Trained functions for providing magnetic field data and applications thereof

Abstract

The invention relates to a method for generating an optimized trained function to provide magnetic field data, a method for providing magnetic field data by means of a trained function, a system control unit, a magnetic resonance apparatus and a computer program product. The method for providing magnetic field data comprises receiving image data as input data of a trained function, applying the trained function to the image data, wherein the trained function is trained based on at least one assumption on a data fidelity of image data corrected from magnetic field data and on at least one characteristic of the magnetic field data, and providing magnetic field data as output data of the trained function.

Description

Technical Field

[0001] The present invention relates to a method for generating optimized trained functions to provide magnetic field data, a method for providing magnetic field data by means of trained functions, a system control unit, a magnetic resonance apparatus, and a computer program product. Background Technology

[0002] In medical technology, magnetic resonance (MR) imaging, also known as magnetic resonance computed tomography (MRI), is characterized by high soft tissue contrast. Here, the subject, especially the patient, is typically positioned within the examination space of the MRI scanner. During MRI measurements, high-frequency (HF) pulses are typically emitted from the transmitter coils of the MRI scanner onto the subject. Additionally, gradient pulses are output from the gradient coils of the MRI scanner, thereby generating a temporary magnetic field gradient in the examination space. These generated pulses excite and trigger position-encoded MRI signals within the patient. The triggered MRI signals are received by the receiver coils of the MRI scanner and used for the reconstruction of MRI imaging (or mapping).

[0003] An alternating magnetic field, known as B1, is generated in the inspection area by transmitting HF pulses. + The patient is situated in the examination area during the magnetic resonance measurement. Therefore, B1 + The field shows the distribution of the launch field. Conversely, B1 - The field describes the distribution of the receiving field. If the transmitting coil and receiving coil are the same, then B1 can generally be considered... + Field and B1 - The shapes of the fields are basically the same. An effective transmit-receive field can also be called a B1 field.

[0004] Furthermore, a strong static main magnetic field, the B0 field, is generated in the examination area. This main magnetic field should also be as homogeneous as possible to avoid artifacts in magnetic resonance imaging. However, in reality, the main magnetic field often exhibits inhomogeneity in some locations. Understanding the spatial distribution of the B0 field, i.e., the static magnetic field, is advantageous so that inhomogeneities in the main magnetic field can be compensated, for example, using shim coils. Summary of the Invention

[0005] However, determining magnetic field data, such as those describing the B1 and / or B0 fields, using common methods is often time-consuming and / or inaccurate. The technical problem this invention aims to solve is to provide a preferred, fast, and / or robust method for providing magnetic field data. More particularly, the technical problem this invention aims to solve is to provide measures capable of performing this method.

[0006] The technical problem is solved according to the present invention by a method for generating an optimized trained function to provide magnetic field data, a method for providing magnetic field data by means of a trained function, a system control unit, a magnetic resonance device, and a computer program product.

[0007] Therefore, a computer-implemented method is proposed for generating optimized, trained function OTFs to provide magnetic field data. The method includes the steps of:

[0008] a) Provide training image data TBD,

[0009] b) Provide corrected training image data KTBD generated based on TBD.

[0010] c) Provide a preliminary trained function VTF.

[0011] d) By applying TBD to VTF, preliminary magnetic field data VMD is generated.

[0012] e) Generate preliminary corrected image data VKBD based on VMD and TBD.

[0013] f) Determine data fidelity information based on VKBD and KTBD.

[0014] g) Based on VMD and the assumptions about the properties of VMD, determine at least one hypothesis fidelity information.

[0015] h) Determine optimization information based on data fidelity information and at least one hypothesis fidelity information.

[0016] i) Generate an OTF by optimizing the VTF based on the optimization information.

[0017] For example, magnetic field data provided by OTF can be used to reduce spatial amplitude fluctuations and / or inhomogeneities in magnetic resonance imaging. These spatial amplitude fluctuations and / or inhomogeneities are advantageously reflected in the magnetic field data.

[0018] Steps a), b), and c) can be performed, in particular, by means of provisioning units and / or interfaces suited to them respectively. Steps d), e), f), g), h), and i) can be performed, in particular, by means of computing units suited to them. Such computing units may in particular include one or more processors and / or memory modules.

[0019] Steps a) to i), especially steps a) to c) and / or steps f) and g) are not necessarily performed in this (alphabetical) order.

