Method and system for generating magnetic resonance images

Abstract

A method of generating magnetic resonance images, comprises acquiring a first set of magnetic resonance images of an organ of a subject. For each acquisition, one or more radiofrequency excitation parameter is varied. The method also comprises accessing a computer readable medium storing a machine learning procedure trained for generating magnetic resonance images, and feeding the procedure with (i) the first set of magnetic resonance images, (ii) a first set of radiofrequency excitation parameters describing variations among the magnetic resonance images in the first set, and (iii) a second set of radiofrequency excitation parameters. The method also comprises receiving from the procedure a synthetic magnetic resonance image corresponding to the second set of radiofrequency excitation parameters.

Description

[0001] METHOD AND SYSTEM FOR GENERATING MAGNETIC RESONANCE IMAGES

[0002] RELATED APPLICATION

[0003] This application claims the benefit of priority of U.S. Provisional Patent Application No. 63 / 682,812 filed on August 14, 2024, the contents of which are incorporated herein by reference in their entirety.

[0004] The project leading to this application has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement no. 101115639).

[0005] FIELD AND BACKGROUND OF THE INVENTION

[0006] The present invention, in some embodiments thereof, relates to Magnetic Resonance Imaging (MRI) and, more particularly, but not exclusively, to the generation of a synthetic magnetic resonance image.

[0007] MRI is a method to obtain an image representing the chemical and physical microscopic properties of materials, by utilizing a quantum mechanical phenomenon, named Nuclear Magnetic Resonance (NMR), in which a system of spins, placed in a magnetic field resonantly absorb energy, when applied with a certain frequency.

[0008] When placed in a magnetic field, a nucleus having a non-zero spin is allowed to be in a discrete set of energy levels, the number of which is determined by the spin, and the separation of which is determined by the gyromagnetic ratio of the nucleus and by the magnetic field. Under the influence of a small perturbation, manifested as a radiofrequency magnetic field, which rotates about the direction of a primary static magnetic field, the nucleus has a time dependent probability to experience a transition from one energy level to another. With a specific frequency of the rotating magnetic field, the transition probability may reach the value of unity. Hence at certain times, a transition is forced on the nucleus, even though the rotating magnetic field may be of small magnitude relative to the primary magnetic field. For an ensemble of nuclei the transitions are realized through a change in the overall magnetization.

[0009] Saturation transfer (ST) MRI is a technology that allows the imaging of molecular tissue properties in living humans. ST-MRI is being investigated for many clinical applications, including tumor detection and grading, early stroke characterization, neurodegenerative disorder imaging, and kidney disease monitoring. SUMMARY OF THE INVENTION

[0010] According to some embodiments of the invention the present invention there is provided a method of generating magnetic resonance images. The method comprises: acquiring a first set of magnetic resonance images of an organ of a subject, wherein for each acquisition, at least one radiofrequency excitation parameter is varied. The method also comprises accessing a computer readable medium storing a machine learning procedure trained for generating magnetic resonance images, and feeding the procedure with (i) the first set of magnetic resonance images, (ii) a first set of radiofrequency excitation parameters describing variations among the magnetic resonance images in the first set, and (iii) a second set of radiofrequency excitation parameters. The method also comprises receiving from the procedure a synthetic magnetic resonance image corresponding to the second set of radiofrequency excitation parameters.

[0011] According to some embodiments of the invention the synthetic magnetic resonance image and the magnetic resonance images correspond to the same MRI protocol.

[0012] According to some embodiments of the invention the synthetic magnetic resonance image and the magnetic resonance images correspond to different MRI protocol.

[0013] According to some embodiments of the invention wherein each of the synthetic magnetic resonance image synthesizes an image corresponding to a MRI protocol selected from the group consisting of Ti-weighted protocol, T2-weighted protocol, fluid-attenuated inversion recovery (FLAIR) protocol, diffusion protocol, and perfusion protocol.

[0014] According to some embodiments of the invention the method comprises concatenating the first set of radiofrequency excitation parameters with the second set of radiofrequency excitation parameters, wherein the feeding comprises feeding the concatenation.

[0015] According to an aspect of some embodiments of the present invention there is provided a method of generating a quantitative MRI (qMRI) map. The method comprises: acquiring a first set of magnetic resonance images of an organ of a subject, wherein for each acquisition, at least one radiofrequency excitation parameter is varied. The method also comprises accessing a computer readable medium storing a machine learning procedure trained for generating qMRI maps describing a biological and / or physical property within a scanned volume; feeding the procedure with (i) the first set of magnetic resonance images, and (ii) a first set of radiofrequency excitation parameters describing variations among the magnetic resonance images in the first set. The method also comprises receiving from the procedure a synthetic qMRI map describing the property. According to some embodiments of the invention the machine learning procedure is trained for generating a plurality of qMRI maps each describing a different property, and the method comprises receiving from the procedure at least two synthetic qMRI maps describing at least two different properties.

[0016] According to some embodiments of the invention the property comprises chemical exchange saturation transfer (CEST) target proton volume fraction.

[0017] According to some embodiments of the invention the property comprises CEST proton exchange rate.

[0018] According to some embodiments of the invention the property comprises at least one property selected from the group consisting of semi-solid magnetization transfer (MT) proton exchange rate, semi-solid MT proton volume fraction, amide effect contribution, guanidinium effect contribution, amine CEST contribution, hydroxyl CEST effect contribution, aliphatic relayed nuclear Overhauser effect (rNOE) contribution, longitudinal relaxation time, transverse relaxation time, effective transverse relaxation time, proton density, diffusion coefficient, perfusion parameter, inhomogeneous magnetization transfer ratio, magnetic susceptibility value, static magnetic field (BO) mapping, radiofrequency field (Bl) mapping, and fat fraction.

[0019] According to some embodiments of the invention a MRI protocol of each of the magnetic resonance images of the first set is selected from the group consisting of Ti-weighted protocol, T2- weighted protocol, fluid-attenuated inversion recovery (FLAIR) protocol, diffusion protocol, and perfusion protocol.

[0020] According to some embodiments of the invention the machine learning procedure comprises a vision transformer encoder.

[0021] According to some embodiments of the invention the machine learning procedure comprises at least one embedding layer for embedding the first set of radiofrequency excitation parameters, at least one embedding layer for embedding the first set of magnetic resonance images, and a concatenation layer for concatenating the embedded first set of radiofrequency excitation parameters with the embedded first set of magnetic resonance images, wherein the vision transformer encoder is fed by the concatenation.

[0022] According to some embodiments of the invention the machine learning procedure comprises a plurality of convolutions layers being fed by the vision transformer encoder.

[0023] According to some embodiments of the invention at least one of the first set of magnetic resonance images is a non-steady-state magnetic resonance image. According to some embodiments of the invention the method comprises defining a plurality of patches over at least one of the magnetic resonance images, wherein the feeding the procedure comprises feeding the plurality of patches.

[0024] According to some embodiments of the invention for each acquisition, at least two radiofrequency excitation parameters are varied.

[0025] According to some embodiments of the invention the at least two radiofrequency excitation parameters comprise a frequency and amplitude of radiofrequency pulses applied during the acquisition.

[0026] According to some embodiments of the invention the method is executed recursively, by redefining, at each recursive execution, the first set of magnetic resonance image to include also the synthetic magnetic resonance image.

[0027] According to some embodiments of the invention the acquisition comprises executing saturation transfer MRI.

[0028] According to some embodiments of the invention the saturation transfer MRI comprises Chemical Exchange Saturation Transfer (CEST) imaging.

[0029] According to some embodiments of the invention the CEST comprises amide proton transfer.

[0030] According to some embodiments of the invention the CEST is selected to target at least one of glucose, glycosaminoglycans, and lipid molecules.

[0031] According to some embodiments of the invention the saturation transfer MRI comprises magnetization transfer imaging.

[0032] According to some embodiments of the invention the magnetization transfer imaging comprises semisolid magnetization transfer imaging.

[0033] According to some embodiments of the invention the saturation transfer MRI comprises saturation transfer difference imaging.

[0034] According to some embodiments of the invention a total duration of the acquisition of the first set of magnetic resonance images is less than one minute.

[0035] According to some embodiments of the invention the organ is selected from the group consisting of a brain, a spinal cord, a heart, a head, a musculoskeletal organ, an abdomen organ, and a pelvis.

[0036] According to some embodiments of the invention at least one of the synthetic magnetic resonance images of the first set is a single-slice image.

[0037] According to some embodiments of the invention at least one of the magnetic resonance images of the first set is a single-slice image. According to some embodiments of the invention at least one of the synthetic magnetic resonance images of the first set is a is a multi-slice image.

[0038] According to some embodiments of the invention at least one of the magnetic resonance images of the first set is a multi-slice image.

[0039] According to an aspect of some embodiments of the present invention there is provided a computer software product. The computer software product comprises a computer-readable medium in which program instructions are stored, which instructions, when read by a data processor, cause the data processor to receive at least the first set of magnetic resonance images, and the first set of radiofrequency excitation parameters, and to execute the method as delineated above and optionally and preferably as further detailed below. In some embodiments of the present invention the data processor also receives the second set of radiofrequency excitation parameters.

[0040] According to an aspect of some embodiments of the present invention there is provided an MRI system for imaging an object. The system comprises: an MRI scanner configured for scanning the object to provide magnetic resonance images, and a data processor configured for executing the program instructions as delineated above and optionally and preferably as further detailed below.

[0041] Unless otherwise defined, all technical and / or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention pertains. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of embodiments of the invention, exemplary methods and / or materials are described below. In case of conflict, the patent specification, including definitions, will control. In addition, the materials, methods, and examples are illustrative only and are not intended to be necessarily limiting.

[0042] Implementation of the method and / or system of embodiments of the invention can involve performing or completing selected tasks manually, automatically, or a combination thereof. Moreover, according to actual instrumentation and equipment of embodiments of the method and / or system of the invention, several selected tasks could be implemented by hardware, by software or by firmware or by a combination thereof using an operating system.

[0043] For example, hardware for performing selected tasks according to embodiments of the invention could be implemented as a chip or a circuit. As software, selected tasks according to embodiments of the invention could be implemented as a plurality of software instructions being executed by a computer using any suitable operating system. In an exemplary embodiment of the invention, one or more tasks according to exemplary embodiments of method and / or system as described herein are performed by a data processor, such as a computing platform for executing a plurality of instructions. Optionally, the data processor includes a volatile memory for storing instructions and / or data and / or a non-volatile storage, for example, a magnetic hard-disk and / or removable media, for storing instructions and / or data. Optionally, a network connection is provided as well. A display and / or a user input device such as a keyboard or mouse are optionally provided as well.

[0044] BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0045] Some embodiments of the invention are herein described, by way of example only, with reference to the accompanying drawings. With specific reference now to the drawings in detail, it is stressed that the particulars shown are by way of example and for purposes of illustrative discussion of embodiments of the invention. In this regard, the description taken with the drawings makes apparent to those skilled in the art how embodiments of the invention may be practiced.

[0046] In the drawings:

[0047] FIG. 1 is a flowchart diagram of a method suitable for generating magnetic resonance images according to some embodiments of the present invention.

[0048] FIG. 2 is a flowchart diagram of a method suitable for generating a quantitative magnetic resonance imaging (qMRI) map according to some embodiments of the present invention.

[0049] FIG. 3 is a schematic illustration of a magnetic resonance imaging system, according to some embodiments of the present invention.

[0050] FIGs. 4A-B are schematic representation of the biophysical-model-free Deep MRI on a Chip (DeepMonC) pipeline for automatic prediction of unseen molecular MRI contrast weighted images (FIG. 4A) and its expansion for the simultaneous quantification of six tissue and scanner parameter maps (FIG. 4B), according to some embodiments of the present invention.

[0051] FIGs. 5A-B shows a comparison between the ground truth (FIG. 5A) and the method-predicted (FIG. 5B) molecular MRI contrast weighted images in the human brain, as obtained in experiments performed according to some embodiments of the present invention.

