Correction of intensity inhomogeneity in a magnetic resonance image

Abstract

The invention relates in particular to a method for generating an intensity-corrected magnetic resonance image. Reference measurement data are provided, wherein at least a portion of the reference measurement data is generated from magnetic resonance signals acquired with at least one receiving antenna or at least one local coil using a magnetic resonance device according to a reference magnetic resonance sequence. Image measurement data is also provided, which is generated from magnetic resonance signals acquired with the at least one receiving antenna or at least one local coil using the magnetic resonance device according to an image magnetic resonance sequence. A function trained by a machine learning algorithm is applied to the reference measurement data and the image measurement data as input data for the trained function.Correction data is generated from the output data of the trained function. This correction data describes a measurement-induced intensity inhomogeneity in the image measurement data. An intensity-corrected magnetic resonance image is then generated based on the image measurement data and the correction data.

Description

[0001] The invention relates to a method for generating an intensity-corrected magnetic resonance image, a method for providing a trained function for outputting correction data, a generation unit for generating an intensity-corrected magnetic resonance image and a computer program product.

[0002] In medical technology, magnetic resonance imaging (MRI), also known as magnetic resonance tomography, is characterized by high soft tissue contrast. In this procedure, the subject, usually a patient, is positioned within the examination chamber of a magnetic resonance scanner. During a magnetic resonance measurement, high-frequency (RF) pulses are typically emitted into the subject using a transmitting coil arrangement within the scanner. Additionally, gradient pulses are generated by a gradient coil within the scanner, creating temporary magnetic field gradients within the examination chamber. These pulses excite nuclear spins in the subject. Relaxation of these excited nuclear spins triggers spatially encoded magnetic resonance signals.The triggered magnetic resonance signals are received by a receiving coil arrangement, which may include one or more receiving antennas, the magnetic resonance device, and used to reconstruct magnetic resonance images.

[0003] Measurement-related, particularly acquisition-related and / or instrument-related, intensity inhomogeneities in magnetic resonance imaging (MRI) are a known yet unsatisfactorily solved problem. These inhomogeneities can be caused, among other things, by spatially varying sensitivities of the receiving antennas used to measure the MRI signals. Such intensity inhomogeneities can complicate accurate diagnosis. The aim is therefore to compensate for such inhomogeneities using suitable methods or at least to reduce them as much as possible.

[0004] The publication Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, Gee JC. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010 Jun;29(6):1310-20. doi: 10.1109 / TMI.2010.2046908. Epub 2010 Apr 8. PMID: 20378467; PMCID: PMC3071855 discloses a method for correcting image inhomogeneities.

[0005] The object of the present invention can be considered to be the generation of magnetic resonance images that exhibit the lowest possible measurement-induced intensity inhomogeneity. This object is achieved by the features of the independent claims. Advantageous embodiments are described in the dependent claims.

[0006] A computer-implemented method for generating an intensity-corrected magnetic resonance image is proposed. Reference measurement data are provided, with at least a portion of the reference data being generated from magnetic resonance signals acquired by at least one receiving antenna and at least one local coil using a magnetic resonance device according to a reference magnetic resonance sequence. Additionally, image measurement data are provided, generated from magnetic resonance signals acquired by the at least one receiving antenna and at least one local coil using the magnetic resonance device according to an image magnetic resonance sequence. A function trained using a machine learning algorithm is applied to the reference measurement data and the image measurement data as input data for the trained function.Correction data is generated from the output data of the trained function. This correction data describes a measurement-induced intensity inhomogeneity in the image measurement data. An intensity-corrected magnetic resonance image is then generated based on the image measurement data and the correction data.

[0007] The recorded magnetic resonance signals are preferably magnetic resonance signals that were recorded using the magnetic resonance device for the purpose of examining an object of study, in particular a patient.

[0008] The portion of the reference measurement data generated from magnetic resonance signals recorded with the at least one receiving antenna of the at least one local coil using the magnetic resonance device according to the reference magnetic resonance sequence is hereinafter also referred to as local coil reference measurement data.

[0009] Preferably, the at least one receiving antenna of the at least one local coil comprises several receiving antennas. The multiple receiving antennas can be distributed across several local coils. Preferably, the at least one local coil comprises several local coils, each with one or more receiving antennas. For example, the at least one local coil comprises two local coils, each with one receiving antenna.

[0010] Especially when multiple receiving antennas are present, stronger measurement-related intensity inhomogeneities can occur, making the correction of the magnetic resonance imaging particularly effective.

[0011] The measurement-related intensity inhomogeneity of the image measurement data can be caused in particular by an inhomogeneous distribution of a B1 field, which is present when recording the magnetic resonance signals according to the reference magnetic resonance sequence and / or the image magnetic resonance sequence.

[0012] During recording, a magnetic alternating field, a so-called B1, is usually generated by an RF pulse, in particular an RF transmission pulse. + -field, generated in an examination area where the patient is located during the magnetic resonance measurement. The B1 + The -field typically represents a distribution of the transmit field. In contrast, a B1--field typically describes a distribution of a receive field. The effective transmit-receive field can also be referred to as the B1-field.

[0013] Preferably, the method takes place in spatial space. The reference measurement data, image measurement data, and / or the magnetic resonance imaging are preferably spatial data (data in spatial space, in particular a spatial imaging representation). Spatial space can also be referred to as image space. Spatial space differs in particular from k-space, which can also be referred to as frequency space.