[0020] The trained function, particularly VTF and OTF, preferably maps input data to output data. Here, the output data may also be related to one or more parameters of the trained function. The trained function (or trained function) is particularly a function trained using machine learning. Other terms for trained function include, for example, trained mapping rule, mapping rule with trained parameters, function with trained parameters, artificial intelligence-based algorithm, or machine learning algorithm. An example of a trained function is an artificial neural network, where the edge weights of the artificial neural network correspond to the parameters of the trained function. The term "neural network" can also be used instead of "neural network." The trained function can also be, in particular, a deep artificial neural network (the technical term in English is "deep neural network" or "deepartificial neural network"). Other examples of trained functions are "support vector machines," and other machine learning algorithms can also be used as trained functions.

[0021] Magnetic field data, especially VMD, can be data describing one or more characteristics of a magnetic field. Here, the magnetic field is preferably the examination area of ​​the magnetic resonance imaging (MRI) device. During MRI measurements using the MRI device, the object to be examined can be positioned within the examination area. The object can be, for example, a patient.

[0022] The TBD may in particular include measurement data acquired by a magnetic resonance imaging (MRI) device, such as raw data, especially k-space data, and / or data derived from, and especially calculated from, these measurement data, such as one or more MRI images reconstructed from the raw data and / or k-space data. These measurement data may in particular have already been acquired during the MRI examination of the subject. The TBD may in particular include information about the anatomical structure of the subject.

[0023] TBD can particularly include combined amplitude maps. Specifically, magnetic resonance signals can be received separately by multiple coil elements of the receiving coil device of a magnetic resonance apparatus, and combined amplitude maps can be created based on these magnetic resonance signals, i.e., the corresponding signals are combined. TBD preferably includes measurement data from different magnetic resonance measurements, especially different objects and / or parts of the object being examined, particularly body parts. For example, VTF is sequentially trained using different raw data and / or magnetic resonance imaging.

[0024] The KTBD is specifically corrected for magnetic field characteristics, such as magnetic field inhomogeneities. To correct TBD and thus provide KTBD, a multitude of conventional, especially image-based, correction methods can be used, such as Statistical Parametric Mapping (SPM), Nonparametric Non-uniformity Intensity Normalization (N3), N4ITK, Uniform Combination Reconstruction (UNICORN), and / or InhomoNet. Because the training of the function, i.e., the generation of the OTF, has no particular time constraint, these correction methods can undoubtedly be computationally and time-consuming. However, the training result, i.e., the OTF, is advantageously designed to provide magnetic field data in a shorter time, especially within the range of common magnetic resonance measurements.

[0025] The proposed method allows for the advantageous generation or training of the OTF to cover multiple anatomical structures. Traditional methods are generally less suitable for this. Therefore, algorithms such as SPM typically work well only for the brain, and the N4 algorithm often requires different parameter sets for each anatomical structure.

[0026] The VTF can be based, in particular, on a neural network with preliminary, unoptimized edge weights as parameters of the VTF. For example, the edge weights can be initially set to 1, and UNET can be used as the architecture (Architektur) in particular. The resulting OTF can be an optimized VTF. For example, the preliminary edge weights of the neural network can be changed in step i) within the range of an optimized VTF.

[0027] Preferably, the TBD is received as input data to the VTF and applied to the VTF. Preferably, the VMD is output as output data to the VTF.

[0028] The generation of the VKBD specifically includes correcting the TBD based on the characteristics of the VMD with respect to the magnetic field, such as the inhomogeneity of the magnetic field.

[0029] Determining data fidelity information includes, for example, applying a loss function and / or a cost function to VKBD and KTBD. The data information preferably describes the similarity and / or data consistency and / or data fidelity between VKBD and KTBD. This data fidelity information can be expressed, for example, as a data fidelity value. Specifically, the greater the similarity and / or data consistency and / or data fidelity between VKBD and KTBD, the smaller the value of the data fidelity value. If, for example, VKBD and KTBD are identical, the data fidelity value can be zero. Advantageously, data fidelity information ensures (to a large extent) data fidelity and / or data consistency between the output data of the trained function, i.e., image data corrected for magnetic field data, and image data corrected by other means.

[0030] In particular, one or more hypothesis fidelity information can be determined, each based on a different hypothesis. Each of these hypotheses can relate to different characteristics of the VMD. At least one hypothesis fidelity information can include at least one hypothesis fidelity value for each hypothesis. Preferably, the OTF is optimized by this method such that the magnetic field data output by the OTF (at least to a large extent) satisfies the hypothesis. The at least one hypothesis is preferably at least one hypothesis that needs to be satisfied, in particular a condition.

