETL) of sampling points in the ky-kzplane. The spiral-like spokes may rotate across different repetition times by the golden angle, facilitating complementary variable-density sampling across different readout blocks. In some cases, the variable-density sampling may provide denser sampling near the center of k-space and sparser sampling in the peripheral regions of k-space. Complete k-space coverage may require NyxNz / ETL repetitions, where Ny and Nz represent the phase encoding number in each direction.

[0044] The accelerated acquisition may be performed at various acceleration factors. In some embodiments, acceleration factors may range from approximately 3 to approximately 12, or other suitable acceleration factors. As non-limiting examples, acceleration factors of approximately 3.3, 6.5, and 11.8 may be employed, though other acceleration factors may be 9QB\125141.05025\102311859.2MGH 2025-095-02125141.05025selected depending on the desired trade-off between acquisition time and image quality. Higher acceleration factors may enable shorter total acquisition times, while lower acceleration factors may provide higher signal-to-noise ratio in the reconstructed images. In some cases, an acceleration factor of up to approximately 11.8-fold may be achieved while preserving parameter estimation accuracy, though the maximum achievable acceleration factor may vary depending on the coil configuration and reconstruction approach.

[0045] In some aspects, parallel imaging techniques may be used to estimate missing k-space data that was not acquired due to undersampling. Parallel imaging techniques may exploit the spatial sensitivity variations of multiple receiver coils in a phased array to recover unsampled k-space locations. In some embodiments, generalized autocalibrating partially parallel acquisitions (GRAPPA) may be employed to estimate missing k-space data. GRAPPA may use a set of autocalibration signal (ACS) lines acquired near the center of k-space to determine coil-specific weighting coefficients. The weighting coefficients may then be applied to estimate the missing k-space data from the acquired undersampled data.

[0046] In some cases, auto-calibration data may be acquired as part of the variable-density sampling scheme. The denser sampling near the center of k-space in the variable-density sampling pattern may provide the auto-calibration data used for parallel imaging reconstruction. In some embodiments, the auto-calibration data may be acquired without requiring a separate calibration scan, as the variable-density sampling may inherently provide sufficient central k-space coverage for calibration purposes.

[0047] In some embodiments, low-rank matrix modeling techniques may be used for image reconstruction from the undersampled k-space data. Low-rank modeling of local k-space neighborhoods (LORAKS) may be employed to jointly reconstruct the acquired multi-contrast images. The LORAKS reconstruction may exploit the low-rank structure of local k-space neighborhoods to recover missing data and reduce noise in the reconstructed images. In some cases, the joint reconstruction of multi-contrast images may leverage the shared information across different contrast weightings to improve reconstruction quality. Additionally or alternatively, a multi-contrast multi-slice zero-shot self-supervised learning (MZS-SSL) approach may be employed for joint reconstruction of the multi -contrast images, which may enable robust reconstruction of highly undersampled data while leveraging information across contrasts and slices.10QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0048] The accelerated data acquisition scheme may enable multiple effective inversion times to be acquired within a single inversion recovery period. By undersampling each readout block and using parallel imaging or low-rank reconstruction techniques to recover the missing data, multiple FLASH readouts and an MGRE readout may be performed during the magnetization recovery period following a single inversion recovery preparation pulse. In some embodiments, three or more readout blocks may be acquired within a single inversion recovery period, with each readout block sampling a different portion of the longitudinal magnetization recovery curve. As non-limiting examples, the combination of undersampled acquisition and advanced reconstruction techniques may enable whole-brain multi-parametric mapping at 1 mm isotropic resolution in approximately 3 minutes at 3T field strength, or at 750 pm isotropic resolution in approximately 13 minutes at 7T field strength, though other combinations of resolution and acquisition time may be achieved depending on the selected acceleration factor, field strength, and coverage requirements.

[0049] FIG. 3 illustrates a flowchart of an example method for multi -parametric quantitative MRI using the MIMOSA pulse sequences described in the present disclosure. The method may be performed by a computing device, a server, or a combination thereof.

[0050] At step 302, multi -echo data may be acquired using a MIMOSA pulse sequence. In some embodiments, the acquisition may include applying a magnetization preparation RF pulse, such as a T2 preparation module or an inversion recovery preparation module, followed by a plurality of readout acquisitions. The readout acquisitions may include one or more FLASH readouts and an MGRE readout, as described above with respect to FIGS. 1 and 2A-2F. The MGRE readout may acquire data at multiple echo times to encode T2* and magnetic susceptibility contrast. The acquisition may employ a variable-density k-space sampling strategy to enable accelerated data acquisition. In some instances, the variable-density sampling may generate spirallike spokes on a Cartesian grid based on a golden angle distribution.

[0051] At step 304, the acquired multi-echo data may be processed to estimate quantitative parameters. In some embodiments, the processing may include reconstructing multi-contrast images from the acquired k-space data.

[0052] In some aspects, the multi-parametric quantitative MRI techniques described herein may employ model-based fitting approaches to jointly estimate relaxation times and susceptibility values from the acquired multi-echo data. The model-based fitting approaches may fit canonical 11QB\125141.05025\102311859.2MGH 2025-095-02125141.05025models of signal decay and recovery to the reconstructed images to calculate Ti, T2, T2*, and susceptibility parameters. As one non-limiting example, Low-rank modeling of local k-space neighborhoods (LORAKS) may be employed to jointly reconstruct the multi-contrast images. Additionally or alternatively, a multi-contrast multi-slice zero-shot self-supervised learning (MZS- SSL) approach may be employed for joint reconstruction of the multi-contrast images, which may enable robust reconstruction of highly undersampled data while leveraging information across contrasts and slices. The images obtained from the MGRE readout may be used to estimate T2* mapping using a variable projection (VAPRO) algorithm or other suitable fitting technique. The corresponding phase images from the MGRE readout may be used to calculate the tissue field (Aft). The estimated T2* mapping may then be used to jointly estimate Ti, T2, and proton density maps by dictionary matching, or the like.

[0053] In some embodiments, the parameter estimation may be performed in a sequential manner. The magnitude and phase information obtained from the images acquired from the MGRE readout may be used to estimate T2* mapping and quantitative susceptibility mapping (QSM) separately. The T2* estimation may be performed using a VAPRO algorithm, which may separate the linear and nonlinear parameters in the signal model to improve fitting efficiency and accuracy. The VAPRO algorithm may fit a mono-exponential decay model to the multi-echo magnitude data to estimate T2* values at each voxel. The magnitude and phase information obtained from the images acquired from the MGRE readout may be used to estimate quantitative susceptibility mapping (QSM). The phase data may be processed to remove background field contributions and to compute the susceptibility distribution. In some cases, susceptibility source separation may be performed to estimate paramagnetic and diamagnetic susceptibility maps from the QSM. The susceptibility source separation may be based on T2' values derived from the T2 and T2* maps. In some aspects, the total magnetic field (Af) may include background field (Afb) induced by background susceptibility sources and tissue field (Aft) induced by tissue susceptibility sources. The tissue field (Aft) may be obtained by applying phase unwrapping and background field removal to the total magnetic field.

