Super resolution for non-rectangular acquisition of raw magnetic resonance data

Abstract

A method for generating high-resolution magnetic resonance imaging (MRI) data of a subject (O) is described. In this method, k-space data (RD) is sampled using a non-perpendicular sampling scheme. Specifically, a non-perpendicular sampling region (TB) of a Cartesian k-space is sampled, while a complementary region (KB) of the Cartesian k-space remains unsampled. Subsequently, initial MRI data (BD) is reconstructed based on the sampled k-space data (RD).Finally, second magnetic resonance imaging (MRI) data sets with increased resolution compared to the reconstructed first MRI data sets are generated by applying a method based on an AI-based model (M). This method is designed to augment the reconstructed first MRI data sets with image information that is transformed into the Fourier space of the reconstructed first MRI data sets and assigned to the complementary region (KB) determined from the k-space sampling. Furthermore, a method for training an AI-based model (M) for a method of augmenting MRI data sets (HRI) with image information is described. An image data generation device (50) for a magnetic resonance imaging system (80) is also described. Finally, a magnetic resonance imaging system (80) is described.

Description

[0001] The invention relates to a method for generating magnetic resonance image data of an examination subject with increased resolution, in which raw magnetic resonance data are acquired using non-rectangular acquisition. Furthermore, the invention relates to a method for training an AI-based model for supplementing magnetic resonance image data with image information. The invention also relates to an image data generation device. Furthermore, the invention relates to a magnetic resonance imaging system.

[0002] Imaging systems based on magnetic resonance imaging, particularly of nuclear spins, so-called magnetic resonance tomographs, have successfully established and proven themselves in a wide variety of applications. In this type of image acquisition, a static basic magnetic field B0, which serves to initially align and homogenize the magnetic dipoles under investigation, is usually superimposed with a rapidly switched magnetic field, the so-called gradient field, to spatially resolve the imaging signal. To determine the material properties of an object to be imaged, the dephasing or relaxation time is determined after a deflection of the magnetization from the initial alignment, allowing the identification of various material-specific relaxation mechanisms and relaxation times.The deflection is usually achieved by a number of RF pulses (the abbreviation RF stands for high frequency), also known as excitation pulses, and the spatial resolution is based on a time-defined manipulation of the deflected magnetization using the gradient field in a so-called measurement sequence or control sequence, which defines a precise temporal sequence of RF pulses, the change in the gradient field (by emitting a switching sequence of gradient pulses) and the acquisition of measured values.

[0003] Typically, an intermediate step is used to map the measured magnetization—from which the aforementioned material properties can be derived—to a spatial coordinate of the measured magnetization in the spatial space in which the object under examination is located. In this intermediate step, acquired raw magnetic resonance data, also referred to as k-space data, are arranged at readout points in so-called "k-space," with the k-space coordinates encoded as a function of the gradient field. The magnitude of the magnetization (in particular the transverse magnetization in a plane perpendicular to the previously described basic magnetic field) at a specific location on the object under examination can be determined from the data at the readout point using a Fourier transformation, which calculates a signal strength of the signal in spatial space from a signal strength (magnitude of magnetization) assigned to a specific frequency (the spatial frequency) or phase position.

[0004] However, k-space is often only partially sampled to save time. This reduced sampling of k-space, however, leads to a reduction in the image information reconstructed based on the sampled k-space data, particularly a reduction in image resolution. To compensate for this loss of information, there are methods that compensate for undersampling. Such methods can be combined with an augmentation of the k-space data or aim to directly increase the resolution in the image data space.

[0005] Super-resolution and partial Fourier reconstruction are means of improving image quality in magnetic resonance imaging by improving image resolution in the image data space.

[0006] Super-resolution can be achieved through machine learning, with the goal often being to increase the resolution of an image by four times or more. Super-resolution has already been achieved for Cartesian sequences. However, there is currently no solution for elliptical scanning and non-Cartesian trajectories such as BLADE / PROPELLER (PROPELLER is an acronym for "Periodically Rotated Overlapping Parallel Lines with Enhanced Reconstruction"—capital letters are part of the acronym). Radial or spiral scanning is also possible.

[0007] Elliptical scanning for BLADE (radial scanning sequence to reduce motion artifacts) and other non-Cartesian acquisitions typically uses k-space filters to avoid truncation artifacts such as Gibbs ringing. However, deep learning solutions are currently not used for these non-Cartesian scanning methods.

[0008] The performance of super-resolution networks in MRI trained on Cartesian images strongly depends on the coverage of k-space. Deep Resolve Sharp is trained on fully sampled images, and performance is greatly reduced when the input image reconstruction includes zero padding (typically with lower sharpening).

[0009] For partial Fourier acquisitions, it has been observed that deep learning-based algorithms can learn to recover unmeasured data from asymmetric acquisitions, similar to conventional partial Fourier algorithms, but also for applications where conventional algorithms fail. Exemplary results were presented in the following paper: Wessling D, Herrmann J, Afat S, Nickel D, Almansour H, Keller G, Othman AE, Brendlin AS, Gassenmaier S. Application of a Deep Learning Algorithm for Combined Super-Resolution and Partial Fourier Reconstruction Including Time Reduction in T1-Weighted Precontrast and Postcontrast Gradient Echo Imaging of Abdominopelvic MR Imaging. Diagnostics (Basel). 2022 Sep 29;12(10):2370. doi: 10.3390 / diagnostics12102370. PMID: 36292057; PMCID: PMC9600324.