[0052] FIGs. 6A-F show a quantitative reconstruction of six molecular MRI, scanner field, and waterproton relaxation quantitative maps, as obtained in experiments performed according to some embodiments of the present invention. FIG. 6A shows images obtained using Ti and T2-mapping, WASABI, and semisolid MT MR-Fingerprinting (MRF) in 8.5 min. FIG. 6B shows the same parameter maps obtained according to some embodiments of the present invention using DeepMonC in merely 28.2 s (94% scan time acceleration). Note the reduced field inhomogeneity (as seen in the Bo and Bi predicted images), which explains the successful noise reduction in the output maps (white arrows). FIG. 6C shows quantitative reconstruction using supervised learning (RF tissue response pretraining excluded), utilizing the same raw input data used in FIG. 6B for comparison. FIGs. 6D-F show statistical analysis of the SSIM, PSNR and NRMSE performance measures, comparing the reconstructed maps shown in b and c, compared to the reference ground truth.****p<0.0001.

[0053] FIGs. 7A-C show statistical analysis of the SSIM, PSNR and NRMSE performance measures obtained in experiments performed according to some embodiments of the present invention, comparing various reconstructed quantitative parameter maps compared to the reference ground truth.

[0054] DESCRIPTION OF SPECIFIC EMBODIMENTS OF THE INVENTION

[0055] The present invention, in some embodiments thereof, relates to Magnetic Resonance Imaging (MRI) and, more particularly, but not exclusively, to the generation of a synthetic magnetic resonance image.

[0056] Before explaining at least one embodiment of the invention in detail, it is to be understood that the invention is not necessarily limited in its application to the details of construction and the arrangement of the components and / or methods set forth in the following description and / or illustrated in the drawings and / or the Examples. The invention is capable of other embodiments or of being practiced or carried out in various ways.

[0057] Some embodiments of the present invention comprise a method for generating synthetic magnetic resonance images, and / or quantitative magnetic resonance imaging (qMRI) maps using machine learning procedures. The method of the present embodiments allow the generation of synthetic images and / or maps from a set of magnetic resonance images that are acquired using a set of radiofrequency excitation parameters.

[0058] FIG. 1 is a flowchart diagram of a method suitable for generating magnetic resonance images according to various exemplary embodiments of the present invention. It is to be understood that, unless otherwise defined, the operations described hereinbelow can be executed either contemporaneously or sequentially in many combinations or orders of execution. Specifically, the ordering of the flowchart diagrams is not to be considered as limiting. For example, two or more operations, appearing in the following description or in the flowchart diagrams in a particular order, can be executed in a different order (e.g., a reverse order) or substantially contemporaneously. Additionally, several operations described below are optional and may not be executed.

[0059] The method begins at 10 and proceeds to 11, at which a first set of magnetic resonance images of an organ of a subject is acquired. The organ being imaged can be selected from various anatomical structures, including, without limitation, the brain (including the cerebrum, cerebellum, brainstem, hippocampus, amygdala, thalamus, hypothalamus, basal ganglia, corpus callosum, ventricles, meninges, and cranial nerves), the spinal cord (including the cervical spine, thoracic spine, lumbar spine, sacral spine, coccygeal spine, spinal canal, nerve roots, and surrounding soft tissues), the heart (including the left ventricle, right ventricle, left atrium, right atrium, myocardium, pericardium, cardiac valves, coronary arteries, aorta, and great vessels), an abdomen organ (the liver, the kidneys, the pancreas, the spleen, the gastrointestinal structures including the stomach, small intestine, large intestine, colon, rectum, appendix, gallbladder, bile ducts, adrenal glands, retroperitoneal structures, peritoneum, omentum, mesentery, abdominal aorta, inferior vena cava, portal vein, and abdominal lymph nodes), pelvis organs (including the bladder, prostate, seminal vesicles, uterus, ovaries, fallopian tubes, cervix, vagina, rectum, pelvic floor muscles, sacroiliac joints, and pelvic lymph nodes), musculoskeletal structures (including bones, joints, cartilage, ligaments, tendons, muscles, bone marrow, synovial fluid, and surrounding soft tissues), head and neck structures (including the orbits, extraocular muscles, optic nerves, sinuses, nasal cavity, nasopharynx, oropharynx, larynx, thyroid gland, parathyroid glands, salivary glands, temporomandibular joints, and cervical lymph nodes), thoracic structures (including the lungs, pleura, mediastinum, esophagus, trachea, bronchi, pulmonary vessels, lymph nodes, and chest wall), vascular structures (including arteries, veins, capillaries, and lymphatic vessels throughout the body), and extremities (including the arms, legs, hands, feet, and associated joints, muscles, and soft tissues).

[0060] Each image in the first set represents a complete two-dimensional or three-dimensional representation of the scanned organ or anatomical region. The images are acquired sequentially or in parallel using modern MRI scanner capabilities, with each acquisition capturing the magnetic resonance signal response under specific imaging conditions.

[0061] The acquisition process of the first set of magnetic resonance images optionally and preferably comprises executing saturation transfer MRI techniques, which represent a class of MRI methods that provide enhanced contrast and quantitative information about tissue composition and molecular environment. Saturation transfer MRI encompasses several related techniques that exploit the exchange of magnetization between different molecular pools within tissue. These techniques involve the application of radiofrequency pulses to selectively saturate specific molecular species, followed by detection of the saturation effect on the water signal.

[0062] Chemical Exchange Saturation Transfer (CEST) imaging represents a particularly important category of saturation transfer MRI. CEST exploits the chemical exchange between labile protons on solute molecules and bulk water protons. In CEST imaging, radiofrequency saturation pulses are applied at the resonance frequency of the exchangeable protons, which reduces their magnetization. This saturation is transferred via chemical exchange to the much larger water proton pool, creating a detectable decrease in the water signal.

[0063] CEST imaging may include amide proton transfer (APT), which specifically targets amide protons found in mobile proteins and peptides. APT imaging is particularly valuable for brain tumor assessment, as tumor tissue typically exhibits elevated mobile protein content compared to normal brain tissue. The amide protons resonate approximately 3.5 ppm downfield from water, and APT contrast reflects both the concentration of mobile proteins and the local pH environment.

[0064] CEST may also be selected to target specific molecules such as glucose, glycosaminoglycans, and lipid molecules. Glucose CEST (glucoCEST) allow non-invasive monitoring of glucose uptake and metabolism, with potential applications in diabetes research and cancer metabolism studies. The method exploits the exchangeable hydroxyl protons on glucose molecules, which resonate at approximately 1-3 ppm from water.

[0065] Glycosaminoglycan CEST (gagCEST) targets the hydroxyl and amide protons on glycosaminoglycan molecules, which are major components of cartilage extracellular matrix. This technique is particularly suitable for osteoarthritis research and cartilage health assessment, as glycosaminoglycan content is an important biomarker of cartilage integrity.

[0066] Lipid CEST techniques target various exchangeable protons associated with lipid molecules, including those on phospholipids and fatty acids. These techniques provide information about membrane composition and lipid metabolism, with applications in cancer research and metabolic studies.

[0067] Other saturation transfer MRI techniques include magnetization transfer imaging, which exploits magnetization exchange between free water protons and protons bound to macromolecules. Unlike CEST, which relies on chemical exchange, magnetization transfer primarily involves dipolar coupling and cross-relaxation mechanisms.

[0068] Semisolid magnetization transfer imaging targets protons associated with macromolecular structures such as lipids, and other large molecules that have restricted mobility. This technique provides information about tissue microstructure and macromolecular content, with applications in neuroimaging, where it can detect subtle changes in white matter integrity and myelination.

[0069] Saturation transfer difference imaging is based on the difference in signal intensity between saturated and non-saturated acquisitions. This imaging enhances the contrast between different tissue types and can highlight regions with specific molecular characteristics. The magnetic resonance imaging protocol for each image in the first set can be selected from various established clinical and research protocols, including, without limitation, a Ti-weighted protocol, a T2-weighted protocol, a Fluid-attenuated inversion recovery (FLAIR) protocol, a diffusion- weighted protocol, and a perfusion- weighted protocol.

[0070] Ti-weighted protocols emphasize differences in longitudinal relaxation times between tissues. These protocols typically use repetition times (TR) and echo times (TE) selected to provide bright signal from tissues with shorter Ti values (such as fat) and darker signal from tissues with longer Ti values (such as cerebrospinal fluid). Ti-weighted imaging is particularly suitable for anatomical visualization and contrast agent studies. T2-weighted protocols emphasize differences in transverse relaxation times between tissues. These protocols use TR and TE values selected to provide bright signal from tissues with longer T2 values (such as cerebrospinal fluid) and darker signal from tissues with shorter T2 values. T2-weighted imaging is sensitive to pathological changes such as edema, inflammation, and tissue damage. FLAIR protocols combine T2-weighted contrast with suppression of cerebrospinal fluid signal. An inversion pulse is timed to null the signal from cerebrospinal fluid, allowing better visualization of periventricular and cortical lesions. FLAIR imaging is particularly valuable in neuroimaging for detecting white matter lesions and cortical abnormalities. Diffusion-weighted protocols measure the microscopic motion of water molecules within tissue, providing information about tissue microstructure and integrity. These protocols use diffusion-sensitizing gradients to encode molecular motion, with the degree of signal attenuation reflecting the extent of diffusion. Diffusion imaging is particularly valuable for stroke detection, white matter tractography, and assessment of tissue cellularity. Perfusion-weighted protocols measure blood flow and hemodynamic parameters within tissue. These may include dynamic contrast-enhanced methods using gadolinium-based contrast agents, or arterial spin labeling techniques that use magnetically labeled blood as an endogenous tracer. Perfusion imaging provides information about tissue viability, vascular permeability, and hemodynamic function.

[0071] During the acquisition process, at least one radiofrequency excitation parameter is optionally and preferably varied for the respective acquisition. The radiofrequency excitation parameters encompass various aspects of the radiofrequency pulses used to excite the nuclear spins in the subject's tissue. Representative examples of radiofrequency excitation parameters which can be varied according to some embodiments of the present invention including, without limitation, frequency, amplitude, waveform, duration, and / or power of radiofrequency pulses applied during the acquisition. The frequency parameter refers to the carrier frequency of the radiofrequency pulse, which determines the specific nuclear spins that will be excited during the pulse application. In magnetic resonance imaging, frequency is typically expressed either as an absolute value in Hertz (Hz) or as a chemical shift offset in parts per million (ppm) relative to a reference frequency, commonly the water proton resonance. For example, in CEST imaging applications, frequency variations are useful since they determine which molecular species is targeted for saturation. In Ti-weighted and T2-weighted imaging, frequency variations may involve off-resonance effects, fat suppression techniques, or multi- spectral imaging approaches. Diffusion imaging protocols may employ frequency variations for fat suppression, Bo shimming compensation, or multi-band excitation techniques for simultaneous multislice imaging. The frequency offsets for multi-band techniques typically range from about 1 kHz to about 5 kHz. Perfusion imaging applications include frequency-selective arterial spin labeling techniques that use frequency offsets to achieve spatial selectivity of the labeling region. Velocity-selective labeling techniques may employ frequency sweeps or multiple frequency offsets to achieve the desired velocity selectivity. The frequency variation pattern may follow systematic approaches such as linear sampling with uniform frequency steps, logarithmic sampling with higher density around specific resonances of interest, or targeted sampling selected for predetermined molecular resonances.

[0072] The amplitude parameter refers to the strength or magnitude of the radiofrequency pulse, typically expressed in microtesla (pT), Gauss (G), or as a percentage of maximum transmitter power. Amplitude controls the degree of saturation or excitation achieved during pulse application and is particularly useful in influences saturation transfer imaging where the amplitude influences the contrast mechanisms. The Amplitude variations may follow systematic patterns such as linear progressions, logarithmic scales, or optimization-based selections. In some embodiments, the amplitude is varied in conjunction with frequency to create a two-dimensional parameter space.