[0054] In some cases, the estimated T2* mapping may serve as a known parameter to jointly estimate Ti, T2, and proton density maps using all acquired images by dictionary matching. The dictionary matching approach may compare the acquired signal evolution to a precomputed dictionary of signal evolutions corresponding to different combinations of Ti, T2, and proton 12QB\125141.05025\102311859.2MGH 2025-095-02125141.05025density values. The dictionary may be created based on Bloch simulation, which may model the evolution of magnetization through the pulse sequence including the magnetization preparation modules and the multiple readout acquisitions. In some embodiments, the dictionary may be generated with parameter ranges such as Ti values from approximately 5 ms to approximately 5000 ms, T2 values from approximately 1 ms to approximately 3000 ms, and T2 values from approximately 1 ms to approximately 500 ms.

[0055] In some embodiments, the dictionary may include corrections for Bi+inhomogeneity and inversion efficiency. Bi+inhomogeneity may cause spatial variations in the flip angles achieved by the RF pulses, which may affect the signal evolution and the accuracy of parameter estimation. In some cases, a Bi+map may be acquired using a turbo-FLASH sequence, and the Bi+values may be quantized within a range of approximately 0.65 to approximately 1.35 and sorted into approximately 50 discrete bins. The corrected flip angle may be calculated as acor= Bi+• a, where a is the nominal flip angle. In some embodiments, inversion efficiency may be estimated by a precomputed look-up table based on Bloch simulation according to Ti and T2 values of each voxel. The inversion efficiency correction may account for imperfect inversion of the longitudinal magnetization by the inversion recovery preparation pulse, where the relationship between magnetizations may be described as Mn+i = -IE • Mn, where IE is the inversion efficiency.

[0056] In some aspects, the Bloch simulation used to generate the dictionary may model the longitudinal magnetization recovery following the magnetization preparation RF pulse. The signal model may describe how the longitudinal magnetization evolves during the recovery period and how the signal intensity at each readout depends on the effective inversion time, the flip angle, and the relaxation parameters of the tissue. By fitting the acquired multi-contrast images to the dictionary entries, the Ti, T2, and proton density values that produce the closest match to the observed signal evolution may be identified.

[0057] In some embodiments, the accelerated data acquisition may improve quantitative accuracy of the parameter estimation. By compressing the acquisition time for each readout block through undersampling and parallel imaging or low-rank reconstruction, the assumption that each readout block samples the magnetization at a single effective inversion time may hold with greater accuracy compared to conventional non-accelerated acquisitions. In conventional acquisitions with longer readout durations, the magnetization may continue to recover during the readout, causing the effective inversion time to vary across different k-space locations within a single 13QB\125141.05025\102311859.2MGH 2025-095-02125141.05025readout block. The accelerated acquisition may reduce the duration of each readout block, thereby reducing the variation in effective inversion time within each block and improving the accuracy of the signal model assumptions.

[0058] In some cases, the compressed acquisition time may enable the readout blocks to sample more discrete points along the longitudinal magnetization recovery curve. The shorter readout durations may allow for tighter temporal spacing between readout blocks, which may provide more sampling points along the recovery curve and may improve the conditioning of the parameter estimation problem. The combination of accelerated acquisition and model-based fitting may enable accurate estimation of multiple relaxation parameters from a single acquisition sequence.

[0059] In some embodiments, susceptibility source separation may be performed following the estimation of T2 and T2* maps. The T2' value, which represents the contribution of magnetic field inhomogeneities to the transverse relaxation, may be derived from the T2 and T2* maps according to the relationship - = - — — R2= The tissue field may be modeled as T2T2T2T2the convolution of a field perturbation kernel and the bulk magnetic susceptibility distribution. As para-magnetic and dia-magnetic susceptibility sources may coexist within individual voxels of brain tissue, the tissue field may be expressed as A / t= D 0 (jpara + XdtaX where Df is the field perturbation kernel, / para is the paramagnetic susceptibility distribution, and %dia is the diamagnetic susceptibility distribution. In some cases, R2' may be represented as the weighted absolute sum of para-magnetic and diamagnetic susceptibility sources according to the relationship R2= Dr■ (I Xpara I +1 Xdia I), where Dr is a relaxometric constant. In some embodiments, the relaxometric constant may be approximately 114 Hz / ppm. In some aspects, QSM inversion may be inherently ill-posed, and incorporating tissue T2 information may improve source separation robustness and fidelity. In some embodiments, paramagnetic and diamagnetic susceptibility may be derived by jointly solving the tissue field equation and the R2' equation given both R2' and Aft. A deep learning model such as / -sepnet may be employed to calculate total QSM from the tissue field and separate the para-magnetic and diamagnetic susceptibility sources.

[0060] At step 306, co-registered parametric maps may be generated for Ti, T2, T2*, and magnetic susceptibility. In some embodiments, the parametric maps may be inherently coregistered because the multiple parameters are derived from a common dataset acquired in a single14QB\125141.05025\102311859.2MGH 2025-095-02125141.05025scan. The co-registered parametric maps may include Ti maps, T2 maps, T2* maps, proton density maps, total susceptibility maps, paramagnetic susceptibility maps, and diamagnetic susceptibility maps. In some cases, the method may generate whole-brain parametric maps at 1 mm isotropic resolution or at higher resolutions such as 750 pm isotropic resolution. In some cases, the simultaneous acquisition of multiple parameters from a single scan may reduce misalignment errors that can arise when combining separately acquired parametric maps. Conventional approaches that acquire different parametric maps in separate scans may introduce spatial misalignment due to patient motion between acquisitions. The single-scan multi-parametric approach described herein may avoid such inter-scan motion artifacts by deriving all parametric maps from the same acquired dataset.

[0061] In some embodiments, the method may include additional steps for quality assessment or post-processing of the parametric maps. For example, the method may include applying spatial filtering, masking, or other image processing operations to the generated parametric maps. In some cases, the method may include synthesizing contrast- weighted images from the estimated parametric maps. The synthesized images may include Ti-weighted images, T2-weighted images, or other contrast weightings that may be computed from the quantitative parameter values.

[0062] In some embodiments, the co-registered parametric maps may provide complementary information about tissue composition and microstructure. For example, Ti relaxation time may be associated with molecular mobility, T2 relaxation time may reflect water molecular environment, and T2* relaxation time may provide insights into tissue microenvironment and iron deposition. The paramagnetic and diamagnetic susceptibility maps may help in characterizing tissue iron and myelin concentration. The combination of multiple coregistered parametric maps may enable comprehensive tissue characterization.