[0010] Similar results have been obtained for diffusion-weighted imaging. However, the aforementioned super-resolution approaches are limited to imaging with a Cartesian or rectangular sampling pattern.

[0011] It is therefore a task in magnetic resonance imaging to improve image quality in a non-rectangular partial scan of k-space.

[0012] This object is achieved by a method for generating magnetic resonance image data of an examination subject with increased resolution according to patent claim 1, by a method for training a trained AI-based model for a method for supplementing magnetic resonance image data with image information according to patent claim 10, by an image data generation device according to patent claim 12 and by a magnetic resonance imaging system according to patent claim 13.

[0013] In the method according to the invention for generating magnetic resonance image data of an examination subject with increased resolution, k-space data are sampled using a non-rectangular sampling pattern. A non-rectangular sampling region of Cartesian k-space is sampled, and a complementary region of Cartesian k-space is not sampled. A non-rectangular sampling pattern is understood to mean a sampling pattern in which a k-space trajectory is tracked that is at least partially non-parallel to the coordinate axes of a Cartesian coordinate system. It should also be taken into account that such a Cartesian coordinate system can be rotated. In particular, curvilinear trajectories or non-parallel, in particular radial, trajectories are non-rectangular.

[0014] Furthermore, initial magnetic resonance image data are reconstructed based on the sampled k-space data. The initial magnetic resonance image data are generated, as usual, based on a Fourier transform from k-space to image data space.

[0015] Finally, second magnetic resonance image data are generated with a resolution increased compared to a resolution of the reconstructed first magnetic resonance image data by applying a method which is based on an AI-based model adapted to the non-rectangular scanning scheme and is designed to supplement the reconstructed first magnetic resonance image data with image information which is transformed into the Fourier space, or whose associated data in the Fourier space associated with the reconstructed first magnetic resonance image data is associated with the complementary region determined from the k-space scanning.

[0016] K-space refers to the space in which the measured multi-channel raw data is sampled or measured. Fourier space is the space in which the "synthetic" raw data arranged by a Fourier transform of the channel-combined image is arranged. A distinction between k-space and Fourier space is important when the k-space data is sampled from a multi-coil system with multiple channels. In this case, k-space has a separate k-space for each channel, in which k-space data is sampled according to a predetermined pulse sequence or trajectory. In contrast, in Fourier space, the transformed data of an overall image, which is based on a reconstruction based on sampled k-space data from all channels of the multi-coil system in question, are displayed together.Nevertheless, the aforementioned sampling areas in k-space, i.e. the non-rectangular sampling area and the complementary area, can also be found in Fourier space, whereby they now result from a Fourier transformation of the respective image data, which were reconstructed by sampling the respective sampling areas in k-space.

[0017] “Magnetic resonance image data” assigned to the complementary domain means that the gain in image information in the form of a higher resolution is achieved by additional image information which is transformed into Fourier space by Fourier transformation, arranged as Fourier space data in the complementary domain.

[0018] An AI-based model adapted to the non-rectangular sampling scheme is an AI-based model that has been trained using image data generated by sampling a non-rectangular sampling region of k-space with a non-rectangular sampling scheme.

[0019] The additional image information is assigned to Fourier-space data that lies outside the scanning range occupied by the trajectory following the non-rectangular scanning pattern. The information obtained through the non-rectangular acquisition of k-space data is thus incorporated unchanged into the second magnetic resonance image data with increased resolution. It is crucial that the AI-based model, after being trained on first magnetic resonance image data generated by scanning k-space using a specific non-rectangular scanning pattern, is also applied to such first magnetic resonance image data. Precisely in such a case, it is possible to generate the second magnetic resonance image data by supplementing the first magnetic resonance image data with image information obtained by Fourier transforming the k-space data of the complementary range.

[0020] Advantageously, the resolution of magnetic resonance image data acquired by scanning k-space with a non-rectangular sampling scheme can be increased. The training of the AI-based model used to increase the resolution is concentrated on image data acquired by applying this specific non-rectangular sampling scheme. It should be ensured that the geometry of the k-space sampling region sampled to reconstruct the training data matches the geometry of the k-space sampling region sampled to generate the first magnetic resonance image data. In this way, a particularly effective and precise increase in the resolution of the first magnetic resonance image data is achieved, while also reducing image artifacts.However, as explained in more detail later, there are also ways to continue using a trained AI-based model without retraining if the field of view changes and thus the distances between the sampling points in k-space change.

[0021] In the method according to the invention for training an AI-based model for supplementing magnetic resonance image data with image information, labeled training data comprising input data and validated result data are generated. The validated result data comprise reconstructed image data reconstructed based on raw data obtained by completely scanning a Cartesian k-space, and the input data comprise image data reconstructed based on a subset of the raw data or k-space data, wherein the subset of k-space data is assigned to a non-rectangular scanning scheme region of k-space, which was generated by setting the k-space data outside the scanning region to the value 0.A “complete” scan of k-space is understood as a scan of a rectangular k-space that includes the two sub-regions defined above, i.e. the non-rectangular scan region and the complementary region.

