Method of classification of magnetic resonance signals, computer program for carrying out said method and system adapted to perform said method

Abstract

The present invention relates to a method of classification of magnetic resonance, MR, signals indicative of one or more portions of a subject or a sample (10) of diagnostic or prognostic interest, realized by means of an NMR scanner (9) and a processing means. The method comprises acquiring, with the NMR scanner (9), a set of MR signal data (2, 2') for each one or more MR signals emitted from the subject or the sample (10). The one or more acquired MR signal data sets (2, 2') consist of one-dimensional time-domain signal data and are processed by the processing means. The processing means classifies the one or more MR signals in one or more diagnostic or prognostic interest groups associated to reference data. The reference data comprise one-dimensional time-domain signal data. The present invention also relates to a system and a computer program adapted for carrying out said method.

Description

[0001] DESCRIPTION

[0002] METHOD OF CLASSIFICATION OF MAGNETIC RESONANCE SIGNALS, COMPUTER PROGRAM FOR CARRYING OUT SAID METHOD AND SYSTEM ADAPTED TO PERFORM SAID METHOD

[0003] FIELD OF THE INVENTION

[0004] The present invention relates to the field of magnetic resonance (MR) diagnostic techniques. More specifically, the invention relates to a method for classifying MR signals indicative of portions of diagnostic or prognostic interest acquired from a subject or a sample, and to a system and a computer program adapted for carrying out said method.

[0005] BACKGROUND OF THE INVENTION

[0006] Magnetic resonance (MR) is a non-invasive technique commonly used to obtain information from an object subjected to non-ionizing electromagnetic (EM) fields. Typically, in this technique, a nuclear MR (NMR) scanner is used to scan a subject or sample, by applying a main magnetic field and executing a sequence of EM fields and acquiring one or more MR signals emitted from said subject or said sample, which are preferably indicative of portions thereof of diagnostic or prognostic interest.

[0007] In the above context, common NMR scanners comprise a main magnet and a RF system, wherein the RF system comprises a number of RF transmitters and RF receivers. The main magnet applies the main magnetic field on the subject or the sample, polarising nuclear spins comprised in a portion of the subject or the sample. As a result, in that portion, a net magnetization of the polarised nuclear spins (that is, a vector sum of the angular momenta of nuclear spins) aligns with the main magnetic field. The RF transmitter is configured to excite the subject or the sample by executing at least one resonant RF field on the subject or the sample and, as a result, the net magnetization is rotated or tipped with respect to the direction of the main magnetic field. When the execution of a resonant RF field ceases, the net magnetization precesses around the main magnetic field, at a frequency known as the Larmor frequency. The precession of the net magnetization leads to the emission of an MR signal in the form of RF radiation, which can be detected by the RF receiver. In general, the RF receiver comprises a resonant circuit to detect the MR signal through an alternating current signal (generally known as free induction decay (FID) signal). Typically, the resonant frequency of the resonant circuit is the aforementioned Larmor frequency. In addition, common NMR scanners are also configured to record the MR signal detected by the RF receiver, so that at least part of the emitted MR signal is acquired.

[0008] Common NMR scanners further comprise a magnetic gradient system, configured to execute one or more inhomogeneous magnetic field gradients on the subject or the sample, superimposed to the main magnetic field. As a consequence, the resonant frequency (i.e. , the Larmor frequency) of the net magnetizations comprised in the subject or the sample depend on the position of said net magnetizations in the subject or the sample. That is to say, the execution of the one or more inhomogeneous magnetic fields, splits the subject or the sample in portions, where each portion comprises a net magnetization associated to a particular Larmor frequency. Only when the radiation frequency of an executed RF field is resonant with said particular Larmor frequency, or when at least said RF field has a radiation frequency bandwidth that includes said particular Larmor frequency, the executed RF field is resonant and rotates or tips the associated net magnetization.

[0009] MR imaging (MRI) methods are remarkably successful to obtain high-quality images of the internal spatial structure of a subject or a sample (typically, a human or animal body, or a part thereof). Said images are reconstructed from the MR signals acquired with NRM scanners. Said high-quality images enable the radiologists to easily and accurately identify the presence / absence of an actual or eventual pathology. In MRI methods, the acquired MR signal data are arranged as k-space data in a k-space, which are used to reconstruct images of the subject or the sample by means of mathematical tools typically involving Fourier transformations.

[0010] Despite their success, common MRI methods present several limitations. On the one hand, the acquired k-space data are constrained by the requirements associated to the reconstruction of a reliable image of the subject / sample. Said requirements typically imply that only little amounts of each emitted MR signals comprising information of the subject / sample are acquired for latter high-quality image reconstruction. Furthermore, large scan times are required to acquire the necessary amount of data and avoid image artifacts to reconstruct said high-quality images. In this context, said large scan times involve the execution of hundreds of EM sequence blocks comprising resonant RF pulses, and repetition times between consecutive EM sequence blocks of the order of 100 ms. Moreover, the quality of the images is directly dependent on the quantity and the quality of the acquired MR signal data arranged in a k-space, so that highly demanding hardware resources are generally required for the NMR scanners. As a result, NMR scanners are generally expensive.

[0011] On the other hand, nowadays the use of data-processing Artificial Intelligence (Al) algorithms has considerably increased in the field of MR methods. For instance, machine learning (ML) has been incorporated into MRI methods, such that the prior knowledge acquired from big datasets leads to a fast and accurate image reconstruction from limited measurements.

[0012] Furthermore, in recent years different Al-based methods and systems for classifying MR signals have been developed, wherein said methods are not based on an image reconstruction of a subject or a sample. These known methods are so-called “imageless methods”.

[0013] An example of an imageless medical imaging method is disclosed in the international patent application WO 2019 / 051411 A1. This patent application discloses a ML-based method for analysing medical image data obtained from a subject with an imaging system. Therein, examples of imaging systems are computed tomography (CT) systems, magnetic resonance (MR) systems, ultrasound (US) systems, ultrasonic CT systems, positron emission tomography (PET) systems, single photon emission computed tomography (SPECT) systems, or x-ray imaging systems. The image data are received by a module comprised in a computing device. This module is based on a trained ML algorithm configured to directly operate on the image data, so that said data are analysed without requiring an image reconstruction from thereof. In particular, the image data received by the ML module are raw data, such as raw k-space data.

[0014] Another example of an imageless method for classifying MR signals and a system implementing said method is disclosed in Singhal et al., "On the Feasibility of Machine Learning Augmented Magnetic Resonance for Point-of-Care Identification of Disease", arXiv preprint arXiv:2301.11962 (2023). An example embodiment therein comprises a prostate-cancer-triaging system used in patients with a high prostate-specific antigen (PSA) score. Said system comprises an NMR scanner, and embedded processing means configured with an ML algorithm. The prostate-cancer-triaging system does not have to produce an image, but just a risk score. In this context, the ML algorithm disclosed in Singhal et al infers the presence / absence of a disease directly from the k-space data obtained by a NMR scanner, skipping the image reconstruction process. Instead, the ML algorithm identifies the subsets of the k-space data that have the largest predictive signal pertaining to the disease being inferred and, subsequently, applies a Deep Leaning (DL) classifier to infer the presence of the disease from said k-space data subsets. To perform the classification task, the ML algorithm is previously trained with sets of k-space data of MR signals obtained from a plurality of subjects.

[0015] A problem present in the prior-art imageless methods for classifying MR signals, such as those disclosed in WO 2019 / 051411 A1 and Singhal et al, is that the processed MR signal data follow from the application of main magnetic fields, the execution of EM sequences and the acquisition of MR signal data as typically realized in MRI methods. Consequently, known imageless methods, as well as systems implementing thereof, suffer from the same limitations of MRI methods. Thus, known imageless methods are constrained by the above high-demanding requirements of MRI methods associated to the reconstruction of a reliable image from the acquired k-space data. In particular, prior-art methods suffer from the use of little amounts of k-space data for each MR signals comprising information of the subject / sample and the need of large scan times. In the end, said constraining requirements compromise the accuracy of prior-art imageless methods for classifying MR signals.

[0016] In view of the above, there is a need in the field of MR techniques for providing alternative imageless methods and systems for classifying MR signals.

[0017] BRIEF DESCRIPTION OF THE INVENTION

[0018] The present invention overcomes the above limitations of the prior art, by providing an improved imageless method for classifying MR signals, and a system adapted to perform said method. The method of the invention is based on the processing, by a processing means, of MR signal data sets acquired with a NMR scanner, wherein said acquired MR signal data sets comprise data that are not suitable for image reconstruction. The acquisition of data that are not suitable for image reconstruction is typically discarded in traditional MRI methods and in known imageless MR methods. However, these data can, in fact, be used to effectively classify MR signals indicative of portions of a subject or sample of diagnostic or prognostic interest, to increase the amount of data acquired per MR signal emitted from the subject or the sample, and to reduce scan times, compared to the known techniques.

[0019] Hence, the present invention relates in a first aspect to a method of classification of magnetic resonance, MR, signals indicative of one or more portions of a subject or a sample of diagnostic or prognostic interest, comprising the realization, by means of an NMR scanner and a processing means, of the following steps: a) applying a main magnetic field and executing a sequence of electromagnetic fields, on the subject or the sample arranged in the NMR scanner; b) acquiring with the NMR scanner, during the application of the main magnetic field and the execution of the EM sequence in step a), at least part of each of the one or more MR signals emitted from the subject or the sample that result from the application of the main magnetic field and the execution of the EM sequence, so that a set of MR signal data is acquired for said each of the one or more emitted MR signals; c) determining, with the processing means, one or more values of similarity between the one or more MR signal data sets of step b) and a plurality of diagnostic or prognostic interest groups associated with reference data, wherein said reference data are associated with one or more reference MR signals associated with one or more reference EM sequences for one or more reference subjects or samples, and wherein said reference data are segmented into the plurality of diagnostic or prognostic interest groups; d) classifying, with the processing means, the one or more MR signals emitted from the subject or the sample in one or more diagnostic or prognostic interest groups of the reference data from the values of similarity determined in step c).

[0020] Advantageously, in the present invention

[0021] - the one or more MR signal data sets consist of one-dimensional time-domain signal data; and

[0022] - the reference data associated with the one or more reference MR signals comprise one-dimensional time-domain signal data.

[0023] The one-dimensional (1 D) time-domain signal data of the one or more data sets and of the reference data are respectively associated with the EM sequence executed in step b) and the one or more reference EM sequences. Herein the EM sequence executed in step b) and each of the one or more reference EM sequences comprise, at least, one resonant RF field. Hence, 1 D time-domain signal data is understood as 1 D time-dependent signal data associated with an MR signal, wherein said data are not rearranged into a 2D or 3D k-space that could produce, by Fourier transformation, an image of the internal spatial structure of the subject or the sample. Accordingly, the 1 D time-domain signal data as defined in the present invention correspond to untransformed 1 D time-dependent MR signals, and not the image raw data in the arranged domain used in the prior art (referred as raw k-space data, for instance, in the patent application WO 2019 / 051411 A1). The acquisition of 1 D time-domain signal data is discarded both in MRI techniques and known imageless methods for classifying MR signals (such as the methods disclosed in the patent application WO 2019 / 051411 A1 , and in Singhal et al.). Furthermore, in known imageless MR methods, such as those based on Al algorithms proposed in WO 2019 / 051411 A1 and in Singhal et al., neither use reference data consisting of 1 D timedomain signal data for classifying MR signals. However, and even though 1 D time-domain signal data are not suitable for reconstructing an image of the subject or the sample, the 1 D time-domain signal data comprise information associated to said subject or said sample that is indeed relevant for the classification of the MR signals. Hence, the accuracy of the method according to present invention is enhanced with respect to known imageless methods for classifying MR signals.

[0024] In a second aspect, the invention relates to a computer program comprising instructions which, when the computer program is executed by a processing means, cause the processing means to carry out a method according to the first aspect of the invention.

[0025] Finally, in a third aspect the present invention relates to a system comprising an NMR scanner and a processing means communicatively coupled with the NMR scanner, wherein the NMR scanner and the processing means comprise hardware and / or software means adapted to perform a method according to the first aspect of the invention.

[0026] DESCRIPTION OF THE DRAWINGS

[0027] The foregoing and other features and advantages will be more fully understood from the detailed description of the invention, as well as from examples referring to the attached figures, which are described in the following paragraphs, wherein:

[0028] Figure 1 schematically represents an EM sequence according to a particular embodiment of the invention, where the MR signals are classified according to two diagnostic groups: Group I) “multiple sclerosis (MS) lesion” and Group II) “no MS lesion”.

