Computer-implemented method for processing a magnetic resonance data set of an imaging area, image processing device, computer program and electronically readable data carrier

Abstract

The invention relates to a computer-implemented method for processing a magnetic resonance data set (23) of an acquisition area, which is based on an acquisition in an examination procedure in a magnetic resonance device with a magnetic resonance sequence in which, in particular, after a common high-frequency excitation pulse (14), several echoes (16) are acquired in an echo train, wherein - a signal progression data set (24) describing the signal progression (17) of the measured magnetic resonance signal over a recording time period, in particular the echo train, in the magnetic resonance sequence is determined in k-space for the magnetic resonance data set (23), and - is passed to a trained image processing function to determine a result data set together with the magnetic resonance data set (23) as input data.

Description

[0001] The invention relates to a computer-implemented method for processing a magnetic resonance data set of an acquisition area, which is based on an acquisition during an examination procedure in a magnetic resonance device with a magnetic resonance sequence in which, in particular, several echoes are acquired in an echo train after a common radio frequency excitation pulse. The invention also relates to an image processing device, a computer program, and an electronically readable data carrier.

[0002] Magnetic resonance imaging (MRI) is now an established tool in medical technology. Since MRI scans often involve longer acquisition times, various approaches have been proposed to reduce these, for example, to increase patient throughput, allow for improved imaging of dynamic processes, and / or increase robustness to motion. To this end, MRI sequences have been developed to enable faster measurements. These include not only MRI sequences with ultrashort echo times but also sequences that utilize multiple readout modules, each capable of reading out k-space lines, following a common radio frequency excitation pulse from a corresponding excitation module. Such a sequence is also referred to as an echo train, with each readout module handling one echo.The respective echo can be generated, for example, by a special preparation module, such as one comprising a high-frequency refocusing pulse.

[0003] A well-known class of such magnetic resonance sequences are turbo spin echo (TSE) sequences, which employ the RARE (Rapid Acquisition with Relaxation Enhancement) technique. Here, a series of refocusing pulses, particularly with a flip angle of 180°, is used to generate an echo train following a high-frequency excitation pulse, for example, with a flip angle of 90°. The refocused echoes of this train are then measured in the corresponding readout modules. Different phase-encoding gradients are used for the echoes to acquire different k-space lines.

[0004] An example of a TSE sequence is the so-called SPACE sequence (Sampling Perfection with Application optimized Contrast using different flip angle Evolution). This is used particularly for 3D imaging (3D-SPACE). It is characterized, among other things, by long echo paths (e.g., more than 100 echoes), short echo spacing, at least partially reduced flip angles (e.g., to avoid tissue heating), and optimized, efficient k-space trajectories.

[0005] In all recording methods with long echo paths, various factors can lead to a significant weighting and thus implicit filtering of the magnetic resonance data in the sampled k-space. In other words, deviations of the actual measurement result from the ideally measured magnetic resonance signal (for example, at the nominal echo time or in a first echo after the radio frequency excitation pulse) or from a reference can occur. This is due to different effects at different times when sampling is performed at a k-space position, particularly (potentially manipulated) relaxation of excited echo signals during the measurement. For example, due to T2 / T2* decay, later echoes in an echo path exhibit lower intensity than earlier echoes.Varying the flip angles across the echo train to achieve signal storage and recovery along the longitudinal magnetization also leads to changes in intensity over time. Depending on the set acquisition parameters, the k-space trajectory can also change (reordering), for example, to achieve a desired echo time. Ultimately, the k-space trajectory defines when a k-space position is sampled.

[0006] However, differences in intensity in k-space do not have a relevant impact on the reconstruction of a magnetic resonance image in image space, particularly through or encompassing a Fourier transformation, from a corresponding magnetic resonance data set, since potential jumps or edges in intensity between neighboring k-space positions within the sampled k-space are rather small.