[0022] The AI-based model to be trained is applied to the labeled input data using the method for supplementing magnetic resonance image data with image information, generating result data.

[0023] The AI-based model is then adapted based on the outcome data and the validated outcome data.

[0024] Finally, the trained AI-based model is preferably provided for the method for generating magnetic resonance image data of an examination subject with increased resolution. Advantageously, the training method generates an AI-based model adapted to a specific scanning scheme, which can be used to improve the resolution of magnetic resonance image data based on k-space data acquired using such a scanning scheme.

[0025] The image data generating device according to the invention has an input interface for receiving sampled k-space data with a non-rectangular sampling scheme, wherein a non-rectangular sampling region of a Cartesian k-space is sampled and a complementary region of the Cartesian k-space is not sampled.

[0026] Part of the image data generation device according to the invention is a reconstruction unit for reconstructing first magnetic resonance image data on the basis of the sampled k-space data.

[0027] The image data generation device according to the invention has, in particular, an image data generation unit for generating second magnetic resonance image data with increased resolution compared to the resolution of the reconstructed first magnetic resonance image data by applying a method based on an AI-based model and designed to supplement the reconstructed first magnetic resonance image data with image information that, transformed into the Fourier space of the reconstructed first magnetic resonance image data, is assigned to the complementary region determined from the k-space scan. The image data processing device according to the invention shares the advantages of the method according to the invention for generating magnetic resonance image data of an examination subject with increased resolution.

[0028] The magnetic resonance imaging system according to the invention comprises the image data generation device according to the invention. The magnetic resonance imaging system according to the invention shares the advantages of the method according to the invention for generating magnetic resonance image data of an examination subject with increased resolution.

[0029] The computer program product according to the invention has program code sections with which all steps of the method according to the invention for generating magnetic resonance image data of an examination subject with increased resolution and of the method for training an AI-based model for a method for supplementing magnetic resonance image data with image information are carried out when the program is executed in a magnetic resonance imaging system or a control device of such a magnetic resonance imaging system.

[0030] A largely software-based implementation has the advantage that existing magnetic resonance imaging systems or their control devices can be easily retrofitted by a software update to operate in the manner according to the invention.

[0031] A majority of the aforementioned components of the image data generation device according to the invention can be implemented entirely or partially in the form of software modules in a processor of a corresponding computing system, e.g., a control device of a magnetic resonance imaging system or a computer used to control such a system. A largely software-based implementation has the advantage that even previously used computing systems can be easily upgraded to operate in the manner according to the invention via a software update.In this respect, the object is also achieved by a corresponding computer program product with a computer program that can be loaded directly into a computer system, with program sections for executing the steps of the inventive method for generating magnetic resonance image data of an examination subject with increased resolution and the method for training an AI-based model for supplementing magnetic resonance image data with image information when the program is executed in the computer system. Such a computer program product may, in addition to the computer program, optionally include additional components such as documentation and / or additional components, including hardware components such as hardware keys (dongles, etc.) for using the software.

[0032] A computer-readable medium, e.g., a memory stick, a hard disk, or another portable or permanently installed data storage device, can be used for transport to the computer system or control device and / or for storage on or in the computer system or control device, on which the program sections of the computer program that can be read and executed by a computer system are stored. For this purpose, the computer system can, for example, have one or more cooperating microprocessors or the like.

[0033] The dependent claims and the following description each contain particularly advantageous embodiments and developments of the invention. In particular, the claims of one claim category can also be developed analogously to the dependent claims of another claim category. Furthermore, within the scope of the invention, the various features of different embodiments and claims can also be combined to form new embodiments.

[0034] In a variant of the method according to the invention for generating magnetic resonance image data of an examination subject with increased resolution, the increased resolution is achieved by supplementing magnetic resonance image data assigned to higher frequencies in Fourier space and k-space. Higher frequencies are typically assigned to regions located further out in k-space. The raw data arranged in this complementary region contain information on the basis of which image data with higher resolution can be acquired. In contrast, the raw data located in the center of k-space serve to determine the basic contours of an examination subject to be imaged at low resolution.

[0035] In a preferred variant of the method according to the invention for generating magnetic resonance image data of an examination subject, the non-rectangular scanning pattern comprises one of the following scanning pattern types: - a radial scan, - an elliptical scan, - a spiral scan, - a BLADE scan or a PROPELLER scan.

[0036] In radial sampling, k-space is sampled using a radial k-space trajectory consisting of multiple radial profiles, each passing through the center of k-space. Radial trajectories are less susceptible to motion artifacts and are therefore often used to record dynamic physiological processes. Furthermore, compared to Cartesian trajectories, radial trajectories are more robust against a rotationally incomplete scan, which is used in time-resolved imaging to improve temporal resolution.

[0037] The PROPELLER technique was developed in the late 1990s as a method for motion reduction. The basic idea was to rotate k-space using a series of radially directed stripes or "rotor blades."