[0029] Figure 2 represents the probability of MS lesion results (left-hand side) and the metrics of said results (right-hand side) according to an embodiment of the invention. Figure 3 shows an EM sequence according to another particular embodiment of the present invention.

[0030] Figure 4 shows an EM sequence according to a particular multifrequency embodiment of the present invention.

[0031] Figure 5 shows a schematic representation of an NMR scanner according to a particular multifrequency embodiment of the present invention. Figure 6 shows a flowchart schematically illustrating a particular embodiment of a method for classifying MR signals according to the present invention.

[0032] Reference numbers used in the figures

[0033] DETAILED DESCRIPTION OF THE INVENTION As explained above, the present invention overcomes the above limitations of the prior art, by providing an improved imageless method for classifying MR signals, and a system adapted to perform said method. The method of the invention is based on the processing, by a processing means, of MR signal data sets acquired with an NMR scanner and consisting of one-dimensional (1 D) time-domain signal data. In said processing, the processing means classifies the above MR signal data sets according to a plurality of diagnostic or prognostic interest groups associated with reference data. Said reference data comprise 1 D time-domain signal data. Herein, 1 D time-domain signal data is understood as 1 D time-dependent signal data associated to an MR signal, wherein said data are not rearranged into a 2D or 3D k-space that could produce, by Fourier transformation, an image of the internal spatial structure of the subject or the sample.

[0034] Hence, the present invention relates, in a first aspect thereof, to a method of classification of magnetic resonance, MR, signals indicative of one or more portions of a subject or a sample of diagnostic or prognostic interest, comprising the realization, by means of a nuclear magnetic resonance, NMR, scanner and a processing means, of the following steps: a) applying a main magnetic field and executing a sequence of electromagnetic, EM, fields, on the subject or the sample arranged in the NMR scanner; b) acquiring with the NMR scanner, during the application of the main magnetic field and the execution of the EM sequence in step a), at least part of each of the one or more MR signals emitted from the subject or the sample that result from the application of the main magnetic field and the execution of the EM sequence, so that a set of MR signal data is acquired for said each of the one or more emitted MR signals. The one or more acquired MR signal data sets consist of 1 D time-domain signal data; c) determining, with the processing means, one or more values of similarity between the one or more MR signal data sets of step b) and a plurality of diagnostic or prognostic interest groups associated with reference data, wherein said reference data are associated with one or more reference MR signals associated with one or more reference EM sequences for one or more reference subjects or samples, and wherein said reference data are segmented into the plurality of diagnostic or prognostic interest groups. In addition, the reference data associated with the one or more reference MR signals comprise 1 D timedomain signal data; d) classifying, with the processing means, the one or more MR signals emitted from the subject or the sample in one or more diagnostic or prognostic interest groups of the reference data from the values of similarity determined in step c). Herein, a sequences of EM fields (herein referred as “EM sequences”) executed by the NMR scanner comprises at least one or more resonant radiofrequency (RF) fields. The term “resonant RF field” refers to a RF field whose radiation frequency is substantially equal to the resonant frequency (i.e., the Larmor frequency) of a net magnetization comprised in the subject or the sample. Herein, the expression “substantially” means that said radiation frequency is equal to said Larmor frequency, or equal to a radiation frequency bandwidth that includes said Lamor frequency.

[0035] The one or more MR signal data sets acquired by the NMR scanner are processed by a processing means, in order to classify the one or more MR signal emitted from the subject or the sample. As used herein, the term “processing means” refers to processing devices, apparatus, programs, circuits, components, systems, and / or subsystems, whether implemented in hardware, tangibly embodied software, or both, and whether or not it is programmable. Hence, according to what is described above, the processing means comprises hardware and / or software means adapted to perform the aforementioned steps c) and d).

[0036] In step c) the processing means determines one or more values of similarity between the one or more acquired MR signal data sets and the diagnostic or prognostic interest groups. Said diagnostic or prognostic interest groups are associated with reference data associated with one or more reference MR signals. As used herein, the term “one or more reference MR signals” refers to one or more MR signals that exemplify and / or model the one or more MR signals acquired from one or more reference subjects or samples, and that result from the execution of one or more reference EM sequences. In addition, the reference data comprise medical information associated with the above one or more reference subjects or samples, so that said reference data are segmented into the plurality of diagnostic or prognostic interest groups. Further, in step d) the processing means classifies the one or more MR signals emitted from the subject or the sample in one or more of the above diagnostic or prognostic interest groups. Hence, the present invention provides an imageless method for classifying MR signals emitted from a subject or a sample. As used herein, the term “imageless” means that said method of classification is not based on an image reconstruction of the subject or the sample.

[0037] A particular classification goal is determined by the particular subject / sample to be scanned, the number of (classification) diagnostic / prognostic interest groups, and the (classification) diagnostic / prognostic interest groups as such. Given a particular classification goal, a person skilled in the art knows the several factors that determine the suitability of the specific executed EM sequence, dimensions of acquired MR signal data sets and of reference data, and configuration of the processing means. For example, the subject / sample where the EM sequence is executed may comprise different tissues and / or compositions, each one with relaxation times that are different form each other. To accurately classify the MR signals, the executed EM sequence and the dimensions of the acquired MR signal data are such that the acquired MR signals entail enough information, so that the processing means is able to distinguish between said tissues and / or compositions. On the other hand, when, for instance, the processing means requires certain spatial information to accurately perform a particular classification task, more than one gradient spoke must be executed in the EM sequence. Concerning the amount of acquired MR signal data, in a particular embodiment of the invention where no magnetic gradient fields are executed in the EM sequence, the time interval At between two adjacent acquired MR signal data is larger than 1% of the minimum among the decay times Ti , T2 among all the species (e.g., tissues) comprised in the one or more portions of the subject / sample. This is compatible with the common physical fact that no dynamical information is present much below (e.g., 100 times) the characteristic time scales of the dynamical system. In a particular embodiment of the invention where magnetic gradient fields are executed, the time interval At between two adjacent acquired MR signal data is larger than 1% of 1 / FoV, and the number of points in a given EM sequence block is similar to the number of voxels with which the spatial information is required. Herein, “FoV” is the spatial extent of the portions of the subject / sample. Notice that the above two conditions of the latter embodiment, in order to be fit inside the time duration (TRn) of an nthEM sequence block, impose constraints on the strength (gn) of the executed gradient field, such that the product yxgnxTRnis similar to 2xN / FoV. Herein, “y” is the is the gyromagnetic ratio (y=2.68xio8rad T'1s’1 , for protons), while “N” is the total number of EM sequence blocks in the executed EM sequence. The latter imposition is compatible with standard Nyquist requirements for image formation, but is not restricted by it, in the sense that signal oversampling (i.e. sampling 10, or 100 times faster than specified by Nyquist criteria) can help in extracting useful extra information from MR signals in ill-conditioned situations, as commonly known in the field of the invention.

[0038] Hence, given a particular classification goal, a person skilled in the art may derive, from the content of the present document and by using routine methods of experimentation or analysis, any suitable combination according to the invention of specific EM sequences, dimensions of acquired MR and of reference data, and processing means configurations. In any case, it should be clarified at this point that the present invention does not claim any particular or advantageous combination of EM sequence, dimensions of acquired MR signal data, dimensions of reference data, and processing means configuration. Instead, the present invention claims an imageless MR method for classifying MR signal that is based on 1 D time-domain signal data. Indeed, contrary to known MRI and imageless MR methods, herein the acquired MR signal data sets consist of 1 D time-domain signal data, and the reference data associated with the classification groups comprise 1 D time-domain signal data. The 1 D time-domain signal data are not transformed into image raw data. Thus, neither the acquired MR signal data sets nor the reference data are subjected to the high- demanding requirements (e.g., large scan times, etc.) of known MRI and imageless MR methods. Consequently, both acquired MR signal data sets and reference data can comprise data that although are typically discarded in said know methods, they comprise information associated to the scanned subject / sample that is indeed relevant for the classification of the MR signals.

[0039] In particular embodiments of the invention, the one or more MR signal data sets and / or the reference data are non-image-reconstructing. As used herein, the term “non-image- reconstructing” means that from none group of the data comprised in the one or more MR signal data sets / reference data, is possible to reconstruct a reliable MRI image reproducing the internal spatial structure of a subject / sample. Hence, even after arranging said group of data into a 2D or 3D k-space and afterwards applying Fourier transformations, such reliable MRI image cannot be produced. In this regard, when at least part of the one or more MR signal data sets / the reference data comprises non-image-reconstructing 1 D time-domain signal data, the whole MR signal data sets / reference data are non-image reconstructing.

[0040] In particular embodiments of the invention, the method further comprises, before step c), the realization of the following step: e) comparing, with the processing means, the one or more MR signal data sets of step b) with the reference data.

[0041] In particular embodiments of the invention, the one or more MR signal data sets comprise one-dimensional time-domain signal data, and the reference data associated with the one or more reference MR signals comprise one-dimensional time-domain signal data.

[0042] In further embodiments the invention provides a method of classification of magnetic resonance, MR, signals indicative of one or more portions of a subject or a sample of diagnostic or prognostic interest, comprising the realization, by means of an NMR scanner and a processing means, of the following steps: a) applying a main magnetic field and executing a sequence of electromagnetic (EM) fields, on the subject or the sample arranged in the NMR scanner; b) acquiring with the NMR scanner, during the application of the main magnetic field and the execution of the EM sequence in step a), at least part of each of the one or more MR signals emitted from the subject or the sample that result from the application of the main magnetic field and the execution of the EM sequence, so that a set of MR signal data is acquired for said each of the one or more emitted MR signals; e) comparing, with the processing means, the one or more MR signal data sets of step b) with reference data, wherein said reference data are associated with one or more reference MR signals associated with one or more reference EM sequences for one or more reference subjects or samples, and wherein said reference data is segmented into a plurality of diagnostic or prognostic interest groups; c) determining, with the processing means, one or more values of similarity between the diagnostic or prognostic interest groups and the one or more MR signal data sets; d) classifying, with the processing means, the one or more MR signals emitted from the subject or the sample in one or more diagnostic or prognostic interest groups of the reference data from the values of similarity determined in step e); wherein

[0043] - the one or more MR signal data sets comprise one-dimensional time-domain signal data; and

[0044] - the reference data associated with the one or more reference MR signals comprise one-dimensional time-domain signal data.

[0045] In particular embodiments of the invention, the reference data comprise patterns and / or relationships comprising medical information, one or more threshold data values, and / or one or more critical data points (e.g., one or more relative and / or absolute maxima / minima), indicating the presence or absence, in the one or more reference subjects or samples, of one or more actual or potential future illnesses, symptoms, hurts, and / or physical features (e.g., amount of a specific fluid, tumor, and / or inflammation in the reference subjects or samples).

[0046] In particular embodiments of the invention, the reference data are correlated to one or more biomarkers indicating the presence or absence, in the one or more reference subjects or samples, of one or more actual or potential future illnesses, symptoms, hurts, and / or physical features. As used herein, the term “biomarker” refers to a measurable indicator of some biological state or condition of the reference subjects or samples.

[0047] In particular embodiments of the invention, the reference data comprise

[0048] - one or more MR signal data sets acquired with an NMR scanner. In a particular embodiment of the invention, said one or more MR signal data sets acquired with NMR scanner comprised in the reference data, are acquired following steps a) and b) of the present method, according to any of the realizations of said steps a) and b) disclosed in this document;

[0049] - one or more simulated MR signal data sets; and / or

[0050] - a plurality of labels, each label corresponding to a group of the plurality of diagnostic or prognostic interest groups in which the reference data are segmented. In particular embodiments, where the processing means are based on machine learning (ML) networks, the labels are the known outcomes that the ML model learns to associate with the input data during training.

[0051] In particular embodiment of the invention, the plurality of diagnostic or prognostic interest groups includes at least one of the following: “MS lesion” and “no MS lesion”; “hydrocephalus lesion” and “no hydrocephalus lesion”; “tumor” and “no tumor”; “hemorrhage” and “no hemorrhage”; “endometriosis” and “no endometriosis”; “synovial fluid” and “no synovial fluid”; “ischemic stroke”, “hemorrhagic stroke” and “no stroke”; “abnormal tissue buildup” and “no abnormal tissue buildup”.

[0052] In particular embodiments of the invention, step c), d) and / or e) is performed with an algorithm configured to recognised data patterns in MR signals data. In particular embodiments of the invention, said algorithm comprises the use of one or more predictive models, so that, based on data patterns recognition, the algorithm outputs one or more predictive outcomes from the one or more MR signal data sets inputted thereof. Furthermore, in particular non-limiting embodiments of the invention,

[0053] - the algorithm comprises the use of one or more predictive models including at least one of the following: one or more regression models, time series models, clustering models, decision trees, ensemble models, artificial intelligence, Al, or machine learning, ML, models; and / or

[0054] - steps c) and d) are performed with the algorithm, wherein the algorithm comprises the use of an ML network trained with training data. Such ML network is an Al or ML model based on a collection of connected units or nodes called artificial neurons, wherein said Al or ML model learns from training data to recognise patterns and / or relationships comprising relevant information.