[0007] Another way to save acquisition time while still providing high-quality magnetic resonance images is to subsequently increase the spatial resolution, known as "superresolution." A lower spatial resolution means that a smaller fraction of k-space around the k-space center needs to be scanned to acquire a magnetic resonance dataset. A resolution enhancement function can then be applied to a base image reconstructed from the magnetic resonance dataset to increase the spatial resolution from the initial spatial resolution of the base image to a second spatial resolution, for example, doubling it.

[0008] Resolution enhancement functions, especially superresolution functions, and other application functions, such as subsampling reconstruction algorithms, are frequently trained using machine learning. For example, neural networks and / or comparable artificial intelligence architectures can be employed.

[0009] When training application functions, especially resolution enhancement functions, training data is often used that is based on a specific signal waveform. For example, it is known to extract the central k-space component from a high-resolution k-space dataset, whose Fourier transform forms the training output dataset (i.e., the fundamental truth). The central k-space component of this component then forms the training input dataset. This k-space dataset was acquired, for example, using single-echo imaging and therefore does not represent the signal waveform in a multi-echo measurement, particularly the T2 decay that occurs during the echo train.If the trained resolution enhancement function is subsequently applied to a magnetic resonance imaging (MRI) dataset acquired using a multi-echo sequence that reads out multiple echoes after a common radio frequency excitation pulse, a different point spread function (PSF) may be present, potentially leading to resolution loss and / or edge exaggeration. In other words, known trained resolution enhancement functions do not model the variable T2 decay effects in multi-echo imaging, particularly TSE sequences, resulting in a degradation of image quality.

[0010] To achieve improvements in this regard, prior art has proposed enhancing the training of the resolution enhancement function to achieve general "T2 deblurring." For example, in an article by Z. Chen et al., "Physics-informed deep learning for T2-deblurred superresolution turbo spin echo MRI," Magn. Reson. Med. 90 (2023), pages 2362 to 2374, it is proposed to train a model, i.e., a function, specifically with physically realistic resolution degradation (asymmetric T2 weighting of the raw high-resolution k-space data). Other methods that use a neural network for image data reconstruction are described in US 2023 / 0021786 A1 or in the article by Nitzan et al., “MA-RECON: Mask-aware deep-neural-network for robust fast MRI k-space interpolation”, Computer Methods and Programs in Biomedicine 244: p. 107942, 2024.

[0011] In an article by Z. Zhou et al., “Neural network enhanced 3D turbo spin echo for MR intracranial vessel wall imaging”, Magnetic Resonance Imaging 78 (2021), pages 7 to 17, it is proposed to reduce relaxation-induced blurring in acquired TSE images using a CNN in order to improve the overall image quality of TSE scans. The invention therefore aims to provide a method for improving the quality of image processing results, particularly at superresolution, of multi-echo magnetic resonance datasets.

[0012] This problem is solved according to the invention by a computer-implemented method, an image processing device, a computer program, and an electronically readable data carrier according to the dependent claims. Advantageous embodiments are described in the sub-claims.

[0013] In a method of the type mentioned at the outset, it is provided according to the invention that - a signal progression data set in k-space describing the signal progression of the measured magnetic resonance signal over a recording time period, in particular the echo train, in the magnetic resonance sequence is determined for the magnetic resonance data set and - is passed to a trained image processing function to determine a result data set together with the magnetic resonance data set as input data.

[0014] The trend data set is determined at least partially by a calculation and / or modeling and / or simulation, using at least one relaxation information about the relaxation in the recording area.

[0015] It is therefore proposed to determine the corresponding signal profile for each magnetic resonance data set to be processed, in particular describing the T2 decay processes and other factors influencing the intensity over the course of the magnetic resonance sequence, and to process it using artificial intelligence in the form of a trained image processing function in such a way that the individual intensity behavior over the specific recording process can be taken into account, which leads to a significant improvement in quality in the final processing result, be it directly in the result data set or in the output data of an application function into which the result data set is fed.Therefore, a deblurring is proposed using decay information individually determined for a magnetic resonance dataset in the form of the trend dataset and artificial intelligence in the form of the trained image processing function in order to obtain improved image quality, a reduction of artifacts and, in particular, improved sharpness when increasing resolution (superresolution).