[0073] The waveform parameter describes the temporal shape or envelope of the radiofrequency pulse, which influences the frequency selectivity, saturation efficiency, and off-resonance effects of the excitation. Different waveforms provide distinct advantages for specific imaging applications and can be optimized for particular contrast mechanisms. For example, rectangular waveforms represent a pulse shape characterized by constant amplitude throughout the pulse duration. Gaussian waveforms feature a bell-shaped amplitude envelope that provides improved frequency selectivity compared to rectangular pulses. Gaussian pulses are typically characterized by their full-width-half-maximum (FWHM) parameter, which determines the frequency selectivity. Sine-shaped waveforms produce approximately rectangular frequency profiles, providing sharp frequency selectivity with reduced off- resonance excitation. Sine pulses are characterized by their time-bandwidth product, which determines the trade-off between pulse duration and frequency selectivity. Higher time -bandwidth products provide better selectivity but require longer pulse durations. Fermi waveforms combine aspects of Gaussian and rectangular shapes, featuring smooth transitions at the pulse edges while maintaining relatively constant amplitude in the central portion. Fermi pulses are characterized by their transition width parameter, which controls the sharpness of the pulse edges. Adiabatic waveforms represent class of pulses which provide uniform excitation despite variations in the Bi field strength. These waveforms typically feature simultaneous modulation of both amplitude and frequency, following adiabatic passages that maintain constant magnetization trajectories regardless of Bi variations. Also contemplated are custom-designed waveforms which can selected for specific applications using numerical optimization techniques. Custom-designed waveforms can be selected to achieve maximum saturation efficiency, minimum SAR deposition, optimal frequency selectivity, robustness to Bo and Bi inhomogeneities, and the like. Sine waveform are suitable for Ti-weighted and T2-weighted MRI, inversion pulse waveforms, such as, but not limited to, adiabatic waveforms (e.g., secant, tanh or tan waveforms) are suitable for FLAIR imaging. Custom-designed waveforms are suitable for diffusion imaging, and perfusion imaging. The variation of the waveform parameter can be variation of one or more parameters for the same type of waveform (e,g,, varying the FWHM parameter for a Gaussian waveform, varying the time -bandwidth product for Sine waveform, etc.), and / or a change in the type of waveform between the respective acquisitions.

[0074] The duration parameter specifies the temporal length of the radiofrequency pulse application, typically expressed in milliseconds (ms) or seconds (s). The pulse duration influences saturation efficiency, frequency selectivity, power deposition, and overall imaging time. The duration parameter At also relates to the flip angle 0, which describes the rotation angle imparted to the magnetization vector by the RF pulse. This flip angle also depends on the gyromagnetic ratio y, and the amplitude of the Bi field, via the relation 0=yBiAt.

[0075] Pulse durations of from about 0.1 ms to about 5 ms are useful for standard gradient echo sequences in Ti -weighted imaging, pulse durations of from about from about 5 ms to about 20 ms are useful for spin echo sequences in Ti -weighted imaging, pulse durations of from about 10 ms to about 100 ms are useful for dynamic imaging applications or when it is desired to shorten the overall imaging time, pulse durations from about 10 ms to about 10 s are useful in saturation transfer MRI, including CEST imaging. In T2-weighted imaging, pulse durations of from about 1 ms to about 5 ms are useful for 90° excitation, pulse durations of from about 2 ms to about 10 ms are useful for 180° excitation, and pulse durations of from about 1 ms to about 3 ms are useful for fast spin echo sequences. In FLAIR protocols suitable pulse durations include durations of from about 5 ms to about 20 ms. In diffusion imaging protocols suitable pulse durations include durations of from about 2 ms to about 8 ms for initial excitation and refocusing pulses. In perfusion imaging protocols, suitable pulse durations include durations of from about 0.5 ms to about 2 ms for rapid gradient echo sequences, and from about 10 ms to about 5 second.

[0076] The power parameter refers to the total radiofrequency energy delivered during the pulse, typically expressed in watts (W) or as a percentage of maximum transmitter power. The power is related to amplitude and duration through the relationship P = A2At / Z, where A is the amplitude, and Z is the impedance of the RF system. Power variations may be implemented through different combinations of amplitude and duration parameters.

[0077] Additional radiofrequency excitation parameters include, without limitation, interpulse delay times, phase cycling parameters, and repetition intervals between successive acquisitions.

[0078] In some embodiments, multiple radiofrequency excitation parameters are varied simultaneously to create multi-dimensional parameter spaces. Two-dimensional parameter variations can combine frequency and amplitude to create saturation transfer spectra as functions of both chemical shift and saturation power. These two-dimensional datasets provide information about exchange kinetics, relaxation properties, and molecular concentrations that cannot be obtained from single-parameter variations. Three-dimensional parameter variations may include frequency, amplitude, and duration, creating datasets that can characterize the saturation transfer properties of the tissue. The parameter variations may follow systematic sampling patterns such as Cartesian grids, radial sampling, spiral trajectories.

[0079] In preferred embodiments, at least two radiofrequency excitation parameters are varied simultaneously. This multi-parameter variation provides richer information content and enables more sophisticated synthetic image generation. The at least two radiofrequency excitation parameters comprise, in specific embodiments, the frequency and amplitude of radiofrequency pulses applied during the acquisition.

[0080] In some embodiments of the present invention one or more of the magnetic resonance images in the first set is a non-steady-state magnetic resonance image. Non-steady-state imaging refers to acquisitions where the magnetization has not reached equilibrium between successive radiofrequency excitations. This condition typically occurs when the repetition time is comparable to or shorter than the Ti relaxation time of the tissue, or when the flip angle and timing parameters create complex magnetization dynamics. Non-steady-state imaging can provide improved contrast and enhanced sensitivity to certain tissue properties. For example, in CEST imaging, non-steady-state conditions may occur during the approach to saturation equilibrium, providing information about exchange kinetics and relaxation parameters that is not available from steady-state measurements.

[0081] The images in the first set may be configured as either single-slice images or multi-slice images. Single-slice images provide higher temporal resolution and are particularly useful for dynamic studies or when a specific anatomical region is of interest. The slice thickness for single-slice acquisitions typically ranges from about 3 mm to about 10 mm, with thinner slices providing better spatial resolution and potentially lower signal-to-noise ratio. Multi-slice images provide broader anatomical coverage and are useful for comprehensive clinical assessment. Multi-slice acquisitions may cover the entire organ of interest or a specific region-of-interest. The number of slices typically ranges from about 10 to about 60, with slice thickness ranging from about 2 mm to about 8 mm and potential gap between slices ranging from 0 mm to about 20% of slice thickness.

[0082] Preferably, the total duration of acquiring the first set of magnetic resonance images is less than about one minute or less than about 50 seconds or less than about 40 seconds or less than about 30 seconds or less than about 20 seconds, e.g., about 10 seconds or less. This is unlike conventional imaging protocols, in which the total duration is typically more than 10 minutes. The advantage of these embodiments is that they improve patient comfort, reducing the likelihood of motion artifacts and claustrophobia-related issues. Another advantage is that it the scanner throughput is increased, allowing more patients to be examined in a given time period. These embodiments are particularly beneficial for pediatric patients, critically ill patients, and others who may have difficulty remaining still for extended periods.

[0083] Referring again to FIG. 1, at 12 a computer readable medium storing a machine learning procedure trained for generating magnetic resonance images is accessed. The computer readable medium may comprise various storage technologies including solid-state drives, hard disk drives, network- attached storage systems, cloud-based storage platforms, or high-performance computing clusters. The storage system preferably has sufficient capacity to store the trained machine learning procedure.

[0084] The machine learning procedure is configured and trained to process magnetic resonance image data and associated radiofrequency excitation parameters to generate synthetic images. This configuration involves input layers that can handle the multi-dimensional nature of MRI data, including spatial dimensions, temporal sequences, and parameter vectors.

[0085] In machine learning, information can be acquired via supervised learning or unsupervised learning. In some embodiments of the invention, the machine learning procedure comprises, or is, a supervised learning procedure. In supervised learning, global or local goal functions are used to optimize the structure of the learning system. In other words, in supervised learning there is a desired response, which is used by the system to guide the learning.

[0086] In some embodiments of the invention, the machine learning procedure comprises, or is, an unsupervised learning procedure. In unsupervised learning there are typically no goal functions. In particular, the learning system is not provided with a set of rules. One form of unsupervised learning according to some embodiments of the present invention is unsupervised clustering in which the data objects are not class labeled a priori.

[0087] Representative examples of machine learning procedures suitable for the present embodiments include, without limitation, clustering, association rule algorithms, feature evaluation algorithms, subset selection algorithms, support vector machines, classification rules, cost-sensitive classifiers, vote algorithms, stacking algorithms, Bayesian networks, decision trees, neural networks, instance-based algorithms, linear modeling algorithms, k-nearest neighbors analysis, ensemble learning algorithms, probabilistic models, graphical models, regression methods, gradient ascent methods, singular value decomposition methods and principal component analysis. In some embodiments of the present invention, the machine learning procedure is a procedure employing decision trees.

[0088] Following is an overview of some machine learning procedures suitable for the present embodiments.

[0089] Support vector machines are algorithms that are based on statistical learning theory. A support vector machine (SVM) according to some embodiments of the present invention can be used for classification purposes and / or for numeric prediction. A support vector machine for classification is referred to herein as "support vector classifier," while a support vector machine for numeric prediction is referred to herein as "support vector regression."

[0090] An SVM is typically characterized by a kernel function, the selection of which determines whether the resulting SVM provides classification, regression or other functions. Through application of the kernel function, the SVM maps input vectors into high dimensional feature space, in which a decision hyper-surface (also known as a separator) can be constructed to provide classification, regression or other decision functions. In the simplest case, the surface is a hyper-plane (also known as linear separator), but more complex separators are also contemplated and can be applied using kernel functions. The data points that define the hyper-surface are referred to as support vectors.

[0091] The support vector classifier selects a separator where the distance of the separator from the closest data points is as large as possible, thereby separating feature vector points associated with objects in a given class from feature vector points associated with objects outside the class. For support vector regression, a high-dimensional tube with a radius of acceptable error is constructed which minimizes the error of the data set while also maximizing the flatness of the associated curve or function. In other words, the tube is an envelope around the fit curve, defined by a collection of data points nearest the curve or surface.

[0092] An advantage of a support vector machine is that once the support vectors have been identified, the remaining observations can be removed from the calculations, thus greatly reducing the computational complexity of the problem. An SVM typically operates in two phases: a training phase and a testing phase. During the training phase, a set of support vectors is generated for use in executing the decision rule. During the testing phase, decisions are made using the decision rule. A support vector algorithm is a method for training an SVM. By execution of the algorithm, a training set of parameters is generated, including the support vectors that characterize the SVM. A representative example of a support vector algorithm suitable for the present embodiments includes, without limitation, sequential minimal optimization.

[0093] In KNN analysis, the affinity or closeness of objects is determined. The affinity is also known as distance in a feature space between objects. Based on the determined distances, the objects are clustered and an outlier is detected. Thus, the KNN analysis is a technique to find distance -based outliers based on the distance of an object from its kth-nearest neighbors in the feature space. Specifically, each object is ranked on the basis of its distance to its kth-nearest neighbors. The farthest away object is declared the outlier. In some cases, the farthest objects are declared outliers. That is, an object is an outlier with respect to parameters, such as a k number of neighbors and a specified distance, if no more than k objects are at the specified distance or less from the object. The KNN analysis is a classification technique that uses supervised learning. An item is presented and compared to a training set with two or more classes. The item is assigned to the class that is most common amongst its k-nearest neighbors. That is, compute the distance to all the items in the training set to find the k nearest, and extract the majority class from the k and assign to item.

[0094] Association rule algorithm is a technique for extracting meaningful association patterns among features. The term "association," in the context of machine learning, refers to any interrelation among features, not just ones that predict a particular class or numeric value. Association includes, but is not limited to, finding association rules, finding patterns, performing feature evaluation, performing feature subset selection, developing predictive models, and understanding interactions between features.