[0063] In some aspects, multiple images with differing contrasts may be combined to generate images with improved accuracy. The multi-echo acquisition may produce images with different levels of Ti contrast corresponding to different effective inversion times, as well as images with different levels of T2* contrast corresponding to different echo times. In some cases, the combination of multiple contrast-weighted images in the parameter estimation process may reduce the effects of noise and may improve the accuracy of the estimated parametric maps compared to single-contrast approaches.15QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0064] In some embodiments, the multiple parametric maps may be used for tissue segmentation. Different tissue types may exhibit characteristic combinations of relaxation times and susceptibility values. For example, white matter, gray matter, and cerebrospinal fluid may have different Ti, T2, and T2* values. The availability of multiple co-registered parametric maps may enable multi-parametric tissue classification, where voxels may be assigned to tissue classes based on the combination of parameter values at each spatial location. Tn some cases, the multiparametric approach may provide improved tissue segmentation compared to segmentation based on a single parameter.

[0065] In some aspects, the estimated parametric maps may be used to synthesize images corresponding to other pulse sequences or protocols. The quantitative parameter values at each voxel may be used as inputs to signal models that describe the signal intensity for different pulse sequence types and acquisition parameters. In some embodiments, synthetic Ti-weighted images may be generated by computing the expected signal intensity at each voxel based on the estimated Ti, T2, and proton density values and a specified set of acquisition parameters such as repetition time, echo time, and flip angle. Similarly, synthetic T2-weighted images or other contrast weightings may be computed from the quantitative parameter values.

[0066] In some cases, the image synthesis capability may enable generation of synthetic images with contrast weightings that were not directly acquired. For example, a synthetic magnetization-prepared rapid gradient echo (MPRAGE) image or a synthetic fluid-attenuated inversion recovery (FLAIR) image may be computed from the estimated parametric maps by applying the appropriate signal model for the desired pulse sequence. The synthetic images may provide clinically familiar contrast appearances without requiring additional scan time.

[0067] In some embodiments, the image synthesis may enable retrospective adjustment of acquisition parameters. The quantitative parametric maps may be used to compute synthetic images corresponding to different repetition times, echo times, inversion times, or flip angles. In some cases, the retrospective parameter adjustment may enable optimization of image contrast for specific clinical applications without repeating the acquisition.

[0068] The combination of co-registered parametric maps and image synthesis capabilities may provide flexibility in clinical workflow. In some aspects, a single multi-parametric acquisition may provide both quantitative parameter values for tissue characterization and synthetic contrast-weighted images for clinical interpretation. The quantitative maps may enable objective 16QB\125141.05025\102311859.2MGH 2025-095-02125141.05025measurement of tissue properties, while the synthetic images may provide familiar contrast appearances for visual assessment.

[0069] At step 308, the co-registered parametric maps generated by the method can then be displayed to a user, stored for later use or further processing, or both. In some embodiments, the displayed or stored parametric maps may include Ti maps, T2 maps, T2* maps, proton density maps, total susceptibility maps, paramagnetic susceptibility maps, and diamagnetic susceptibility maps. In some cases, the parametric maps may be provided to a clinician for diagnostic assessment, such as for evaluating tissue abnormalities, monitoring disease progression, or assessing treatment response. Additionally or alternatively, the parametric maps may be used as input for further analysis, such as tissue segmentation, lesion detection, or image synthesis. In some aspects, the parametric maps may be stored in a database or archive for longitudinal comparison with future acquisitions or for retrospective analysis.

[0070] Referring now to FIG. 4, a flowchart is illustrated as setting forth the steps of an example method for reconstructing multi-contrast images using a suitably trained neural network or other machine learning algorithm implementing an MZS-SSL reconstruction. As will be described, the neural network or other machine learning algorithm takes undersampled multicontrast k-space data as input data and generates reconstructed multi-contrast images as output data. As an example, the reconstructed multi-contrast images can be used for subsequent parameter estimation to generate quantitative parametric maps.

[0071] An overview of this MZS-SSL reconstruction is illustrated in FIG. 5. In (A), the undersampling k-space masks are partitioned into three disjoint subsets: validation, loss, and training masks. In (B), the slices from all contrasts are randomly shuffled, and one slice is selected and multiplied with the generated training and loss k-space mask. Subsequently, the k-space data is converted into images where different contrasts are incorporated into the channel dimension, serving as input to the unrolled network. The loss mask is used to compute the training loss and update the network parameters. In (C), validation is conducted in each training epoch as a stopping criterion by computing the validation loss of the middle slice using the validation mask. In (D), after training, whole-brain volume with multi-contrast can be reconstructed slice by slice.

[0072] The method includes accessing undersampled multi-contrast k-space data with a computer system, as indicated at step 402. Accessing the undersampled k-space data may include retrieving such data from a memory or other suitable data storage device or medium. Additionally 17QB\125141.05025\102311859.2MGH 2025-095-02125141.05025or alternatively, accessing the undersampled k-space data may include acquiring such data (e.g., using the MIMOSA pulse sequence and acquisition techniques described in the present disclosure) with an MR1 system and transferring or otherwise communicating the data to the computer system, which may be a part of the MRI system.

[0073] A trained neural network (or other suitable machine learning algorithm) is then accessed with the computer system, as indicated at step 404. In general, the neural network is trained, or has been trained, on training data in order to reconstruct multi-contrast images from undersampled k-space data.

[0074] Accessing the trained neural network may include accessing network parameters (e g., weights, biases, or both) that have been optimized or otherwise estimated by training the neural network on training data. In some instances, retrieving the neural network can also include retrieving, constructing, or otherwise accessing the particular neural network architecture to be implemented. For instance, data pertaining to the layers in the neural network architecture (e.g., number of layers, type of layers, ordering of layers, connections between layers, hyperparameters for layers) may be retrieved, selected, constructed, or otherwise accessed.

[0075] In some embodiments, the neural network may implement an unrolled network architecture. An unrolled network may include a series of iterative blocks, where each block includes one or more convolutional layers followed by a data consistency layer. The data consistency layer may enforce consistency between the reconstructed images and the acquired k-space data. In some cases, the data consistency layer may be based on a conjugate gradient algorithm. The unrolled network architecture may include an input layer that receives multicontrast k-space data or image data. The input layer may connect to one or more hidden layers that perform image reconstruction operations. In some embodiments, the multi-contrast data from different contrast weightings may be concatenated along the channel dimension and provided as input to the neural network. The concatenation of multi -contrast data along the channel dimension may enable the neural network to leverage shared information across different contrast weightings during reconstruction.