[0055] In particular embodiments of the invention, the training data of the trained ML network comprises data associated with the above one or more reference MR signals. Consequently, in these particular embodiments the reference data comprise patterns and / or relationships learnt by the trained ML network from said data associated with the one or more reference MR signals. The learnt patterns and / or relationships comprises medical information associated with the above one or more reference subjects or samples.

[0056] Artificial neurons are commonly comprised in a plurality of layers. In particular embodiments of the invention, the trained ML network comprises an input layer, one or more intermediate hidden layers, and / or one output layer. In particular embodiments of the invention, the one or more MR signal data sets are inputted to an input layer comprised in the trained ML network. In particular embodiments of the invention, the one or more values of similarity determined in step c) comprise one or more predictive outcomes comprised in an output layer comprised in the trained ML network. In particular embodiments of the invention, said one or more predictive outcomes weight the probability that the one or more emitted MR signals are associated with one or more of the above diagnostic or prognostic interest groups. In particular embodiments of the invention, the classification of the one or more MR signals emitted from the subject or the sample is derived from one or more predictive outcomes comprised in an output layer of the trained ML network.

[0057] According to the present invention, a trained ML network includes, but is not limited to, one or more multi-layer perceptron (MLPs), convolutional neural networks (CNNs), recurrent neural network (RNNs), and combinations of any of the foregoing.

[0058] In particular embodiments of the invention, the classification performed by the processing means in step d) comprises

[0059] - one or more binomial parameters that qualitatively indicate that the one or more MR signals emitted from the subject or the sample are associated to one or more of the above diagnostic or prognostic interest groups; and / or

[0060] - one or more scores that quantitatively indicate the degree of probability that the one or more MR signals emitted from the subject or the sample are associated to one or more of the above diagnostic or prognostic interest groups. In particular embodiments of the invention, the EM sequence executed in step a) is selected according to a symptoms or condition database, wherein a specific EM sequence or combination of EM sequences are executed in step a) depending on a prescribed symptom or condition input information. In particular embodiments of the invention, said input information is inputted to the NMR scanner, or to the processing means or another means communicatively coupled with the NMR scanner.

[0061] In particular embodiments of the invention, the EM sequence executed in step a) is designed by an algorithm. This algorithm is configured to design said EM sequence according to the specific hardware resources of the NMR scanner used for realising the method, to the specific subject or specific sample arranged thereof, and / or to a specific a prescribed symptom or condition input information. Accordingly said designed EM sequence is optimal for the MR signals classification, given the above specific hardware resources of the NMR scanner, the specific subject or specific the sample, and / or specific a prescribed symptom or condition input information. In particular embodiments, said algorithm is an optimization algorithm that includes a gradient based minimization algorithm, a population-based optimization algorithm (e.g., a genetic algorithm, a differential evolution algorithm, swarm, annealing, or any other evolutionary algorithm), or Al based algorithm.

[0062] In particular embodiments of the invention, the one or more portions of the subject or the sample of diagnostic or prognostic interest are indicative of one or more clinical features, said clinical features comprising at least one of the following: amount of MS lesion, amount of hydrocephalus liquid, amount of tumor, amount of hemorrhage, amount of endometriosis, amount of bodily fluid leak (e.g., synovial fluid), amount of abnormal tissue buildup; and wherein the method further comprise performing the following step: f) determining, with the processing means, the one or more clinical features, from the one or more MR signal data sets of step b) and based on information associated with the reference data. In particular embodiments of the invention, said processing means is configured with a regression algorithm for determining said clinical features.

[0063] In particular embodiments of the invention, the EM sequence executed in step a) comprises a single EM sequence block comprising a single resonant RF field, and a set of MR signal data is acquired from a single MR signal emitted from the subject or the sample in step b). In yet other particular embodiments of the invention, the above single EM sequence block further comprises one or more inhomogeneous magnetic field gradients. In the context of the present invention, the term “EM sequence block” refers to a part of the EM sequence comprising one or more resonant RF fields with same radiation frequency. Thus, in these particular embodiments of the invention, it is sufficient for the processing means to process the MR signal data set of a single emitted MR signal for classifying thereof. Advantageously, the scan times are considerably reduced with respect to MRI methods and known imageless MR methods, wherein typically hundreds of EM sequence blocks are executed on the subject or the sample, and the subsequent MR signal data sets are acquired. In particular embodiments of the invention, the above single resonant RF field executed in step a) is a continuous and / or time-modulated RF field executed during all step a), contrary to RF pulses commonly executed in known MR techniques.

[0064] In particular embodiments of the present invention, the EM sequence executed in step a) comprises a plurality of EM sequence blocks, wherein each EM sequence block comprises one or more resonant RF fields, so that the subject or the sample is repeatedly re-excited during step a). This allows to eventually avoid the inherent limitations (e.g., noise effects) of the short decaying time T2* characterizing the MR signals (e.g., the decaying time T2* is around 1 ms or less in hard tissues such as bones, teeth, etc.).

[0065] An inhomogeneous main magnetic field may be applied with an NMR scanner on a subject or a sample, so that the value of the resonant frequency (i.e., the Larmor frequency) of a net magnetization comprised in said subject or said sample, depends on the position of said net magnetization in the subject or the sample. In this scenario, an EM sequence comprising a plurality of EM sequence blocks may be executed with the above NMR scanner on the subject or the sample, so that said subject or said sample is repeatedly re-excited during said execution. Each EM sequence block comprises one or more resonant RF fields that only excite a portion of the subject or the sample. This means that the radiation frequency of said one or more resonant RF fields is resonant with the Larmor frequency of the net magnetization comprised in the above portion of the subject or the sample. On the contrary, the remaining portions of the of the subject or the sample are not excited by the one or more resonant RF fields comprised in the above EM sequence block, or they return to a relaxed state if resonant RF fields comprised in previously executed EM sequence blocks have already excited them. The situation described in the present paragraph is herein referred to as “multifrequency” or “high bandwidth” implementation. Multifrequency implementations can be realized in imageless methods for classifying MR signals, whether or not the acquired MR signal data sets consist of 1 D time-domain signal data and the reference data comprise 1 D time-domain signal data. Hence, according to the above, in particular embodiments of the invention (hereinafter referred as “multifrequency embodiments” or “high bandwidth embodiments”), aside from executing in step a) an EM sequence comprising a plurality of EM sequence blocks, where each EM sequence block comprises one or more resonant RF fields, so that the subject or the sample is repeatedly re-excited during step a):

[0066] - the main magnetic field applied in step a) is a non-uniform main magnetic field, so that a position-dependent Larmor frequency is imposed along the spatial extent of the subject or the sample along at least one direction in said subject or said sample, wherein the values of said position-dependent Larmor frequency along the spatial extent of the subject or the sample along said at least one direction are delimited by a bandwidth range (AWL) given by

[0067] (Y B0) / (2TT), wherein ABo is the difference (ABo) between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample, and y is the protonic gyromagnetic ratio;

[0068] - the EM sequence is divided in a plurality of sets of EM sequence blocks;

[0069] - the one or more RF fields of each of the EM sequence blocks of a set EM sequence blocks has a radiation frequency, wherein said radiation frequency is the same for EM sequence blocks of the same set of EM sequence blocks;

[0070] - the radiation frequency of a set of EM sequence blocks is different for different sets of EM sequence blocks and is resonant with a single value of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample.

[0071] In particular multifrequency embodiments, the difference (ABo) between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample, is equal or larger than 0.5 times the maximum amplitude (Bo, max) of the non-uniform main magnetic field.

[0072] In addition, in particular multifrequency embodiments, the ratio between the bandwidth range (AWL) delimiting the values of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample, and the maximum value of the position-dependent Larmor frequency in said spatial extent of the subject or the sample along the at least one direction in said subject or said sample is larger than 0.1 , larger than 0.2, larger than 0.3, larger than 0.4, or larger than 0.5; and / or equal to 0.1 , 0.2, 0.3, 0.4, or 0.5.

[0073] In particular multifrequency embodiments, the bandwidth range ( wL) delimiting the values of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample, is equal or larger than 100 kHz. In particular multifrequency embodiments, the bandwidth range delimiting the values of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample is between 100 kHz and 1000 kHz, between 150 kHz and 1000 kHz, between 200 kHz and 1000 kHz, between 250 kHz and 1000 kHz, between 300 kHz and 1000 kHz, between 350 kHz and 1000 kHz, between 400 kHz and 1000 kHz, between 450 kHz and 1000 kHz, between 500 kHz and 1000 kHz, between 550 kHz and 1000 kHz, between 600 kHz and 1000 kHz, between 650 kHz and 1000 kHz, between 700 kHz and 1000 kHz, between 750 kHz and 1000 kHz, between 800 kHz and 1000 kHz, between 850 kHz and 1000 kHz, between 900 kHz and 1000 kHz, or between 950 kHz and 1000 kHz; and / or is equal to 100 kHz, 150 kHz, 200 kHz, 250 kHz, 300 kHz, 350 kHz, 400 kHz, 450 kHz, 500 kHz, 550 kHz, 600 kHz, 650 kHz, 700 kHz, 750 kHz, 800 kHz, 850 kHz, 900 kHz, 950 kHz, or 1000 kHz.

[0074] In particular embodiments of the invention, step b) comprises the record of the at least part of each of the one or more emitted MR signals detected by the NMR scanner. In particular embodiments of the invention, said NMR scanner detects the one or more emitted MR signals by means of a resonant circuit (for instance, a resonant circuit comprised in a RF receiver arranged in the NMR scanner). Typically, a resonant circuit is characterised by a quality factor or Q factor, given by the quotient of the resonant frequency of said resonant circuit and the frequency bandwidth of said resonant circuit. The Q factor measures the signal-to-noise-ratio (SNR) detected by the resonant circuit: the larger the Q factor, the larger the SNR of the detected emitted MR signal.

[0075] In the above context, in particular multifrequency embodiments,

[0076] - step b) comprises the record of the at least part of each of the one or more emitted MR signals detected by the NMR scanner by means of resonant circuit; - the resonant circuit has a retunable resonant frequency, so that, for the detection of the emitted MR signal resulting from the execution of each set of EM sequence blocks, the resonant frequency of said resonant circuit coincides with the radiation frequency of the one or more resonant RF fields comprised in said set of EM sequence blocks; and

[0077] - the resonant circuit has a frequency bandwidth smaller than the bandwidth range delimiting the values of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample. In particular embodiments of the invention, the frequency bandwidth of the resonant circuit is at least two times smaller than said bandwidth range.

[0078] According to the above, by sweeping through frequencies comprised in the bandwidth range delimiting the above values of the position-dependent Larmor frequency, the NMR scanner is configured to successively excite the subject or the sample and acquire MR signal data from different portions of said subject or said sample. At the same time, the Q factor of the resonant circuit detecting the emitted MR signals is relatively large, as the frequency bandwidth of said resonant circuit is smaller than the bandwidth range delimiting the above values of the position-dependent Larmor frequency. Consequently, in these multifrequency embodiments, the SNR of the detected MR signals is relatively large. Moreover, the above Q factor (and therefore the above SNR), is particularly enhanced in multifrequency embodiments wherein the difference (ABo) between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample, is equal or larger than 0.5 times the maximum amplitude (Bo, max) of the non-uniform main magnetic field.

[0079] In addition, in particular multifrequency embodiments, wherein the NMR scanner is configured to detect the one or more emitted MR signals by means of a resonant circuit with a retunable resonant frequency, said NMR scanner is configured to steppedly execute, on the subject or the sample, the sets of EM sequence blocks comprised in the EM sequence of step a), so that all radiation frequencies of the one or more resonant RF fields comprised in the sets of EM sequence blocks, cover the whole bandwidth range delimiting the values of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample.

[0080] In further embodiments of the invention, wherein the EM sequence executed in step a) comprises a plurality of EM sequence blocks: - the total number of EM sequence blocks comprised in the EM sequence is less than 100;

[0081] - in at least one pair of two consecutive EM sequence blocks comprised in the EM sequence, the repetition time between each of the first resonant RF fields of said at least one pair of two consecutive EM sequence blocks is less than 100 ms;

[0082] - the EM sequence blocks comprised in the EM sequence are different in strengths, directions, flip angles and / or repetition times; and / or

[0083] - the EM sequence blocks comprised in the EM sequence are randomly varied during step a). In particular embodiments of the invention, the EM sequence blocks comprised in the EM sequence are randomly varied by randomly varying the strengths, directions, flip angles and / or repetition times.