[0095] The term "association rules" refers to elements that co-occur frequently within the datasets. It includes, but is not limited to, association patterns, discriminative patterns, frequent patterns, closed patterns, and colossal patterns.

[0096] A usual primary step of association rule algorithm is to find a set of items or features that are most frequent among all the observations. Once the list is obtained, rules can be extracted from them.

[0097] The aforementioned self-organizing map is an unsupervised learning technique often used for visualization and analysis of high-dimensional data. Typical applications are focused on the visualization of the central dependencies within the data on the map. The map generated by the algorithm can be used to speed up the identification of association rules by other algorithms. The algorithm typically includes a grid of processing units, referred to as "neurons." Each neuron is associated with a feature vector referred to as observation. The map attempts to represent all the available observations with optimal accuracy using a restricted set of models. At the same time, the models become ordered on the grid so that similar models are close to each other and dissimilar models far from each other. This procedure enables the identification as well as the visualization of dependencies or associations between the features in the data.

[0098] Feature evaluation algorithms are directed to the ranking of features or to the ranking followed by the selection of features based on their impact.

[0099] Information gain is one of the machine learning methods suitable for feature evaluation. The definition of information gain requires the definition of entropy, which is a measure of impurity in a collection of training instances. The reduction in entropy of the target feature that occurs by knowing the values of a certain feature is called information gain. Symmetrical uncertainty is an algorithm that can be used by a feature selection algorithm, according to some embodiments of the present invention. Symmetrical uncertainty compensates for information gain's bias towards features with more values by normalizing features to a [0,1] range.

[0100] Subset selection algorithms rely on a combination of an evaluation algorithm and a search algorithm. Similarly to feature evaluation algorithms, subset selection algorithms rank subsets of features. Unlike feature evaluation algorithms, however, a subset selection algorithm suitable for the present embodiments aims at selecting the subset of features with the highest impact, while accounting for the degree of redundancy between the features included in the subset. The benefits from feature subset selection include facilitating data visualization and understanding, reducing measurement and storage requirements, reducing training and utilization times, and eliminating distracting features to improve classification.

[0101] Two basic approaches to subset selection algorithms are the process of adding features to a working subset (forward selection) and deleting from the current subset of features (backward elimination). In machine learning, forward selection is done differently than the statistical procedure with the same name. The feature to be added to the current subset in machine learning is found by evaluating the performance of the current subset augmented by one new feature using cross-validation. In forward selection, subsets are built up by adding each remaining feature in turn to the current subset while evaluating the expected performance of each new subset using cross-validation. The feature that leads to the best performance when added to the current subset is retained and the process continues. The search ends when none of the remaining available features improves the predictive ability of the current subset. This process finds a local optimum set of features.

[0102] Backward elimination is implemented in a similar fashion. With backward elimination, the search ends when further reduction in the feature set does not improve the predictive ability of the subset. The present embodiments contemplate search algorithms that search forward, backward or in both directions. Representative examples of search algorithms suitable for the present embodiments include, without limitation, exhaustive search, greedy hill-climbing, random perturbations of subsets, wrapper algorithms, probabilistic race search, schemata search, rank race search, and Bayesian classifier.

[0103] A decision tree is a decision support algorithm that forms a logical pathway of steps involved in considering the input to make a decision.

[0104] The term "decision tree" refers to any type of tree-based learning algorithms, including, but not limited to, model trees, classification trees, and regression trees.

[0105] A decision tree can be used to classify the datasets or their relation hierarchically. The decision tree has a tree structure that includes branch nodes and leaf nodes. Each branch node specifies an attribute (splitting attribute) and a test (splitting test) to be carried out on the value of the splitting attribute, and branches out to other nodes for all possible outcomes of the splitting test. The branch node that is the root of the decision tree is called the root node. Each leaf node can represent a classification or a value. The leaf nodes can also contain additional information about the represented classification such as a confidence score that measures a confidence level in the represented classification (i.e., the likelihood of the classification being accurate). For example, the confidence score can be a continuous value ranging from 0 to 1, with a score of 0 indicating a very low confidence (e.g., the indication value of the represented classification is very low) and a score of 1 indicating a very high confidence (e.g., the represented classification is almost certainly accurate).

[0106] Regression techniques which may be used in accordance with some embodiments of the present invention include, but are not limited to, linear regression, multiple regression, logistic regression, probit regression, ordinal logistic regression, ordinal probit regression, Poisson regression, negative binomial regression, multinomial logistic regression (MLR) and truncated regression.

[0107] A logistic regression or logit regression is a type of regression analysis used for predicting the outcome of a categorical dependent variable (a dependent variable that can take on a limited number of values, whose magnitudes are not meaningful but whose ordering of magnitudes may or may not be meaningful) based on one or more predictor variables. Logistic regression may also predict the probability of occurrence for each data point. Logistic regressions also include a multinomial variant. The multinomial logistic regression model is a regression model which generalizes logistic regression by allowing more than two discrete outcomes. That is, it is a model that is used to predict the probabilities of the different possible outcomes of a categorically distributed dependent variable, given a set of independent variables (which may be real-valued, binary-valued, categorical-valued, etc.). For binaryvalued variables, a cutoff between the 0 and 1 associations is typically determined using the Youden Index.

[0108] A Bayesian network is a model that represents variables and conditional interdependencies between variables. In a Bayesian network, variables are represented as nodes, and nodes may be connected to one another by one or more links. A link indicates a relationship between two nodes. Nodes typically have corresponding conditional probability tables that are used to determine the probability of a state of a node given the state of other nodes to which the node is connected. In some embodiments, a Bayes optimal classifier algorithm is employed to apply the maximum a posteriori hypothesis to a new record in order to predict the probability of its classification, as well as to calculate the probabilities from each of the other hypotheses obtained from a training set and to use these probabilities as weighting factors for future predictions. An algorithm suitable for a search for the best Bayesian network includes, without limitation, global score metric -based algorithm. In an alternative approach to building the network, Markov blanket can be employed. The Markov blanket isolates a node from being affected by any node outside its boundary, which is composed of the node's parents, its children, and the parents of its children. Instance -based techniques generate a new model for each instance, instead of basing predictions on trees or networks generated (once) from a training set.

[0109] The term "instance," in the context of machine learning, refers to an example from a dataset.

[0110] Instance -based techniques typically store the entire dataset in memory and build a model from a set of records similar to those being tested. This similarity can be evaluated, for example, through nearest-neighbor or locally weighted methods, e.g., using Euclidean distances. Once a set of records is selected, the final model may be built using several different techniques, such as the naive Bayes.

[0111] Neural networks are a class of algorithms based on a concept of inter-connected "neurons." In a typical neural network, neurons contain data values, each of which affects the value of a connected neuron according to connections with pre-defined strengths, and whether the sum of connections to each particular neuron meets a pre-defined threshold. By determining proper connection strengths and threshold values (a process also referred to as training), a neural network can achieve efficient recognition of images and characters. Oftentimes, these neurons are grouped into layers in order to make connections between groups more obvious and to ease computation of values. Each layer of the network may have differing numbers of neurons, and these may or may not be related to particular qualities of the input data.

[0112] In one implementation, called a fully-connected network, each of the neurons in a particular layer is connected to and provides input values to each of the neurons in the next layer. These input values are then summed and this sum is used as an input for an activation function (such as, but not limited to, ReLU or Sigmoid). The output of the activation function is then used as an input for the next layer of neurons. This computation continues through the various layers of the neural network, until it reaches a final layer. At this point, the output of the fully-connected network can be read from the values in the final layer, representing the synthetic magnetic resonance images corresponding to the desired radiofrequency excitation parameters.

[0113] Convolutional neural networks (CNNs) include one or more convolutional layers in which the transformation of a neuron value for the subsequent layer is generated by a convolution operation. The convolution operation includes applying a convolutional kernel (also referred to in the literature as a filter) multiple times, each time to a different patch of neurons within the layer. The kernel typically slides across the layer until all patch combinations are visited by the kernel. The output provided by the application of the kernel is referred to as an activation map of the layer. Some convolutional layers are associated with more than one kernel. In these cases, each kernel is applied separately, and the convolutional layer is said to provide a stack of activation maps, one activation map for each kernel. Such a stack is oftentimes described mathematically as an object having D+l dimensions, where D is the number of lateral dimensions of each of the activation maps. The additional dimension is oftentimes referred to as the depth of the convolutional layer. For example, in CNNs that are configured to process two-dimensional magnetic resonance image data, a convolutional layer that receives the two- dimensional image data provides a three-dimensional output, with two-dimensional activation maps and one depth dimension.

[0114] The advantage of using CNN for magnetic resonance image synthesis is the use of layers of kernels providing the ability to learn different levels of complexity of anatomical and contrast features within the magnetic resonance images, enabling the network to capture both local tissue characteristics and global anatomical structures necessary for generating high-quality synthetic images with appropriate radiofrequency parameter-dependent contrast.

[0115] The machine learning procedure used according to some embodiments of the present invention is a trained machine learning procedure. A machine learning procedure can be trained according to some embodiments of the present invention by feeding a machine learning training program with magnetic resonance imaging data of a cohort of subjects, wherein the training data comprises sets of magnetic resonance images acquired with varying radiofrequency excitation parameters and corresponding target synthetic images or quantitative maps. The training dataset can include multiple examples of input image sets with their associated radiofrequency parameter variations, paired with desired output images corresponding to specific radiofrequency excitation parameter combinations. Once the data are fed, the machine learning training program generates a trained machine learning procedure of a selected type which can then be used to generate synthetic magnetic resonance images corresponding to input radiofrequency excitation parameters without the need to re-train it.

[0116] In some embodiments of the present invention the machine learning procedure comprises a vision transformer encoder, which represents an architecture for image processing tasks. Vision transformers adapt the transformer architecture, originally developed for natural language processing, to handle visual data by treating image patches as sequences of tokens. The vision transformer encoder can divide input images into fixed-size patches, which can then be embedded (e.g., linearly) into a higher-dimensional space. Position embeddings can be added to retain spatial information, and the resulting sequence can be processed through multiple transformer encoder layers.

[0117] Each transformer encoder layer of the present embodiments comprises multi-head self-attention mechanisms and feed-forward networks. The self-attention mechanism allows the model employed by the machine learning procedure to select or prefer relevant parts of the image when processing each patch, allowing it to capture long-range dependencies and complex spatial relationships that are useful for medical image analysis.

[0118] The machine learning procedure of the present embodiments can comprise embedding layers for different types of input data. At least one embedding layer is optionally and preferably dedicated to embedding the first set of radiofrequency excitation parameters. This embedding layer transforms the numerical parameter values into high-dimensional vector representations that can be effectively processed by the machine learning procedure.

[0119] On or more of the embedding layers is dedicated to embedding the first set of magnetic resonance images. Image embedding optionally and preferably comprises convolutional neural network layers and / or patch-based processing that extracts relevant features from the input images. The image embedding layers are optionally and preferably designed to capture both local features (such as edges, textures, and intensity patterns) and global features (such as anatomical structures and overall image characteristics).

[0120] In some embodiments of the present invention a concatenation layer combines the embedded radiofrequency excitation parameters with the embedded magnetic resonance images. This concatenation operation creates a unified representation that contains both image information and parameter information, allowing subsequent processing layers to understand the relationships between imaging conditions and image appearance. The concatenation may be implemented at multiple levels within the architecture of the machine learning procedure. Early concatenation combines low-level features, allowing the network to learn how parameters influence basic image characteristics. Late concatenation combines high-level features, indicative of how parameters affect complex image patterns and structures.

[0121] In some embodiments of the present invention the vision transformer encoder is fed by the concatenated representation, processing the combined image and parameter information through its attention mechanisms and feed-forward networks. This processing allows the machine learning procedure to generate contextually appropriate synthetic images that reflect the desired imaging parameters.