[0031] The characteristics of VMD can be, in particular, the performance, specificity, and / or features and / or structure of VMD. This assumption can be, in particular, general and / or fundamental and / or non-specific. Specifically, this assumption does not depend on the specific TBD provided as input data to VTF. This assumption can, in particular, include assumptions about at least one general and / or fundamental and / or non-specific characteristic of the magnetic field typically present in a magnetic resonance apparatus. This characteristic can, in particular, relate to the performance and / or specificity and / or features and / or structure of such magnetic field. This assumption can, in particular, be based on prior knowledge of such magnetic field.

[0032] Preferably, in order to optimize the VTF, steps c) to i) are performed iteratively, in particular repeatedly, wherein the OTF generated in step i) is provided as the VTF in step c) for subsequent iterations, in particular repetitions. The method may include one or more such iterations.

[0033] For example, the method involves iteratively minimizing at least one optimal value of the optimization information. Ideally, the optimal value will thus tend to become smaller and smaller with each iteration.

[0034] Preferably, the iteration includes analysis of the optimization information determined in step h) and / or the hypothesis fidelity information determined in step g) and / or the data fidelity information determined in step f) to optimize the VTF in a targeted manner and generate the OTF in subsequent iterations. In particular, the optimization information determined in step h) and / or the hypothesis fidelity information determined in step g) and / or the data fidelity information determined in step f) are compared with the optimization information determined in step h) and / or the hypothesis fidelity information determined in step g) and / or the data fidelity information determined in step f) in previous iterations to optimize the VTF in a targeted manner in subsequent iterations.

[0035] For example, if the optimized value of the optimization information determined in step h) is larger in iteration N-1 than in iteration N, and if the goal is to minimize the optimized value, then in iteration N+1, the optimization of the VTF relative to iteration N is advantageously performed in the same manner as the optimization relative to iteration N-1 in iteration N. Conversely, if, for example, the optimized value of the optimization information determined in step h) is smaller in iteration N-1 than in iteration N, and if the goal is to minimize the optimized value, then in iteration N+1, the optimization of the VTF relative to iteration N is advantageously performed in the opposite manner to the optimization relative to iteration N-1 in iteration N.

[0036] Preferably, steps d) to i) are performed iteratively, especially repeatedly, at such a frequency until the optimization information meets a preset condition. Such a preset condition can be understood as an interruption condition.

[0037] For example, the optimization information includes optimization values, and steps d) to i) are performed iteratively, especially repeatedly, at such a frequency until the optimization values ​​are below a preset limit value. This limit value is preferably selected or preset such that the VMD has sufficiently good quality below this limit value, or the VTF is optimized to such an extent that a VMD with sufficiently good quality can be output.

[0038] Preferably, the VMD describes a B1 field, especially a B1 map, and / or a B0 field, especially a B0 map, and / or a bias field, especially a bias map. In particular, the VMD can be a B1 map and / or a B0 map and / or a bias map. The bias field can especially be a field describing intensity disturbances and / or intensity variations caused by inhomogeneities in B1 and / or B0.

[0039] The B1 plot can be particularly regarded as representative of the two-dimensional or three-dimensional spatial distribution of the B1 field. The B0 plot can particularly be regarded as representative of the two-dimensional or three-dimensional spatial distribution of the B0 field. The distortion plot can particularly be regarded as representative of the two-dimensional or three-dimensional spatial distribution of the distortion field.

[0040] Preferably, VMD describes the B1 map and / or distortion map, and the assumptions upon which step g) is based relate to the smoothness of the B1 map or distortion map. Preferably, it is assumed that the B1 map or distortion map is as smooth as possible, especially as edgeless as possible or with the weakest possible edges and / or with low spatial resolution. For example, a metric for spatial resolution can be determined empirically, which can be used as a (target) assumption. Advantageously, the OTF is trained by this assumption to suppress possible anatomical information contained in the input data, i.e., image data, in the output data, i.e., magnetic field data.

[0041] Preferably, determining the optimization information involves relatively weighting the data fidelity information and at least one hypothesis fidelity information. The weighting, and in particular the possible weighting coefficients describing the weighting, can be determined empirically, for example.

[0042] In particular, the optimization information may include an optimization value, which is a weighted sum of a data fidelity value and at least one assumed fidelity value, wherein the coefficients of the sum represent weighting coefficients, for example, OW = DTW + ∑ n c n ATW n Where OW is the optimization value, DTW is the data fidelity value, and C is the data fidelity value. n It is a weighted coefficient and ATW n It is the assumed fidelity value.

[0043] The iterative optimization of the VTF to generate the OTF can thus be represented, for example, as an optimization problem min(OW), which is solved by iteratively adjusting the VTF, especially its weights or edge weights, to minimize the optimization value.