[0038] Each "rotor blade" consists of several parallel phase-encoded lines that can be acquired using fast spin-echo or gradient-echo methods. Typically, 8 to 32 rotor blade lines are acquired in a single acquisition. The rotor blades are then rotated by a small angle (10° to 20°), acquiring a second data set. This process continues until image data from the entire k-space circle has been collected.

[0039] The PROPELLER trajectory through k-space offers several unique advantages. The center of k-space (which contains the highest signal amplitude and contributes most to image contrast) is oversampled, meaning the signal-to-noise and contrast-to-noise ratios are high. Oversampling in this region also provides redundancy of information, meaning that the data for each new blade can be compared to data from previous blades for consistency. As the patient moves between acquisitions of k-space data from different blades, the data for the next blade can be corrected (or even discarded entirely) depending on how anomalous its central information appears.

[0040] Particularly preferably, the generation of magnetic resonance image data with an increased resolution compared to a resolution of the reconstructed magnetic resonance image data comprises a sub-step for generating data consistency.

[0041] The process, which is based on an AI-based model, initially generates preliminary second magnetic resonance imaging data.

[0042] To ensure data consistency, the preliminary second magnetic resonance image data are transformed into Fourier space, generating Fourier space data associated with the preliminary second magnetic resonance image data. Furthermore, modified Fourier space data are generated, and the Fourier space data associated with the preliminary second magnetic resonance image data are modified within the non-rectangular scanning range by projecting the sampled k-space data or Fourier space data associated with the first magnetic resonance image data onto the Fourier space data associated with the second magnetic resonance image data. During this "projection," the first magnetic resonance image data are transformed into Fourier space, generating Fourier space data associated with the sampled k-space data.These Fourier space data then replace the Fourier space data associated with the second magnetic resonance image data in a sub-area of ​​the Fourier space or are combined with them in that sub-area. The modified Fourier space data are finally transformed into the image data space, generating data-consistent second magnetic resonance image data. Advantageously, the image information from the first magnetic resonance image data is incorporated into the second magnetic resonance image data.

[0043] In the method according to the invention, a dedicated model for the specific scanning pattern type used is particularly preferably used as the AI-based model. Crucial for reducing artifacts and ensuring good data consistency is that the AI-based model is trained using training data generated for a specific scanning pattern type. The image quality of the second magnetic resonance image data is advantageously particularly good by applying a dedicated model to the first magnetic resonance image data.

[0044] In a particularly preferred embodiment of the method according to the invention for generating magnetic resonance image data of an examination subject with increased resolution, in which raw magnetic resonance data is acquired by non-Cartesian acquisition, a plurality of image subregions formed by a subset of the magnetic resonance image data are defined depending on the geometry of the scanning pattern. The AI-based model is applied to the respective image subregions. The high-resolution magnetic resonance image data generated on the basis of the magnetic resonance image data of the respective image subregions are combined to form a high-resolution overall image. The AI-based model can advantageously be applied to magnetic resonance image data associated with images with different geometries. It should be noted here that the geometry of the scanning pattern depends on the image geometry or the geometry of the field of view.To apply an AI-based model trained on a specific geometry of a sampling pattern to different geometries of a field of views, the field of view is essentially divided into potentially overlapping sub-areas whose geometry matches the sampling pattern of the AI-based model. The high-resolution partial images or image sub-areas generated by the AI-based model are then combined into a high-resolution overall image. The application range of an AI-based model trained on a specific geometry of a sampling pattern is advantageously extended to different geometries of a field of views.

[0045] During the process of composing the partial images or image subregions into a high-resolution overall image, the high-resolution overall image is preferably composed or combined by adding the high-resolution magnetic resonance image data from overlapping sections of the respective image subregions based on a fade-out. Such a "fade-out" comprises a function gradually decreasing towards the image edges between the values ​​0 and 1, which is applied to edge regions of overlapping image subregions and with which a smooth transition between the image subregions is generated. Advantageously, a sudden transition between the image subregions is avoided, thus improving the image quality of the high-resolution overall image.

[0046] In the method according to the invention for generating magnetic resonance image data of an examination subject, k-space data is preferably sampled as part of a subsampling using a non-rectangular sampling scheme with a multi-coil system. Experiments have shown that AI-based models trained on non-rectangular trajectories also efficiently increase the resolution of image data based on k-space data acquired with a multi-coil system.

[0047] If an imaging method for the simultaneous reading of several slices is used in the method according to the invention, the number N of slices to be read out simultaneously is equal to 2 in a particularly effectively realizable embodiment of the method. Advantageously, the acquisition time can be further reduced by a simultaneous reading of several slices, so that a higher patient throughput and improved comfort for the patients to be examined is achieved.

[0048] Alternatively, the number N of slices to be read simultaneously can be 3. In principle, it would be optimal to read as many slices as possible simultaneously. However, the number of slices that can be read simultaneously is limited by the following circumstances: The energy delivered to the patient per pulse is proportional to the number of slices N. However, the energy delivery per time is limited. Furthermore, the more slices are read simultaneously, the more difficult separation in image reconstruction becomes, as the noise in the separated images increases. This increase is nonlinear. Therefore, in practice, it is only possible to read a few slices simultaneously.