[0122] The machine learning procedure may optionally and preferably comprise a one or more convolution layers that are fed by the vision transformer encoder. These convolutional layers serve as decoders that transform the high-level features produced by the transformer into pixel-level image outputs. For example, the convolutional layers can have progressively increasing spatial resolution. The architecture of these convolutional layers may comprise residual connections, attention mechanisms, and normalization layers to ensure stable training and high-quality outputs. At 13, the machine learning procedure is fed with inputs that collectively provide sufficient information for synthetic image generation. One input can comprise the first set of magnetic resonance images. These images can optionally and preferably undergo preprocessing that may include intensity normalization, spatial alignment, artifact correction, and format conversion. Intensity normalization ensures that images acquired under different conditions have comparable intensity ranges, and spatial alignment can correct for patient motion between acquisitions.

[0123] Another input can comprise a first set of radiofrequency excitation parameters describing the variations among the magnetic resonance images in the first set. These parameters can be organized as numerical vectors or matrices that encode the specific imaging conditions used for each image acquisition. The parameter encoding may include direct numerical values (such as frequency offsets in Hz, amplitude values in pT, and duration values in milliseconds) or derived features that capture the characteristics of the radiofrequency excitation conditions.

[0124] An additional input comprises a second set of radiofrequency excitation parameters representing imaging conditions for the synthetic image to be generated. This second set is different from the first set of parameters. The second set of radiofrequency excitation parameters comprises target parameters for which no corresponding acquired image exists, but for which a synthetic image is desired. The second set of radiofrequency excitation parameters may correspond to imaging conditions that were not included in the original acquisition protocol. These parameters may comprise interpolated values between acquired conditions, extrapolated values beyond the acquired range, or entirely different parameter combinations that leverage the learned relationships within the model employed by the machine learning procedure.

[0125] In some embodiments, the method comprises concatenating the first set of radiofrequency excitation parameters with the second set of radiofrequency excitation parameters, wherein the feeding comprises feeding this concatenation to the machine learning procedure. This concatenation creates a combined parameter representation that includes both the reference conditions (from the acquired images) and the target conditions (for the desired synthetic image). The concatenation operation may involve simple vector concatenation, weighted combination, or other fusion techniques.

[0126] In some embodiments, the method comprises defining a plurality of patches over at least one of the magnetic resonance images, wherein feeding the procedure comprises feeding the plurality of patches. These embodiments allow more detailed processing of image content and can improve the quality and accuracy of synthetic image generation. The patches are typically rectangular regions extracted from the original images. The patch size can be selected to improve computational efficiency while preserving anatomical and pathological features. Generally, smaller patches allow finer detail processing, while larger patches preserve more contextual information but may require greater computational resources. The patch can be defined using an overlapping sampling scheme, or by an adaptive scheme based on a defined region of interest or information content. Overlapping patches are advantageous since they improve continuity in the final synthetic images by providing multiple predictions for boundary regions, which can then be averaged or blended to reduce artifacts. Each patch can be processed independently through the initial stages of the machine learning procedure. This allows parallel processing and efficient utilization of computational resources. The use of patches also allows handling of images with varying sizes and resolutions.

[0127] At 14, a synthetic magnetic resonance image corresponding to the second set of radiofrequency excitation parameters is received from the machine learning procedure. This synthetic image provides clinically relevant information that would otherwise require additional scan time to acquire. When patching is employed, the machine learning procedure may produce a synthetic patch for each input patch. In these embodiments, the synthetic patches are combined to reconstruct the full synthetic image. This reconstruction process may involve simple concatenation, weighted averaging, or use of blending techniques that account for patch boundaries and overlapping regions.

[0128] The synthetic magnetic resonance image exhibits image characteristics that correspond to what would theoretically be observed if an actual acquisition were performed using the second set of radiofrequency excitation parameters. These characteristics can include tissue contrast, signal intensities, spatial resolution, and noise characteristics that are consistent with the input imaging conditions.

[0129] The synthetic image may correspond to the same magnetic resonance imaging protocol as the acquired images, representing an interpolation or extrapolation within the same contrast mechanism. Alternatively, the synthetic image may correspond to a different magnetic resonance imaging protocol, representing a cross-contrast synthesis that generates images with different tissue contrast mechanisms from the input images.

[0130] When the synthetic image corresponds to the same protocol, it typically represents imaging conditions that were not included in the original acquisition (typically due to some practical constraints). For example, in CEST imaging, the synthetic image can correspond to a specific frequency offset that was not sampled in the original protocol, or to a different saturation amplitude that provides optimal contrast for a particular molecular species. In some embodiments of the present invention the input set of parameters also comprises the imaging protocol by which the input images are acquired. When the synthetic image corresponds to a different protocol, referred to herein as a target protocol, the machine learning procedure is optionally and preferably fed by the target protocol and provides a synthetic image that corresponds to the target protocol. For example, input images acquired using a Ti-weighted protocol can be fed into the machine learning procedure, and machine learning procedure can use these images to generate synthetic image corresponding to a target protocol once fed with the respective target protocol.

[0131] The synthetic magnetic resonance image may synthesize an image corresponding to various magnetic resonance imaging protocols including, without limitation, Ti-weighted protocol, T2- weighted protocol, FLAIR protocol, diffusion protocol, and perfusion protocol.

[0132] In some embodiments, the method is executed recursively, by redefining, at each recursive execution, the first set of magnetic resonance images to include also the synthetic magnetic resonance image generated in the previous iteration. Typically, the recursive execution begins with the original acquired images and generates a first synthetic image. In the subsequent iteration, this synthetic image is added to the input dataset, effectively expanding the available information for generating additional synthetic images. This process can continue for multiple iterations, with each iteration potentially generating synthetic images corresponding to different parameter combinations. The advantage of the recursive execution is that it allows the generation of multiple synthetic images from a single set of acquired images, effectively multiplying the information content of the original acquisition. Another advantage is that it can improve the accuracy of subsequent synthetic images by providing additional reference points and interpolation anchors. Another advantage is that it allows exploring parameter spaces that extend beyond the original acquisition range. Each iteration can push further into unexplored parameter combinations, guided by the learned relationships from previous iterations.

[0133] The method ends at 15.

[0134] The present embodiments also comprise a technique that can generate a quantitative magnetic resonance imaging (qMRI) map, which provides pixel-wise measurements of specific biological and / or physical properties within a volume scanned by MRI.

[0135] FIG. 2 is a flowchart diagram of a method suitable for generating a quantitative magnetic resonance imaging (qMRI) map according to various exemplary embodiments of the present invention.

[0136] The method begins at 20 and continues to 11 at which at which a first set of magnetic resonance images of an organ of a subject is acquired, wherein for each acquisition, one or more radiofrequency excitation parameter is varied as further detailed hereinabove. At 21, a computer readable medium storing a machine learning procedure trained specifically for generating qMRI maps is accessed. This machine learning procedure can be similar to the procedure described above with respect to method 10, except that instead of being trained to synthesize images, it is trained to generate qMRI maps describing a biological and / or physical property within scanned volumes. In some embodiments of the present invention the initialization point for training this machine learning procedure can be the trained network parameters obtained at 12.

[0137] This machine learning procedure can be trained according to some embodiments of the present invention by feeding a machine learning training program with magnetic resonance imaging data of a cohort of subjects, wherein the training data comprises sets of magnetic resonance images acquired with varying radiofrequency excitation parameters and corresponding ground truth quantitative parameter maps. The training dataset can include multiple examples of input image sets with their associated radiofrequency parameter variations, paired with validated quantitative maps describing specific biological and physical properties such as relaxation times, exchange rates, volume fractions, diffusion coefficients, or other tissue parameters. The ground truth quantitative maps can be obtained by quantitative MRI techniques (including non-lH MRI), physical phantom measurements with known parameter values, other imaging modalities (such as PET, SPECT, etc.), histological data, or theoretical simulations based on established biophysical models. Once the training data are fed, the machine learning training program generates a trained machine learning procedure of a selected type which can then be used to generate synthetic quantitative MRI maps describing tissue properties from input magnetic resonance images and radiofrequency excitation parameters without the need to re-train it.

[0138] At 22, the machine learning procedure is fed with the first set of magnetic resonance images and the first set of radiofrequency excitation parameters describing variations among the magnetic resonance images in the first set. Unlike the procedure described above with respect to method 10, the procedure of method 20 need not be fed by the second set of target parameters, because this procedure extracts intrinsic tissue properties without generating synthetic images corresponding to a set of parameters which was not applied during the acquisition.

[0139] At 23, a synthetic qMRI map describing a specific biological or physical property is received from the machine learning procedure. This map provides pixel-wise quantitative measurements that can be used for clinical diagnosis, treatment planning, and research applications.

[0140] The machine learning procedure of method 20 can be trained for generating a specific qMRI map for a specific property. Alternatively, the machine learning procedure of method 20 can be trained for generating a plurality of qMRI maps, each describing a different property, and the method may comprise receiving from the procedure at least two synthetic qMRI maps describing at least two different properties. Still alternatively, the machine learning procedure of method 20 can receive the property for which it is desired to obtain a qMRI map, and the procedure can generate a qMRI map for the specific input property.

[0141] Representative examples of biological and / or physical properties suitable according to some embodiments of the present invention include, without limitation, parameters related to CEST, magnetization transfer, relaxation, diffusion, perfusion, and other biophysical processes.

[0142] CEST-related properties can include CEST target proton volume fraction, which quantifies the concentration of exchangeable protons associated with specific molecular species. This parameter provides information about molecular content and can be used to assess tissue composition, metabolic activity, and pathological changes. CEST-related properties can alternatively or additionally include CEST proton exchange rate, which is a property that represents the kinetic parameter describing the rate of chemical exchange between target protons and water protons. This parameter reflects the local chemical environment, including pH, temperature, and molecular interactions, providing insights into tissue physiology and pathology. Other contemplated properties include semi-solid magnetization transfer (MT) exchange rate and semi-solid MT proton volume fraction, which describe the magnetization transfer characteristics between water protons and protons bound to macromolecular structures. These parameters provide information about tissue microstructure, macromolecular content, and structural integrity.

[0143] The qMRI map may in some embodiments of the present invention quantify properties related to specific molecular targets, including contribution from amide proton transfer (APT) effects reflecting mobile protein and peptide content, contribution from guanidinium CEST effects indicating creatine and arginine concentrations, amine CEST contributions from various biomolecules, contribution from hydroxyl CEST effects from glucose and glycosaminoglycan content, and aliphatic relayed nuclear Overhauser effect (rNOE) contributions reflecting lipid and macromolecular interactions.

[0144] The qMRI map may in some embodiments of the present invention quantify MRI parameters including, without limitation, Ti, T2, T2*, proton density, and the like. The qMRI map may in some embodiments of the present invention be at least one of: a diffusion coefficient map, which quantifies molecular motion and tissue microstructure, a perfusion parameter map, which quantifies blood flow, blood volume, and hemodynamic characteristics, an inhomogeneous magnetization transfer ratio map, a quantitative susceptibility map, which quantifies local magnetic susceptibility variations, a static magnetic field (Bo) map, which quantifies field homogeneity, a radiofrequency field (Bi) map, which quantifies transmit field efficiency, and a fat fraction map, which quantifies lipid content relative to water.

[0145] The method ends at 24.

[0146] The methods described herein can be useful in various clinical scenarios. Neurological applications include brain tumor assessment, stroke evaluation, multiple sclerosis monitoring, and neuro- degenerative disease characterization. The ability to generate multiple contrast types and quantitative maps from a single acquisition is particularly valuable in emergency settings. Oncological applications benefit from the multi-parametric information provided by the synthetic images and quantitative maps. Tumor characterization, treatment response monitoring, and radiation therapy planning can all benefit from the comprehensive tissue characterization provided by these methods. Musculoskeletal applications include cartilage assessment, muscle characterization, and joint evaluation. The methods of the present embodiments are particularly beneficial for pediatric patients and those with limited mobility or pain.

[0147] Reference is now made to FIG. 3 which is a schematic illustration of a magnetic resonance imaging system 40 for imaging a body 42, according to some embodiments of the present invention. System 40 comprises a static magnet system 44 which generating a substantially homogeneous and stationary magnetic field Bo in the longitudinal direction, a gradient assembly 46 which generates instantaneous magnetic field gradient pulses to form a non-uniform superimposed magnetic field, and a radiofrequency transmitter system 48 which generates and transmits radiofrequency pulses to body 42.