[0044] Preferably, the TBD and / or KTBD is compressed, particularly reduced-sampled image data. Therefore, the amount of data associated with the image data can be reduced, which in turn reduces the computational cost of generating the OTF. Image data compression may, in particular, include reducing the spatial resolution of the image data.

[0045] For example, obtaining the original training image data, i.e., TBD i TBD is generated from these original training image data through compression. Furthermore, it can be derived from the TBD. i Generate the original, corrected training image data, i.e., KTBD. i Then, KTBD can be generated by compressing the original corrected training image data.

[0046] Preferably, the VTF is based on neural networks, especially convolutional neural networks (CNNs), and especially U-Net.

[0047] U-Net preferably includes a network with contraction and expansion paths. In the contraction path, spatial information is typically reduced while feature information is increased, and in the expansion path, feature and spatial information are combined. This can advantageously improve the resolution of the output data.

[0048] Preferably, in order to optimize the VTF, especially in order to generate the OTF, the weights, especially the edge weights, of the neural network on which the VTF is based are changed.

[0049] Preferably, the generation of the VKBD involves multiplying the B1 map described by the VMD with the magnetic resonance image described by the TBD. Advantageously, this multiplication can correct for image artifacts that may exist in the magnetic resonance imaging caused by inhomogeneities in the B1 field (described by the B1 map).

[0050] Preferably, determining the data fidelity information includes applying a cost function, particularly an L1 cost function, to VKBD and KTBD. Determining the data fidelity information specifically includes calculating a comparison value between VKBD and KTBD, particularly the L1 loss. For example, DTW can be represented as follows, particularly as a penalty term (Strafterm) in an optimization method for generating OTF:

[0051] DTW = VKBD - KTBD1

[0052] Preferably, the determination of the at least one hypothesis fidelity information includes generating k-space data kRD based on the B1 graph and / or distortion graph described by the VMD, wherein the generation of the kRD includes Fourier transforming the B1 graph and / or distortion graph to k-space, wherein comparative k-space data is generated based on the kRD, and wherein the determination of the at least one hypothesis fidelity information includes applying a cost function, especially an L1 cost function, to the kRD and the comparative k-space data. For example, ATW can be represented as follows, especially as another penalty term in an optimization method for generating OTF:

[0053] ATW = kRD - VkRD1

[0054] The B1 diagram can be advantageously transformed into the k-space using a Fourier transform. The k-space can, in particular, be a position-frequency space. The k-space can be described, in particular, using a two-dimensional data model.

[0055] The Fourier transform can be, in particular, a two-dimensional Fourier transform. The Fourier transform can also be, in particular, a Fast Fourier Transform (FFT).

[0056] Preferably, the generation of VkRD includes setting the value of kRD in at least one segment of the k-space to zero, wherein the at least one segment is located outside a predetermined, particularly central, segment of the k-space.

[0057] Advantageously, VkRD generated in this way is suitable for determining the smoothness of B1 plots or distortion plots by applying a cost function, especially the L1 cost function, as the assumed fidelity value ATW. This allows VTF to be trained in particular to suppress interfering information about the anatomical structure of the subject being examined to provide TBD during VMD generation.

[0058] For example, the k-space includes 128x128 points, wherein the preset segment includes 32x32 points located at the center of the k-space, and the values ​​of the k-space points located outside the central segment are set to zero.

[0059] Furthermore, a computer-implemented method is proposed for providing magnetic field data, particularly B1 and / or B0 and / or distortion maps, using a trained function, the method comprising:

[0060] - Receive image data as input data for a trained function.

[0061] - Apply the trained function to the image data,

[0062] The trained function is based on

[0063] i) Data fidelity of image data corrected based on magnetic field data

[0064] and

[0065] ii) At least one assumption about at least one characteristic of the magnetic field data

[0066] Trained

[0067] - Provides magnetic field data as output data for a trained function.

[0068] Advantageously, the trained function can provide magnetic field data quickly and reliably. Furthermore, the typical computation time for the entire (three-dimensional) 3D volume is only a few seconds, while traditional algorithms typically require several minutes.

[0069] Preferably, the trained function is an optimized trained function OTF, which is generated using the previously described method for generating OTFs. The potential advantages of this method can be transferred to methods for applying the trained function.

[0070] The reception of image data can be performed, in particular, by a suitable receiving unit and / or interface. The application of OTF to the input data can be performed, in particular, by a suitable computing unit. Such a computing unit may include one or more processors and / or memory modules. Such a computing unit may, in particular, be part of the system control unit of a magnetic resonance imaging (MRI) device.