[0014] The spatial space preferably describes a space in which a spatial distribution of a nuclear magnetic resonance is represented. The spatial space can be viewed as the result of a transformation of the measured magnetic resonance signals in k-space (raw data), which contain the frequency components of the signals. A transformation from spatial space to k-space (and vice versa) can be performed, for example, using a Fourier transform.

[0015] Preferably, the spatial space is a real, in particular geometric, space. The spatial space can, for example, be defined by a two- or three-dimensional coordinate system whose coordinate axes are spatial axes, such as an x-, y-, and / or z-axis. These are, for example, orthogonal axes. The spatial space can be described by a coordinate system that specifies the position and orientation of the object under investigation in the magnetic field of the magnetic resonance device.

[0016] Reference measurement data and image measurement data are, for example, data that have been transformed from the magnetic resonance signals (as raw data) into spatial space by means of a transformation, such as a Fourier transform. The reference measurement data can, in particular, include at least one reference image and / or at least one reference image. The image measurement data can, in particular, include at least one (diagnostic) image and / or at least one (diagnostic) image.

[0017] Preferably, the reference measurement data or the underlying magnetic resonance signals are acquired during one or more preliminary measurements prior to the (actual) main measurement. Preferably, the image measurement data or the underlying magnetic resonance signals are acquired during the (actual) main measurement.

[0018] The reference magnetic resonance sequence is preferably a magnetic resonance sequence different from the image magnetic resonance sequence. A magnetic resonance sequence can, in particular, comprise a temporal sequence of magnetic resonance pulses generated by the magnetic resonance device. Possible magnetic resonance pulses are RF pulses and / or gradient pulses. The image magnetic resonance sequence is, for example, a T2-weighted spin-echo sequence.

[0019] The at least one local coil is preferably part of a receiving coil arrangement of the magnetic resonance device. The at least one local coil can comprise several local coils. The receiving antennas comprise, for example, electrically conductive loops suitable for receiving the magnetic resonance signals. A local coil can, for example, comprise one, two, four, eight, 16, or 32 receiving antennas. The local coil is designed to be positioned directly next to the subject of the examination, in particular a patient, from whom the magnetic resonance image is generated. Advantageously, the patient-proximity arrangement of the local coil allows for the reception of magnetic resonance signals with a high signal-to-noise ratio.

[0020] If the at least one receiving antenna comprises multiple receiving antennas, these are preferably spatially distributed across the at least one local coil, i.e., they are located at different points within the at least one local coil. The multiple receiving antennas typically exhibit different sensitivities, particularly sensitivity distributions, for receiving the magnetic resonance signals. The sensitivity of a receiving antenna can depend, in particular, on its position and / or orientation relative to the patient and / or magnetic field. The different sensitivities, and especially sensitivity distributions, of the multiple receiving antennas can lead, in particular, to measurement-induced intensity inhomogeneity in the image measurement data. Advantageously, this measurement-induced intensity inhomogeneity in the image measurement data can be corrected using correction data.

[0021] The measurement-induced intensity inhomogeneity is preferably an intensity inhomogeneity caused by specific conditions during the acquisition of the magnetic resonance signals according to the reference magnetic resonance sequence and / or the image magnetic resonance sequence. These conditions may be due, in particular, to the patient being examined, who is depicted by the magnetic resonance imaging. Specifically, the measurement-induced intensity inhomogeneity may include an anatomically caused intensity inhomogeneity, particularly due to the patient's anatomy. The measurement-induced intensity inhomogeneity may, in particular, be caused by a specific loading of the at least one receiving antenna, especially due to the patient's specific anatomy, during the acquisition of the magnetic resonance signals according to the reference magnetic resonance sequence and / or the image magnetic resonance sequence.Furthermore, the measurement-related intensity inhomogeneity may be caused by the measurement technique used for recording the magnetic resonance signals according to the reference magnetic resonance sequence and / or the image magnetic resonance sequence, in particular the type and arrangement of the local coil on the patient and / or other measurement-related factors.

[0022] The correction data can, in particular, include a correction map. The correction map can, in particular, include a spatial distribution of correction values, especially correction factors. Advantageously, the measurement-related intensity inhomogeneity can be compensated for using the correction data.

[0023] Preferably, the correction data is in the form of a matrix, in particular a correction matrix. Preferably, the image measurement data is also in the form of a matrix, in particular an image matrix. Preferably, each element of the image matrix corresponds to a voxel of the intensity-corrected magnetic resonance image. Preferably, the size of the correction matrix is ​​equal to the size of the image matrix. Preferably, each element of the correction matrix has a value, in particular a scaling factor, which is multiplied by the corresponding value of the image matrix when generating the intensity-corrected magnetic resonance image.

[0024] The machine learning algorithm is preferably based on statistical algorithms, in particular learning algorithms. Preferably, the machine learning algorithm maps predefined training data to a mathematical model and adapts it to the training data in such a way that it can generalize from it to new cases. This process can be called training. After training, the solution path found is advantageously stored in the model. The trained model can make predictions for new data, in particular reference measurement data and / or image measurement data as input data, and in particular generate correction data as output data.