[0084] In particular embodiments of the invention, the total number of EM sequence blocks comprised in an EM sequence comprising a plurality of EM sequence blocks is less than 90, less than 80, less than 70, less than 60, less than 50, less than 40, less than 30, less than 25, less than 20, less than 15, less than 10, less than 5, or less than 3; is more than 2, more than 5, more than 10, more than 25, more than 20, more than 25, more than 30, more than 40, more than 50, more than 60, more than 70, more than 80, or more than 90; and / or is equal to 90, 80, 70, 60, 50, 40, 30, 25, 20, 15, 10, 9, 8, 7, 6, 5, 4, 3, or 2.

[0085] In the embodiments of the invention wherein the total number of EM sequence blocks comprised in the EM sequence is less than 100, the scan times are reduced and the efficiency of the method for classifying MR signals is enhanced, with respect to MRI methods and known imageless methods for classifying MR signals. Said known MR methods typically comprise the execution of hundreds of different EM sequence blocks on a subject or the sample. In addition, said hundreds of different EM sequence blocks of known MR methods, also typically comprised a plurality of inhomogeneous magnetic field gradients, so that the acquired MR signal data sets are associated to k-space data sets of a k-space. These k-space data sets fill said k-space in different directions. In this sense, the execution of hundreds of different EM sequence blocks on a subject or a sample is particularly relevant in MRI methods to reconstruct a reliable image thereof. On the other hand, in particular embodiments of the invention, the repetition time between each of the first resonant RF fields of at least one pair of two consecutive EM sequence blocks is less than 200 ms, less than 180 ms, less than 160 ms, less than 140 ms, less than 120 ms, less than 100 ms, 90 ms, less than 80 ms, less than 70 ms, less than 60 ms, less than 50 ms, less than 40 ms, less than 30 ms, less than 25 ms, less than 20 ms, less than 15 ms, less than 10 ms, less than 5 ms, or less than 1 ms; and / or is equal to 200 ms, 180 ms, 160 ms, 140 ms, 120 ms, 100 ms, 90 ms, 80 ms, 70 ms, 60 ms, 50 ms, 40 ms, 30 ms, 25 ms, 20 ms, 15 ms, 10 ms, 5 ms, or 1 ms.

[0086] In particular embodiments of the invention, the value of any repetition time is larger than 1 % of the minimum among the decay times Ti, T2 among all the species (e.g., tissues) comprised in the one or more portions of the subject / sample.

[0087] In the embodiments of the invention where the repetition time between each of the first resonant RF fields of at least one pair of two consecutive EM sequence blocks is less than 100 ms, scan times are also reduced compared to MRI methods and known imageless methods. Indeed, in MRI methods repetition times of the order of 100 ms or larger are required, in order to prevent artifacts in the reconstructed images. This is because said order of repetition times in MRI methods is sufficient to relax all the excited net magnetizations comprised in the excited portions of the subject / sample, as the excitation only takes place for each EM sequence block within a relatively narrow bandwidth range in said MRI methods (equal or smaller than 10 kHz).

[0088] As mentioned above, in particular embodiments of the invention, the EM sequence blocks comprised in the EM sequence are different in strengths, directions, flip angles and / or repetition times. Advantageously, the MR signal data sets acquired during the execution of the EM sequence look different from each other. These MR signal data sets that look different from each other may provide information about the different resonant species comprised in the subject or the sample (e.g., muscle, bone, blood, sclerosis lesions, etc). As used herein, the term “strength” refers to the maximum magnitude value of at least one inhomogeneous magnetic field gradient comprised in each of the EM sequence blocks. As used herein, the term “direction” refers to the direction with respect to a predefined reference frame of the subject or the sample of at least one inhomogeneous magnetic field gradient comprised in each of the EM sequence blocks. As used herein, the term “flip angle” refers to the angle to which a net magnetization is rotated or tipped via the execution of a resonant RF field comprised in an EM sequence block, relative to the direction of the main magnetic field applied in step a).

[0089] In particular embodiments of the present invention, step b) is performed

[0090] - at full or quasi-full duty cycle; and / or

[0091] - during the execution of at least one resonant RF field comprised in the EM sequence executed in step a).

[0092] As used herein, the term “full duty cycle” refers to acquiring the at least part of each of the one or more emitted MR signals during a percentage of 100% of the time duration of the execution of the EM sequence executed in step a). On the other hand, as used herein, the term “quasi-full duty cycle” refers to acquiring the at least part of each of the one or more emitted MR signals during a percentage of less than 100% but at least 80% of the time duration of the execution of the EM sequence executed in step a).

[0093] When step b) is performed at full or quasi-full duty cycle, it is possible to measure the time evolution of the one or more MR signals during at least 80% of the time duration of the execution of the EM sequence executed in step a). In said embodiments of the invention, the NMR scanner operates at a duty cycle (between 100% and 80% duty cycle) considerably larger than in MRI methods and known imageless methods for classifying MR signals. Indeed, in said known MR methods the NMR scanners seldom operate at 50% duty cycle (even worse, in some cases they operate at duty cycles as low as 0.1 %). Advantageously, when step b) is performed at full or quasi-full duty cycle, the amount of acquired MR signal data is enhanced, such that the SNR of the acquired MR signal data sets is increased with respect to above prior-art MR methods. Accordingly, the accuracy of the classification method according to said embodiments of the invention is enhanced, with respect to known imageless methods for classifying MR signals.

[0094] In addition, in particular embodiments of the invention:

[0095] - the EM sequences executed in step a) comprises one or more inhomogeneous magnetic field gradients, so that a position-dependent Larmor frequency is imposed along at least one direction in the subject or the sample. In particular embodiments of the invention, said one or more inhomogeneous magnetic field gradients comprise one or more inhomogeneous magnetic field gradient pulses;

[0096] - the EM sequence executed in step a) comprises one or more resonant RF fields comprising one or more RF pulses; - the main magnetic field applied in step a) is a time-modulated main magnetic field. In particular embodiments of the invention, the time-modulated main magnetic field is a cycling main magnetic field, such as, in a non-limiting example, a pre-polarization main magnetic field; and / or

[0097] - step b) is performed while the subject or the sample moves relative to the NMR scanner, or to a part thereof.

[0098] In a second aspect the present invention relates to a computer program comprising instructions which, when the computer program is executed by a processing means, cause the processing means to carry out a method according to the first aspect of the invention.

[0099] Finally, a third aspect the present invention relates to a system comprising an NMR scanner and a processing means communicatively coupled with the NMR scanner, wherein the NMR scanner and the processing means comprise hardware and / or software means adapted to perform a method according to the first aspect of the invention.

[0100] In particular embodiments of the invention, the processing means

[0101] - is communicatively coupled with the NMR scanner in wirelessly way and / or by means of one or more wires; and / or

[0102] - is integrated in the NMR scanner.

[0103] In particular embodiments of the invention, the NMR scanner comprises at least a main magnet, a RF transmitter, a RF receiver, and a data acquisition unit, wherein

[0104] - the main magnet is configured to apply a main magnetic field for polarising nuclear spins comprised in a subject or a sample, when said subject or said sample is arranged in the NMR scanner;

[0105] - the RF transmitter is configured to execute one or more resonant RF fields on said subject or said sample;

[0106] - the RF receiver is arranged to detect one or more MR signals emitted from said subject or said sample; and

[0107] - the data acquisition unit is configured to acquire at least part of each of the one or more MR signals detected by the RF receiver, by recording a set of one or more MR signal data for each of the one or more MR signals detected by the RF receiver.

[0108] In particular embodiments of the invention, the RF receiver comprises a resonant circuit. In these embodiments, the RF receiver detects MR signals through alternating currents induced by said MR signals in the resonant circuit. In particular embodiments of the invention, the RF receiver comprises a resonant circuit comprising at least one coil. In particular embodiments of the invention, the RF transmitter and the RF receiver are integrated in a same physical element (e.g., one or more coils).

[0109] Moreover, in particular embodiments of the invention the NMR scanner is adapted to realize the aforementioned steps a) and b) of a method according to any of the multifrequency embodiments described above. Thus, in particular embodiments of the invention, the NMR scanner comprises the above main magnet, RF transmitter and RF scanner, wherein

[0110] - the main field is a non-uniform main magnetic field that imposes a positiondependent Larmor frequency along the spatial extent of the subject or the sample along at least one direction in said subject or said sample, wherein the values of said positiondependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample are delimited by a bandwidth range (AWL) given by

[0111] (Y B0) / (2TT), wherein ABo is the difference between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample, and y is the protonic gyromagnetic ratio;

[0112] - the RF transmitter is configured to execute, on the subject or the sample and during the application of the non-uniform main magnetic field, a plurality of resonant RF fields comprised in EM sequence blocks, wherein each EM sequence block is comprised in a set of EM sequence blocks and each EM sequence block comprises one or more RF fields with a radiation frequency, wherein the radiation frequency is the same for EM sequence blocks of the same set of EM sequence blocks, and wherein the radiation frequency of a set of EM sequence blocks is different for different sets of EM sequence blocks and resonant with a single value of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample;

[0113] - the RF receiver comprises a resonant circuit;

[0114] - said resonant circuit has a retunable resonant frequency, so that, for the detection of the emitted MR signals resulting from the execution of each set of EM sequence blocks executed by the RF transmitter, the resonant circuit is configured to change the resonant frequency to the radiation frequency of the one or more resonant RF fields comprised in said set of EM sequence blocks executed by the RF transmitter; and - said resonant circuit has a frequency bandwidth smaller than the bandwidth range delimiting the values of the position-dependent Larmor frequency along the spatial extent of the subject or the sample along the at least one direction in said subject or said sample.

[0115] Since the above non-uniform main magnetic field already spatially encodes the subject or the sample, in the multifrequency embodiments it is not necessarily required a magnetic gradient system configured to execute one or more inhomogeneous magnetic field gradients for said spatial encoding of the subject or the sample. Hence, in particular multifrequency embodiments, the NMR scanner does not comprise said magnetic gradient system.

[0116] In particular multifrequency embodiments, the above RF transmitter is configured to steppedly execute the aforementioned sets of EM sequence blocks, so that all radiation frequencies of the one or more resonant RF fields comprised in said sets cover the whole the bandwidth range delimiting the values of the above position-dependent Larmor frequency imposed by the main magnetic field. On the order hand, in particular multifrequency embodiments, the main magnet applying the non-uniform main magnetic field is a single-sided magnet. In further embodiments of the invention, the retunable resonant frequency of the resonant circuit is retunable by means one or more capacitors comprised in said resonant circuit, by changing the capacitance of each of the one or more capacitators. In non-limiting examples, the capacitance of a capacitor can be changed by switching said capacitor, and / or by adjusting voltages on, for example, varactor diodes.

[0117] Furthermore, in particular embodiments of the invention

[0118] - the NMR scanner comprises a magnetic gradient system configured to execute one or more inhomogeneous magnetic field gradients on a subject or a sample, when said subject or said sample is arranged in said NMR scanner, so that a position-dependent Larmor frequency is imposed along at least one direction in the subject or the sample. In particular embodiments of the invention, the magnetic gradient system comprises one or more coils;

[0119] - the NMR scanner is a resource-relaxed NMR scanner; and / or

[0120] - the NMR scanner is a non-image-reconstructing NMR scanner.

[0121] In the context of present invention, the term “resource-relaxed NMR scanner” refers to an NMR scanner with relaxed hardware resources, compared to common MRI scanners. Typically, common MRI scanners are configured to apply a highly homogeneous main magnetic fields and execute highly linear inhomogeneous magnetic field gradients on a subject or a sample, in order to reconstruct, from the subsequently acquired MR signal data sets, a high-quality image. Concretely, a resource-relaxed NMR scanner comprises:

[0122] - a main magnet configured to apply, on a subject or a sample, a main magnetic field with an amplitude (Bo) on the subject or the sample smaller than 1 T and / or with a difference (ABo) between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) equal or larger than 0.5 times the maximum amplitude (Bo, max), when said subject or said sample is arranged in the NMR scanner; and / or

[0123] - a magnetic gradient system configured to execute, on a subject or a sample, at least one non-linear inhomogeneous magnetic field gradient, when said subject or said sample is arranged in the NMR scanner. As used herein, the term “non-linear inhomogeneous magnetic field gradient” means that, in the point of maximum deviation in the field of view of the subject or the sample, the magnetic field amplitude of said at least one inhomogeneous magnetic field gradient deviates a 20% or more, with respect to the maximum magnetic field amplitude of said at least one inhomogeneous magnetic field gradient. The term “field of view” or “FoV” refers to one or more portions of the subject or the sample from which the MR signal data are acquired.