[0071] Advantageously, the trained function only requires image data as input, thus eliminating the need for potential additional magnetic resonance measurements, especially predictors and / or calibration measurements. Furthermore, the application of the OTF is decoupled from the reconstruction of the magnetic resonance imaging.

[0072] Advantageously, control data for performing magnetic resonance measurements is determined based on the provided magnetic field data. Preferably, magnetic resonance measurements are performed using the determined control data.

[0073] Control data can be applied, for example, to output pTx pulses via a magnetic resonance device. Control data for multiple transmitting coils in a transmitting coil device specifically includes, respectively, the shape and / or amplitude and / or phase of the partial pulses and / or the time delay between partial pulses and / or the number of partial pulses. For example, an emissible HF transmitting pulse consists of multiple partial pulses that are distinct from each other and can be emitted separately by the transmitting coils of a multi-channel transmitting coil device; that is, the emissible HF transmitting pulse is a pTx pulse. Furthermore, a pTx pulse may also include at least one gradient pulse. The sum of such partial pulses, or at least one gradient pulse, of a multi-channel pulse can be particularly described by the control data.

[0074] The generated B1 field can be advantageously controlled more precisely by pTx pulses; this control may be particularly advantageous in applications with a narrowed field of view, shaped saturation bands, or for reducing specific absorption rate (SAR). In particular, pTx pulses can compensate for magnetic field inhomogeneities (e.g., in the range of "HF homogenization"), which may be especially advantageous at higher main magnetic field strengths above 7 Tesla.

[0075] At least one shape and / or amplitude and / or phase of an HF transmit pulse or a portion of a pulse may correspond, for example, to the shape and / or amplitude and / or phase of a voltage pulse applied to the transmitting coil device and / or a current pulse flowing through the transmitting coil device.

[0076] At least one shape and / or amplitude and / or duration of the gradient pulse may correspond, for example, to the shape and / or amplitude and / or duration of a voltage pulse applied to the gradient coil unit and / or a current pulse flowing through the gradient coil unit.

[0077] Non-uniformity in the B1 field can particularly cause interfering signal and contrast variations. Advantageously, these variations are corrected during magnetic resonance measurements. For this purpose, an appropriate pTx pulse is advantageously used, determined by means of magnetic field data determined via OTF.

[0078] Furthermore, a system control unit is proposed, designed to perform the previously described method for providing magnetic field data. A magnetic resonance apparatus having such a system control unit is also proposed.

[0079] The advantages of the proposed system control unit and magnetic resonance device substantially correspond to the advantages of the proposed method for providing magnetic field data, which have been detailed above. The features, advantages, or alternative embodiments mentioned herein can also be applied to other claimed technical solutions and vice versa.

[0080] Furthermore, a computer program product is proposed, comprising a program capable of being directly loaded into the memory of the (programmable) system control unit of a magnetic resonance imaging (MRI) device, and having program means, such as libraries and auxiliary functions, to perform the proposed method when the computer program is executed in the system control unit of the MRI device. Here, the computer program product may include software with source code that still needs to be compiled and bound, or only needs to be interpreted, or may include executable software code that only needs to be loaded into the system control unit for execution.

[0081] This computer program product enables the rapid, repeatable, and robust execution of the suggested methods. The computer program product is configured to allow the system control unit to execute the suggested method steps. Advantageously, the system control unit possesses preconditions such as corresponding working memory, a corresponding graphics card, or a corresponding logic unit, thus enabling efficient execution of the corresponding method steps. The computer program product is stored, for example, on a computer-readable medium, or on a network or server, where it can be loaded into the processor of the local system control unit, which may be directly connected to the magnetic resonance imaging (MRI) device or designed as part of the MRI device.

[0082] Furthermore, control information for a computer program product can be stored on an electronically readable data carrier. This control information on the electronically readable data carrier can be designed such that, when the data carrier is used in the system control unit of the magnetic resonance imaging (MRI) device, the control information performs the suggested methods. Examples of electronically readable data carriers are DVDs, magnetic tapes, or USB flash drives, on which electronically readable control information, particularly software, is stored. When this control information is read from the data carrier and stored in the system control unit of the MRI device, all embodiments of the invention described above can be performed. Therefore, the invention can also be carried out from the computer-readable medium and / or the electronically readable data carrier. Attached Figure Description

[0083] Other advantages, features, and details of the invention arise from the embodiments described below and from the accompanying drawings. Corresponding parts are provided with the same reference numerals in all the drawings.

[0084] In the attached diagram:

[0085] Figure 1 A schematic diagram of a magnetic resonance device is provided.

[0086] Figure 2 The method for generating an optimized, trained function to provide magnetic field data is illustrated in the diagram.