[0025] By using image measurement data in addition to reference measurement data as input data for the trained function, the trained function is provided with additional information. This information advantageously allows for a better differentiation between anatomically caused and measurement-related signal inhomogeneities. This allows measurement-related inhomogeneities to be corrected while leaving anatomically caused inhomogeneities unaffected. It also reduces the risk of, for example, spatially slow fluctuations in image intensity, which are actually anatomically caused, being incorrectly identified and (incorrectly) corrected as inhomogeneities caused by the imaging technique.

[0026] One embodiment of the method provides that at least a (further) part of the reference measurement data is generated from magnetic resonance signals acquired with at least one receiving antenna of a body coil, in particular a whole-body coil, permanently installed in the magnetic resonance device. In contrast to the local coil reference measurement data, the reference measurement data acquired with the at least one receiving antenna of the body coil permanently installed in the magnetic resonance device will hereinafter also be referred to as body coil reference measurement data.

[0027] Advantageously, body coils permanently installed in the magnetic resonance device exhibit a largely homogeneous reception field. Advantageously, the size, geometry, and / or arrangement of the body coil's antennas and / or the placement of such a body coil in the magnetic resonance device result in a largely homogeneous reception field.

[0028] The body coil reference measurement data can advantageously be used as reference data for a homogeneous reception field. A sensitivity distribution of the at least one receiving antenna of the at least one local coil can advantageously be derived from a comparison of the local coil reference measurement data with the body coil reference measurement data. In particular, information about the respective reception field of the at least one receiving antenna of the at least one local coil can be obtained from this.

[0029] One embodiment of the method provides that the reference magnetic resonance sequence is designed to produce a neutral contrast, in particular image contrast, of the reference measurement data.

[0030] Neutral contrast, in this context, refers specifically to a non-relaxation-dominated contrast, particularly image contrast. Specifically, neutral contrast is an image contrast with minimal special weighting, such as T1 or T2 weighting. Advantageously, measurement data acquired with neutral contrast exhibits the characteristic that intensity differences in images acquired with different receiving antennas are primarily attributable to antenna sensitivities and / or other measurement-related effects.

[0031] Preferably, the neutral contrast is a contrast that exhibits proton density weighting. Advantageously, in a proton density-weighted magnetic resonance imaging (MRI) image, the contrast is primarily influenced by the proton density of the imaged tissue. T1 and T2 effects are advantageously suppressed.

[0032] One embodiment of the method provides that the image measurement data includes an image coordinate system (particularly in spatial space) and that initial reference measurement data, in particular initial local coil reference measurement data and / or body coil reference measurement data, are provided, which also includes a reference coordinate system (particularly in spatial space). To provide, and in particular generate, the reference measurement data, the initial reference measurement data are transformed from the reference coordinate system into the image coordinate system. Preferably, this is a transformation between two spatial spaces.

[0033] In particular, the reference measurement is transformed into the same spatial arrangement as the image measurement. Advantageously, the reference measurement data covers a comparatively large volume, and any slices of the reference measurement data are often aligned along the grandial axes. For the image magnetic resonance sequence (the actual imaging sequence), a specific slice orientation and / or geometry tailored to the patient's anatomy and position is advantageously used. To correct the image measurement data acquired in this way, the reference measurement data is advantageously transformed into the corresponding coordinate system of the image measurement data. Such a transformation can, for example, include matrix multiplication, interpolation of the reference measurement data, and / or, in particular, volumetric cropping of the reference measurement data.

[0034] One embodiment of the method provides that the reference measurement data, in particular the local coil reference measurement data, are combined data derived from individual measurement data, each acquired (individually) by the magnetic resonance device according to the reference magnetic resonance sequence using the at least one receiving antenna of the at least one local coil. Advantageously, combined local coil reference measurement data, in particular a combined local coil reference image, enable a simpler, and in particular less computationally intensive, generation of the correction data and / or the intensity-corrected magnetic resonance image.

[0035] One embodiment of the method provides that the trained function is based on a neural network, in particular a convolutional neural network (CNN). Advantageously, convolutional neural networks are particularly well suited for processing image data, such as the reference measurement data and / or the image measurement data.

[0036] In particular, the trained function is based on a U-Net. A U-Net preferably comprises a network with a contraction path and an expansion path. In the contraction path, spatial information is typically reduced while feature information is increased, and in the expansion path, feature information and spatial information are recombined. Advantageously, this can increase the resolution of the output data.

[0037] Furthermore, a computer-implemented method for providing a trained function for outputting correction data is proposed. Training reference measurement data is provided, wherein at least a portion of the training reference measurement data, in particular local coil training reference measurement data, is generated from magnetic resonance signals acquired with at least one receiving antenna of at least one local coil using a magnetic resonance device. Optionally, body coil training reference measurement data is also provided as a further portion of the training reference measurement data, wherein the body coil training reference measurement data is generated from magnetic resonance signals acquired with at least one receiving antenna of a body coil permanently installed in the magnetic resonance device.Furthermore, training image measurement data is provided, wherein the training image measurement data is generated from magnetic resonance signals acquired by the magnetic resonance device using the at least one receiving antenna of the at least one local coil. Additionally, training target image measurement data (ground truth data) is provided, wherein the training target image measurement data is training image measurement data corrected by a correction algorithm. The trained function is trained based on the training reference measurement data, in particular the local coil training reference measurement data and optionally the body coil training reference measurement data, and the training image measurement data (as training input data) and the training target image measurement data (as training output data). The function for outputting correction data, thus trained, is then provided.