[0124] In particular embodiments, the main magnet of the resource-relaxed NMR scanner is configured to apply a main magnetic field with an amplitude (Bo) on the subject or the sample smaller than 0.1 T.

[0125] The above undemanding hardware resources of resource-relaxed NMR scanners imply a considerable reduction of costs and an availability increase, compared to common MRI scanners. Advantageously, a system according to the invention comprising a resource- relaxed NMR scanner, may be sited in small clinics, outpatient departments, rural regions, low-income areas, etc.

[0126] On the other hand, as used herein, the term “non-image-reconstructing NMR scanner” refers to an NMR scanner that does not produce an image of a subject or a sample arranged thereof.

[0127] In particular embodiments of the invention, the processing means are communicatively coupled with the aforementioned data acquisition unit, so that the data acquisition unit provides to the processing means the one or more MR signal data sets acquired by the data acquisition unit. In particular embodiments of the invention, the system comprises a display that outputs the classification of the one or more emitted MR signals performed by the processing means.

[0128] All the terms and embodiments described anywhere in this document are equally applicable to all aspects of the invention. It should be noted that, as used in the specification and in the appended claims, the singular forms “a”, “an”, and “the” include their plural referents unless the context clearly indicates otherwise. Similarly, the term “comprises” or “comprising”, referred to any element of the present description, is understood as “one or more” of such elements (i.e., at least such element), and also includes the exclusive presence of such element (i.e., consisting of such element).

[0129] EXAMPLES

[0130] The following invention is hereby described by way of the following examples, which are to be construed as merely illustrative and not limitative of the scope of the invention.

[0131] Example 1 : classification of MR signals associated to multiple sclerosis

[0132] Herein it is described a non-limiting example of an imageless MR method for classification of MR signals according to the invention. In particular, the MR signals are acquired from a 2D human brain slice, and the goal is classifying said MR signal according to two diagnostic groups: Group I) “multiple sclerosis (MS) lesion”, and Group II) “no MS lesion”. To this end, the 2D brain human slice is subjected, by means of a NMR scanner, to a main magnetic field and to an EM sequence. At the same time, at least part of each of the MR signals emitted from the 2D human brain slice is acquired with the NMR scanner. The executed EM sequence and the dimensions of the MR signal data sets acquired from the 2D brain human slice are suitable for classification of said MR signals according to the above Groups l-ll. In this regard, the set of acquired MR signal data for each emitted MR signal is processed by a processing means so that said classicisation is performed. A detailed description of this particular embodiment is provided in the following. Said description is divided in two parts. The first part, A, is devoted to the description of the particular EM sequence and acquisition of the MR signal data sets. The second part, B, is devoted to the description of the particular processing means configuration and of the classification results.

[0133] A.- Executed EM sequence and acquisition of MR signal data Figure 1 schematically represents the executed EM sequence (1) and acquired MR signal data sets (2, 2’) with the NMR scanner in the present example. Figure 1 only shows the initial part of the EM sequence (1), while the remaining part of the EM sequence (1) is represented by dots in the right-hand side of the figure. A horizontal arrow indicates the time (t) flow of the EM sequence (1). The MR signals result from the execution of the EM sequence (1), and the application, with the NMR scanner, of a main magnetic field (not shown in Figure 1) that polarises nuclear spins comprised in the subject / sample, i.e., 2D human brain slice. The MR signal data sets (2, 2’) advantageously consist of 1 D timedomain signal data which are not transformed into image raw data of the arranged domain. In particular, the MR signal data sets (2, 2’) are voltage values recorded by a data acquisition unit comprised in the NMR scanner. These voltage values are associated to the alternating currents induced by the MR signals in a resonant circuit comprised in the NMR scanner. According to the above, the executed EM sequence (1) and the dimensions of acquired MR signal data sets (2, 2’) are suitable for the present classification task, so that the acquired MR signals entail enough information to accurately classified thereof. At this point, it should be clarified that, rather than claiming specific EM sequences or dimensions of acquired MR signal data sets, the present invention defines that the suitability of EM sequences (1) and of the acquisition of MR signal data sets (2, 2’) is not subjected to the high-demanding requirements (e.g. large scan times, etc.) of known MRI or imageless methods. Said requirements are imposed in said known methods, since therein the acquired MR signal data are required to allow for the reconstruction of a reliable imaged of the internal spatial structure of the scanned subject / sample.

[0134] In the present example, the executed EM sequence (1) is derived according to a symptoms or condition database. In particular, the database consists of 935 2D brain digital phantom slices derived from real human patients. The 2D brain slices comprise one or more types of tissues (white / grey matter, cerebrospinal fluid, and MS lesion), wherein half of said 2D brain slices comprise at least one MS lesion. In this context, the EM sequence (1) is “tailored” according to the above database, by determining the values of parameters charactering the EM fields of the EM sequence (1) and dimensions of acquired MR signal data sets (2, 2’) that are suitable for the specific classification task. Without loss of generality, said determination follows a magnetic resonance fingerprinting-like (MRF-like) methodology.

[0135] It should be mentioned that, when applied in known MRI methods, the MRF methodology provides an estimation of the required number of EM sequence blocks for accurate distinguishability between target tissue(s) and the rest of tissues. In said known methods, the number of EM sequence blocks is determined on a basis of pixel-wise distinguishability of tissues, which can be ascertained. However, the above estimation relies on highly- undersampled image reconstruction, where pixel assignation is not exact, and it is the dictionary-based look up which corrects for pixel miss-assignation that depends on the reconstruction algorithm, the noise realization, etc. In other words, an MRF methodology in known MRI methods provides an estimation of the EM sequence characteristics that allows to accurately distinguish target tissue(s) and the rest of tissues. Said estimation requires a further improvement to determine the final suitable characterisation of the suitable EM sequence (1).

[0136] In the above context, the MRF-like determination of the present example provides an estimation of the characterisation of the suitable EM sequence that requires a further improvement analogous to that of known MRI methods (although the imageless method of the invention does not rely on image pixels as such). In the present example the required improvement is further associated with the lack of a mathematical theorem / connection relating the minimal amount of EM blocks, of gradient spokes, and of acquired data for performing the imageless classification with a particular processing means configuration. In this regard, the MRF-like determination of the present example is improved by performing thereof together with the development of the specific configuration of the processing means that carries out the classification task. When the processing means is based on ML networks, the minimal suitable number of EM sequence blocks, of gradient spokes, and of acquired MR signal data is determined at the end of the ML network development. On the other hand, regarding the suitable number of gradient spokes, in the present example the presence / absence of MS lesion does not require spatially local information on tissues, but only global characteristics (global time decay by Ti, T2 of each kind of tissue). Accordingly, in the present example the suitable number of gradient spokes is smaller than in known MRI methods.

[0137] In the present example, the processing means is based on a ML algorithm, whose specific architecture is described later on. The MLF-like determination of the suitable characterisation of the EM fields of the EM sequence (1) and dimensions of the acquired MR signal data sets (2, 2’) is described in the following.

[0138] From each 2D brain digital phantom slice of the database, a plurality of eventual EM sequences and the respective acquired MR signals data sets are simulated, wherein each simulated EM sequence has a number of EM blocks. Each of said EM blocks is characterised by a flip angle-repetition time pair. All the simulated EM sequences and acquired MR signal data sets are imputed into a differential-evolution optimizationalgorithm. No gradient pulse is included in the simulated EM sequences, since spatial information is not required in the present example, as explained above. The differentialevolution optimization algorithm is configured to determine the best combination of flip angles and repetition times based in the distinguishability between MS lesions and other tissues. In said determination, the algorithm compares the simulated MR signal associated to each simulated EM sequence, with the expected MR signal resulting from the application of said simulated EM sequence to tissues only consisting of one type of tissue (i.e. , only costing of MS lesions, only consisting of white matter, only consisting of grey matter, and only consisting of cerebrospinal fluid). Said comparison focuses on the echoes of the MR signals. Furthermore, the discriminant of the comparison is the dot product of each simulated MR signal and the expected MR signal for each only-one-type-of tissue case. In addition, in the present MRF-like determination the first EM sequence block of the EM sequence (1) is set to have a flip angle, 01, equal to 180°, so that the first resonant RF pulse of the EM sequence (1) is equivalent to an inversion recovery-like preparation pulse. The first repetition time, TRi, is also set, so that the first EM sequence block prepares the magnetisation of each tissue for optimal distinguishability, as commonly done in MRF methodologies.

[0139] The MLF-like determination and the development of the ML algorithm leads to an EM sequence (1) and dimensions of the acquired MR signal data sets (2, 2’) that are suitable for classifying MR signals emitted from a 2D brain slice into the Groups l-ll. It should be clarified in relation to this point that the methods for the determination of the values characterising an EM sequence and of dimensions of the acquired MR signal data sets suitable for a particular MR classification task is generally known. Therefore, the present invention does not claim any particular or advantageous MRF-like, Cramer-Rao or equivalent methodology combined with any particular or advantageous development of a specific processing means configuration. Instead, the invention claims the general application thereof to determine EM sequences (1) and dimensions of acquired MR signal data sets (2, 2’) not subjected to the high-demanding requirements of known MRI and imageless MR methods. Indeed, the simulated EM sequences and acquired MR signal data sets are not subjected to said requirements. In fact, the EM sequence (1) of the present example is set to comprise a single inhomogeneous magnetic field gradient pulse fixed at a single direction in space. This can be seen in Figure 1 , wherein the EM sequence (1) comprises an inhomogeneous magnetic field gradient pulse (3) (represented with dashed lines) that rises up to a maximum strength value. The inhomogeneous magnetic field gradient pulse (3) is kept at that maximum strength value and at the same direction in space until the end of the EM sequence (1). Thus, said inhomogeneous magnetic field gradient pulse (3) imposes a position-dependent Larmor frequency along a single direction in the subject / sample until the end of the EM sequence (1). In the present example, the above maximum strength is set to a value, such that the standard Nyquist requirements of known MRI reconstruction methods is satisfied for all the repetition times of the EM sequence (1). In this regard, the number of acquired MR signal data per repetition time is set to 200, i.e. the amount of MR acquired data per repetition time for reconstructing a 100x100 pixels image with oversampling 2 in an MRF-based MRI method.

[0140] Thus, the EM sequence (1) of the present example has N=30 EM sequence blocks (4, 4’). Figure 1 only shows the first and second EM sequence blocks (4, 4’) of said N blocks. Dotted vertical lines separate the EM sequences blocks (4, 4’) of the EM sequence (1). Each EM sequence block (4, 4’) comprises a first resonant RF pulse (5, 5’) that sets a flip angle (01, 02) in the subject / sample. That is to say, the net magnetization comprised in the subject / sample, resulting from the application of the main magnetic field and aligned with the direction of said main magnetic field, tips by said flip angle (0i, ©2) with respect to the direction of the main magnetic field. The N EM sequence blocks (4, 4’) are different from each other with respect to the flip angles (01, ©2). In particular, in the present example the values of the different flip angles (0i, ©2) are equal or lager than 10° and equal or smaller than 150°. Each EM sequence block (4, 4’) further comprises a second resonant RF pulse (6, 6’) executed later than the first RF pulse (5, 5’). This second resonant RF pulse (6, 6’) is a TT resonant RF pulse, that reverses by a TT angle the tipped net magnetization. Hence, a MR signal is emitted from the subject / sample due to the execution of each EM sequence block (4, 4’). Each MR signal comprises a first part associated with the first resonant RF pulse (5, 5’), and a second part associated with the second resonant RF pulse (6, 6’). This second part corresponds to a spin echo of the MR signal. The absolute maximum of each MR signal data set (2, 2’) in Figure 1 is associated to said spin echo. On the other hand, the N EM sequence blocks (4, 4’) are also different from each other with respect to the repetition times (TR1, TR2). In particular, in the present example the values of the different repetition times (TR1, TR2) are equal or lager than 10 ms and equal or smaller than 200 ms. The time durations of the repetition times (TR1, TR2) between first resonant RF pulses (5, 5’) of the first three EM sequence blocks (4, 4’) of the EM sequence (1) are indicated with two double arrows in the bottom side of Figure 1 . As described above, when the EM sequence (1) is executed on a 2D brain slice, the slice emits the MR signals. Figure 1 represents the acquisition time windows (7) of the MR signal data sets (2, 2’) as rectangles represented below the horizontal time-flow arrow. Contrary to common MRI and imageless MR methods, at each EM sequence block (4, 4’) a first acquisition time window (7) is implemented out of the region around spin echoes. Additionally, a second acquisition time window (7) in each EM block (4, 4’) finishes right at the absolute maximum associated to the respective spin echo. The MR signal data sets (2, 2’) are acquired almost throughout the whole-time duration of the EM sequence (1). In particular, the acquisition step of the method is performed at quasi-full duty cycle, so that MR signal data sets (2, 2’) are acquired during a percentage of less 100% but at least 80% of the execution time duration of the EM sequence (1). In the best-case scenarios of known MRI and imageless methods, the MR signal data sets are acquired during a percentage of seldom more than 50% of the execution time duration of the EM sequence. Advantageously, in the present example the amount of acquired MR signal data is enhanced so that the signal-to-noise ratio of the acquired MR signal data sets (2, 2’) is increased, with respect to said prior-art methods. Accordingly, the accuracy of the classification method is enhanced compared to known imageless MR methods.