[0087] Figure 3 This illustrates different possible aspects of methods for generating optimized, trained functions;

[0088] Figure 4 The diagram illustrates a method for providing magnetic field data via a trained function. Detailed Implementation

[0089] exist Figure 1 The image schematically shows a magnetic resonance imaging (MRI) device 10. The MRI device 10 includes a magnet unit 11 having a main magnet 12 for generating a strong and, particularly, time-constant, main magnetic field 13. The main magnetic field 13 can also be referred to as the BO field. Furthermore, the MRI device 10 includes a patient receiving area 14 for accommodating a patient 15. The examination area of ​​the MRI device 10 is located at the center of the patient receiving area 14, in which the main magnetic field 13 has particularly high homogeneity. In this embodiment, the patient receiving area 14 is designed as cylindrical and is cylindrically surrounded by the magnet unit 11 in the circumferential direction. However, in principle, a different design for the patient receiving area 14 is always conceivable. The patient 15 can be pushed into the patient receiving area 14 via a patient support device 16 of the MRI device 10. For this purpose, the patient support device 16 has a patient table 17 movably designed within the patient receiving area 14.

[0090] The magnet unit 11 also includes a gradient coil unit 18 for generating a magnetic field gradient, which is used for position encoding during imaging. The gradient coil unit 18 is controlled by a gradient control unit 19 of the magnetic resonance apparatus 10. The magnet unit 11 also includes a high-frequency antenna unit 20, which in this embodiment is designed as a body coil fixedly integrated into the magnetic resonance apparatus 10. The high-frequency antenna unit 20 is controlled by a high-frequency antenna control unit 21 of the magnetic resonance apparatus 10. The high-frequency antenna unit 20 includes a transmitting coil assembly and a receiving coil assembly, which are identical in this case; that is, the coil that emits HF transmit pulses is also the coil that receives magnetic resonance signals. The transmitting field is commonly referred to as B1. + The field and the receiving field are called B1 - A magnetic field 13, generated by the main magnet 12, is radiated into the inspection area. The magnetic resonance signal is generated through the relaxation of the excited nuclei. The high-frequency antenna element 20 is designed to receive the magnetic resonance signal.

[0091] Especially in magnetic resonance imaging (MRI) devices with a high-intensity main magnetic field 13, such as 7 Tesla, the transmitting and receiving coil devices are typically not fixedly integrated into the body coils of the MRI device 10, but rather arranged directly on the patient 15 as localized transmitting and receiving coil devices (not shown here). Such transmitting and receiving coil devices particularly include multiple transmitting and / or receiving channels, thus they are adapted for parallel transmitting and / or receiving.

[0092] To control the main magnet 12, the gradient control unit 19, and the high-frequency antenna control unit 21, the magnetic resonance apparatus 10 has a system control unit 22. The system control unit 22 centrally controls the magnetic resonance apparatus 10, such as executing a predetermined imaging gradient echo sequence. The system control unit 22 is preferably designed to, according to... Figure 4 A method for providing magnetic field data is executed using a trained function. Furthermore, the system control unit 22 includes an analysis unit (not shown in detail) for analyzing the magnetic resonance signals acquired during the MRI examination. Additionally, the MRI apparatus 10 includes a user interface 23 connected to the system control unit 22. Control information, such as imaging parameters, and reconstructed MRI images can be displayed to a medical operator on a display unit 24 of the user interface 23, for example, on at least one monitor. Furthermore, the user interface 23 has an input unit 25 through which the medical operator can input information and / or parameters during measurement.

[0093] Figure 2This diagram schematically illustrates a computer-implemented method for generating an optimized, trained function OTF to provide magnetic field data. In S110, training image data TBD is provided. In S120, corrected training image data KTBD generated based on TBD is provided. In S130, a preliminary trained function VTF is provided. In S140, preliminary magnetic field data VMD is generated by applying TBD to the VTF. For example, VMD can describe a B1 field, especially a B1 diagram, and / or a B0 field, especially a B0 diagram, and / or a distortion field. In S150, preliminary corrected image data VKBD is generated based on VMD and TBD. In S160, data fidelity information is determined based on VKBD and KTBD. In S170, at least one hypothesis fidelity information is determined based on VKBD and assumptions about VMD. In S180, optimization information is determined based on the data fidelity information and at least one hypothesis fidelity information. In S185, it is checked whether the optimization information satisfies preset conditions. If the preset conditions are met, the method ends in END, thus completing the optimization of VTF. VTF is then the OTF. If the preset conditions are not yet met, the optimization continues by changing VTK in S190. The VTK changed in S190 is the (new) VTK from S130. In further iterations, steps S130 to S180 are executed again to check if the preset conditions are met again in S185. Therefore, multiple iterations are performed as needed to generate the OTF.