[0038] Preferably, training the trained function includes, in particular iteratively, minimizing a loss function depending on the provided training reference measurement data, the provided training image measurement data, and the provided training target image measurement data.

[0039] The loss function can, for example, take the following form: Loss function = Mean (estimated correction data * training image measurement data - training target image measurement data) 2 The loss function, for example, represents the squared mean of estimated correction data multiplied by the training image measurement data minus the training target image measurement data.

[0040] The estimated correction data can, for example, include an estimated correction map. Preferably, the neural network first estimates a correction map. The training image measurement data preferably represent an uncorrected magnetic resonance image.

[0041] Preferably, this method for providing the trained function is supervised machine learning. The machine learning algorithm is presented with a dataset containing the training target image measurement data as already known target variables. Advantageously, the algorithm learns relationships and / or dependencies in the data that explain these target variables.

[0042] The provision of the training reference measurement data can be analogous to the provision of the reference measurement data. The provision of the training image measurement data can be analogous to the provision of the image measurement data. The features disclosed for the method for generating an intensity-corrected magnetic resonance image can be transferred accordingly to the training reference measurement data and / or the training image measurement data.

[0043] Possible correction algorithms include, for example, image filtering algorithms that determine inhomogeneities based on the measured training image data, or signal processing processes in which overview measurements are carried out with different receiving antennas and inhomogeneities and corresponding correction factors are determined from this.

[0044] For example, training target image measurement data is provided using the correction algorithm disclosed in the publication Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, Gee JC. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010 Jun;29(6):1310-20. doi: 10.1109 / TMI.2010.2046908. Epub 2010 Apr 8. PMID: 20378467; PMCID: PMC3071855. This is an image filtering algorithm that can determine and compensate for inhomogeneities based on the measured training image measurement data. The N4ITK algorithm replaces the B-spline smoothing component of the N3 algorithm with an alternative B-spline approximation method that is faster, more robust, and more flexible. Furthermore, they modify the iterative optimization scheme to allow for incremental updates of a bias field estimator.A multitude of other possible correction algorithms are known to those skilled in the art, so that the invention is not limited to the use of the above-mentioned correction algorithm.

[0045] Furthermore, a (previously described) computer-implemented method for generating an intensity-corrected magnetic resonance image is proposed, wherein the trained function is provided according to a (previously described) computer-implemented method for providing the trained function.

[0046] Furthermore, a generation unit for producing an intensity-corrected magnetic resonance image is proposed. This unit comprises a first input interface for receiving reference measurement data, wherein at least a portion of the reference measurement data is generated from magnetic resonance signals acquired by at least one receiving antenna and at least one local coil using a magnetic resonance device. The generation unit also comprises a second input interface for receiving image measurement data, wherein the image measurement data is generated from magnetic resonance signals acquired by the magnetic resonance device using at least one receiving antenna and at least one local coil. Finally, the generation unit includes an output interface for outputting correction data.Furthermore, the generation unit includes a calculation unit for applying a trained function to the reference measurement data and the image measurement data, whereby the correction data are generated as output data.

[0047] Optionally, the generating unit includes a third input interface for receiving additional reference measurement data generated from magnetic resonance signals recorded with at least one receiving antenna of a body coil permanently installed in the magnetic resonance device.

[0048] The advantages of the generation unit for producing an intensity-corrected magnetic resonance image essentially correspond to the advantages of the computer-implemented method for generating an intensity-corrected magnetic resonance image, which have been described in detail above. Features, advantages, or alternative embodiments mentioned here can also be transferred to the other claimed items and vice versa.

[0049] In other words, the claims in question can also be further developed with features described or claimed in connection with a method. The corresponding functional features of the method are thereby realized by corresponding physical modules, in particular by hardware modules.

[0050] Furthermore, a computer program product is proposed that comprises a program and can be directly loaded into the memory of a programmable generating unit for generating an intensity-corrected magnetic resonance image. The computer program product includes program resources, such as libraries and auxiliary functions, to execute a proposed method when run in the generating unit. The computer program product may comprise software with source code that still needs to be compiled and bound or interpreted, or executable software code that only needs to be loaded into the system control unit for execution.

[0051] The proposed method can advantageously be executed quickly, identically, and robustly by the computer program product. The computer program product is preferably configured such that it can execute the proposed method steps using the generation unit. The system control unit possesses the necessary prerequisites, such as sufficient main memory, a suitable graphics card, or a suitable logic unit, so that the respective method steps can be executed efficiently.

[0052] The computer program product is stored, for example, on a computer-readable medium or on a network or server, from where it can be loaded into the processor of a local generation unit. Furthermore, control information for the computer program product can be stored on an electronically readable data carrier. Examples of electronically readable data carriers include a hard drive, an SSD, a DVD, a magnetic tape, or a USB flash drive, on which electronically readable control information, in particular software, is stored. If this control information is read from the data carrier and stored in a generation unit, all proposed embodiments of the previously described methods for generating an intensity-corrected magnetic resonance image can be carried out.