[0049] In a preferred embodiment of the method according to the invention for training an AI-based model for a method for supplementing magnetic resonance image data with image information to generate magnetic resonance image data with increased resolution, the model comprises a conventional super-resolution network. Super-resolution networks are used to increase the resolution of image data by modifying the image data in the image data space.

[0050] In the method according to the invention for training an AI-based model for a method for supplementing magnetic resonance image data with image information, the input data and / or result data preferably comprise magnitude images or complex-valued images, preferably with real and imaginary parts as different channels.

[0051] In neural networks, images are treated as tensors. Typically, such tensors have the dimensions Batch, Channel, Depth, Height, and Width (abbreviated BCDHW). The last three are spatial dimensions. Channel is a single dimension, with channels also correlated with each other through convolutions. Furthermore, not all AI frameworks can handle complex numbers (a few years ago, this was completely impossible, but it is slowly becoming more common). Therefore, it is common to view complex-valued images as tensors with two channels. The real and imaginary parts are "stacked" (arranged in a stack structure).

[0052] Particularly preferably, the AI-based model includes a hard projection to establish data consistency, so that only frequencies outside the sampling range can be changed. In this context, a "hard projection" should be understood to mean that, in Fourier space, the Fourier space data generated by a Fourier transformation of the first magnetic resonance image data into Fourier space is fully adopted for the second magnetic resonance image data, and the additional image information is limited to data whose associated Fourier space data lie outside this Fourier space range, i.e., outside the Fourier space range covered by Fourier space data generated by a Fourier transformation of the first magnetic resonance image data into Fourier space.Figuratively speaking, a type of mask is used to remove the low-frequency Fourier space data from the preliminary second magnetic resonance image data and replace it with the Fourier space data associated with the first magnetic resonance image data. The modified Fourier space data are then transformed into image data space, creating second magnetic resonance image data with increased resolution.

[0053] Alternatively, the AI-based model includes a soft projection, such that frequencies within the sampling range are shifted toward the frequencies in the input image data. A soft projection is understood to be a projection in which the low-frequency Fourier-space data of the preliminary second magnetic resonance image data in Fourier space is only partially removed. Instead, a blend of the Fourier-space data associated with the first magnetic resonance image data and the Fourier-space data associated with the second magnetic resonance image data occurs in the inner region of Fourier space, which is covered by Fourier-space data generated by transforming the first magnetic resonance image data into Fourier space.

[0054] The invention is explained in more detail below with reference to exemplary embodiments in the accompanying figures. They show: Fig. 1 a diagram illustrating a magnetic resonance image in image data space and associated Fourier space data in Fourier space, Fig. 2 is a flowchart illustrating a method for generating magnetic resonance image data of an examination subject with increased resolution according to an embodiment of the invention, Fig. 3 is a diagram illustrating an architecture of an artificial neural network used to Fig. 2 illustrated procedure can be implemented, Fig. 4 is a flowchart illustrating a method for training an AI-based model to augment magnetic resonance image data, Fig. 5 a schematic representation of an image data generation device according to an embodiment of the invention, Fig. 6 a schematic representation of a relationship between a field of view and a parameterization of an elliptical scanning scheme in k-space, Fig. 7 a schematic representation of a section-wise generation of high-resolution magnetic resonance image data by acquiring k-space data with a circular scanning scheme, Fig. 8 a magnetic resonance imaging system according to an embodiment of the invention.

[0055] In Fig. 1 shows a diagram 10 which shows magnetic resonance image data BD in the image data space (see left in Fig. 1) and the Fourier spatial data FD associated with the magnetic resonance image data BD (see right in Fig. 1) in Fourier space. On the left side of Figure 10, image data BD in the image data space of a patient's skull is shown, and on the right side of Figure 10, corresponding Fourier space data FD of the image data BD shown on the left are shown. The Fourier space data FD is generated by a Fourier transform of the image data BD into Fourier space. The Fourier space data FD shows a circular region. A spatial region in k-space corresponding to the Fourier space is sampled, for example, by a radial trajectory. As shown on the right side in Fig. As shown in Figure 1, the subregion TB of Fourier space occupied by the transformation of the image data BD into Fourier space does not correspond to the entire square-represented Fourier space, but only to a circular subregion TB thereof. A complementary region KB forms the region of Fourier space not occupied by the Fourier space data FD associated with the image data BD.

[0056] In Fig. 2 shows a flowchart 200 which illustrates a method for generating magnetic resonance image data HBD of an examination subject with increased resolution according to an embodiment of the invention.

[0057] In step 2.I, k-space data RD are sampled in a subsampling process using a non-rectangular sampling scheme, as described, however, in Fourier space and not in k-space, for example, in Fig. 1 on the right side of Figure 10. As shown in Fig. As illustrated in Figure 1, not the entire square k-space is sampled, but only a circular subregion TB of it. A complementary region KB of k-space, however, remains unsampled.

[0058] In step 2.II, magnetic resonance image data BD are reconstructed based on the subsampled k-space data RD of the circular sub-area TB.