[0094] Reference Figure 3 The different aspects of possible embodiments should be described in more detail. Raw training image data TBD is provided. i D1. These TBDs i D1 can be derived, for example, by magnetic resonance measurement recordings using magnetic resonance device 10, or by data recorded using magnetic resonance device 10, especially by calculation.

[0095] By TBD i D1 generates training image data TBD by compression, or more precisely, by reducing the resolution. D2. Here, the resolution is reduced to 128x128 pixels (downsampling). i Compression of D1 can reduce the computational load required in subsequent processes (especially S140, S150, S160, S170 and S180) because less data needs to be processed; however, compression is only optional.

[0096] TBDD2 is provided according to S110 and serves as input data for VTFD4 provided according to S130. For example, VTFD4 is a deep learning network based on U-Net. By applying VTFD4 to TBD, VMD is generated according to S140. These VMDs are plots of the distorted field D5 describing the inhomogeneity of the B1 field.

[0097] According to S150, VKBDD6 is generated based on VMD and TBDD2. To do this, the graph of the distortion field D5 is multiplied by TBDD2 to compensate for the inhomogeneity of the B1 field in VKBDD6. According to S160, data fidelity information D9 is determined based on VKBDD6 and KTBDD8. To generate KTBDD8, TBD is first corrected for inhomogeneity in the B1 field. i D1, for example, uses correction methods such as Statistical Parametric Mapping (SPM), Nonparametric Non-uniform Intensity Normalization (N3), N4ITK, and / or Uniform Combination Reconstruction (UNICORN) to obtain the original corrected training image data KTBD. i D7. Then with TBD i Similarly, compress KTBD i D7, which is compressed to a resolution of 128x128 pixels to obtain KTBDD8. (Alternatively, compression could be performed first, followed by correction). KTBD i Compression is also optional.

[0098] Here, data fidelity information D9 is determined by applying a cost function to VKBD and KTBD, where the cost (or loss) determined in this process is the data fidelity information D9. The L1 cost function is particularly suitable for this purpose. The closer VKBDD6 corresponds to KTBDD8, the lower the cost value. Therefore, KTBDD8 is used as a reference standard or true value data relative to VKBDD6.

[0099] On the other hand, according to S170, at least one hypothesis fidelity information D12 is determined based on VKBD and the assumptions about VMD. To this end, a Fast Fourier Transform (FFT) is performed on the graph of the distortion field D5 to obtain k-space data kRD, D10. Therefore, the graph of the distortion field D5 is transformed from image space to k-space.

[0100] From kRDD10, we derive the comparative k-space data VkRD, D11. This is achieved here by setting all k-space values ​​outside the central region of the k-space to zero. In this case, the size of the central region is a 32x32 k-space value. Assume the fidelity information D12 is the cost of the cost function applied to kRDD10 and VkRDD11. The L1 cost function is particularly applicable here. Therefore, the high information content in the periphery of the k-space of kRD D10 is penalized.

[0101] The periphery of the k-space typically contains information about the edges in the associated image space, i.e., in this case, the edges in the graph of the distortion field D5. However, it is assumed that the distortion field D5 should not have edges; rather, edges are typical of the anatomical features of the object being examined and should not be reflected in the distortion field D5. Therefore, smoothness, i.e., "no edges," is an appropriate assumption for the distortion graph.

[0102] In this example, only one assumption was made about the distortion map, or only one assumption fidelity was determined. However, it is also possible to determine multiple assumption fidelity information.

[0103] According to S180, optimization information D13 is determined based on data fidelity information D9 and assumption fidelity information D12. Optimization information D13 can be, for example, a weighted sum of the cost of data fidelity information D9 and the cost of assumption fidelity information D12. For example, the weighting coefficients can be determined empirically.

[0104] The optimization objective can be, in particular, to minimize the optimization information D13 (or the optimized value describing the optimization information D13). Now, for example, we can check according to S185 whether the optimization information D13 (or the optimized value describing the optimization information D13) is now small enough. If so, the optimization is complete, and the VKF used at the end is the result of the optimization, i.e., the OTF.

[0105] If the optimization information D13 (or the optimization value describing the optimization information D13) is not small enough, then modify VKF and perform one or more other iterations with the modified VFK.