[0053] Furthermore, a training system is proposed comprising a first training interface configured to receive training reference measurement data, wherein at least a portion of the training reference measurement data is generated from magnetic resonance signals acquired with at least one receiving antenna and at least one local coil using a magnetic resonance device. The training system further comprises a second training interface configured to receive training image measurement data, wherein at least a portion of the training image measurement data is generated from magnetic resonance signals acquired with the at least one receiving antenna and at least one local coil using the magnetic resonance device. The training system further comprises a third training interface configured to receive training target image measurement data, wherein the training target image measurement data is training image measurement data corrected by a correction algorithm.Furthermore, the training system includes a training computational unit configured to train the function based on the training reference measurement data and the training image measurement data as training input data, and the training target image measurement data as training output data. The training system also includes a fourth training interface configured to provide the trained function with the output of the correction data.

[0054] Further advantages, features, and details of the invention will become apparent from the exemplary embodiments described below and from the drawings. Corresponding parts are designated with the same reference numerals in all figures.

[0055] They show: Fig. 1 a magnetic resonance device with a generating unit for generating an intensity-corrected magnetic resonance image, Fig. 2 a diagram of a method for generating an intensity-corrected magnetic resonance image, Fig. 3 a diagram of an extended method for generating an intensity-corrected magnetic resonance image, Fig. 4. A diagram of a convolutional neural network for providing correction data. Fig. 5 a diagram of a procedure for providing a trained function for outputting correction data Fig. 6 a generation unit for generating an intensity-corrected magnetic resonance image.

[0056] In Fig. Figure 1 schematically depicts a magnetic resonance imaging (MRI) device 10. The MRI device 10 comprises a magnetic unit 11, which includes a main magnet 12 for generating a strong and, in particular, time-constant main magnetic field 13. The MRI device 10 also includes a patient receiving area 14 for receiving a patient 15. In the present embodiment, the patient receiving area 14 is cylindrical and is cylindrically surrounded in one circumferential direction by the magnetic unit 11. However, a different configuration of the patient receiving area 14 is conceivable. The patient 15 can be moved into the patient receiving area 14 by means of a patient positioning device 16 of the MRI device 10. For this purpose, the patient positioning device 16 has a patient table 17 that is movably designed within the patient receiving area 14.

[0057] The magnet unit 11 further comprises a gradient coil 18 for generating magnetic field gradients, which are used for spatial encoding during imaging. The gradient coil 18 is controlled by a gradient control unit 19 of the magnetic resonance device 10. The magnet unit 11 also includes a body coil 20, which is permanently integrated into the magnetic resonance device 10. The body coil 20 is controlled by a high-frequency antenna control unit 21 of the magnetic resonance device 10 and transmits high-frequency magnetic resonance sequences into an examination space, which is essentially formed by a patient acquisition area 14 of the magnetic resonance device 10. This excites atomic nuclei in the main magnetic field 13 generated by the main magnet 12. Relaxation of the excited atomic nuclei generates magnetic resonance signals. The body coil 20 is configured to receive the magnetic resonance signals.Furthermore, the magnetic resonance device comprises a local coil 26, which is arranged directly next to the patient 15. It includes several receiving antennas 27, which are designed to receive the magnetic resonance signals. The receiving antennas 27 are spatially distributed over the local coil and / or over the patient 15, and each has its own sensitivity for receiving the magnetic resonance signals. In particular, each of the several receiving antennas 27 can be assigned a receiving channel.

[0058] The magnetic resonance device 10 includes a system control unit 22 for controlling the main magnet 12, the gradient control unit 19, and the high-frequency antenna control unit 21. The system control unit 22 centrally controls the magnetic resonance device 10, for example, by performing a predetermined imaging gradient echo sequence. The system control unit 22 also includes a generation unit 100 for generating an intensity-corrected magnetic resonance image. The generation unit 100 can thus be part of the magnetic resonance device 100. However, it is also conceivable that the generation unit 100 is designed independently of the magnetic resonance device 10. For example, it is conceivable that the generation unit 100 is connected to the magnetic resonance device 10 via a network.

[0059] Furthermore, the magnetic resonance imaging (MRI) device 10 includes a user interface 23, which is connected to the system control unit 22. Control information, such as imaging parameters, as well as reconstructed MRI images, can be displayed on a display unit 24, for example, on at least one monitor, of the user interface 23 for medical personnel. The user interface 23 also has an input unit 25, by means of which information and / or parameters can be entered by the medical personnel during a measurement procedure. The x, y, and z axes define the spatial area in which the MRI device 10 and the patient 15 are located.

[0060] In Fig. 2 is a computer-implemented method for generating an intensity-corrected magnetic resonance image, which can be carried out in particular with the generating unit 100.

[0061] Reference measurement data are provided in S10. At least part of the reference measurement data is generated from magnetic resonance signals recorded by the multiple receiving antennas 27 of the local coil 26 using the magnetic resonance device 10 according to a reference magnetic resonance sequence. Preferably, the reference magnetic resonance sequence is designed to ensure a neutral contrast of the reference measurement data.

[0062] Image measurement data is provided in S20. The image measurement data is generated from magnetic resonance signals that were recorded with the multiple receiving antennas 27 of the local coil 26 using the magnetic resonance device 10 according to an image magnetic resonance sequence.