[0016] In general, a trained function replicates cognitive functions that people associate with other human brains. Through training based on training data (machine learning), the trained function is able to adapt to new circumstances and detect and extrapolate patterns. Another term for "trained function" is "trained machine learning model."

[0017] Generally speaking, the parameters of a trained function can be adjusted through training. Specifically, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be used. Furthermore, representational learning (also known as feature learning) can be employed. The parameters of the trained function can be adjusted iteratively through multiple training steps. In particular, a specific cost function can be minimized during training. For example, the backpropagation algorithm can be used when training a neural network.

[0018] A trained function can, for example, comprise a neural network, a support vector machine (SVM), a decision tree, and / or a Bayesian network, and / or the trained function can be based on k-means clustering, Q-learning, genetic algorithms, and / or assignment rules. Specifically, a neural network can be a deep neural network, a convolutional neural network (CNN), or a deep CNN. Furthermore, the neural network can be an adversarial network, a deep adversarial network, and / or a generative adversarial network (GAN).

[0019] A particular advantage of the trained image processing function is that it includes at least one CNN. A Convolutional Neural Network (CNN) is a neural network that uses a convolution operation instead of general matrix multiplication in at least one of its layers, the so-called convolution layer. Specifically, a convolution layer can perform a scalar product of one or more convolution kernels on the incoming data / images of the convolution layer, where the entries of the one or more convolution kernels are the parameters or weights that are adjusted through training. In particular, the inner Frobenius product and the ReLu activation function can be used. A CNN can include additional layers, such as pooling layers, fully connected layers, and normalization layers. CNNs can process input images or...Datasets can be processed extremely efficiently because a convolution operation based on different cores can extract a wide variety of data features. By adjusting the weights of the convolution cores, the relevant data features can be identified during training. Furthermore, because weights are shared among the convolution layer cores, fewer parameters need to be trained, thus avoiding overfitting during the training phase and allowing for faster training or a greater number of layers in the CNN, thereby increasing the network's performance.

[0020] It should be noted at this point that, within the scope of this description, a signal waveform should be understood as meaning that the data of the waveform dataset describe a comparable measure of the signal intensity, in particular taking into account the T2 signal decay and possibly other effects such as signal recovery through flip angles, over the recording time period, especially the echo train, and thus, if the order of sampling the k-space information, i.e. the k-space trajectory, is known, possibly after a reordering, also in k-space.In other words, the time series dataset assigns, for example, k-space positions, and in particular at least a subset of the sampled k-space positions, a measure of signal intensity relative to a reference. This measure is derived from the temporal signal evolution over the recording time period, for example, the echoes, including at least their T2 relaxation and optionally at least one other effect (e.g., varying flip angles), and the respective sampling time within the recording time period, particularly in the echo train, due to the sampling order. Specifically, the time series dataset can, for example, be a decay matrix in which each sampled k-space position is assigned a decay value as a measure of signal intensity, for example, in the form of a weighting with respect to a reference curve and / or one of the echoes, in particular the first echo (which would correspond to a single-echo recording as the reference).

[0021] The invention is described below primarily with reference to a magnetic resonance sequence in which, particularly after a common high-frequency excitation pulse, several echoes are recorded in an echo train (as the acquisition time interval), since particularly advantageous results have been observed here. However, it is also quite possible to use the invention in other fields of application where a signal waveform may occur that differs significantly from other signal waveforms of other magnetic resonance sequences, but for which the same image processing measures, especially those of artificial intelligence, are to be applied. An example is magnetic resonance sequences with ultrashort echo times (UTE sequences), in which a strong decay of the signal can already be observed along a single k-space line to be sampled. This will also be briefly discussed in some examples.