[0059] In step 2.III, magnetic resonance image data HBD are generated with an increased resolution compared to a resolution of the reconstructed magnetic resonance image data BD by applying an AI-based model to the reconstructed magnetic resonance image data BD.

[0060] In Fig. 3 shows a diagram 30 which illustrates an architecture of an artificial neural network with which the Fig. 2 can be implemented. The neural network started with an input of magnetic resonance image data BD. The magnetic resonance image data BD is processed in a total of 10 layers, each of which has a convolution function F and an activation function RELU. Following each three or four of these 10 layers, a concatenation CC of the intermediate results is performed, followed by another convolution F. Subsequently, a pixel interleaving PV is performed, followed by three further filtering F or convolutions. Finally, data consistency DC is generated. Finally, magnetic resonance image data HBD is output with increased resolution.

[0061] In Fig. 4 shows a flowchart 400 illustrating a method for training an AI-based model M for a method for supplementing magnetic resonance image data BD with image information.

[0062] In step 4.I, labeled training data L-TD is generated, which comprises input data ID and validated result data V-RS. The validated result data V-RS comprises image data VBD reconstructed by a complete scan of k-space. The input data ID comprises image data BD reconstructed based on k-space data acquired using a non-rectangular sampling scheme. This k-space data was generated by setting the frequencies of the k-space data outside the sampling range associated with the image data VBD of the validated result data V-RS to zero.

[0063] In step 4.II, the AI-based model M to be trained is applied to the labeled input data L-ID, generating result data RS.

[0064] In step 4.III, the AI-based model M is adapted based on the result data RS and the validated result data V-RS.

[0065] In step 4.IV, the trained AI-based model AI-M is output. Steps 4.II to 4.III can be repeated more often with the respective adapted model AI-M as the AI-based model M to be trained in order to further adapt the adapted model AI-M. The iteration implemented in this way is terminated when the adapted model AI-M has met a predetermined quality criterion. For example, a result vector RS deviates from the validated result data V-RS by less than a predetermined threshold from a result vector of the validated result data V-RS in order for the quality criterion to be met. The deviation can be determined, for example, by calculating a vector norm of the difference between the two vectors.

[0066] In Fig. 5 is a schematic representation of an image data generation device 50 according to an embodiment of the invention.

[0067] The image data generation device 50 comprises an input interface 51 for receiving k-space data RD which were sampled in the context of a subsampling with a non-rectangular sampling scheme.

[0068] Furthermore, the control device comprises a reconstruction unit 52 for reconstructing first magnetic resonance image data BD on the basis of the subsampled k-space data RD.

[0069] Part of the image data generation device 50 is also an image data generation unit 53 for generating second magnetic resonance image data HBD with an increased resolution compared to a resolution of the reconstructed magnetic resonance image data BD by applying an AI-based model to the reconstructed magnetic resonance image data BD.

[0070] In Fig. Figure 6 shows a schematic representation of the relationship between a field of view (FoV), i.e., an examination area to be imaged, and a parameterization of an elliptical scanning pattern (AS) in k-space. Fig. Figure 6 shows a rectangular field of view (FoV) in the image data space. Corresponding to this field of view (FoV), an elliptical sampling pattern (AS) is formed in k-space. This means that the eccentricity of the ellipse of the elliptical sampling pattern (AS) is matched to the dimensions of the field of view (FoV).

[0071] In Fig. Figure 7 shows a schematic representation of a section-wise generation of high-resolution magnetic resonance image data by acquiring k-space data with a circular scanning scheme ASK. While the field of view FoV has the same dimensions as in Fig. 6, a circular scanning scheme ASK is now used to acquire k-space data instead of an elliptical scanning scheme AS. To ensure that the process of increasing the resolution works correctly despite the different scanning geometry and the field of view (FoV), the field of view (FoV) is Fig. 7, the image is divided into overlapping square sub-areas TB1, TB2. The method for generating magnetic resonance image data HBD with increased resolution is applied separately to the two square sub-areas TB1, TB2, which are shown in Fig. 7 each extend beyond the vertical center line ML of the field of view FoV. The overlapping partial images TB1, TB2 resulting from this process are then merged with increased resolution, with the overlapping area being added or superimposed with a so-called "fade-out."

[0072] In Fig. Figure 8 shows a roughly schematic representation of a magnetic resonance system 80 according to the invention (hereinafter referred to as "MR system"). It comprises, on the one hand, the actual magnetic resonance scanner 102 with an examination room 103 or patient tunnel, into which a patient O, or in this case a patient or test subject, in whose body, for example, a specific organ to be imaged is located, can be moved on a couch 108.

[0073] The magnetic resonance scanner 102 is equipped in the usual way with a basic field magnet system 104, a gradient system 106 as well as an RF transmit antenna system 105 and an RF receive antenna system 107. In the illustrated embodiment, the RF transmit antenna system 105 is a whole-body coil permanently installed in the magnetic resonance scanner 102, whereas the RF receive antenna system 107 consists of local coils to be arranged on the patient or subject (in Fig. 8 is symbolized only by a single local coil). In principle, however, the whole-body coil 105 can also be used as an RF receiving antenna system and the local coils 107 as an RF transmitting antenna system, provided that these coils can each be switched to different operating modes.