[0106] Figure 4 A computer-implemented method for providing magnetic field data is illustrated schematically. In S210, image data is received. In S220, the received image data is applied to a trained function, which is trained based on the fidelity of the image data corrected according to the magnetic field data and at least one assumption about at least one characteristic of the magnetic field data. Preferably, the trained function is based on... Figure 2The OTF of the generation method of 3 and / or 4 is used. In S230, magnetic field data is output and provided. In S240, control data for performing magnetic resonance measurements is determined based on the provided magnetic field data.

[0107] In S240, pTx pulses are designed, for example, using distortion maps (as output data of the OTF). These pTx pulses can be used, for example, for T1-weighted, T2-weighted, and / or diffusion-weighted magnetic resonance measurements. In particular, HF shimming can be performed using magnetic field data output from the OTF.

[0108] Finally, it should be reiterated that the methods described in detail above, as well as the system control unit and magnetic resonance device shown, are merely embodiments, and those skilled in the art can modify the embodiments in various ways without departing from the scope of the invention. Furthermore, the use of the indefinite article "a" does not preclude the possibility that a particular feature may exist multiple times. Similarly, the term "unit" does not preclude the possibility that a component may consist of multiple interacting sub-components, which may, if necessary, be spatially distributed.

Claims

1. A computer-implemented method for generating an optimized, trained function OTF to provide magnetic field data about the magnetic field in an examination region of a magnetic resonance apparatus, comprising the steps of: a) Provide training image data TBD, b) Provide KTBD training image data, corrected for the characteristics of the magnetic field generated according to TBD. c) Provide a preliminary trained function VTF. d) By applying TBD to VTF, preliminary magnetic field data VMD is generated. e) Generate preliminary VKBD image data on the characteristics of the magnetic field based on VMD and TBD. f) Determine data fidelity information based on VKBD and KTBD. in, The data fidelity information describes the similarity and / or data consistency and / or data fidelity between VKBD and KTBD. g) Based on VMD and at least one assumption about a characteristic of VMD, determine at least one hypothesis fidelity information. Wherein, the assumptions underlying step g) involve the smoothness of the B1 graph and / or distortion graph described by VMD. h) Determine optimization information based on data fidelity information and at least one hypothesis fidelity information. The determination of optimization information includes relatively weighting data fidelity information and at least one hypothesis fidelity information. i) Generate an OTF by optimizing the VTF based on the optimization information.

2. The method according to claim 1, in, Iteratively execute steps c) through i). The OTF generated in step i) is provided as the VTF in step c) for subsequent iterations.

3. The method according to claim 2, in, Steps d) through i) are executed iteratively until the optimization information meets the preset conditions.

4. The method according to claim 1, in, The TBD and / or KTBD are compressed image data.

5. The method according to claim 1, in, The VTF is based on a neural network. In order to optimize the VTF, the weights of the neural network on which the VTF is based are changed.

6. The method according to claim 5, in, The VTF is based on a convolutional neural network.

7. The method according to claim 6 above, in, The VTF is based on U-Net.

8. The method according to claim 1, in, The generation of the VKBD involves multiplying the B1 map described by VMD with the magnetic resonance image described by TBD.

9. The method according to claim 1, in, Determining the data fidelity information includes applying a cost function to VKBD and KTBD.

10. The method according to claim 9, in, The cost function is the L1 cost function.

11. The method according to claim 1, in, Determining the at least one hypothetical fidelity information includes generating k-space data kRD based on the B1 graph and / or distortion graph described by the VMD. The generation of kRD includes Fourier transforming the B1 map and / or the distorted map into the k-space. Specifically, comparison k-space data VkRD is generated based on the kRD. The determination of at least one hypothetical fidelity information includes applying a cost function to kRD and comparing k-space data.

12. The method according to claim 11, in, The cost function is the L1 cost function.

13. The method according to claim 11, in, The generation of the comparison k-space data includes setting the value of at least one segment of kRD to zero, wherein the at least one segment is located outside a preset segment of kRD.

14. A computer-implemented method for providing magnetic field data using a trained function, comprising: - Receive image data as input data to a trained function. - Apply the trained function to the image data. in, The trained function is based on i) Data fidelity of image data corrected based on magnetic field data and ii) At least one assumption about at least one characteristic of the magnetic field data Trained - Provides magnetic field data as output data for an optimized, trained function. The trained function is an optimized trained function OTF, which is generated by the method according to any one of claims 1 to 13.

15. A system control unit for a magnetic resonance apparatus, the system control unit being designed to perform the method according to claim 14.

16. A computer program product comprising a program and capable of being directly loaded into the memory of a programmable system control unit of a magnetic resonance imaging apparatus, having program instructions for performing the method of claim 14 when the program is executed in the system control unit of the medical imaging apparatus.

Images