[0022] To determine the time series dataset, a particularly advantageous refinement provides that the time series dataset is derived at least partially from measurement data of a measurement of the examination process. A measurement has the particular advantage of actually capturing the individual measurement circumstances and characteristics of the recording process within the recording area. It has been recognized that in many recording processes, especially when using multiple magnetic resonance sequences with a single echo train, comparable partial measurements of the measured signal intensity are either already being performed at different times along the echo train anyway, or can at least be easily performed, particularly as an additional measure. From this, the signal profile over time can then be determined along the echo train, and, taking into account the sampling sequence during the acquisition of the magnetic resonance data, the signal profile can also be determined in k-space.

[0023] For example, in a particularly advantageous embodiment, magnitude data from a phase-corrected scan of the magnetic resonance sequence, acquired without phase-encoding gradients, can be used as the measurement data. Within the scope of the present invention, it was recognized that, for example, in TSE sequences, a phase-corrected scan is often acquired prior to the actual measurement of the magnetic resonance data to correct eddy current effects. This partial measurement of the acquisition process comprises a complete echo train and is performed without phase encoding, so that the individual echoes are directly comparable with respect to signal decay. Typically, only the phase data from such a phase-corrected scan are used.According to the invention, it is now proposed to extract the signal profile from the magnitude data recorded in the phase correction scan, which describe the signal progression during the echo train.

[0024] Additionally or alternatively, it may be provided that supplementary measurement data of certain echoes, in particular comprising the first and the last echo of an echo train, which were recorded without phase-encoding gradients, and / or repeat measurement data of at least one k-spatial segment that was measured several times during the echo train, are used as measurement data. In magnetic resonance sequences or recording processes in which no phase correction scan is performed, for example in the HASTE sequence, it may therefore be provided, for example, that non-phase-encoded echoes are measured in at least one echo train in which, in particular, magnetic resonance data are measured.Since the k-space center is measured without phase coding anyway, the first and last echoes, for example, can also be recorded without phase coding. This allows an exponential decay profile to be calculated from these at least three data points, enabling signal intensities to be interpolated for other k-space positions and thus determining a signal waveform. Alternatively or additionally, it is also conceivable to use the same phase coding multiple times, for example, by measuring the same k-space line multiple times, in order to obtain support points for interpolation and / or extrapolation.

[0025] It should be noted here that in such magnetic resonance sequences, which utilize echo trains with multiple echoes, a k-space line is typically acquired for each echo. In principle, other partial k-space trajectories for acquisition modules of individual echoes are conceivable, but less preferred. It is also conceivable to record the decay along k-space lines (usually k xTo resolve the loss, however, the differences in the weighting resulting from the loss are often assumed to be smaller, so that in the trend data set, the k-space positions of a k-space line can be assigned a common signal trend value (decay value), in particular in the sense of a weighting applicable to the entire k-space line. It is possible, for example, to determine signal intensities averaged over the k-space lines, i.e., the readout direction, and to use the averaged signal intensities over the echo path to derive the signal trend along at least one phase-encoding direction.

[0026] If only individual echoes are considered, for example in the aforementioned case of UTE sequences, the trend data set can be determined, for instance, by measuring a free induction decay (FID). This means recording a magnetic resonance signal without encoding in the readout direction. The decay can then be determined from the signal envelope.

[0027] The trend dataset will be determined at least partially by calculation and / or modeling and / or simulation, using at least one relaxation information about the relaxation in the recording area. A calculation or similar process is preferably performed based on the measurement data, so that, as already explained above, the trend dataset can be extrapolated and / or interpolated from support points provided by the additional measurement data together with measurement data of the k-space center and / or by the repeated measurement data, using a decay model. As already mentioned, an exponential decay model, for example, can be used here.

[0028] It is conceivable, though less preferred, to use more distant measurement data or even to make assumptions about the recording range in order to determine the progression dataset based on a calculation, a model, and / or a simulation. For example, the relaxation information may include at least a T2 and / or T2* relaxation time that can be used as the basis for the decay.