[0141] In the present example, each acquired MR signal data set (2, 2’) consist of 200 data points. As already explained, said MR signal data sets (2, 2’) consist of 1 D time-domain signal data. More precisely, the MR signal data sets (2, 2’) are non-image-reconstructing. Indeed, a reliable MRI image reproducing the internal spatial structure of the 2D brain slice cannot be reconstructed from the acquired MR signal data sets (2, 2’), neither separately from each data set (2, 2’), nor from all the data sets (2, 2’) taken together. This is because all the MR signals result from the application of a magnetic field gradient pulse (3) in a single direction of the space. Consequently, the acquired MR signals prevent from distinguishing spatial structures that are subjected to the same gradient strength.

[0142] In alternative embodiments of the invention, the acquired MR signal data sets (2, 2’) may comprise at least one group of the data from which it is possible to reconstruct a reliable MRI image reproducing the internal spatial structure of the subject / sample. In any case, in said alternative embodiments the acquired MR signal data sets (7, 7’) also consists of 1 D time-domain signal data. It should be clarified at this point that the present invention does not claim the generic acquisition of M R signal data sets, but rather the fact that said acquired MR signal data sets (2, 2’) are not subjected to the high-demanding requirements of known MR and imageless MR methods. Furthermore, this characteristic feature extends to the data that are involved in the classification procedure performed by the processing means.

[0143] B.- Configuration of the processing means and classification results

[0144] The acquired MR signal data sets (2, 2’) are processed by a processing means, so that the classification of MR signals emitted from the subject / sample (i.e., the 2D brain slice) is performed without transforming the acquired data into image raw data of the arranged domain. The processing means performs the classification according to a plurality of diagnostic or prognostic interest groups (i.e., Groups l-ll). The diagnostic / prognostic interest groups are associated with reference data. Contrary to prior-art imageless MR methods, in the present invention the performance of the processing means is advantageously based on 1 D time-domain signal data. Indeed, the acquired MR signal data sets (2, 2’) to be classified consist of 1 D time-domain signal data, and the reference data defining the diagnostic / prognostic interest groups (on which said classification is based on) comprise 1 D time-domain signal data. The 1 D time-domain signal data are not rearranged into a 2D or 3D k-space that could produce, by Fourier transformation, a reliable image of the internal spatial structure of the subject / sample.

[0145] The reference data are associated with reference MR signals associated with EM sequences for reference subjects / samples. In the present example, reference subjects / samples are 2D brain digital phantom slices of the aforementioned database, and reference EM sequences and reference MR signals are analogous to those of Figure 1. Moreover, the reference data are segmented into the above Group l-ll. In this context, the processing means determines one or more values of similarity between the acquired MR signal data sets (2, 2’) and Groups l-ll associated with the reference data. From said values of similarity, the processing means classifies the MR signals emitted from the 2D brain slice in of Groups l-ll.

[0146] In the present example, the determination of the similarity values and the classification derived from thereof is performed with an algorithm. More precisely, to performed said determination and classification steps, the processing means is configured with an algorithm based on a ML network. The ML network is trained prior to use in the method, so that the ML network learns to recognise patterns and / or relationships comprising relevant medical information from training data. In particular, the training data are the above reference data, so that the learnt patterns and / or relationships comprises medical information associated with the above Groups l-ll. Without loss of generality, the ML network of the present example includes a convolutional neural network (CNN) comprising a first set of layers, and a multi-layer perceptron (MLP) comprising a second set of layers. In addition, the classification performed by the processing means includes providing a score that quantitatively indicates the degree of probability that the MR signals emitted from 2D brain slice are associated with MS lesions. The ML network is trained and tested with, respectively, 825 and 110 2D brain digital phantom slices of the aforementioned database. A 10-fold Cross-Validation (CV) is implemented with the training 2D brain digital phantom slices to optimize the CNN’s architecture (i.e., number of layers, kernels, kernels’ sizes, etc.) and the training hyperparameters. The optimised architecture setup consists of a first set of layers consisting of two 1 D convolutional layers, and a second set of layers consisting of two dense layers whose output is passed through a sigmoid activation function yielding the predicted probability of having MS lesion in each 2D brain digital phantom slice. A threshold of 0.5 determines MS lesion presence in the MR signals. Concerning the loss-function, the ML network is trained by binary cross-entropy between its output and the label of the input signal (presence / absence of MS lesion). Alternative and equivalently optimal ML architectures, with more or less convolutional and dense layers, may be used according to the invention. In the following, the performance of the ML network of the present example is explained in detailed.

[0147] The MR signal data sets (2, 2’) are submitted to the first set of layers, wherein each of the acquired MR signal data sets (2, 2’) is associated with one of the aforementioned thirty repetition times (TRi, TR2). Each of said data sets (2, 2’) consists of 200 data points, wherein each data point includes a real part and an imaginary part of the MR signals.

[0148] The first set of layers is configured to “translate” the acquired MR signal data sets (2, 2’) into representative features, through the application of convolutional kernels. Accordingly, the first set of layers consists of an input layer wherein the acquired MR signal data sets (2, 2’) are imputed, and an output layer that outputs the representative features extracted from the MR signal data sets (2, 2’). In the input layer, sixteen convolutional kernels are applied to the imputed MR signal data sets (2, 2’). As a result, the input layer outputs a dimensionally reduced result: from the 60x200 imputed MR signal data, 64x196 features are outputted by the input layer. These features are then inputted to the output layer, so that features are similarly extracted at the output layer, such that the output layer outputs a set of representative features for the MR signals. The processing means is configured to vectorise the set of representative features outputted by the first set of (convolutional) layers, and to feed them in the second set of (dense) layers. The second set of layers is configured to receive the representative features and to provide, from said representative features, a score indicating the probability of MS lesion in the 2D brain slice. In this particular example, the second set of layers comprises two layers (i.e. , an input layer and an output layer), which are densely connected to each other. The output layer comprises an output neuron that, through the application of a sigmoid function, provides the aforementioned score.

[0149] Figure 2 represents in a first plot (left-hand side) the probability outputted by the ML network of the present example for the 110 2D test-set brain slices mentioned above. In addition, Figure 2 represents in a second plot (right-hand side) different metrics of the results of the first plot. The MR signals used in Figure 2 are 110 simulated MR signals representing real MR signals resulting from the application of the EM sequence (1) of Figure 1 to the 110 2D brain slices. In the first plot, the probability of MS lesion is represented as a function of the amount of MS lesions of the 2D brain digital phantom slices. The amount of MS lesion is represented as the tissue area covered by the MS lesions in a 2D brain slice. Notice that the first plot only shows the true positives and the false negatives obtained out of said 110 2D brain slices. Herein, the false positives and the true negatives cover 0 cm2. Thus, for the sake of clarity of the figure, the false positives and the true negatives are not plotted in the first plot. The number of true negatives is equal to 75, while the number of false positives is equal to 0. With regard to the represented metrics, the second plot represents the Accuracy, the Area Under the Curve (AUC), the F1-Score, the True Positive Rate (TPR) and the False Positive Rate (FPR). It should be mentioned at this point that, while herein a single inhomogeneous magnetic field gradient pulse (3) is executed, an MRF-based MRI classification method with thirty EM sequence blocks would require tens of gradient spokes (e.g., equal or around 40 spokes) to similarly classify MS lesions. The signal to noise ratio (SNR) of the results is around 25. In alternative embodiments with less noise wherein the number N of blocks is reduced to 10, MS lesions can be classified for SNR above 60.

[0150] In another embodiment of the invention, after classifying the MR signals in the above Group I “MS lesion”, the processing means is further configured to determine the amount of MS lesion comprised in the scanned 2D brain slice, from the acquired MR signal data sets (2, 2’) and based on information associated to the reference data. In particular, to perform said determination, the processing means are configured with a regression algorithm. In this regard, in a non-limiting example the processing means is configured with a network similar to that described above, which is optimised for MS lesion amount detection. In particular, the first set of layers is the same as above, whereas the second set of layers consists of a single dense layer, a normalisation layer and a 10% dropout before the last neuron, the latter having a ReLLI activation. This non-limiting example network is similarly trained as describe above for the classifying network, wherein in this case the loss-function is the Mean Squared Error between predicted output (MS amount) and the reference label of the input MR signal data (MS amount). Thus, though the ReLLI activation, the amount of MS lesion is herein outputted in cm2.

[0151] Other embodiments of the method of the invention are deducible by a person skilled in the art from the content of the present document. A particular combination of EM sequence, dimensions of data or processing means configurations suitable for performing the method depends on the scanned sample / subject and the diagnostic / prognostic interest groups. For instance, if a target lesion may be comprised in the subject / sample and is expected to be relatively small compared to the rest of the subject / sample, the larger the number of EM sequence blocks (4, 4’) the easier to comprise information of said small lesion in the acquired data. On the other hand, if a target disease requires that certain spatial information must be entailed in the acquired data, more than one gradient spoke must be executed. The larger the number of gradient spokes, the more precise the spatial information entailed in the acquired MR signal data. Nevertheless, in particular embodiments of the invention said number of gradient spokes is smaller than in known MRI methods.

[0152] According to the above, in another embodiment of the invention the scanned subject / sample is a 3D brain, the diagnostic / prognostic interest groups are again Group I) “MS lesion” and Group II) “no MS lesion”, and the executed EM sequence (1) is analogous to that of Figure 1. In a particular embodiment, said analogous EM sequence (1) comprises a single or a plurality of different (in strength and direction) inhomogeneous magnetic field gradient pulses (3), and more EM blocks (4, 4’) than the EM sequence (1) of Figure 1. In another embodiment wherein the scanned subject / sample is a 3D brain and the diagnostic / prognostic interest groups are the above Groups l-ll, the executed EM sequence (1) is segmented in a number M of parts. Each part is configured to produce MR signals emitted from a 2D slice of the scanned 3D brain, through the execution of an initial sliceselection pulse before each of the M parts. Each part of the executed EM sequence (1) is analogous to the EM sequence (1) of Figure 1. In another embodiment of the invention, the scanned subject / sample is a 3D brain, the diagnostic / prognostic interest groups are Group I) “hydrocephalus lesion” and Group II) “no hydrocephalus lesion”, and the executed EM sequence (1) is analogous to that of Figure 1. In a particular embodiment, said analogous EM sequence (1) comprises a single or a plurality of different (in strength and direction) inhomogeneous magnetic field gradient pulses (3), and more EM blocks (4, 4’) than the EM sequence (1) of Figure 1. In another embodiment of the invention, the method additionally or alternatively classifies the MR signals emitted from the 3D brain according to other two groups: Group A “hydrocephalus lesion in the left hemisphere” and Group B “hydrocephalus lesion in the right hemisphere”. In another embodiment, in order to classify the MR signals in the latter two Groups A-B, the execution of the EM sequence (1) and the acquisition of MR signal data is performed with a hand-held NMR scanner (similar to that shown in Figure 5 below), adapted to execute two inhomogeneous magnetic field gradient pulses (3) that are perpendicular to each other. This allows to acquire MR signal data comprising spatial information regarding the left and right hemispheres of the 3D brain. In this embodiment, the EM sequence is divided in two parts, wherein each part is analogous to the EM sequence (1) of Figure 1. The first / second part of said EM sequence is executed by arranging the NMR scanner towards the left-right / back- front sides of the head.

[0153] In another embodiment of the invention, the scanned subject / sample is a 3D brain, the diagnostic / prognostic interest groups are Group I) “ischemic stroke”, Group II) “hemorrhagic stroke” and Group III) “no stroke”, and the executed EM sequence (1) is analogous to that of Figure 1. In a particular embodiment, said analogous EM sequence (1) comprises more EM blocks (4, 4’) than the EM sequence (1) of Figure 1 , for better distinguishability between ischemic strokes and hemorrhagic strokes.