[0076] Each hidden layer may perform a different function. For example, some hidden layers can be convolutional hidden layers which can, in some instances, extract features from the input images. Other hidden layers can perform statistical functions such as batch normalization, activation functions, or residual connections. In some of the hidden layers each node is connected 18QB\125141.05025\102311859.2MGH 2025-095-02125141.05025to each node of the next hidden layer, which may be referred to then as dense layers. The data consistency layers interspersed within the unrolled network may project the intermediate reconstructed images back onto the space of images consistent with the acquired k-space data.

[0077] The last hidden layer in the neural network is connected to the output layer. The output layer may generate reconstructed multi -contrast images corresponding to the different contrast weightings acquired by the multi-parametric pulse sequence. In some embodiments, the output layer may generate reconstructed images for each of the FLASH readouts and each echo of the MGRE readout. The reconstructed multi-contrast images may then be used for subsequent parameter estimation to generate Ti, T2, T2*, proton density, and magnetic susceptibility maps.

[0078] The undersampled k-space data are then input to the one or more trained neural networks, generating output as reconstructed multi-contrast images, as indicated at step 406. In some embodiments, the k-space data may be converted into images where different contrasts are incorporated into the channel dimension, serving as input to the unrolled network. The neural network may process the multi-contrast input data through the series of iterative blocks, with each block applying learned convolutional operations followed by data consistency enforcement.

[0079] In some cases, the neural network may perform slice-by-slice reconstruction of whole-brain volumes. During inference, the trained model may reconstruct whole-brain volumes by processing each slice independently through the neural network. The slice-by-slice reconstruction approach may enable efficient processing of three-dimensional volumetric data while leveraging the trained network parameters.

[0080] Additionally or alternatively, transfer learning may be employed to reduce reconstruction time for new subject data. In some embodiments, a pre-trained network may be used as a starting point for reconstruction of new datasets. The transfer learning approach may reduce the reconstruction time compared to training a new network from scratch for each subject. In some cases, the reconstruction time may be reduced from approximately 18 hours to approximately 3 hours by utilizing a pre-trained network and implementing transfer learning.

[0081] The reconstructed multi-contrast images generated by inputting the undersampled k-space data to the trained neural network(s) can then be displayed to a user, stored for later use or further processing, or both, as indicated at step 408. In some embodiments, the reconstructed images may be provided to a parameter estimation module for generation of quantitative parametric maps.19QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0082] Referring now to FIG. 6, a flowchart is illustrated as setting forth the steps of an example method for training one or more neural networks (or other suitable machine learning algorithms) on training data using a zero-shot self-supervised learning approach, such that the one or more neural networks are trained to receive undersampled multi -contrast k-space data as input data in order to generate reconstructed multi -contrast images as output data. In some embodiments, a multi -contrast multi-slice zero-shot self-supervised learning (MZS-SSL) approach may be employed for joint reconstruction of all whole-brain contrasts.

[0083] In general, the neural network(s) can implement any number of different neural network architectures. For instance, the neural network(s) could implement an unrolled network architecture that includes iterative blocks with convolutional layers and data consistency layers. In some embodiments, the neural network may adopt an architecture similar to Zero-DeepSub or other suitable unrolled network architectures. The data consistency layers may be based on a conjugate gradient algorithm to enforce consistency with the acquired k-space data.

[0084] The method includes partitioning the undersampling mask into disjoint subsets with a computer system, as indicated at step 602. In some embodiments, the undersampling mask may be partitioned into three disjoint subsets using a randomly uniform distribution: a training mask, a validation mask, and a loss mask. The partitioning of the undersampling mask into disjoint subsets may enable self-supervised training without requiring fully sampled reference data.

[0085] In some cases, a set of slice-specific training and loss mask pairs may be generated based on slice numbers to promote data incoherence. The slice-specific mask pairs may provide different sampling patterns for different slices, which may improve the training of the neural network by exposing the network to diverse undersampling patterns. In some embodiments, the training mask may be used to enforce data consistency constraints within the unrolled network, while the loss mask may be used to compute the training loss.

[0086] The method can include assembling training data from the undersampled k-space data using a computer system. In some embodiments, during training, a single random slice may be selected per epoch. The multi -contrast data from the selected slice may be concatenated along the channel dimension and fed into the neural network. The random selection of slices during training may expose the network to different portions of the volumetric data across training epochs.

[0087] One or more neural networks (or other suitable machine learning algorithms) are trained on the training data, as indicated at step 604. In general, the neural network can be trained 20QB\125141.05025\102311859.2MGH 2025-095-02125141.05025by optimizing network parameters (e.g., weights, biases, or both) based on minimizing a loss function. In some embodiments, the training mask may be applied to enforce data consistency constraints within the unrolled network. The paired loss mask may be used to compute the training loss and update network parameters.

[0088] Training a neural network may include initializing the neural network, such as by computing, estimating, or otherwise selecting initial network parameters (e.g., weights, biases, or both). During training, the neural network receives the multi-contrast k-space data for a selected slice and generates reconstructed images. The loss mask is then applied to the reconstructed images to compute the training loss at k-space locations that were not used for data consistency enforcement. The current neural network can then be updated based on the calculated loss (e.g., using backpropagation methods based on the calculated loss). For instance, the current neural network can be updated by updating the network parameters (e.g., weights, biases, or both) in order to minimize the loss according to the loss function. In some embodiments, validation may be conducted in each training epoch as a stopping criterion by computing the validation loss of a middle slice using the validation mask. The training continues until a training condition is met, such as when the validation loss reaches a minimum or when a predetermined number of training epochs have been completed. When the training condition has been met, the current neural network and its associated network parameters represent the trained neural network. Different types of training processes can be used to adjust the network parameters based on the training data. The training processes may include, for example, gradient descent, Adam optimizer, or other suitable optimization algorithms.

[0089] The zero-shot self-supervised learning approach may enable training of the neural network without requiring fully sampled reference data. In some embodiments, the neural network may be trained using only the undersampled k-space data acquired from a single subject. The selfsupervised training approach may leverage the redundancy in the multi -contrast data and the structure of the undersampling pattern to learn an effective reconstruction mapping. In some cases, the zero-shot approach may provide subject-specific reconstruction that adapts to the particular characteristics of each dataset.

[0090] The one or more trained neural networks are then stored for later use, as indicated at step 606. Storing the neural network(s) may include storing network parameters (e.g., weights, biases, or both), which have been computed or otherwise estimated by training the neural 21QB\125141.05025\102311859.2MGH 2025-095-02125141.05025network(s) on the training data. Storing the trained neural network(s) may also include storing the particular neural network architecture to be implemented. In some embodiments, the stored trained neural network may be used for transfer learning to accelerate reconstruction of new subject data. The pre-trained network parameters may serve as initialization for training on new datasets, which may reduce the training time compared to training from scratch.