[0148] System 40 further comprises an acquisition system 50 which acquires magnetic resonance signal from the body, and a control system 52 which is configured for implementing MRI protocols, in order to acquire sets of magnetic resonance images, wherein for each acquisition, at least one radiofrequency excitation parameter is varied, as further detailed hereinabove.

[0149] In various exemplary embodiments of the invention system 40 further comprises an image producing system 54 which produces magnetic resonance images from the signals of each acquisition. Image producing system 54 typically implements a Fourier transform so as to transform the data into an array of image data.

[0150] The gradient pulses and / or whole body pulses can be generated by a generator module 64 which is typically a part of control system 52. Generator module 64 produces data which indicates the timing, strength and shape of the radiofrequency pulses which are to be produced, and the timing of and length of the data acquisition window. Gradient assembly 46 typically comprises Gx, Gyand G- coils each producing the magnetic field gradients used for position encoding acquired signals. Radiofrequency transmitter system 48 is typically a resonator which is used both for transmitting the radiofrequency signals and for sensing the resulting signals radiated by the excited nuclei in body 42. The sensed magnetic resonance signals can be demodulated, filtered, digitized etc. in acquisition system 50 or control system 52.

[0151] System 40 also comprises a data processor 60 and a computer-readable medium 62. Computer- readable medium 62 stores program instructions and also one or more of the machine learning procedures. The instructions and machine learning procedures are read by data processor 60, and cause data processor 60 to receive the first set of magnetic resonance images, the first set of radiofrequency excitation parameters, and optionally also the second set of radiofrequency excitation parameters, and to execute method 10 or method 20.

[0152] As used herein the term “about” refers to ±10 %

[0153] The terms "comprises", "comprising", "includes", "including", “having” and their conjugates mean "including but not limited to".

[0154] As used herein, the singular form "a", "an" and "the" include plural references unless the context clearly dictates otherwise. For example, the term "a compound" or "at least one compound" may include a plurality of compounds, including mixtures thereof.

[0155] Throughout this application, various embodiments of this invention may be presented in a range format. It should be understood that the description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the invention. Accordingly, the description of a range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range. For example, description of a range such as from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6 etc., as well as individual numbers within that range, for example, 1, 2, 3, 4, 5, and 6. This applies regardless of the breadth of the range.

[0156] Whenever a numerical range is indicated herein, it is meant to include any cited numeral (fractional or integral) within the indicated range. The phrases “ranging / ranges between” a first indicate number and a second indicate number and “ranging / ranges from” a first indicate number “to” a second indicate number are used herein interchangeably and are meant to include the first and second indicated numbers and all the fractional and integral numerals therebetween. It is appreciated that certain features of the invention, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable sub-combination or as suitable in any other described embodiment of the invention. Certain features described in the context of various embodiments are not to be considered essential features of those embodiments, unless the embodiment is inoperative without those elements.

[0157] Various embodiments and aspects of the present invention as delineated hereinabove and as claimed in the claims section below find experimental support in the following examples.

[0158] EXAMPLES

[0159] Reference is now made to the following examples, which together with the above descriptions illustrate some embodiments of the invention in a non limiting fashion.

[0160] Example 1

[0161] Decoding Human Brain Tissue Response to Radiofrequency Excitation

[0162] MRI relies on radiofrequency (RF) excitation in controlling proton spins. For clinical diagnosis, a comprehensive biophysical information is collated via multiple MRI contrasts, acquired using a series of RF sequences. The Inventors appreciated that this leads to lengthy examinations. This example describes a transformer-based framework that captures the spatiotemporal magnetic signal evolution and decodes the brain tissue response to RF excitation, optionally and preferably constituting an MRI on a chip. Following a per-subject rapid calibration scan (28.2 s), a wide variety of image contrasts can be automatically generated, including fully quantitative molecular, water relaxation, and magnetic field maps. The method of the present embodiments was validated using in vivo healthy human volunteers, enabling 94% acceleration compared to alternatives. The deep MRI on a chip (hereinafter DeepMonC) framework allows revealing the molecular composition of the human brain tissue in a wide range of pathologies, while offering clinically attractive scan times.

[0163] Introduction

[0164] MRI is one of the most powerful diagnostic tools in present-day clinical healthcare1. One of its most compelling advantages is wide versatility, which facilitates the acquisition of different biophysical information types using a single imaging modality2. This trait is rooted in the ability to program MRI scans for emphasizing a particular tissue property of interest. Specifically, a series of radiofrequency (RF) pulses are designed and applied to initiate a cascade of interactions with the tissue proton spins. The particular waveform, duration, power, and frequency of each RF pulse, as well as the characteristics of the entire RF train ensemble affect the resulting contrast, which is tailored to reflect microstructure, water content, cellularity, blood flow, molecular composition, or even functional characteristics3.

[0165] As a single MRI pulse sequence is often incapable to determine the detailed tissue state and diagnosis with sufficient certainty, standard clinical MRI exams include a serial acquisition of several pulse sequences4. For example, brain cancer MRI protocols are typically comprised of Ti-weighted, T2-weighted, fluid-attenuated inversion recovery (FLAIR), diffusion, and potentially also perfusion, and MR-spectroscopy imaging5. The Inventors realized that while multi-sequence acquisition provides rich information, it translates into exceedingly long examination times (e.g., 20-60 min) that increase patient discomfort, healthcare costs, and waiting lines6. Moreover, the image contrast depends not only on the acquisition parameters, but also on the particular tissue characteristics, which are highly variable across subjects. Therefore, while a given protocol may be sufficient for differentiating between several tumor components for one patient, it may be insufficient for another, and fine tuning of the RF pulses used may be needed (e.g., choosing a different flip angle, saturation pulse power, etc.). As radiological image analysis is commonly performed after the acquisition is completed and the subject has left, rescheduling a subject scan is either impractical or results in increased costs and prolonged waiting lines7.

[0166] In recent years, several new mechanisms were increasingly studied for further enriching the biochemical information portfolio provided by MRI scans. Saturation transfer (ST) MRI constitute one such example8, due to its RF-tunable sensitivity for various molecular properties, such as mobile protein and peptide volume fraction, intracellular pH, and glutamate concentration9. ST-MRI has shown promise for a variety of clinical applications10, such as tumor detection and grading11 12, early stroke characterization13 14, neurodegenerative disorder imaging15 16, and kidney disease monitoring17 18. The Inventors appreciate that the associated scan times of ST-MRI are relatively long. The Inventors also appreciate that multi-contrast ST imaging has heretofore been less practical, because each ST target compound is characterized by a distinct proton exchange rate, requiring acquisition of a separate pulse sequence for each application of interest.

[0167] Quantitative imaging of biophysical tissue properties offers improved reproducibility, sensitivity, and consistency across sites and scanners, compared to contrast-weighted imaging19. Recently, imaging techniques that combine biophysical (differential-equation-based) models with artificial intelligence (Al) were suggested for accelerating water-pool relaxometry20 24and quantitative ST-MRI acquisition and reconstruction25 31. However, the Inventors found that the complexity of the multiproton pool in-vivo environment and the challenge in accurately modeling the large number of free tissue parameters limit the accuracy of this approach for molecular MRI. The Inventors found that this leads to substantial variability between the biophysical values reported by various groups (each incorporating different model assumptions)32 34and / or to increased acquisition times, as water pool and magnetic field parameters may need to be separately estimated via additional pulse sequences, to reduce the model complexity26,35.

[0168] This Example describes a deep learning framework which is preferably free of any biophysicalmodel. The framework (FIGs. 4A and 4B) was developed for providing rich biological information in-vivo, while circumventing the need for lengthy multi-pulse sequence MRI acquisition. The framework presented herein was able to capture the spatiotemporal magnetic signal evolution dynamics in living humans and decode the brain tissue response to RF excitation, constituting a deep MRI on a chip. When employed on unseen subjects, the method of the present embodiments was able to accurately emulate the spin evolution dynamics, and to generate a variety of new image contrasts, as well as fully quantitative molecular, water relaxation, and magnetic field maps. Preferably, a per-subject calibration scan is employed. This example described a rapid per-subject calibration scan of 28.2 seconds or less.

[0169] Results

[0170] Deep Eearning

[0171] The DeepMonC core module was designed to capture the spatiotemporal dynamics of MRI signal propagation as a response to RF excitation, enabling the generation of unseen image contrast. A vision transformer36,37was assembled and incorporated with dual-domain input, comprised of RF excitation information and real-world tissue response image counterparts. An extension module was also designed, which further quantifies six biophysical tissue parameters across the entire 3D brain, without the need for any additional input.

[0172] This Example describes two modules for the deep MRI on a chip technique of the present embodiments. A first module is a core module (FIG. 4A) for which the inputs are a sequence of six non-steady-state MRI calibration images and an RF excitation parameter tensor. The tensor includes two concatenated parts: a first set of m pairs of acquisition parameters used for obtaining a first set of m calibration images, and a second set of m pairs parameters for which it is desired to have a subsequent (virtually acquired) image output. Separate embeddings for the real-image-data and the physical RF properties are then learned, using a vision transformer and a fully connected layer, respectively.

[0173] A second module for the deep MRI on a chip technique of the present embodiments is a quantification module (FIG. 4B), in which a transfer learning strategy was implemented. In this module, the core module weights were plugged-in, the last layer was removed, and two new convolutional layers augmentation were realized. Quantification-oriented learning was performed using ground-truth reference data. Note that while the input to the quantification module is shown to be the same as the input to the core module this need not necessarily be the case, since, for some applications, it may not be necessary for the input to the quantification module to include the second set of parameters, as these parameters do not affect the quantification maps.

[0174] Biophysical-model-free prediction of the tissue response to RF excitation

[0175] The DeepMonC core module was validated for generating unseen semisolid MT and amide CEST contrast images, using a dataset comprising 68 different brain slices, each acquired using 30 pseudo-random RF excitations from a new human subject, not used for training. The first six images were used for per-subject calibration, followed by DeepMonC activation for predicting the multi-contrast images associated with the next six response images. The process was continued recursively, until the entire 30-long sequence was predicted based on merely the first six calibration images. A representative example of DeepMonC output compared to the ground truth is shown in FIGs. 5A (ground truth) and 5B (prediction). Note that the tissue response requested for prediction was associated with RF parameters that are extrapolated far beyond the range of the calibration training parameter, creating a highly challenging settings (dashed lines compared to solid lines in FIG. 4A, bottom). Nevertheless, an excellent visual agreement between the generated and ground-truth image-sets was obtained.

[0176] Tables 1 and 2, below, provide statistical analysis (p<0.0001) of similarity index measure (SSIM), peak signal-to-noise ratio (PSNR) and normalized mean square error (NRMSE) performance measures, comparing the reconstructed molecular contrast-weighted images compared to the reference ground truth. The data in Table 1 correspond to the prediction of M new contrasts based on M calibrations, and the data in Table 2 corresponds to the recursive prediction of 4M new contrasts based on M calibration images Table 1

[0177] Table 2 As demonstrated in Tables 1 and 2, the quantitative estimation of image and pixel wise value differences yielded SSIM > 0.96, PSNR > 32, and NRMSE < 3.5%. The inference time required for reconstructing whole brain six unseen contrasts was 32.69 sec / 58.67 sec, using an Nvidia RTX 3060 GPU or a desktop CPU (Intel I9-12900F), respectively. Rapid quantification of biophysical tissue parameters

[0178] The extended quantification module was validated using a separate dataset, comprised of 68 slice images form a different healthy human volunteer. The resulting images were visually and perceptually similar to the ground truth reference (FIGs. 3A, 3B, 4), yielding an SSIM score of 0.918, PSNR = 30.14, and NRMSE = 0.049. To further analyze the contribution of the decoded tissue re- sponse information captured by DeepMonC core module for the quantification task, a separate ablation study was performed. The same quantification architecture (FIG. IB) was trained to receive the same inputs (including the RF parameter information), and output the same six quantitative biophysical parameter maps, yet without using the pre-trained DeepMonC weights. This resulted in a statistically significant lower SSIM (0.805, p<0.0001, n=68 image pairs) and PSNR (25.733, p<0.0001, n=68 image pairs), and higher NRMSE (0.084, p<0.0001, n=68 image pairs) (FIG. 3D) and a clear deterioration in the parameter map visual resemblance to the ground truth (FIG. 3C).