[0154] Example 2: executed EM sequence and acquisition of MR signal data

[0155] Similarly to Figure 1 , Figure 3 shows a schematic representation according to an embedment of the invention of an EM sequence (1) executed with an NMR scanner on a subject / sample and the MR signal data sets (2, 2’) acquired with said NMR scanner. Indeed, in both Figures 1 and 3, elements that are the same or equivalent are indicated by the same reference signs (numbers, Greek letters, etc.). Like in Figure 1 , in Figure 3 it is not shown the main magnetic field applied with the NMR scanner during the execution of the EM sequence (1). The EM sequence (1) of Figure 3 has N EM sequence blocks (4, 4’), although only the first two EM sequence blocks (4, 4’) are shown in the figure. Similar to Figure 1 , each EM sequence block (4, 4’) comprises a first resonant RF pulse (5, 5’) that sets an initial first flip angle (0i, 62) in the subject / sample, and a second resonant TT RF pulse (6, 6’). The two flip angles (01, 62) of the EM sequence (1) are different from each other. In addition, in the present example each EM sequence block (4, 4’) comprises a third resonant RF pulse (8, 8’), which is also a TT RF pulse. In this way, the time signal after each first resonant pulse (5, 5’) starts from a situation where all spins are refocused. The second TT RF pulse (8, 8’) of each EM sequence block (4, 4’) enhances SNR and further improves signal data acquisition at maximum height of the signal, since the echo does not coincide with the execution of an RF pulse. Similarly to Figure 1 , the repetition times (TR1, TR2) are indicated with a double arrows in Figure 3.

[0156] On the other hand, contrary to Figure 1 , in the present example the EM sequence (1) comprises different inhomogeneous magnetic field gradient pulses (3, 3’). Figure 3 shows a first inhomogeneous magnetic field gradient pulse (3) that rises up to a certain strength value and direction before the execution of the first resonant RF pulse (5) of the first EM sequence block (4). A second inhomogeneous magnetic field gradient pulse (3’) is similarly executed on the second EM sequence block (4’). The inhomogeneous magnetic field gradient field pulse (3’) of the second EM sequence block (4’) rises up to a strength value and direction different to those of the inhomogeneous magnetic field gradient pulse (3) of the first EM sequence block (4). Thus, the two EM sequences blocks (4, 4’) are different in strengths, directions, and flip angles (01, 62). Consequently, the MR signal data sets (2, 2’) associated to the emitted MR signals look different from each other.

[0157] Similarly to Figure 1 , the acquired MR signal data set (2, 2’) consist of 1 D time-domain signal data that are acquired at quasi-full duty cycle. However, in Figure 3 the second acquisition time window (7) of each EM sequence block (4, 4’) extends beyond the first spin echo maximum.

[0158] Example 3: multifrequency embodiment

[0159] Herein it is described a non-limiting example of multifrequency embodiment according to the present invention. Similarly to Figure 3, Figure 4 shows a schematic representation of an EM sequence (1) and the acquired MR signal data sets (2, 2’). However, in the present example the NMR scanner applies a non-uniform main magnetic field (not shown in Figure 4) on the subject / sample. Consequently, a position-dependent Larmor frequency is imposed along the spatial extent of the subject / sample along one direction in said subject / sample. Without loss of generality, herein said one direction is set to the z direction.

[0160] The values of the above position-dependent Larmor frequency are delimited by a bandwidth range (AWL) given by

[0161] (YAB0) / (2TT), wherein “ABo” is the difference between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent of the subject / sample along the z direction, and “y” is the protonic gyromagnetic ratio (y=2.68x108rad T'1s-1).

[0162] The EM sequence (1) of Figure 4 comprises a number N of sequence blocks (4, 4’). Figure 4 shows the nthEM block (4) and the (n+ 1 )thEM block (4’) comprised in said EM sequence (1). The dots represented in both left- and right-handed sides of Figure 4 indicate that more EM sequence blocks are comprised in the EM sequence (1) aside from those shown therein. Similarly to the EM sequence (1) of Figure 3, both nthand (n+1)thEM bocks (4, 4) comprise a first resonant RF pulses (5, 5’) setting an initial flip angle (0n, 0n+i), a second resonant TT RF pulse (6, 6), and a third resonant TT pulse (8, 8’). However, none of the EM sequence blocks (4, 4’) of the entire EM sequence (1) comprise magnetic gradient fields, as the spatial encoding of the subject / sample is already performed by the non-uniform main magnetic field.

[0163] In the present embodiment, the NMR scanner applying the non-uniform main magnetic field further comprises a resonant circuit. In particular, the resonant circuit has a retunable resonant frequency. Hence, to detect the MR signal resulting from the execution of an EM sequence block (4, 4’), the resonant frequency of said resonant circuit coincides with the radiation frequency of the one or more resonant RF pulses (5, 5’, 6, 6’, 8, 8’) comprised in said EM sequence block (4, 4’). In addition, the resonant circuit has a frequency bandwidth (Awc) smaller than the delimiting bandwidth range (AWL) imposed on the subject / sample by the non-uniform main magnetic field. Advantageously, in this particular embodiment the EM sequence blocks (4, 4’) are contiguously grouped by groups of same radiation frequency each, so that the EM sequence (1) is divided in a number M<N of sets of EM sequence blocks (4, 4’). Thus, in each of the sets the radiation frequency of the RF pulses (5, 5’, 6, 6’, 8, 8’) is the same for all the EM sequence blocks (4, 4’) of the set. In addition, the frequency bandwidth (Awc) of the resonant circuit is given by

[0164] Awc= AWL / M, wherein “M” is the aforementioned number of sets of EM sequence blocks (4, 4’) comprised in the EM sequence (1).

[0165] In the above context, the spatial extent encoded by the non-uniform magnetic field along the z direction in the subject / sample can be seen as split into M different portions. The EM sequence (1) steppedly sweeps the resonant frequency of the resonant circuit across the whole delimiting bandwidth range (AWL) imposed by the non-uniform main magnetic field. In this respect, between two consecutive sets of EM sequence blocks (4, 4’), the difference between the radiation frequency of the resonant pulses RF pulses (5, 5’, 6, 6’, 8, 8’) of the later set of EM sequence blocks (4, 4’), and the radiation frequency of the resonant pulses RF pulses (5, 5’, 6, 6’, 8, 8’) of the earlier set EM sequence blocks (4, 4’), is equal to frequency bandwidth (Awc) of the resonant circuit. In addition, the associated resonant frequency of the resonant circuit changes by an amount equal to the frequency bandwidth (Awc) of the resonant circuit, when the EM sequence (1) passes from one set of EM sequence blocks (4, 4’) to the next set of EM sequence blocks (4, 4’).

[0166] According to the above, for instance, if the nthEM sequence block (4) of Figure 4 belongs to the mthset of EM sequence blocks, the resonant frequency of the first, second and third resonant RF pulses (5, 6, 8) is given by

[0167] WRF(n)= WRF(m)= WL. max - m Awc, and the associated resonant frequency of the resonant circuit in the detection of the subsequent nthemitted MR signal is given by,

[0168] Wc(n)= Wc(m)= WL, max - m Awc.

[0169] Herein, “wi_,max” is the maximum value of the position-dependent Larmor frequency along the spatial extent of the subject / sample along the z direction. On the other hand, if the (n+1)thEM sequence block (4’) belongs to the (m+1)thset of EM sequence blocks, the resonant frequency of the first, second, and third resonant RF pulses (5’, 6’, 8’), and the respective associated resonant frequency of the resonant circuit, increase by an amount equal to the frequency bandwidth (Awc) of the resonant circuit, compared to the previous mthset of EM sequence (4, 4’) blocks. Hence, in the embodiment of Figure 4, the M sets of EM sequence blocks (4, 4’) comprised in the EM sequence (1) cover the whole delimiting bandwidth range (AWL) imposed by the non-uniform main magnetic field. In this sense, the NMR scanner applies the non-uniform main magnetic field, executes the EM sequence (1), and acquires the MR signal data sets (2, 2’), so that it successively excites different portions of the subject / sample and acquires the MR signals from said different portions.

[0170] Figure 5 shows a schematic representation of an NMR scanner (9) in a non-limiting multifrequency embodiment. Said NMR scanner (9) is a hand-held NMR scanner configured to apply a main magnetic field and execute EM sequences on a subject (10), and to acquire the subsequent MR signals resulting thereof. To this end, NMR scanner (9) comprises a main magnet (11), a RF system (12), and a data acquisition unit (not shown in Figure 5).

[0171] The main magnet (11) is a single-sided magnet configured to apply a non-uniform main magnetic field, so that a position-dependent Larmor frequency is imposed along the spatial extent of the subject (10) in the z direction. The values of said position-dependent Larmor frequency are delimited by a bandwidth range (AWL) given by

[0172] (Y AB0) / (2TT), wherein ABo is the difference (ABo) between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent of the subject (10) in the z direction.

[0173] Figure 5 shows a graphic superimposed over the subject (10). Said graphic indicates how the position-dependent Larmor frequency changes along the z direction. Therein, both the spatial extent (Az) of the subject (10) in the z direction and the delimiting bandwidth range (AWL) along said spatial extent (Az), are highlighted by two double arrows.

[0174] Concerning the RF system (12) of Figure 5, it comprises a RF transmitter and a RF receiver. In particular, the RF transmitter and the RF receiver are integrated in a same resonant circuit comprising a coil. The resonant circuit has a frequency bandwidth smaller than the above delimiting bandwidth range (AWL). Advantageously, the Q factor of said resonant circuit is relatively large, given the above delimiting bandwidth range (AWL).

[0175] The RF system (12) is configured to work in a transmitter mode and in a receiver mode. Hence, when working in the transmitter mode, the RF system (12) is a RF transmitter configured to execute, on the subject (10) and during the application of the non-uniform main magnetic field, a plurality of resonant RF fields comprised in EM sequence blocks (4, 4’). Each EM sequence block (4, 4’) forms part of a set of EM sequence blocks (4, 4’) as defined above, so that the radiation frequency of the RF pulses (5, 5’, 6, 6’, 8, 8’) is the same for all the EM sequence blocks (4, 4’) of the set. The radiation frequency is different for different sets of EM sequence blocks (4, 4’) and resonant with a single value of the position-dependent Larmor frequency along the z direction of the subject (10).

[0176] On the other hand, when working in the receiver mode, the RF system (12) is a RF receiver configured to retune the resonant frequency of the resonant circuit comprised thereof. Accordingly, to detect the MR signal resulting from each set of EM sequence blocks (4, 4’) executed by the RF transmitter, the resonant frequency of said resonant circuit changes to the radiation frequency of the one or more resonant RF fields comprised in said set of EM sequence blocks (4, 4’).

[0177] Finally, the data acquisition unit of the NMR scanner (9), is configured to acquire at least part of each of the one or more MR signals detected by the RF receiver, by recording a set of MR signal data for each of said one or more detected MR signals. According to the present invention, the one or more MR signal data sets advantageously consist of 1 D timedomain signal data, contrary to known imageless MR methods. In addition, the data acquisition unit is communicatively coupled to a processing means, so that the data acquisition unit provides to said processing means the one or more MR signal data sets. The processing means processes said one or more MR signal data sets, to classify the above one or more MR signals in one or more diagnostic or prognostic interest groups.

[0178] Example 4: Method flowchart

[0179] Figure 6 shows a flowchart comprising five stage-blocks (13)-(17) that schematically illustrate a particular embodiment of an imageless method for classifying MR signals according to present invention. In this example embodiment, a user that is a non-medical expert has a question of medical relevance that wants to address. In particular, the question of medical relevance is: “I have pain in the knee: is my anterior cruciate ligament torn?”. To address said question, the user uses a system comprising a resource-relaxed NMR scanner and a processing means. The processing means is communicatively coupled with and integrated in the resource-relaxed NMR scanner. In the above context, in the first stage-block (13) of Figure 6, the user inputs the question of medical relevance as a prescribed symptom input information to the system.

[0180] Then, in the second stage-block (14), the system, based on the prescribed symptom input information inputted in the first block (13), selects from a database an EM sequence to address the above question of medical relevance.

[0181] The method proceeds to the third stage-block (15), wherein the EM sequence selected in the second stage-block (14) is executed by the resource-relaxed NMR scanner on the user’s pained knee arranged thereof. In addition, during the execution of the EM sequence, the resource-relaxed NMR scanner also applies a main magnetic field on the user’s pained knee. Consequently, the user’s pained knee emits MR signals. Furthermore, during the execution of the above EM sequence, a set of MR signal data is acquired with the resource- relaxed NMR scanner for at least part of each of said emitted MR signals. Contrary to priorart MRI methods and imageless MR methods, said MR signal data sets acquired with the resource-relaxed NMR scanner consist of 1 D time-domain signal data that are not transformed into image raw data of the arranged domain. Advantageously, the acquired data comprise information associated with said user’s knee that is indeed relevant for the classification of the above MR signals. Hence, the accuracy of the method according to present invention is enhanced with respect to prior-art imageless classification methods of MR signals.