[0029] The relaxation information can be determined, at least partially, from a previously measured map and / or relaxometry measurement of the imaging area, particularly during the examination procedure. Relaxometry and / or mapping methods can therefore be used. For example, if a T2 map of the imaging area is available, a mean decay curve for the entire k-space can be calculated from the layer-wise T2 value distribution, or a spatially resolved determination, for example, layer-wise, can be performed.

[0030] Furthermore, it can also be provided that the relaxation information is determined from a mapping rule that assigns at least part of the relaxation information to information about the anatomical location and / or the material composition of the acquisition area. Such a mapping rule can also use at least one sequence parameter of the magnetic resonance sequence as an input value, for example, to determine the radiofrequency excitation. For example, based on a body region of the acquisition area and set sequence parameters, a signal waveform can be determined from a lookup table with typical decay values.

[0031] The trend data set is determined with resolution in at least one direction of k-space, for example, in the at least one phase-encoding direction. In general, a preferred embodiment of the invention may provide that the trend data set, particularly from the measurement data, is determined along a readout direction of the magnetic resonance sequence and / or for different local coil elements of a local coil arrangement used. It is therefore possible to provide resolution along the readout and / or coil dimension, at least in an intermediate step, in order to provide the most comprehensive information possible as a trend data set.

[0032] In general, it can be advantageous to provide the historical dataset in the same dimensionality and / or size as the magnetic resonance dataset as input data for the trained image processing function, since this allows for simpler joint processing and / or structure of the trained image processing function.

[0033] In a particularly preferred first embodiment of the present invention, the trained image processing function can be a preparatory function that outputs as output data (and thus result data set) a corrected magnetic resonance data set whose signal waveform corresponds to a reference waveform, in particular the signal waveform of a single-echo acquisition. In this case, a preparatory function is used to correct the magnetic resonance data set with respect to the signal waveform, in particular the signal decay across the echo path or the resulting weighting, such that the corrected magnetic resonance data set provided as output data corresponds to one that would be obtained with a reference waveform (for example, single-echo imaging). In particular, the reference waveform is determined by the signal waveform during acquisition or...other determination of application function training data that was used for training a trained application function to be used below.

[0034] In other words, it can be provided that the corrected magnetic resonance dataset is used as input data for a trained application function, in particular a resolution enhancement function, where the application function is trained with application function training datasets whose application function training input sets exhibit the reference curve. For example, if a resolution enhancement function, in particular a superresolution function, is trained with application function training data based on single-echo imaging, in which, therefore, T2 signal decay and possibly other effects, such as those that lead to weighting effects in multi-echo imaging, do not occur, the effect of signal decay can be factored out by means of the trained preparation function and thus has no negative impact on the performance of the application function, in this case the resolution enhancement function. This results in better image quality and a reduced number of artifacts.

[0035] It should be noted here that a trained preparatory function is used to reduce or eliminate the weights in the magnetic resonance (MRI) dataset that arise from the signal waveform. At least in theory, it would be possible to correct the MRI dataset by dividing it by the waveform dataset containing the weights. However, in practice, MRI data exhibits noise, with the noise level remaining constant across k-space, while the signal (or its weights) varies. A direct modification of the MRI dataset with the waveform dataset would therefore result in a significant increase in the noise level. It is therefore proposed to use machine learning to generate a corrected MRI dataset without increasing the noise.

[0036] The preparation function can conveniently include a Convolutional Neural Network, in particular a U-Net. A U-Net is a CNN that was originally developed for image segmentation (see the article by O. Ronneberger et al., "U-Net: Convolutional Networks for Biomedical Image Segmentation", arXiv:1505.04597) and is also excellently suited for various other image-to-image processing tasks, such as the modification presented here with regard to signal propagation or signal decay, insofar as it leads to a weighting of different echoes or k-space points along a k-space line.In a specific implementation, for example, the original magnetic resonance data set (in k-space) and the subsequent data set, adapted to the same matrix size, particularly as a decay matrix, can be passed as input data to the preparation function comprising a U-Net for the interference process. The output layer then outputs a k-space matrix corrected for the decay, namely the corrected magnetic resonance data set. The k-space thus corrected to a reference profile can then be passed to the application function, which may comprise at least one further network, optionally after at least a partial Fourier transform into image space. Such a transformation is particularly useful if the application function is a resolution enhancement function. The application function can also be, for example, a reconstruction function of the parallel imaging.Of course, other network architectures are also conceivable for the preparation function.