[0074] The MR system 80 further comprises a central control device 113 used to control the MR system 80. This central control device 113 comprises a sequence control unit 114 for pulse sequence control. This unit controls the temporal sequence of radio-frequency pulses (RF pulses) and gradient pulses as a function of a selected imaging sequence PS according to a pulse sequence scheme PSS. Such an imaging sequence PS or the pulse sequence scheme PSS underlying the imaging sequence PS can be specified, for example, within a measurement or control protocol P. Typically, different control protocols P for different measurements are stored in a memory 119 and can be selected by an operator (and modified if necessary) and then used to perform the measurement.

[0075] To output the individual RF pulses, the central control device 113 has a radio-frequency transmission device 115, which generates the RF pulses, amplifies them, and feeds them into the RF transmission antenna system 105 via a suitable interface (not shown in detail). To control the gradient coils of the gradient system 106, the control device 113 has a gradient system interface 116. The sequence control unit 114 communicates in a suitable manner, e.g., by transmitting sequence control data SD, with the radio-frequency transmission device 115 and the gradient system interface 116 for transmitting the pulse sequences PS. The control device 113 also has a radio-frequency reception device 117 (which also communicates in a suitable manner with the sequence control unit 114) for acquiring magnetic resonance signals received from the RF transmission antenna system 107 in a coordinated manner.

[0076] The central control device 113 also comprises an image data generation device 50 according to the invention, which Fig. 5 has a structure illustrated in detail.

[0077] Part of the image data generation device 50 is also a reconstruction unit 52 (see Fig. 5), which receives the acquired data after demodulation and digitization as raw data or k-space data RD and reconstructs magnetic resonance image data from it. These magnetic resonance image data are subsequently optimized for their resolution, and the resulting magnetic resonance image data HBD with increased resolution can then be stored, for example, in a memory 119.

[0078] The central control unit 113 can be operated via a terminal with an input unit 111 and a display unit 109, via which the entire MR system 80 can thus also be operated by an operator. MR images can also be displayed on the display unit 109, and using the input unit 111, possibly in combination with the display unit 109, measurements can be planned and started, and in particular, suitable control protocols with suitable measurement sequences can be selected and modified if necessary, as explained above.

[0079] The MR system 80 according to the invention and in particular the control device 113 can also have a large number of further components, not shown in detail here, but usually present on such devices, such as a network interface in order to connect the entire system to a network and to be able to exchange raw data RD and / or image data BD or parameter maps, but also other data, such as patient-relevant data or control protocols.

[0080] How suitable raw data RD can be acquired and MR images BD reconstructed from them by irradiating RF pulses and generating gradient fields is generally known to the person skilled in the art and will not be explained in more detail here.

[0081] From the above description, it is clear that the invention effectively provides possibilities for improving a method for controlling a magnetic resonance imaging system for generating magnetic resonance image data with regard to the required time duration.

[0082] It should be noted that the features of all embodiments or developments disclosed in figures can be used in any combination.

[0083] Finally, it is pointed out once again that the detailed methods and structures described above are exemplary embodiments and that the basic principle can be varied widely by a person skilled in the art without departing from the scope of the invention, insofar as it is defined by the claims. For the sake of completeness, it is also pointed out that the use of the indefinite articles “a” or “an” does not exclude the possibility that the features in question may be present multiple times. Likewise, the term “unit” does not exclude the possibility that it consists of several components, which may also be spatially distributed. Regardless of the grammatical gender of a particular term, persons with male, female or other gender identity are included. QUOTES CONTAINED IN THE DESCRIPTION

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

[0000] Wessling D, Herrmann J, Afat S, Nickel D, Almansour H, Keller G, Othman AE, Brendlin AS, Gassenmaier S. Application of a Deep Learning Algorithm for Combined Super-Resolution and Partial Fourier Reconstruction Including Time Reduction in T1-Weighted Precontrast and Postcontrast Gradient Echo Imaging of Abdominopelvic MR Imaging. Diagnostics (Basel). 2022 Sep 29;12(10):2370. doi: 10.3390 / diagnostics12102370. PMID: 36292057; PMCID: PMC9600324

[0009]