[0182] In the fourth stage-block (16) of Figure 6, the resource-relaxed NMR scanner provides the MR signal data sets acquired in the third stage-blocks (15) to the processing means. The processing means is based on an algorithm configured to recognise data patterns in MR signal data information. In particular, said algorithm comprises the uses of a trained ML network. The training data of said trained ML network comprises data associated with one or more reference MR signals. More precisely, said training data are MR signal data sets previously acquired in response to a plurality of reference EM sequences executed on a plurality of reference knees. The training data are segmented into two groups:

[0183] - Group I: “knees with a torn anterior cruciate ligament”;

[0184] - Group II: “knees without a torn anterior cruciate ligament”.

[0185] Contrary to prior-art imageless MR methods, the above training data comprise 1 D timedomain signal data that are not transformed into image raw data of the arranged domain. In this context, in the fourth stage-block (16), the processing means recognises patterns in the one or more MR signal data sets acquired in the third stage-block (15), so that it determines one or more values of similarity between the above two Groups l-ll and the one or more acquired MR signal data sets. Finally, in the fifth stage-block (17) of Figure 6, from the values of similarity determined in the fourth stage-block (16), the processing means classifies the MR signals emitted from the user’s pained knee in either the Group I or Group II. Said classification comprises a binomial answer to the above question of medical relevance inputted in the first stage-block (13). The binomial answer is outputted by means of a display comprised in the system. In particular, the display outputs a “yes” answer when the MR signals are classified in the above Group I, or a “no” answer when the MR signals are classified in the above Group II.

[0186] In other non-limiting example embodiments of the invention, a question of medical relevance inputted to the system may be “is it the amount of fluid in the brain according to normal / expected levels in a subject with hydrocephalus”, “which is the type of stroke in the brain of a subject”, or “does the brain of a subject suffer from multiple sclerosis”.

Claims

AMENDED CLAIMS[received by the International Bureau on 2 April 2025 (02.04.2025)]1 A method of classification of magnetic resonance, MR, signals indicative of one or more portions of a subject or a sample (10) of diagnostic or prognostic interest, comprising the realization, by means of an NMR scanner (9) and a processing means, of the following steps: a) applying a main magnetic field and executing a sequence of electromagnetic (EM) fields (1), on the subject or the sample (10) arranged in the NMR scanner; b) acquiring with the NMR scanner (9), during the application of the main magnetic field and the execution of the EM sequence (1) in step a), at least part of each of the one or more MR signals emitted from the subject or the sample (10) that result from the application of the main magnetic field and the execution of the EM sequence (1), so that a set of MR signal data (2, 2’) is acquired for said each of the one or more emitted MR signals; c) determining, with the processing means, one or more values of similarity between the one or more MR signal data sets (2, 2’) of step b) and a plurality of diagnostic or prognostic interest groups associated with reference data, wherein said reference data are associated with one or more reference MR signals associated with one or more reference EM sequences for one or more reference subjects or samples, and wherein said reference data are segmented into the plurality of diagnostic or prognostic interest groups; d) classifying, with the processing means, the one or more MR signals emitted from the subject or the sample (10) in one or more diagnostic or prognostic interest groups of the reference data from the values of similarity determined in step c); the method being characterising in that- the one or more MR signal data sets (2, 2’) consist of one-dimensional time-domain signal data; and- the reference data associated with the one or more reference MR signals comprise one-dimensional time-domain signal data.2.- The method according to claim 1 , wherein the one or more MR signal data sets (2, 2’) and the reference data are non-image-reconstructing.3.- The method according to any one of claims 1 to 2, wherein the method further comprises, before step c), the realization of the following step: e) comparing, with the processing means, the one or more MR signal data sets (2, 2’) of step b) with the reference data.4.- The method according to any one of claims 1 to 3, wherein step c), d) and / or e) is performed with an algorithm configured to recognise data patterns in MR signals data.5.- The method according to claim 4, wherein- the algorithm comprises the use of one or more predictive models including at least one of the following: one or more regression models, time series models, clustering models, decision trees, ensemble models, artificial intelligence (Al) or machine learning (ML) models; and / or- steps c) and d) are performed with the algorithm, wherein the algorithm comprises the use of an ML network trained with training data.6.- The method according to any one of claims 1 to 5, wherein the EM sequence (1) executed in step a) is selected according to a symptoms or condition database, wherein a specific EM sequence (1) or combination of EM sequences (1) are executed in step a) depending on a prescribed symptom or condition input information.7.- The method according to any one of claims 1 to 6, wherein the one or more portions of the subject or the sample (10) of diagnostic or prognostic interest are indicative of one or more clinical features, said clinical features comprising at least one of the following: amount of MS lesion, amount of hydrocephalus liquid, amount of tumor, amount of hemorrhage, amount of endometriosis, amount of bodily fluid leak (e.g., synovial fluid), amount of abnormal tissue buildup; and wherein the method further comprise performing the following step: f) determining, with the processing means, the one or more clinical features, from the one or more MR signal data sets of step b) and based on information associated with the reference data.8.- The method according to any one of claims 1 to 7, wherein the EM sequence (1) executed in step a) comprises a plurality of EM sequence blocks (4, 4’), wherein each EM sequence block (4, 4’) comprises one or more resonant RF fields (5, 5’, 6, 6’, 8, 8’), so that the subject or the sample (10) is repeatedly re-excited during step a).9.- The method according to claim 8, wherein- the main magnetic field applied in step a) is a non-uniform main magnetic field, so that a position-dependent Larmor frequency is imposed along the spatial extent (Az) of the subject or the sample (10) along at least one direction in said subject or said sample (10),wherein the values of said position-dependent Larmor frequency along the spatial extent (Az) of the subject or the sample (10) along said at least one direction are delimited by a bandwidth range ( WL) given by(Y AB0) / (2TT), wherein ABo is the difference between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent (Az ) of the subject or the sample (10) along the at least one direction in said subject or said sample (10), and y is the protonic gyromagnetic ratio;- the EM sequence (1) is divided in a plurality of sets of EM sequence blocks (4, 4’);- the one or more RF fields (5, 5’, 6, 6’, 8, 8’) of each of the EM sequence blocks (4, 4’) of a set of EM sequence blocks (4, 4’) has a radiation frequency, wherein said radiation frequency is the same for EM sequence blocks (4, 4’) of the same set of EM sequence blocks (4, 4’);- the radiation frequency of a set of EM sequence blocks (4, 4’) is different for different EM sequence blocks (4, 4’) and resonant with a single value of the positiondependent Larmor frequency along the spatial extent (Az) of the subject or the sample (10) along the at least one direction in said subject or sample (10).10.- The method according to claim 9, wherein- step b) comprises the record of the at least part of each of the one or more emitted MR signals detected by the NMR scanner (9) by means of resonant circuit;- the resonant circuit has a retunable resonant frequency, so that, for the detection of the emitted MR signal resulting from the execution of each set of EM sequence blocks (4, 4’), the resonant frequency of said resonant circuit coincides with the radiation frequency of the one or more resonant RF fields (5, 5’, 6, 6’, 8, 8’) comprised in said set of EM sequence blocks (4, 4’); and- the resonant circuit has a frequency bandwidth smaller than the bandwidth range (AWL) delimiting the values of the position-dependent Larmor frequency along the spatial extent (Az) of the subject or the sample (10) along the at least one direction in said subject or said sample (10).11.- The method according to any one of claims 7 to 10, wherein:- the total number of EM sequence blocks (4, 4’) comprised in the EM sequence (1) is less than 100;- in at least one pair of two consecutive EM sequence blocks (4, 4’) comprised in the EM sequence (1), the repetition time (TRI , TR2) between each of the first resonant RF fields(5, 5’) of said at least one pair of two consecutive EM sequence blocks (4’, 4’) is less than 100 ms;- the EM sequence blocks (4, 4’) comprised in the EM sequence (1) are different in strengths, directions, flip angles (0i, ©2) and / or repetition times (TRI, TR2); and / or- the EM sequence blocks (4, 4’) comprised in the EM sequence (1) are randomly varied during step a).12.- The method according to any one of claims 1 to 11 , wherein step b) is performed- at full or quasi-full duty cycle; and / or- during the execution of at least one resonant RF field (5, 5’) comprised in the EM sequence (1) executed in step a).13.- A computer program comprising instructions which, when the computer program is executed by a processing means communicatively coupled with an NMR scanner (9), cause the processing means and the NMR scanner (9) to perform a method according to any one of claims 1 to 12.14.- A system comprising an NMR scanner (9) and a processing means communicatively coupled with the NMR scanner (9), wherein the NMR scanner (9) and the processing means comprise hardware and / or software means adapted to perform a method according to any one of claims 1 to 12.15.- The system according to claim 14, wherein the NMR scanner (9) comprises at least a main magnet (11), a RF transmitter (12), a RF receiver (12), and a data acquisition unit, wherein- the main magnet (11) is configured to apply a main magnetic field for polarising nuclear spins comprised in a subject or a sample (10), when said subject or said sample (10) is arranged in the NMR scanner (9);- the RF transmitter (12) is configured to execute one or more resonant RF fields (5, 5’, 6, 6’) on said subject or said sample (10);- the RF receiver (12) is arranged to detect one or more MR signals emitted from said subject or said sample (10); and- the data acquisition unit is configured to acquire at least part of each of the one or more MR signals detected by the RF receiver, by recording a set of one or more MR signal data (2, 2’) for each of the one or more MR signals detected by the RF receiver.16.- The system according to claim 15, wherein- the main field is a non-uniform main magnetic field that imposes a positiondependent Larmor frequency along the spatial extent (Az) of the subject or the sample (10) along at least one direction in said subject or said sample (10), wherein the values of said position-dependent Larmor frequency along the spatial extent (Az) of the subject or the sample (10) along the at least one direction in said subject or said sample (10) are delimited by a bandwidth range (AWL) given by(Y AB0) / (2TT), wherein ABo is the difference between the maximum amplitude (Bo, max) and the minimum amplitude (Bo, min) of the non-uniform main magnetic field along the spatial extent (Az) of the subject or the sample (10) along the at least one direction in said subject or said sample (10), and y is the protonic gyromagnetic ratio;- the RF transmitter (12) is configured to execute, on the subject or the sample (10) and during the application of the non-uniform main magnetic field, a plurality of resonant RF fields (5, 5’, 6, 6’, 8, 8’) comprised in EM sequence blocks (4, 4’), wherein each EM sequence block (4, 4’) comprised in a set of EM sequence blocks (4, 4’) and each EM sequence block (4, 4’) comprises one or more RF fields (5, 5’, 6, 6’) with a radiation frequency, wherein the radiation frequency is the same for EM sequence blocks (4, 4’) of the same set of EM sequence blocks (4, 4’), and wherein the radiation frequency of a set of EM sequence blocks (4, 4’) is different for different sets of EM sequence blocks (4, 4’) and resonant with a single value of the position-dependent Larmor frequency along the spatial extent (Az) of the subject or the sample (10) along the at least one direction in said subject or said sample (10);- the RF receiver (12) comprises a resonant circuit;- said resonant circuit has a retunable resonant frequency, so that, for the detection of the emitted MR signals resulting from the execution of each set of EM sequence blocks (4, 4’) executed by the RF transmitter (12), the resonant circuit is configured to change the resonant frequency to the radiation frequency of the one or more resonant RF fields (5, 5’, 6, 6’, 8, 8’) comprised in said set of EM sequence blocks (4, 4’) executed by the RF transmitter (12); and- said resonant circuit has a frequency bandwidth smaller than the bandwidth range (AWL) delimiting the values of the position-dependent Larmor frequency along the spatial extent (Az) of the subject or the sample (10) along the at least one direction in said subject or said sample (10).17.- The system according to any one of claims 14 to 16, wherein- the NMR scanner (9) comprises a magnetic gradient system configured to execute one or more inhomogeneous magnetic field gradients (3, 3’) on a subject or a sample (10), when said subject or said sample (10) is arranged in said NMR scanner (9), so that a position-dependent Larmor frequency is imposed along at least one direction in the subject or the sample (10);- the NMR scanner (9) is a resource-relaxed NMR scanner; and / or- the NMR scanner (9) is a non-image-reconstructing NMR scanner.

Images