[0063] In S30, a trained function is applied to the reference measurement data and the image measurement data as input data for the trained function. The trained function is based on a machine learning algorithm.

[0064] In S40, correction data is provided as output data of the trained function. The correction data describes an intensity inhomogeneity of the image measurement data caused by measurement technology, in particular device-related (especially by the magnetic resonance device 10) and / or by the loading of the receiving antennas 27.

[0065] In S50, an intensity-corrected magnetic resonance image is generated based on the image measurement data and the correction data. The generated magnetic resonance image can be displayed, for example, by display unit 24.

[0066] In Fig. Figure 3 presents an extended method, and the additional possible aspects of the method are discussed below. In S1, magnetic resonance signals are recorded using the multiple receiving elements 27 of the local coil 26, from which local coil reference measurement data are generated as part of the reference measurement data. These magnetic resonance signals are recorded using a local coil reference magnetic resonance sequence.

[0067] In S2, 20 magnetic resonance signals are recorded using the body coil, from which local coil reference measurement data are generated as part of the reference data. These magnetic resonance signals are recorded using a body coil reference magnetic resonance sequence.

[0068] In S3, magnetic resonance signals are recorded using the multiple receiving elements 27 of the local coil 26, from which image measurement data are generated.

[0069] The generation of the measurement data (i.e., the reference measurement data, in particular the local coil reference measurement data and / or the body coil reference measurement data, and / or the image measurement data) can, in particular, comprise a transformation of the magnetic resonance signals (as raw data) in k-space into measurement data in spatial space. Preferably, the measurement data are in the form of image data. In particular, the measurement data each have several pixels, each of which is assigned an image value, in particular an intensity value. In particular, the pixels are located in spatial space.

[0070] Preferably, the magnetic resonance signals received by the respective receiving antennas 27 of the local coil are available as separate individual measurement data. Preferably, the individual measurement data are combined to form the reference measurement data, in particular the local coil reference measurement data; i.e., the reference measurement data, in particular the local coil reference measurement data, are combined data. If the body coil also has several receiving coils, the individual measurement data recorded separately by these coils can likewise be combined before further processing.

[0071] The acquisition of magnetic resonance signals in S1 and / or S2 can be performed, for example, as part of calibration measurements (prior to the acquisition of magnetic resonance signals in S3). Advantageously, the method for generating an intensity-corrected magnetic resonance image uses measurement data that is already being acquired during the magnetic resonance examination.

[0072] The provision of the reference measurement data in S10 includes the provision of the local coil reference measurement data in S11 and an (optional) provision of the body coil reference measurement data in S12.

[0073] The image measurement data provided in S20 have an image coordinate system. In particular, the reference measurement data provided in S11 and S12 represent initial (preliminary) reference measurement data that have a reference coordinate system. Advantageously, in S13 the initial reference measurement data are transformed from the reference coordinate system to the image coordinate system.

[0074] In Fig. Figure 4 shows an example of a convolutional neural network (CNN) on which the trained function used in S30 is based. In the illustrated embodiment, the convolutional neural network 200 comprises an input layer 210, a convolutional layer 211, a pooling layer 212, a fully connected layer 213, and an output layer 214. Alternatively, the convolutional neural network 200 can include multiple convolutional layers 211, multiple pooling layers 212, and multiple fully connected layers 213, as well as other layer types. The order of the layers can be chosen arbitrarily; typically, fully connected layers 213 are used as the last layers before the output layer 214.

[0075] In particular, within a convolutional neural network 200, the nodes 220, ..., 224 of a layer 210, ..., 214 can be considered arranged as a d-dimensional matrix or as a d-dimensional image. Specifically, in the two-dimensional case, the value of the node 220, ..., 224 indexed by i and j in the nth layer 210, ..., 214 can be denoted as x(n)[i,j]. However, the arrangement of the nodes 220, ..., 224 of a layer 210, ..., 214 typically has no influence on the computations performed within the convolutional neural network 200 itself, since these are determined solely by the structure and the weights of the edges.

[0076] In particular, a convolution layer 211 is characterized in that the structure and weights of the incoming edges form a convolution operation based on a certain number of kernels. Specifically, the structure and weights of the incoming edges are chosen such that the values ​​x(n) k of the nodes 221 of the convolution layer 211 can be expressed as a convolution x (n) k = K k * x (n-1) based on the values ​​x (n-1) The node 220 of the preceding layer 210 is calculated, where the convolution * in the two-dimensional case is defined as follows: x(n)k[i,j]=(Kk∗x(n-1))[i,j]=∑i'∑j'Kk[i',j']⋅x(n-1)[i−l, j−j'].

[0077] Here is the k-th kernel K ka d-dimensional matrix (in this embodiment a two-dimensional matrix) that is normally small compared to the number of nodes 220, ..., 224 (e.g., a 3x3 matrix or a 5x5 matrix). This implies, in particular, that the weights of the incoming edges are not independent, but are chosen such that they generate the said convolution equation. In particular, for a kernel that is a 3x3 matrix, there are only 9 independent weights (each entry of the kernel matrix corresponds to an independent weight), regardless of the number of nodes 220, ..., 224 in the respective layer 210, ..., 214. In particular, for a convolution layer 211, the number of nodes 221 in the convolution layer is equal to the number of nodes 220 in the preceding layer 210 multiplied by the number of kernels.