Claims

[1] Method for generating magnetic resonance image data (HBD) of an examination object (O) with increased resolution, comprising the steps: - Sampling k-space data (RD) with a non-rectangular sampling scheme, wherein a non-rectangular sampling region (TB) of a Cartesian k-space is sampled and a complementary region (KB) of the Cartesian k-space is not sampled, - Reconstructing first magnetic resonance image data (BD) based on the sampled k-space data (RD), - generating second magnetic resonance image data (HBD) with a resolution increased compared to a resolution of the reconstructed first magnetic resonance image data (BD) by applying a method which is based on an AI-based model (M) and is designed to supplement the reconstructed first magnetic resonance image data (BD) with image information which, when transformed into Fourier space, is assigned to the complementary region (KB) determined from the k-space scan. [2] Method according to claim 1, wherein the complementary region (KB) has higher sampling frequencies compared to the sampling frequencies of the non-rectangular sampling region (TB) and the increased resolution is achieved by supplementing image information associated with the higher sampling frequencies in k-space located in the complementary region (KB). [3] The method of claim 1 or 2, wherein the non-rectangular scanning scheme comprises one of the following scanning scheme types: - an elliptical scan, - a radial scan, - a spiral scan, - a BLADE scan or a PROPELLER scan. [4] Method according to one of the preceding claims, wherein the generation of the second magnetic resonance image data (HBD) comprises a sub-step for generating data consistency (DC), wherein the method, which is based on an AI-based model (M), initially generates preliminary second magnetic resonance image data and for generating the data consistency (DC) - the preliminary second magnetic resonance image data are transformed into Fourier space, whereby Fourier space data associated with the preliminary second magnetic resonance image data are generated, - modified Fourier spatial data are generated, wherein the Fourier spatial data associated with the preliminary second magnetic resonance image data are modified within the non-rectangular scanning range (TB) by a projection of the Fourier spatial data (FD) associated with the first magnetic resonance image data (BD) onto the Fourier spatial data associated with the preliminary second magnetic resonance image data, - the modified Fourier space data are transformed into the image data space, generating data-consistent second magnetic resonance image data (HBD). [5] The method according to claim 4, wherein the projection comprises a hard projection configured such that the Fourier space data associated with the preliminary second magnetic resonance image data within the non-rectangular scanning range (TB) are replaced by the Fourier space data (FD) associated with the first magnetic resonance image data (BD). [6] The method of claim 4, wherein the projection comprises a soft projection configured such that the Fourier space data associated with the preliminary second magnetic resonance image data within the non-rectangular scanning range (TB) are combined with the Fourier space data (FD) associated with the first magnetic resonance image data (BD). [7] Method according to one of the preceding claims, wherein the AI-based model (M) used is a dedicated model for the scanning scheme type specifically used for the non-rectangular scanning scheme. [8] Method according to one of the preceding claims, wherein, depending on the geometry of the non-rectangular scanning scheme, a plurality of image sub-regions (TB1, TB2) formed by a subset of the first magnetic resonance image data (BD) are defined and the method, which is based on an AI-based model (M), is applied separately to the respective image sub-regions (TB1, TB2) and the second magnetic resonance image data (HBD) generated on the basis of the first magnetic resonance image data (BD) of the respective image sub-regions (TB1, TB2) are combined to form an overall image with increased resolution. [9] Method according to claim 8, wherein the overall image is combined by a fade-out-based addition of the second magnetic resonance image data (HBD) of mutually overlapping portions of the respective image sub-regions (TB1, TB2). [10] Method for training an AI-based model (M) for a method for supplementing magnetic resonance image data (BD) with image information, comprising the steps: - Generating labeled training data (L-TD) containing input data (ID) and validated result data (V-RS), where - the validated result data (V-RS) comprise reconstructed magnetic resonance image data (VBD) of an examination subject (O) which were reconstructed on the basis of k-space data (RD) obtained by a complete scan of a Cartesian k-space, and - the input data (ID) comprise reconstructed magnetic resonance image data (BD) of the examination object (O), which were reconstructed on the basis of a subset of the k-space data (RD), wherein the subset of the k-space data (RD) is assigned to a non-rectangular scanning range (TB) of the k-space, which was generated by setting the k-space data (RD) outside the non-rectangular scanning range (TB) to the value 0, - Applying the AI-based model (M) to be trained using the method for supplementing magnetic resonance image data (BD) with image information to the input data (ID), whereby result data (RS) are generated, - Generating a trained AI-based model (AI-M) by adapting the AI-based model (M) based on the result data (RS) and the validated result data (V-RS). [11] Method according to claim 10, wherein the trained AI-based model (AI-M) is based on a conventional super-resolution network and / or wherein the input data (ID) and / or result data (RS) comprise magnitude images or complex-valued images. [12] Image data generating device (50), comprising: - an input interface (51) for receiving sampled k-space data (RD) with a non-rectangular sampling scheme, wherein a non-rectangular sampling area (TB) of a Cartesian k-space is sampled and a complementary area (KB) of the Cartesian k-space is not sampled, - a reconstruction unit (52) for reconstructing first magnetic resonance image data (BD) on the basis of the sampled k-space data (RD), - an image data generation unit (53) for generating second magnetic resonance image data (HBD) with an increased resolution compared to a resolution of the reconstructed first magnetic resonance image data (BD) by applying a method which is based on an AI-based model (M) and is designed to supplement the reconstructed first magnetic resonance image data (BD) with image information which, when transformed into Fourier space, is assigned to the complementary region (KB) determined from the k-space scan. [13] A magnetic resonance imaging system (60) comprising an image data generating device (50) according to claim 12. [14] Computer program product comprising a computer program which is loadable directly into a memory of a control device (113) of a magnetic resonance imaging system (80), with program code sections to carry out all steps of a method according to one of claims 1 to 11 when the program is executed in the control device (113). [15] Computer-readable medium on which program sections executable by a computer unit are stored in order to carry out the steps of a method according to one of claims 1 to 11 when the program sections are executed by the computer unit.

Images