[0037] Several approaches are conceivable for training the preparation function. In a first possible embodiment, the training of the preparation function can be carried out by recording preparation function training input data sets in a first training measurement and preparation function training output data sets in a second training measurement using the reference curve, in particular as a single-echo measurement. For example, it can be provided that a training pair consisting of preparation function training input data set and preparation function training output data set is determined by performing a first training measurement with an echo train recording and a second training measurement with a single-echo recording, but with the same timing, i.e., the same echo time, on the same, as still and unchanged, as possible recording area, in particular directly consecutively.For example, a TSE sequence can first be performed with a turbo factor of 8, followed by the same magnetic resonance sequence as a single-echo measurement with the same echo time (essentially with a turbo factor of 1). A particularly useful advanced training approach involves using stationary objects, especially phantoms and / or fruits and / or vegetables and / or dead tissue, to obtain training input and output data in successive measurement runs. In other words, to avoid physiological effects in subjects, objects with relaxation times and structures comparable to the human body can be used. Examples include anatomical phantoms, fruits, and / or vegetables.Additionally or alternatively, but less preferably, it may also be provided that at least part of the preparatory function training data is determined by simulation, in particular Bloch simulation using the reference curve.

[0038] Regarding training measurements, it should be noted that, as with other methods for acquiring training data, the reference curve of a single-echo measurement does not necessarily have to correspond to that of a multi-echo measurement. For example, other training output datasets can be generated that simply exhibit less decay than the corresponding training input datasets. In other words, imperfect target k-spaces can also be generated, for example, by reducing the turbo factor in TSE measurements. For instance, in the first training measurement, training input data with a turbo factor of 16 can be acquired, and in the second training measurement, training output data with a turbo factor of 4 (which then define the reference curve) can be acquired. For HASTE sequences, parallel imaging can also be used to reduce the echo train length when acquiring training output data.For example, in the first training measurement, training input presets can be acquired using parallel imaging with a 2x acceleration, while in the second training measurement, training output presets (target k-spaces) can be acquired using a 4x parallel imaging acceleration. Subsequently, the presets are reconstructed using parallel imaging reconstruction functions to obtain the corresponding training dataset. In other words, the reconstructed k-spaces then serve as the input and output k-spaces for training the preparatory function.

[0039] Regarding the UTE sequences mentioned, training of the preparation function can be performed, for example, using recordings with a low field strength of the background magnetic field, such as 0.5 to 0.6 T, where a correspondingly lower, and therefore slower, decay occurs. A simulated signal curve for higher field strengths can then be superimposed on the measurement results.

[0040] In a particularly advantageous further development of the present invention with regard to the determination of preparatory function training datasets, it can be provided that * a synthesized and / or at least essentially noise-free basic dataset of the reference curve in k-space is provided, * at least one training history dataset is applied to the base dataset to determine a modification dataset, and * the modification data set and the basic data set are provided with noise, in particular Gaussian noise, to obtain a preparation function training input data set and a preparation function training output data set, respectively.

[0041] This approach therefore starts with an existing, low-noise or even noise-free basic dataset of the reference waveform in k-space, whereby such a basic dataset can be synthesized, for example, particularly through simulation, but can also be obtained in other ways, for example by using denoising methods and the like. The basic dataset can then preferably be modified, in particular multiplied, with a multitude of different training waveform datasets describing signal waveforms over k-space, in particular decay matrices, in order to subsequently superimpose the basic dataset and the modification result with noise, in particular Gaussian noise, with the same distribution to obtain pairs of pre-function training input datasets and pre-function training output datasets. In this way, a simple basic generation or...Training data can be expanded. It should be noted that the preparation function can also be trained in this way with regard to denoising if noise of lower amplitude or lower noise levels is used for the modification dataset.