[0078] If the nodes 220 of the preceding layer 210 are arranged as a d-dimensional matrix, the use of multiple kernels can be interpreted as adding another dimension (referred to as the "depth" dimension), so that the nodes 221 of the convolution layer 221 are arranged as a (d+1)-dimensional matrix. If the nodes 220 of the preceding layer 210 are already arranged as a (d+1)-dimensional matrix with one depth dimension, the use of multiple kernels can be interpreted as extending along the depth dimension, so that the nodes 221 of the convolution layer 221 are also arranged as a (d+1)-dimensional matrix, with the size of the (d+1)-dimensional matrix being larger with respect to the depth dimension by a factor of the number of kernels than in the preceding layer 210.

[0079] One advantage of using convolutional layers 211 is that a spatially local correlation of the input data can be exploited by enforcing a local connectivity pattern between nodes of neighboring layers, in particular by connecting each node only to a small range of the nodes of the preceding layer.

[0080] In the illustrated embodiment, the input layer 210 comprises 36 nodes 220 arranged as a two-dimensional 6x6 matrix. The convolution layer 211 comprises 72 nodes 221 arranged as two two-dimensional 6x6 matrices, each of which is the result of convolution of the values ​​of the input layer with a kernel. Accordingly, the nodes 221 of the convolution layer 211 can be interpreted as arranged as a three-dimensional 6x6x2 matrix, the last dimension being the depth dimension.

[0081] A pooling layer 212 can be characterized by the structure and weights of the incoming edges and the activation function of its nodes 222, which form a pooling operation based on a nonlinear pooling function f. For example, in the two-dimensional case, the values ​​x(n) of the nodes 222 of the pooling layer 212 can be determined based on the values ​​x (n-1) The node 221 of the preceding layer 211 is calculated as x(n)[i,j]=f(x(n-1)[id1, jd2],…,x(n-1)[id1+d1-1, jd2+d2-1]).

[0082] In other words, by using a pooling layer 212, the number of nodes 221, 222 can be reduced by replacing a number d1·d2 of neighboring nodes 221 in the preceding layer 211 with a single node 222, which is computed as a function of the values ​​of this number of neighboring nodes in the pooling layer. Specifically, the pooling function f can be the max function, the average, or the L2 norm. In particular, for a pooling layer 212, the weights of the incoming edges are fixed and are not changed by training.

[0083] The advantage of using a pooling layer 212 is that the number of nodes 221, 222 and the number of parameters are reduced. This leads to a reduction in the computational overhead in the network and to the control of overfitting.

[0084] In the illustrated embodiment, the pooling layer 212 is a max-pooling, in which four adjacent nodes are replaced by only one node whose value is the maximum of the values ​​of the four adjacent nodes. The max-pooling is applied to each d-dimensional matrix of the previous layer; in this embodiment, the max-pooling is applied to each of the two two-dimensional matrices, thereby reducing the number of nodes from 72 to 18.

[0085] A fully meshed layer 213 can be characterized by the fact that a majority, in particular all edges, are present between node 222 of the previous layer 212 and node 223 of the fully meshed layer 213, and the weight of each of the edges can be adjusted individually.

[0086] In this embodiment, the nodes 222 of the preceding layer 212 of the fully meshed layer 213 are displayed both as two-dimensional matrices and additionally as unconnected nodes (shown as a node line, with the number of nodes reduced for better clarity). In this embodiment, the number of nodes 223 in the fully meshed layer 213 is equal to the number of nodes 222 in the preceding layer 212. Alternatively, the number of nodes 222 and 223 can be different.

[0087] Furthermore, in this embodiment, the values ​​of the nodes 224 of the output layer 214 are determined by applying the softmax function to the values ​​of the nodes 223 of the preceding layer 213. Applying the softmax function results in the sum of the values ​​of all nodes 224 of the output layer being 1, and all values ​​of all nodes 224 of the output layer are real numbers between 0 and 1. In particular, if the convolutional neural network 200 is used to categorize input data, the values ​​of the output layer can be interpreted as the probability that the input data falls into one of the various categories.

[0088] A convolutional neural network 200 can also include a ReLU layer (acronym for "Rectified Linear Units"). Specifically, the number of nodes and the structure of the nodes in a ReLU layer are the same as the number of nodes and the structure of the nodes in the preceding layer. In particular, the value of each node in the ReLU layer is calculated by applying a rectification function to the value of the corresponding node in the preceding layer. Examples of rectification functions are f(x) = max(0,x), the hyperbolic tangent function, and the sigmoid function.

[0089] In particular, convolutional neural networks can be trained using the backpropagation algorithm. To prevent overfitting, regularization methods can be used, such as dropping nodes 220, ..., 224, stochastic pooling, using artificial data, weight reduction based on L1 or L2 norms, or max-norm constraints.

[0090] Based on Fig. Section 5 illustrates an exemplary procedure for providing a trained function for outputting correction data. In S110, training reference measurement data are provided, wherein at least a part of the training reference measurement data is generated from magnetic resonance signals recorded by several receiving antennas 27 of the local coil 26 using the magnetic resonance device 10.

[0091] In S120, training image measurement data is provided, wherein the training image measurement data is generated from magnetic resonance signals that are recorded with the multiple receiving antennas 27 of the local coil 27 by means of the magnetic resonance device 10.