[0091] FIG. 7 shows an example of a system 700 for multi -parametric quantitative MRI in accordance with some embodiments described in the present disclosure. As shown in FIG. 7, a computing device 750 can receive one or more types of data (e.g., undersampled multi-contrast k-space data) from data source 702. In some embodiments, computing device 750 can execute at least a portion of a multi-parametric quantitative MRI system 704 to generate co-registered parametric maps from data received from the data source 702.

[0092] Additionally or alternatively, in some embodiments, the computing device 750 can communicate information about data received from the data source 702 to a server 752 over a communication network 754, which can execute at least a portion of the multi -parametric quantitative MRI system 704. In such embodiments, the server 752 can return information to the computing device 750 (and / or any other suitable computing device) indicative of an output of the multi-parametric quantitative MRI system 704.

[0093] In some embodiments, computing device 750 and / or server 752 can be any suitable computing device or combination of devices, such as a desktop computer, a laptop computer, a smartphone, a tablet computer, a wearable computer, a server computer, a virtual machine being executed by a physical computing device, and so on. The computing device 750 and / or server 752 can also reconstruct images from the data. In some cases, the computing device 750 and / or server 752 may apply low-rank modeling of local k-space neighborhoods (LORAKS) or a multi-contrast multi-slice zero-shot self-supervised learning (MZS-SSL) approach to jointly reconstruct multicontrast images from the undersampled k-space data.

[0094] In some embodiments, data source 702 can be any suitable source of data (e.g., undersampled multi-contrast k-space data), such as an MRI system, another computing device (e.g., a server storing k-space data), and so on. In some embodiments, data source 702 can be local to computing device 750. For example, data source 702 can be incorporated with computing device 750 (e.g., computing device 750 can be configured as part of a device for measuring, recording, estimating, acquiring, or otherwise collecting or storing data). As another example, data source 22QB\125141.05025\102311859.2MGH 2025-095-02125141.05025702 can be connected to computing device 750 by a cable, a direct wireless link, and so on. Additionally or alternatively, in some embodiments, data source 702 can be located locally and / or remotely from computing device 750, and can communicate data to computing device 750 (and / or server 752) via a communication network (e.g., communication network 754).

[0095] In some embodiments, communication network 754 can be any suitable communication network or combination of communication networks. For example, communication network 754 can include a Wi-Fi network (which can include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e.g., a 3G network, a 4G network, etc., complying with any suitable standard, such as CDMA, GSM, LTE, LTE Advanced, WiMAX, etc.), other types of wireless network, a wired network, and so on. In some embodiments, communication network 754 can be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi-private network (e.g., a corporate or university intranet), any other suitable type of network, or any suitable combination of networks. Communications links shown in FIG. 7 can each be any suitable communications link or combination of communications links, such as wired links, fiber optic links, Wi-Fi links, Bluetooth links, cellular links, and so on.

[0096] Referring now to FIG. 8, an example of hardware 800 that can be used to implement data source 702, computing device 750, and server 752 in accordance with some embodiments of the systems and methods described in the present disclosure is shown.

[0097] As shown in FIG. 8, in some embodiments, computing device 750 can include a processor 802, a display 804, one or more inputs 806, one or more communication systems 808, and / or memory 810. In some embodiments, processor 802 can be any suitable hardware processor or combination of processors, such as a central processing unit (CPU), a graphics processing unit (GPU), and so on. In some embodiments, display 804 can include any suitable display devices, such as a liquid crystal display (LCD) screen, a light-emitting diode (LED) display, an organic LED (OLED) display, an electrophoretic display (e.g., an “e-ink” display), a computer monitor, a touchscreen, a television, and so on. In some embodiments, inputs 806 can include any suitable input devices and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, and so on.

[0098] In some embodiments, communications systems 808 can include any suitable hardware, firmware, and / or software for communicating information over communication network 23QB\125141.05025\102311859.2MGH 2025-095-02125141.05025754 and / or any other suitable communication networks. For example, communications systems 808 can include one or more transceivers, one or more communication chips and / or chip sets, and so on. In a more particular example, communications systems 808 can include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and so on.

[0099] In some embodiments, memory 810 can include any suitable storage device or devices that can be used to store instructions, values, data, or the like, that can be used, for example, by processor 802 to present content using display 804, to communicate with server 752 via communications system(s) 808, and so on. Memory 810 can include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 810 can include random-access memory (RAM), read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), other forms of volatile memory, other forms of non-volatile memory, one or more forms of semi-volatile memory, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, and so on. In some embodiments, memory 810 can have encoded thereon, or otherwise stored therein, a computer program for controlling operation of computing device 750. In such embodiments, processor 802 can execute at least a portion of the computer program to present content (e.g., images, user interfaces, graphics, tables), receive content from server 752, transmit information to server 752, and so on. For example, the processor 802 and the memory 810 can be configured to perform the methods described herein (e.g., the method of FIG. 3, the method of FIG. 4, or the method of FIG. 6). In some cases, the processor 802 may execute a trained neural network to reconstruct multi-contrast images from undersampled k-space data, and may perform parameter estimation via dictionary matching based on Bloch simulation to generate Ti, T2, T2*, proton density, and magnetic susceptibility maps.

[0100] In some embodiments, server 752 can include a processor 812, a display 814, one or more inputs 816, one or more communications systems 818, and / or memory 820. In some embodiments, processor 812 can be any suitable hardware processor or combination of processors, such as a CPU, a GPU, and so on. In some embodiments, display 814 can include any suitable display devices, such as an LCD screen, LED display, OLED display, electrophoretic display, a computer monitor, a touchscreen, a television, and so on. In some embodiments, inputs 816 can24QB\125141.05025\102311859.2MGH 2025-095-02125141.05025include any suitable input devices and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, and so on.[00101J In some embodiments, communications systems 818 can include any suitable hardware, firmware, and / or software for communicating information over communication network 754 and / or any other suitable communication networks. For example, communications systems 818 can include one or more transceivers, one or more communication chips and / or chip sets, and so on. In a more particular example, communications systems 818 can include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and so on.

[0102] In some embodiments, memory 820 can include any suitable storage device or devices that can be used to store instructions, values, data, or the like, that can be used, for example, by processor 812 to present content using display 814, to communicate with one or more computing devices 750, and so on. Memory 820 can include any suitable volatile memory, nonvolatile memory, storage, or any suitable combination thereof. For example, memory 820 can include RAM, ROM, EPROM, EEPROM, other types of volatile memory, other types of nonvolatile memory, one or more types of semi-volatile memory, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, and so on. In some embodiments, memory 820 can have encoded thereon a server program for controlling operation of server 752. In such embodiments, processor 812 can execute at least a portion of the server program to transmit information and / or content (e.g., data, images, a user interface) to one or more computing devices 750, receive information and / or content from one or more computing devices 750, receive instructions from one or more devices (e.g., a personal computer, a laptop computer, a tablet computer, a smartphone), and so on.