[0179] Discussion

[0180] The past few decades have seen an increasing reliance on MRI examinations for clinical diagnosis. In parallel to the discovery of new contrast mechanisms, and their exploitation for improving the diagnosis certainty, the long waiting lines associated with the relatively long scan times, that further build-up when using several imaging protocols at the same session, have prevented the effective reach of this imaging modality to the general public. The present embodiments provide an intelligent system for decoding the human brain tissue response to RF excitation, where the complex interactions between the MR scanner and the biological target are captured and stored. The system of the present embodiments allows the generation of a variety of on-demand image contrasts, that faithfully recapitulates the true physical counterparts obtained from physical imaging.

[0181] In this Example, the same system was expanded and modified to facilitate fully quantitative mapping of various tissue biophysical parameters, including water-proton -relaxation, semisolid macromolecule information, and magnetic field homogeneity. The only input that was used in this Example for the inference time was a per-subject calibration, achieved within a very short (28.2 sec) scan. While the architecture was designed for image translation of m-to-m size, it can be recursively applied according to some embodiments of the present invention on the first model output, preserving attractive performance over up to m-to-4m translations (Tables 1 and 2). As known multi-contrast CEST imaging techniques (achieved via multi-saturation-pulse-power excitation) required a 15 minute-long-scan38, the technique of the present embodiments constitute an improvement in the performance as it is based on faster calibration. Moreover, DeepMonC can compensate for unknown spin history, so that non- fully-relaxed acquired images can serve as the calibration input. In all cases, a satisfactory reconstruction was obtained (SSIM > 0.96, PSNR > 36, NRMSE > 2%), even though the performance is not independent of the particular acquisition parameters used for calibrating the system and specific acquisition parameters demanded by the user.

[0182] The quantitative mapping module (FIG. IB) demonstrated excellent agreement with the ground truth, as well as an ability to mitigate hardware -related artifacts (white arrows in FIGs. 3A and 3B), accompanied by a drastic reduction in whole-brain scan time (28.2 sec compared to 8.5 min using alternative state of the art). The ground-truth reference used in this case, was obtained via standard water proton relaxometry, magnetic field-mapping and semioslid MT MRF. This flexible extension module can be trained using alternative reference modalities, such as31P imaging (for reconstructing pH maps), MR-spectroscopy (for quantifying other compounds and metabolites), or non-MRI images.

[0183] In addition to the quantification demonstrated in this Example, or as an alternative, the present embodiments contemplate also other quantification tasks.

[0184] It is to be understood that that while the technique demonstrated in this Example focused on semisolid MT and CEST-based acquisition protocols, the present embodiments also contemplate use of other MRI protocols and contrast mechanisms, such as, but not limited to, as diffusion, perfusion, susceptibility mapping, and others.

[0185] It is to be understood that that while the technique demonstrated in this Example focused on brain imaging, the present embodiments contemplate application of the technique for any other organ. Preferably, the machine learning procedure is trained specifically for each type of organ for which MRI prediction is to be performed according to some embodiments of the present invention.

[0186] Known in the art are synthetic MRI method that are based on a mandatory parameter quantification step, obtained via biophysical signal models20,39. The Inventors found that this approach is less than being optimal and may even be not suitable for complex multi-proton-pool imaging where there is an enormous number of tissue parameters, which cannot be simultaneously and accurately detected. In distinction from these methods, the technique of the present embodiments provides a fast biophys- ical-model-free strategy to learn the complicated interaction between the RF excitation and the underlying tissue biophysical properties.

[0187] Methods

[0188] Datasets

[0189] All subject were scanned at Tel Aviv University, using a 64-channel 3T MRI (Prisma, Siemens Healthineers). The research protocol was approved by the Tel Aviv University Institutional Ethics Board (study no. 0007572-2) and the Chaim Sheba Medical Center Ethics Committee (0621-23-SMC).

[0190] Ten healthy volunteers were recruited and signed an informed consent form. The training cohort was separated into 80% training and 20% validation.

[0191] Core module

[0192] In this model the Inventors used as an input a sequence of images x G RCXHXWand a tensor of the corresponding parameter values Bi and mryof the current and predicted scans peR2xl-*C)5where c is the number of images in the sequence. The model predicts the desired sequence of scans according to the given parameters y G RCx / / xl' . This architecture utilizes the unique property of transformer encoders by taking into consideration both the MRI images, their corresponding scan parameters and the predicted parameter values. This way, the model learns the mathematical relations between the scans and their parameters.

[0193] Every sequence of scans was patched, projected and embedded according to a vision transformer architecture. The input shape of the images was 144 by 144, and the number of input images per sample was 6. For the transformer, the following hyper-parameters were used: embedding dimension of 768, dropout of 0, mlp size of 3072, number of transformer layers of 3 and number of attention heads is 4.

[0194] Each image was patched into 9 pieces. A linear projection of flattened patches was created for each of the patches. The scan parameters tensor was embedded using a fully connected layer to match the dimensions of the embedded images. Before transferring the data to the transformer encoder, both of the embedded data was concatenated to one tensor containing information about the images and about the parameters.

[0195] After the transformer encoder, a sequential set of convolutions layers is being applied in order to create a sequence of images at the output of the model. The sequential convolutional layers contained three blocks. A first and a second blocks contained convolutional layer of 3x3, Batch normalization, ReLU activation function and Max pooling. A third block contains up-sample layer, con layer of 3x3, and sigmoid activation function in order to keep the predicted image values in the range of [0,1]. The output sequence possesses the same number of images as the input sequence.

[0196] Each of the healthy human volunteers was scanned using whole-brain 3D semisolid MT and amide, non-steady-state and fast acquisition protocols32. The input tensor (FIG. 1 A) was a combination of an MRI images sequence, their current scan parameter values and the desired future scan parameters. The total samples used for training and validation was 456,000 and 114,000, respectively. The Inventors tested the accuracy of prediction one sequence compared to the ground truth sequence and the accuracy of predicting the whole slice images in an iterative way compared to the ground truth.

[0197] Quantification module

[0198] After gaining knowledge about the connection between the MRI scans and their corresponding parameters, the Inventors applied this knowledge on a second model in order to predict six different quantitative images fss, Bo, Bi, Ti, T2). The weights of the previous model were applied as an initiative weights to the second model. The last layers in the architecture were also changed. The last block in the sequential convolutional layers of the first model was modified by removing the convolutional layer of 3x3 and the sigmoid and adding to it a new convolutional layer of 3x3, Batch normalization layer and ReLU activation. A fourth block, including a convolutional layer of 3x3 and a sigmoid activation function, was added a to ensure pixel values in the range of [0,1]. The output dimensions in the second model equivalent to the first model. The whole network was trained using new labels due to the fact that the domain of the data was different.

[0199] The input to the second model was the same as the first model. The last layers of the first model were modified by removing the last convolutional layer and adding two new convolutional layers. Loss Function

[0200] In both models, the loss function was defined as a combination of Structural similarity index measure and LI, according to the following: where y" is the predicted sequence and y is the ground truth sequence. This loss takes into consideration both the statistics and the amplitude of the result sequence.

[0201] Data Evaluation

[0202] For the first model, 9 different healthy volunteers’ scans were used for building the training dataset. Every volunteer contains scans using MT and Amide protocols. Every given scan was of length 30 and contained n different slices. Given a desired sequence length of k, current k images were paired with the next k images. This way for every volunteer there were (30 - 2k + 1 ) * n different pairs. This process was repeated for the parameters of every sequence in a same manner. Eventually there were (30 - 2k + l) * n * nvoisdifferent pairs for the training set. The test set for the first model consisted of data taken from a different volunteer.

[0203] For the second model, data from 8 different healthy volunteers was used. Each volunteer had 6 different quantitative images per slice. All the volunteers were scanned using only MT protocol. For the input data, both the scan sequences and the corresponding scan parameters were arranged in the same manner like the first model. For the output data, all the pairs in a given slice were matched to the relevant slice of quantitative images. Thus, for each slice there were (30 - 2k +1) pairs with the same label. Training Properties

[0204] For the first model, a machine with 5 GPUs of RTX 5000 was used. The model was trained using total batch size of 64, learning rate of 0.0004 and was trained for 259 epochs. The total training time took 5 days. For the second model, a single GPU of RTX 5000 was used. The model was trained using total batch size of 16, learning rate of 0.002 and was trained for 348 epochs. The total training time took 3 days. Both of the models were implemented using PyTorch.

[0205] Although the invention has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, it is intended to embrace all such alternatives, modifications and variations that fall within the spirit and broad scope of the appended claims.

[0206] It is the intent of the applicant(s) that all publications, patents and patent applications referred to in this specification are to be incorporated in their entirety by reference into the specification, as if each individual publication, patent or patent application was specifically and individually noted when referenced that it is to be incorporated herein by reference. In addition, citation or identification of any reference in this application shall not be construed as an admission that such reference is available as prior art to the present invention. To the extent that section headings are used, they should not be construed as necessarily limiting. In addition, any priority document(s) of this application is / are hereby incorporated herein by reference in its / their entirety.

[0207] REFERENCES

[0208] [1] van Beek, E. J. et al. Value of mri in medicine: More than just another test? Journal of Magnetic Resonance Imaging 49, el4-e25 (2019).

[0209] [2] Yousaf, T., Dervenoulas, G. & Politis, M. Advances in mri methodology. International review of neurobiology 141, 31-76 (2018).

[0210] [3] Bernstein, M. A., King, K. F. & Zhou, X. J. Handbook of MRI pulse sequences (Elsevier, 2004).

[0211] [4] Ellingson, B. M. et al. Consensus recommendations for a standardized brain tumor imaging protocol in clinical trials. Neuro-oncology 17, 1188-1198 (2015).

[0212] [5] Kaufmann, T. J. et al. Consensus recommendations for a standardized brain tumor imaging protocol for clinical trials in brain metastases. Neuro-oncology 22, 757-772 (2020).

[0213] [6] Edelstein, W. A., Mahesh, M. & Carrino, J. A. Mri: time is dose — and money and versatility. Journal of the American College of Radiology: JACR 7, 650 (2010).

[0214] [7] Sreekumari, A. et al. A deep learning-based approach to reduce rescan and recall rates in clinical mri examinations. American Journal of Neuroradiology 40, 217-223 (2019).

[0215] [8] Van Zijl, P. C., Lam, W. W., Xu, J., Knutsson, L. & Stanisz, G. J. Magnetization transfer contrast and chemical exchange saturation transfer mri. features and analysis of the fielddependent saturation spectrum. Neuroimage 168, 222-241 (2018).

[0216] [9] Liu, G., Song, X., Chan, K. W. & McMahon, M. T. Nuts and bolts of chemical exchange saturation transfer mri. NMR in Biomedicine 26, 810-828 (2013).

[0217]

[0010] Jones, K. M., Pollard, A. C. & Pagel, M. D. Clinical applications of chemical exchange saturation transfer (cest) mri. Journal of Magnetic Resonance Imaging 47, 11-27 (2018).

[0218]

[0011] Zhou, J. et al. Differentiation between glioma and radiation necrosis using molecular magnetic resonance imaging of endogenous proteins and peptides. Nature medicine 17, 130-134 (2011).

[0219]

[0012] Zhou, J., Heo, H.-Y., Knutsson, L., van Zijl, P. C. & Jiang, S. Apt-weighted mri: techniques, current neuro applications, and challenging issues. Journal of Magnetic Resonance Imaging 50, 347-364 (2019).