[0092] In S130, training target image measurement data is generated by applying a correction algorithm to the training image measurement data. In S140, the training target image measurement data is made available.

[0093] In S150, the trained function is trained based on the training reference measurement data and the training image measurement data as training input data and the training target image measurement data as training output data.

[0094] In S160, the trained function is provided for outputting the correction data.

[0095] In Fig. 6 is a generating unit 100 for generating an intensity-corrected magnetic resonance image, shown schematically.

[0096] It comprises a first input interface 101 for receiving reference measurement data, wherein at least part of the reference measurement data is generated from magnetic resonance signals acquired by the magnetic resonance device 10 using multiple receiving antennas 27 of the local coil 26. The generation unit 100 comprises a second input interface 102 for receiving image measurement data, wherein the image measurement data is generated from magnetic resonance signals acquired by the multiple receiving antennas 27 of the local coil 26 using the magnetic resonance device 10. Furthermore, the generation unit 100 comprises a processing unit 103 for applying a trained function to the reference measurement data and the image measurement data, generating the correction data as output data, and an output interface 104 for outputting the correction data.

[0097] Finally, it should be noted once again that the methods described in detail above, as well as the illustrated production unit, are merely exemplary embodiments which can be modified in various ways by a person skilled in the art without departing from the scope of the invention. Furthermore, the use of the indefinite articles "a" or "an" does not preclude the possibility that the features in question may be present multiple times. Likewise, the term "unit" does not preclude the possibility that the components in question consist of several interacting sub-components, which may also be spatially distributed. Regardless of the grammatical gender of a particular term, persons of male, female, or other gender identities are included. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited non-patent literature

[0000] Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, Gee JC. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010 Jun;29(6):1310-20. doi: 10.1109 / TMI.2010.2046908. Epub 2010 Apr 8. PMID: 20378467; PMCID: PMC3071855 [0004, 0044]

Claims

[1] Computer-implemented method for generating an intensity-corrected magnetic resonance image comprising - Providing reference measurement data, wherein at least part of the reference measurement data is generated from magnetic resonance signals recorded with at least one receiving antenna or at least one local coil using a magnetic resonance device according to a reference magnetic resonance sequence, - Providing image measurement data, wherein the image measurement data are generated from magnetic resonance signals recorded with the at least one receiving antenna of the at least one local coil by means of the magnetic resonance device according to an image magnetic resonance sequence, - Applying a function trained using a machine learning algorithm to the reference measurement data and the image measurement data as input data of the trained function, - Providing correction data as output data of the trained function, where the correction data describes a measurement-related intensity inhomogeneity of the image measurement data, - Generation of the intensity-corrected magnetic resonance image based on the image measurement data and the correction data. [2] Method according to claim 1 wherein at least part of the reference measurement data are generated from magnetic resonance signals recorded with at least one receiving antenna of a body coil fixed in the magnetic resonance device. [3] Method according to any of the preceding claims, wherein the reference magnetic resonance sequence is designed to produce a neutral contrast of the reference measurement data. [4] Method according to any of the preceding claims, where the image measurement data have an image coordinate system, the procedure further includes - Providing initial reference measurement data that includes a reference coordinate system, wherein the provision of reference measurement data includes - Transforming the initial reference measurement data from the reference coordinate system into the image coordinate system. [5] Method according to any of the preceding claims, wherein the reference measurement data are combined data which are combined from individual measurement data which are each recorded with the at least one receiving antenna of the at least one local coil by means of the magnetic resonance device according to the reference magnetic resonance sequence. [6] Method according to any of the preceding claims, wherein the trained function is based on a neural network, in particular a convolutional neural network, in particular a U-Net. [7] Computer-implemented method for providing a trained function for outputting correction data comprising - Providing training reference measurement data, wherein at least part of the training reference measurement data is generated from magnetic resonance signals recorded with the at least one receiving antenna or at least one local coil using a magnetic resonance device, - Providing training image measurement data, wherein the training image measurement data are generated from magnetic resonance signals recorded with the at least one receiving antenna of the at least one local coil using the magnetic resonance device, - Providing training target image measurement data, wherein the training target image measurement data are training image measurement data corrected by means of a correction algorithm, - Training the trained function based on the training reference measurement data, the training image measurement data, and the training target image measurement data. - Provision of the trained function for outputting the correction data. [8] Computer-implemented method according to any one of claims 1 to 6, wherein the trained function is provided according to a method according to claim 7. [9] Generating unit for generating an intensity-corrected magnetic resonance image comprising, - a first input interface for receiving reference measurement data, wherein at least part of the reference measurement data is generated from magnetic resonance signals recorded with at least one receiving antenna or at least one local coil using a magnetic resonance device, - a second input interface for receiving image measurement data, wherein the image measurement data are generated from magnetic resonance signals recorded with the at least one receiving antenna of the at least one local coil by means of the magnetic resonance device, - an output interface for outputting correction data, - a computing unit for applying a trained function to the reference measurement data and the image measurement data, whereby the correction data is generated as output data. [10] Computer program product comprising a program that can be directly loaded into a memory of a generating unit for generating an intensity-corrected magnetic resonance image, comprising program means for executing a method according to any one of claims 1 to 6 when the program is executed in the generating unit.

Images