[0103] In some embodiments, the server 752 is configured to perform the methods described in the present disclosure. For example, the processor 812 and memory 820 can be configured to perform the methods described herein (e.g., the method of FIG. 3, the method of FIG. 4, or the method of FIG. 6). In some cases, the server 752 may perform joint reconstruction of multi-contrast images using a trained neural network and may estimate quantitative parameters including Ti, T2, T2*, proton density, and magnetic susceptibility values from the reconstructed images.25QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0104] In some embodiments, data source 702 can include a processor 822, one or more data acquisition systems 824, one or more communications systems 826, and / or memory 828. In some embodiments, processor 822 can be any suitable hardware processor or combination of processors, such as a CPU, a GPU, and so on. In some embodiments, the one or more data acquisition systems 824 are generally configured to acquire data, images, or both, and can include an MRI system. Additionally or alternatively, in some embodiments, the one or more data acquisition systems 824 can include any suitable hardware, firmware, and / or software for coupling to and / or controlling operations of an MRI system. In some embodiments, one or more portions of the data acquisition system(s) 824 can be removable and / or replaceable.

[0105] Note that, although not shown, data source 702 can include any suitable inputs and / or outputs. For example, data source 702 can include input devices and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, a trackpad, a trackball, and so on. As another example, data source 702 can include any suitable display devices, such as an LCD screen, an LED display, an OLED display, an electrophoretic display, a computer monitor, a touchscreen, a television, etc., one or more speakers, and so on.

[0106] In some embodiments, communications systems 826 can include any suitable hardware, firmware, and / or software for communicating information to computing device 750 (and, in some embodiments, over communication network 754 and / or any other suitable communication networks). For example, communications systems 826 can include one or more transceivers, one or more communication chips and / or chip sets, and so on. In a more particular example, communications systems 826 can include hardware, firmware, and / or software that can be used to establish a wired connection using any suitable port and / or communication standard (e.g., VGA, DVI video, USB, RS-232, etc.), Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and so on.

[0107] In some embodiments, memory 828 can include any suitable storage device or devices that can be used to store instructions, values, data, or the like, that can be used, for example, by processor 822 to control the one or more data acquisition systems 824, and / or receive data from the one or more data acquisition systems 824; to generate images from data; present content (e.g., data, images, a user interface) using a display; communicate with one or more computing devices 750; and so on. Memory 828 can include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 828 can include RAM, ROM,26QB\125141.05025\102311859.2MGH 2025-095-02125141.05025EPROM, EEPROM, other types of volatile memory, other types of non-volatile memory, one or more types of semi-volatile memory, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, and so on. In some embodiments, memory 828 can have encoded thereon, or otherwise stored therein, a program for controlling operation of data source 702. In such embodiments, processor 822 can execute at least a portion of the program to generate images, transmit information and / or content (e.g., data, images, a user interface) to one or more computing devices 750, receive information and / or content from one or more computing devices 750, receive instructions from one or more devices (e.g., a personal computer, a laptop computer, a tablet computer, a smartphone, etc.), and so on.

[0108] In some embodiments, any suitable computer-readable media can be used for storing instructions for performing the functions and / or processes described herein. For example, in some embodiments, computer-readable media can be transitory or non-transitory. For example, non-transitory computer-readable media can include media such as magnetic media (e.g., hard disks, floppy disks), optical media (e.g., compact discs, digital video discs, Blu-ray discs), semiconductor media (e.g., RAM, flash memory, EPROM, EEPROM), any suitable media that is not fleeting or devoid of any semblance of permanence during transmission, and / or any suitable tangible media. As another example, transitory computer-readable media can include signals on networks, in wires, conductors, optical fibers, circuits, or any suitable media that is fleeting and devoid of any semblance of permanence during transmission, and / or any suitable intangible media.

[0109] As used herein in the context of computer implementation, unless otherwise specified or limited, the terms “component,” “system,” “module,” “framework,” and the like are intended to encompass part or all of computer-related systems that include hardware, software, a combination of hardware and software, or software in execution. For example, a component may be, but is not limited to being, a processor device, a process being executed (or executable) by a processor device, an object, an executable, a thread of execution, a computer program, or a computer. By way of illustration, both an application running on a computer and the computer can be a component. One or more components (or system, module, and so on) may reside within a process or thread of execution, may be localized on one computer, may be distributed between two or more computers or other processor devices, or may be included within another component (or system, module, and so on).27QB\125141.05025\102311859.2MGH 2025-095-02125141.05025

[0110] In some implementations, devices or systems disclosed herein can be utilized or installed using methods embodying aspects of the disclosure. Correspondingly, description herein of particular features, capabilities, or intended purposes of a device or system is generally intended to inherently include disclosure of a method of using such features for the intended purposes, a method of implementing such capabilities, and a method of installing disclosed (or otherwise known) components to support these purposes or capabilities. Similarly, unless otherwise indicated or limited, discussion herein of any method of manufacturing or using a particular device or system, including installing the device or system, is intended to inherently include disclosure, as embodiments of the disclosure, of the utilized features and implemented capabilities of such device or system.

[0111] Referring particularly now to FIG. 9, an example of an MRI system 900 that can implement the methods described here is illustrated. The MRI system 900 includes an operator workstation 902 that may include a display 904, one or more input devices 906 (e.g., a keyboard, a mouse), and a processor 908. The processor 908 may include a commercially available programmable machine running a commercially available operating system. The operator workstation 902 provides an operator interface that facilitates entering scan parameters into the MRI system 900. The operator workstation 902 may be coupled to different servers, including, for example, a pulse sequence server 910, a data acquisition server 912, a data processing server 914, and a data store server 916. The operator workstation 902 and the servers 910, 912, 914, and 916 may be connected via a communication system 940, which may include wired or wireless network connections.

[0112] The pulse sequence server 910 functions in response to instructions provided by the operator workstation 902 to operate a gradient system 918 and a radiofrequency (“RF”) system 920. Gradient waveforms for performing a prescribed scan are produced and applied to the gradient system 918, which then excites gradient coils in an assembly 922 to produce the magnetic field gradients Gx, Gv, and Gzthat are used for spatially encoding magnetic resonance signals. The gradient coil assembly 922 forms part of a magnet assembly 924 that includes a polarizing magnet 926 and a whole-body RF coil 928.