[0042] The training methods for the preparation function (or other functions) described here can also be the subject of a separate deployment procedure, which can be carried out independently of the generally described processing procedure, in which the respective trained function is deployed upon completion. For example, a deployment procedure for a trained preparation function, as introduced above, is conceivable, in which the preparation function is trained using... - to determine preparatory function training datasets * a synthesized and / or at least essentially noise-free basic dataset of the reference curve in k-space is provided, * at least one training history dataset is applied to the base dataset to determine a modification dataset, and * the modification data set and the basic data set are provided with noise, in particular Gaussian noise, to obtain a preparation function training input data set and a preparation function training output data set respectively, and / or - Preparation function training input datasets are recorded in a first training measurement and preparation function training output datasets are recorded with a second training measurement using the reference curve, in particular as a single echo measurement, and / or - at least a portion of the preparation function training data is determined by simulation, in particular Bloch simulation using the reference curve, after which the preparation function is trained using the preparation function training data sets thus determined, and the trained preparation function is made available, for example via an interface. Such a provisioning procedure can be provided by a provisioning system and / or stored as a provisioning computer program on an electronically readable provisioning medium.

[0043] In a second embodiment of the present invention, which uses an alternative preparation function, the trained image processing function can be an application function, in particular a resolution enhancement function, for determining a result data set as output data. It is therefore also possible to pass the trend data set, for example as a decay matrix in k-space, which can specify a decay value, in particular as a weighting, for each sampled k-space position, as additional input data to an architecture existing for a specific purpose, i.e., the trained application function.In other words, an existing superresolution functional architecture, or any other existing trained resolution enhancement functional architecture in machine learning, can be modified to accept a historical dataset, such as a decay matrix, as additional input data in the input layer. Such an architecture can operate in k-space, hybrid space, or image space. For example, the historical dataset could reside in k-space, while the magnetic resonance data might reside in image space or a hybrid space.

[0044] Regarding the training of such an application function, in particular the acquisition of corresponding application function training datasets, comprising application function training input datasets and application function training output datasets, approaches comparable to those described for the preparation function can be used. In particular, training measurements can then also be performed, preferably using stationary objects, especially phantoms and / or fruits and / or vegetables and / or dead tissue, to carry out training measurements for acquiring training input data and training output data in successive measurement processes.

[0045] The method according to the invention can generally be used for both two-dimensionally scanning magnetic resonance sequences and three-dimensionally scanning magnetic resonance sequences.

[0046] In addition to the processing method, the present invention also relates to an image processing device comprising a computing device with at least one processor and at least one storage means, wherein the computing device is for processing a magnetic resonance data set of an imaging area, which is based on an imaging procedure in a magnetic resonance device with a magnetic resonance sequence in which, in particular, several echoes are recorded in an echo train after a common high-frequency excitation pulse: - an investigation unit for determining a trend data set in k-space describing the signal progression of the measured magnetic resonance signal over a recording time period, in particular the echo train, in the magnetic resonance sequence, and - a processing unit for determining a result data set using a trained image processing function that uses the trend data set together with the magnetic resonance data set as input data.

[0047] All statements relating to the method according to the invention can be applied analogously to the image processing device according to the invention and vice versa, so that the advantages already mentioned can also be obtained with the image processing device.

[0048] In the computing device, functional units are formed by hardware and / or software to carry out steps of the processing method according to the invention. Further functional units may be provided for implementing further preferred embodiments of the processing method. For example, the computing device may also include an application unit for applying an application function and / or a training unit. The training unit may include a subunit for determining training data sets.

[0049] The image processing device according to the invention can be advantageously integrated into a magnetic resonance imaging (MRI) system. In this case, the computing unit can be provided as part of a control unit for the MRI system. The control unit can also include a sequence unit that can control the acquisition of the MRI data set as well as the acquisition of measurement data in general, and in particular also additional measurement data and / or repeated measurement data. The sequence unit can also be configured to perform first and second training measurements to determine training data sets.