[0220]

[0013] Zhou, J., Payen, J.-F., Wilson, D. A., Traystman, R. J. & Van Zijl, P. C. Using the amide proton signals of intracellular proteins and peptides to detect ph effects in mri. Nature medicine 9, 1085— 1090 (2003).

[0221]

[0014] Sun, P. Z., Zhou, J., Sun, W., Huang, J. & Van Zijl, P. C. Detection of the ischemic penumbra using ph-weighted mri. Journal of Cerebral Blood Flow & Metabolism 27, 1129-1136 (2007).

[0222]

[0015] Cai, K. et al. Magnetic resonance imaging of glutamate. Nature medicine 18, 302-306 (2012).

[0016] Cember, A. T., Nanga, R. P. R. & Reddy, R. Glutamate-weighted cest (glucest) imaging for mapping neurometabolism: An update on the state of the art and emerging findings from in vivo applications. NMR in Biomedicine 36, e4780 (2023).

[0223]

[0017] Stabinska, J., Keupp, J. & McMahon, M. T. Cest mri for monitoring kidney diseases. In Advanced Clinical MRI of the Kidney: Methods and Protocols, 345-360 (Springer, 2023).

[0224]

[0018] Longo, D. L., Cutrin, J. C., Michelotti, F., Irrera, P. & Aime, S. Noninvasive evaluation of renal ph homeostasis after ischemia reperfusion injury by cest-mri. NMR in Biomedicine 30, e3720 (2017).

[0225]

[0019] Seiberlich, N. et al. Quantitative magnetic resonance imaging (Academic Press, 2020).

[0226]

[0020] Ma, D. et al. Magnetic resonance fingerprinting. Nature 495, 187-192 (2013).

[0227]

[0021] Fujita, S. et al. Mr fingerprinting for contrast agent-free and quantitative characterization of focal liver lesions. Radiology: Imaging Cancer 5, e230036 (2023).

[0228]

[0022] Gaur, S. et al. Magnetic resonance fingerprinting: a review of clinical applications. Investigative Radiology 58, 561-577 (2023).

[0229]

[0023] Cohen, O., Zhu, B. & Rosen, M. S. Mr fingerprinting deep reconstruction network (drone). Magnetic resonance in medicine 80, 885-894 (2018).

[0230]

[0024] Fyrdahl, A., Seiberlich, N. & Hamilton, J. I. Magnetic resonance fingerprinting: The role of artificial intelligence. In Artificial Intelligence in Cardiothoracic Imaging, 201-215 (Springer, 2022).

[0231]

[0025] Perlman, O., Farrar, C. T. & Heo, H.-Y. Mr fingerprinting for semisolid magnetization transfer and chemical exchange saturation transfer quantification. NMR in Biomedicine 36, e4710 (2023).

[0232]

[0026] Perlman, O. et al. Quantitative imaging of apoptosis following oncolytic virotherapy by magnetic resonance fingerprinting aided by deep learning. Nature biomedical engineering 6, 648- 657 (2022).

[0233]

[0027] Perlman, O., Zhu, B., Zaiss, M., Rosen, M. S. & Farrar, C. T. An end-to-end ai-based framework for automated discovery of rapid cest / mt mri acquisition protocols and molecular parameter quantification (autocest). Magnetic Resonance in Medicine 87, 2792-2810 (2022).

[0234]

[0028] Cohen, O. et al. Cest mr fingerprinting (cest-mrf) for brain tumor quantification using epi readout and deep learning reconstruction. Magnetic resonance in medicine 89, 233-249 (2023).

[0029] Nagar, D., Vladimirov, N., Farrar, C. T. & Perlman, O. Dynamic and rapid deep synthesis of chemical exchange saturation transfer and semisolid magnetization transfer mri signals. Scientific Reports 13, 18291 (2023).

[0235]

[0030] Kang, B., Kim, B., Park, H. & Heo, H.-Y. Learning-based optimization of acquisition schedule for magnetization transfer contrast mr fingerprinting. NMR in Biomedicine 35, e4662 (2022).

[0236]

[0031] Vladimirov, N. et al. Quantitative molecular imaging using deep magnetic resonance fingerprinting. Protocol Exchange Preprint (2024).

[0237]

[0032] Weigand- Whittier, J. et al. Accelerated and quantitative three-dimensional molecular mri using a generative adversarial network. Magnetic Resonance in Medicine 89, 1901-1914 (2023).

[0238]

[0033] Heo, H.-Y. et al. Quantifying amide proton exchange rate and concentration in chemical exchange saturation transfer imaging of the human brain. Neuroimage 189, 202-213 (2019).

[0239]

[0034] Carradus, A. J., Bradley, J. M., Gowland, P. A. & Mougin, O. E. Measuring chemical exchange saturation transfer exchange rates in the human brain using a particle swarm optimisation algorithm. NMR in Biomedicine 36, e5001 (2023).

[0240]

[0035] Kim, B., Schar, M., Park, H. & Heo, H.-Y. A deep learning approach for magnetization transfer" contrast mr fingerprinting and chemical exchange saturation transfer imaging. Neuroimage 221, 117165 (2020).

[0241]

[0036] Hatamizadeh, A. et al. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE / CVF winter conference on applications of computer vision, 574-584 (2022).

[0242]

[0037] Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (2020).

[0243]

[0038] Fabian, M. S. et al. Comprehensive 7 t cest: A clinical mri protocol covering multiple exchange rate regimes. NMR in Biomedicine e5096 (2024).

[0244]

[0039] Ji, S. et al. Synthetic mri: technologies and applications in neuroradiology. Journal of Magnetic Resonance Imaging 55, 1013-1025 (2022).

Claims

WHAT IS CLAIMED IS:

1. A method of generating magnetic resonance images, comprising: acquiring a first set of magnetic resonance images of an organ of a subject, wherein for each acquisition, at least one radiofrequency excitation parameter is varied; accessing a computer readable medium storing a machine learning procedure trained for generating magnetic resonance images; feeding said procedure with (i) said first set of magnetic resonance images, (ii) a first set of radiofrequency excitation parameters describing variations among said magnetic resonance images in said first set, and (iii) a second set of radiofrequency excitation parameters; and receiving from said procedure a synthetic magnetic resonance image corresponding to said second set of radiofrequency excitation parameters.

2. The method according to claim 1, wherein said synthetic magnetic resonance image and said magnetic resonance images correspond to the same magnetic resonance imaging protocol.

3. The method according to claim 1, wherein said synthetic magnetic resonance image and said magnetic resonance images correspond to different magnetic resonance imaging protocol.

4. The method according to any of claims 2 and 3, wherein each of said synthetic magnetic resonance images synthesizes an image corresponding to a magnetic resonance imaging protocol selected from the group consisting of Ti-weighted protocol, T2-weighted protocol, fluid- attenuated inversion recovery (FLAIR) protocol, diffusion protocol, and perfusion protocol.

5. The method according to any of claims 1-4, comprising concatenating said first set of radiofrequency excitation parameters with said second set of radiofrequency excitation parameters, wherein said feeding comprises feeding said concatenation.

6. A method of generating a quantitative magnetic resonance imaging (qMRI) map, comprising: acquiring a first set of magnetic resonance images of an organ of a subject, wherein for each acquisition, at least one radiofrequency excitation parameter is varied;accessing a computer readable medium storing a machine learning procedure trained for generating qMRI maps describing a biological and / or physical property within a scanned volume; feeding said procedure with (i) said first set of magnetic resonance images, and (ii) a first set of radiofrequency excitation parameters describing variations among said magnetic resonance images in said first set; and receiving from said procedure a synthetic qMRI map describing said property.

7. The method according to claim 6, wherein said machine learning procedure is trained for generating a plurality of qMRI maps each describing a different property, and the method comprises receiving from said procedure at least two synthetic qMRI maps describing at least two different properties.

8. The method according to any of claims 6 and 7, wherein said property comprises chemical exchange saturation transfer (CEST) target proton volume fraction.

9. The method according to any of claims 6-8, wherein said property comprises chemical exchange saturation transfer (CEST) proton exchange rate.

10. The method according to any of claims 6-9, wherein said property comprises at least one property selected from the group consisting of semi-solid magnetization transfer (MT) proton exchange rate, semi-solid MT proton volume fraction, amide effect contribution, guanidinium effect contribution, amine CEST contribution, hydroxyl CEST effect contribution, aliphatic relayed nuclear Overhauser effect (rNOE) contribution, longitudinal relaxation time, transverse relaxation time, effective transverse relaxation time, proton density, diffusion coefficient, perfusion parameter, inhomogeneous magnetization transfer ratio, magnetic susceptibility value, static magnetic field (Bo) mapping, radiofrequency field (Bi) mapping, and fat fraction.

11. The method according to any of claims 1-9, wherein a magnetic resonance imaging protocol of each of said magnetic resonance images of said first set is selected from the group consisting of Ti-weighted protocol, T2-weighted protocol, fluid-attenuated inversion recovery (FLAIR) protocol, diffusion protocol, and perfusion protocol.

12. The method according to any of claims 1-11, wherein said machine learning procedure comprises a vision transformer encoder.

13. The method according to claim 12, wherein said machine learning procedure comprises at least one embedding layer for embedding said first set of radiofrequency excitation parameters, at least one embedding layer for embedding said first set of magnetic resonance images, and a concatenation layer for concatenating said embedded first set of radiofrequency excitation parameters with said embedded first set of magnetic resonance images, wherein said vision transformer encoder is fed by said concatenation.

14. The method according to claim 13, wherein said machine learning procedure comprises a plurality of convolutions layers being fed by said vision transformer encoder.

15. The method according to any of claims 1-14, wherein at least one of said first set of magnetic resonance images is a non-steady-state magnetic resonance image.

16. The method according to any of claims 1-15, comprising defining a plurality of patches over at least one of said magnetic resonance images, wherein said feeding said procedure comprises feeding said plurality of patches.

17. The method according to any of claims 1-16, wherein for each acquisition, at least two radiofrequency excitation parameters are varied.

18. The method according to claim 17, wherein said at least two radiofrequency excitation parameters comprise a frequency and amplitude of radiofrequency pulses applied during said acquisition.

19. The method according to any of claims 1-18, being executed recursively, by redefining, at each recursive execution, said first set of magnetic resonance image to include also said synthetic magnetic resonance image.

20. The method according to any of claims 1-19, wherein said acquiring comprises executing saturation transfer MRI.

21. The method according to claim 20, wherein said saturation transfer MRI comprises Chemical Exchange Saturation Transfer (CEST) imaging.

22. The method according to claim 21, wherein said CEST comprises amide proton transfer.

23. The method according to claim 21, wherein said CEST is selected to target at least one of glucose, glycosaminoglycans, and lipid molecules.

24. The method according to claim 20, wherein said saturation transfer MRI comprises magnetization transfer imaging.

25. The method according to claim 24, wherein said magnetization transfer imaging comprises semisolid magnetization transfer imaging.

26. The method according to claim 20, wherein said saturation transfer MRI comprises saturation transfer difference imaging.

27. The method according to any of claims 1 -26, wherein a total duration of said acquisition of said first set of magnetic resonance images is less than one minute.

28. The method according to any of claims 1-26, wherein said organ is selected from the group consisting of a brain, a spinal cord, a heart, a head, a musculoskeletal organ, an abdomen organ, and a pelvis.

29. A computer software product, comprising a computer-readable medium in which program instructions are stored, which instructions, when read by a data processor, cause the data processor to receive at least a first set of magnetic resonance images, and a first set of radiofrequency excitation parameters, and to execute the method according to any of claims 1-28.

30. A magnetic resonance imaging (MRI) system for imaging an object, the system comprising: an MRI scanner configured for scanning the object to provide magnetic resonance images; and a data processor configured for executing the program instructions according to claim 29.

31. The method, computer software product or system according to any of claims 1-30, wherein at least one of said magnetic resonance images of said first set is a single-slice image.

32. The method, computer software product or system according to any of claims 1-31, wherein at least one of said magnetic resonance images of said first set is a multi-slice image.

Images