[0113] RF waveforms are applied by the RF system 920 to the RF coil 928, or a separate local coil to perform the prescribed magnetic resonance pulse sequence. Responsive magnetic28QB\125141.05025\102311859.2MGH 2025-095-02125141.05025resonance signals detected by the RF coil 928, or a separate local coil, are received by the RF system 920. The responsive magnetic resonance signals may be amplified, demodulated, filtered, and digitized under direction of commands produced by the pulse sequence server 910. The RF system 920 includes an RF transmitter for producing a wide variety of RF pulses used in MRI pulse sequences. The RF transmitter is responsive to the prescribed scan and direction from the pulse sequence server 910 to produce RF pulses of the desired frequency, phase, and pulse amplitude waveform. The generated RF pulses may be applied to the whole-body RF coil 928 or to one or more local coils or coil arrays.

[0114] In some embodiments, the RF system 920 may be configured to generate and transmit RF pulses for magnetization preparation and signal excitation as part of the multiparametric acquisition methods described herein. For example, the RF system 920 may apply a T2 preparation module that includes a 90-degree tip-down RF pulse followed by a series of adiabatic RF refocusing pulses and a 90-degree tip-up RF pulse. The RF system 920 may also apply an inversion recovery preparation pulse to invert the longitudinal magnetization. Following the magnetization preparation, the RF system 920 may apply a series of excitation RF pulses at specified flip angles for the FLASH readouts and the MGRE readout. The RF system 920 may receive timing and amplitude instructions from the pulse sequence server 910 to coordinate the application of RF pulses with the gradient waveforms.

[0115] In some aspects, the gradient system 918 may be configured to establish spatial encoding gradients during the acquisition sequence. In some embodiments, the gradient system 918 may apply frequency-encoding gradients, phase-encoding gradients, and partition-encoding gradients to spatially encode the MRI signals. The gradient system 918 may apply gradient waveforms that correspond to a variable-density k-space sampling pattern, generating spiral-like spokes on a Cartesian grid based on a golden angle distribution. For the MGRE readout, the gradient system 918 may apply a series of readout gradients with alternating polarity to generate multiple echoes at different echo times. The gradient system 918 may receive gradient waveform specifications from the pulse sequence server 910 and may synchronize gradient application with the RF pulse timing.

[0116] The RF system 920 also includes one or more RF receiver channels. An RF receiver channel includes an RF preamplifier that amplifies the magnetic resonance signal received by the coil 928 to which it is connected, and a detector that detects and digitizes the I and Q quadrature 29QB\125141.05025\102311859.2MGH 2025-095-02125141.05025components of the received magnetic resonance signal. The magnitude of the received magnetic resonance signal may, therefore, be determined at a sampled point by the square root of the sum of the squares of the 1 and Q components:M=yll2+Q2',

[0117] and the phase of the received magnetic resonance signal may also be determined according to the following relationship:

[0118] The pulse sequence server 910 may receive patient data from a physiological acquisition controller 930. By way of example, the physiological acquisition controller 930 may receive signals from a number of different sensors connected to the patient, including electrocardiograph (ECG) signals from electrodes, or respiratory signals from a respiratory bellows or other respiratory monitoring devices. These signals may be used by the pulse sequence server 910 to synchronize, or “gate,” the performance of the scan with the subject’s heart beat or respiration.

[0119] The pulse sequence server 910 may also connect to a scan room interface circuit 932 that receives signals from various sensors associated with the condition of the patient and the magnet system. Through the scan room interface circuit 932, a patient positioning system 934 can receive commands to move the patient to desired positions during the scan.

[0120] In some cases, the pulse sequence server 910 may be configured to control the acquisition sequence and coordinate the acquired data. In some embodiments, the pulse sequence server 910 may store and execute pulse sequence instructions that specify the timing, amplitude, and duration of RF pulses and gradient waveforms. The pulse sequence server 910 may coordinate the RF system 920 and the gradient system 918 to execute the magnetization preparation modules, the FLASH readouts, and the MGRE readout in the specified order. The pulse sequence server 910 may also control the data acquisition hardware to sample the MRI signals at the appropriate times during each readout. In some aspects, the data processing server 914 may process the acquired k-space data to reconstruct multi-contrast images and estimate quantitative parameters, including Ti, T2, T2*, proton density, and magnetic susceptibility values from the reconstructed images.30QB\125141.05025\1023118592MGH 2025-095-02125141.05025

[0121] The digitized magnetic resonance signal samples produced by the RF system 920 are received by the data acquisition server 912. The data acquisition server 912 operates in response to instructions downloaded from the operator workstation 902 to receive the real-time magnetic resonance data and provide buffer storage, so that data is not lost by data overrun. In some scans, the data acquisition server 912 passes the acquired magnetic resonance data to the data processor server 914. In scans that require information derived from acquired magnetic resonance data to control the further performance of the scan, the data acquisition server 912 may be programmed to produce such information and convey it to the pulse sequence server 910. For example, during pre-scans, magnetic resonance data may be acquired and used to calibrate the pulse sequence performed by the pulse sequence server 910. As another example, navigator signals may be acquired and used to adjust the operating parameters of the RF system 920 or the gradient system 918, or to control the view order in which k-space is sampled. In still another example, the data acquisition server 912 may also process magnetic resonance signals used to detect the arrival of a contrast agent in a magnetic resonance angiography (“MRA”) scan. For example, the data acquisition server 912 may acquire magnetic resonance data and processes it in real-time to produce information that is used to control the scan.

[0122] The data processing server 914 receives magnetic resonance data from the data acquisition server 912 and processes the magnetic resonance data in accordance with instructions provided by the operator workstation 902. Such processing may include, for example, reconstructing two-dimensional or three-dimensional images by performing a Fourier transformation of raw k-space data, performing other image reconstruction algorithms (e.g., iterative or backproj ection reconstruction algorithms), applying filters to raw k-space data or to reconstructed images, generating functional magnetic resonance images, or calculating motion or flow images.

[0123] Images reconstructed by the data processing server 914 are conveyed back to the operator workstation 902 for storage. Real-time images may be stored in a data base memory cache, from which they may be output to operator display 902 or a display 936. Batch mode images or selected real time images may be stored in a host database on disc storage 938. When such images have been reconstructed and transferred to storage, the data processing server 914 may notify the data store server 916 on the operator workstation 902. The operator workstation 90231QB\125141.05025\102311859.2MGH 2025-095-02125141.05025may be used by an operator to archive the images, produce fdms, or send the images via a network to other facilities.[00124J The MR1 system 900 may also include one or more networked workstations 942. For example, a networked workstation 942 may include a display 944, one or more input devices 946 (e.g., a keyboard, a mouse), and a processor 948. The networked workstation 942 may be located within the same facility as the operator workstation 902, or in a different facility, such as a different healthcare institution or clinic.

[0125] The networked workstation 942 may gain remote access to the data processing server 914 or data store server 916 via the communication system 940. Accordingly, multiple networked workstations 942 may have access to the data processing server 914 and the data store server 916. In this manner, magnetic resonance data, reconstructed images, or other data may be exchanged between the data processing server 914 or the data store server 916 and the networked workstations 942, such that the data or images may be remotely processed by a networked workstation 942.

[0126] The present disclosure has described one or more preferred embodiments, and it should be appreciated that many equivalents, alternatives, variations, and modifications, aside from those expressly stated, are possible and within the scope of the disclosure.32QB\125141.05025\102311859.2

Claims

MGH 2025-095-02125141.05025CLAIMS1. A method for multi-parametric quantitative magnetic resonance imaging (MRI), comprising:applying a magnetization preparation radio frequency (RF) pulse to a subject using an MRI system;following application of the magnetization preparation RF pulse, performing a plurality of readout acquisitions as longitudinal magnetization recovers back to equilibrium, wherein the plurality of readout acquisitions comprises:one or more single-echo readouts; anda multi-echo gradient echo (MGRE) readout that acquires data at a plurality of echo times;sampling k-space data during the plurality of readout acquisitions using a variabledensity k-space sampling strategy; andreconstructing multi -contrast images from the sampled k-space data.

2. The method of claim 1, wherein the magnetization preparation RF pulse comprises an inversion recovery RF pulse that inverts the longitudinal magnetization.

3. The method of claim 2, wherein the inversion recovery RF pulse comprises an adiabatic inversion recovery pulse.

4. The method of claim 1, further comprising applying a T2 preparation module prior to the one or more single-echo readouts, wherein the T2 preparation module comprises a 90-degree tip-down RF pulse followed by a series of adiabatic RF refocusing pulses and a 90-degree tip-up RF pulse.

5. The method of claim 1, wherein the MGRE readout comprises four or more echoes spanning a range of echo times.33QB\125141.05025\1023118592MGH 2025-095-02125141.050256. The method of claim 1, wherein the variable-density k-space sampling strategy comprises a Variable-Density sampling and Radial view ordering (VDRad) Cartesian trajectory that generates spiral-like spokes on a Cartesian grid.

7. The method of claim 6, wherein the spiral-like spokes rotate across different repetition times by a golden angle.

8. The method of claim 1, wherein the variable-density k-space sampling strategy implements denser sampling near a center of k-space and sparser sampling in peripheral regions of k-space.

9. The method of claim 1, wherein the one or more single-echo readouts employ a center-out k-space sampling order, and wherein the MGRE readout employs an out-center k-space sampling order.

10. The method of claim 1, wherein each of the plurality of readout acquisitions samples a different portion of a longitudinal magnetization recovery curve, such that data is acquired from a plurality of different effective inversion times.

11. The method of claim 1, wherein reconstructing the multi-contrast images comprises applying low-rank modeling of local k-space neighborhoods (LORAKS) to jointly reconstruct the multi-contrast images.

12. A method for reconstructing multi-contrast images using a neural network, comprising:accessing, with a computer system, undersampled multi-contrast k-space data acquired using a multi-parametric MRI pulse sequence;accessing, with the computer system, a trained neural network, wherein the trained neural network implements an unrolled network architecture comprising a plurality of iterative blocks, each iterative block comprising one or more convolutional layers followed by a data consistency layer;inputting the undersampled multi-contrast k-space data to the trained neural network; and 34QB\125141.05025\102311859.2MGH 2025-095-02125141.05025generating, as output from the trained neural network, reconstructed multi-contrast images.

13. The method of claim 12, wherein the undersampled multi -contrast k-space data comprises data acquired from a plurality of FLASH readouts and a plurality of echoes from an MGRE readout.

14. The method of claim 12, wherein multi-contrast data from different contrast weightings is concatenated along a channel dimension and provided as input to the trained neural network.

15. The method of claim 12, wherein the trained neural network performs slice-by-slice reconstruction of whole-brain volumes.

16. The method of claim 12, further comprising estimating quantitative parameters from the reconstructed multi-contrast images, wherein the quantitative parameters comprise Ti, T2, T2*, and magnetic susceptibility values.

17. The method of claim 16, wherein estimating the quantitative parameters comprises:estimating T2* mapping from magnitude information of images acquired from the MGRE readout using a variable projection (VAPRO) algorithm; andestimating Ti, T2, and proton density maps by dictionary matching using the estimated T2* mapping.

18. The method of claim 17, wherein the dictionary matching comprises comparing acquired signal evolution to a precomputed dictionary of signal evolutions generated based on Bloch simulation.

19. The method of claim 16, wherein estimating the quantitative parameters further comprises performing susceptibility source separation to estimate paramagnetic and diamagnetic susceptibility maps.35QB\125141.05025\102311859.2MGH 2025-095-02125141.0502520. The method of claim 19, wherein the susceptibility source separation is based on T2' values derived from the T2 and T2* maps.

21. The method of claim 12, further comprising generating co-registered parametric maps for Ti, T2, T2*, and magnetic susceptibility, wherein the parametric maps are inherently coregistered because multiple parameters are derived from a common dataset acquired in a single scan.

22. A method for training a neural network for multi-contrast image reconstruction, comprising:accessing, with a computer system, undersampled multi-contrast k-space data; partitioning an undersampling mask into a plurality of disjoint subsets comprising a training mask, a validation mask, and a loss mask;training a neural network on the undersampled multi-contrast k-space data using a zeroshot self-supervised learning approach, wherein training comprises:applying the training mask to enforce data consistency constraints within the neural network;applying the loss mask to compute a training loss and update network parameters; andapplying the validation mask to compute a validation loss as a stopping criterion; andstoring the trained neural network for later use.

23. The method of claim 22, wherein the neural network implements an unrolled network architecture comprising a plurality of iterative blocks, each iterative block comprising one or more convolutional layers followed by a data consistency layer.

24. The method of claim 23, wherein the data consistency layer is based on a conjugate gradient algorithm.36QB\125141.05025\102311859.2MGH 2025-095-02125141.0502525. The method of claim 22, wherein partitioning the undersampling mask comprises partitioning using a randomly uniform distribution.

26. The method of claim 22, further comprising generating a set of slice-specific training and loss mask pairs based on slice numbers to promote data incoherence.

27. The method of claim 22, wherein training the neural network comprises selecting a single random slice per epoch, and concatenating multi-contrast data from the selected slice along a channel dimension.

28. The method of claim 22, wherein the neural network is trained using the undersampled multi-contrast k-space data acquired from a single subject.

29. The method of claim 22, wherein training the neural network comprises: initializing the neural network with initial network parameters;receiving multi-contrast k-space data for a selected slice;generating reconstructed images from the multi -contrast k-space data;applying the loss mask to the reconstructed images to compute the training loss at k-space locations not used for data consistency enforcement; andupdating the network parameters based on the computed training loss.

30. The method of claim 29, wherein updating the network parameters comprises using backpropagation methods based on the computed training loss.37QB\125141.05025\102311859.2