[0050] A (processing) computer program according to the invention can be directly loaded into a storage medium of a computing unit or image processing unit and comprises program elements such that, when the computer program is executed on the computing unit, the latter is caused to carry out the steps of a (processing) method according to the invention. The computer program can be stored on an electronically readable data carrier according to the invention, which therefore includes control information stored thereon, comprising at least one computer program according to the invention and designed such that, when the data carrier is used in a computing unit or image processing unit, the latter is configured to carry out a method according to the invention. The data carrier can, in particular, be a non-transient data carrier, for example, a CD-ROM.

[0051] Further advantages and details of the present invention will become apparent from the exemplary embodiments described below and from the drawings. These show: Fig. 1 a flowchart of a first embodiment of the method according to the invention, Fig. 2 excerpts from the course of an examination procedure using a multi-echo magnetic resonance sequence, Fig. 3 an exemplary signal waveform via an echo train, Fig. 4 a sequence diagram of a HASTE echo train with additional measurements, Fig. 5. Schematic representation of the structure of a preparatory function, Fig. 6. A procedure for training a preparatory function, Fig. 7 a flowchart of a second embodiment of the method according to the invention, and Fig. 8 the functional structure of an image processing device according to the invention.

[0052] The following describes exemplary embodiments of the present invention with respect to two-dimensionally scanning magnetic resonance sequences. In these embodiments, to record a magnetic resonance data set in at least one echo train, k-space is scanned along k-space lines according to a scanning sequence, specifically a k-space trajectory, wherein in each echo a k-space line is scanned along different k-space positions. However, the process described here can also be applied to three-dimensionally scanning magnetic resonance sequences and / or magnetic resonance sequences scanning along other scanning patterns.

[0053] When multiple echoes are read out in multiple readout modules within an echo path, a certain signal pattern is present along the echo path, primarily due to T2 decay. In other words, a different magnetic resonance signal is measured depending on which echo a particular k-space line is actually sampled as along the echo path, according to the sampling order. This effect can be interpreted as a weighting. Besides T2 decay, relevant effects can also arise due to variable flip angles along the echo path. Different signal intensities depending on when sampling occurs along an echo path do not occur with single-echo recordings. Simultaneously, the signal patterns differ between different specific magnetic resonance sequences, particularly depending on the sampling order in k-space.This can affect the quality of processing results obtained through the use of application functions trained, in particular, by machine learning, especially if these functions are designed or trained for other, particularly essentially constant or uniform, signal waveforms across the sampled k-space. Two exemplary implementations are presented below to demonstrate improvements in this regard.

[0054] In the first embodiment of the Fig. In step S1, magnetic resonance (MRI) data of a scan area, for example, a patient, are acquired during an examination procedure. A MRI sequence is used, and after a common radio frequency excitation pulse, several echoes are read out in an echo train. Measurement data are also acquired, particularly using the MRI sequence, which can be used in step S2 to determine a trend dataset that describes the signal evolution during sampling in k-space. The measurement data describes the signal evolution, especially the T2 decay, during the echo train. From this, and given knowledge of the respective sampling order in the at least one echo train used to acquire the MRI dataset, a signal evolution across k-space can also be derived.

[0055] Fig. Section 2 explains this in more detail using a first example, namely a TSE sequence as a magnetic resonance sequence. In TSE sequences, phase correction scans for eddy current correction are often performed before the actual magnetic resonance data is acquired. Fig. Figure 2 shows acquisition sections 1 and 2, in which a phase-corrected scan is performed for a first and a second layer, respectively. The continuation points indicate that further layers may follow. In the sequence diagram 3 associated with acquisition section 1, the top graph 4 shows the radio frequency activity, the second graph 5 shows readout gradients, the third graph 6 shows phase-encoding gradients, and the fourth graph 7 shows readout time windows 8 for acquiring the respective echoes 16. The structure known for multi-echo acquisitions is recognizable, in which, after an initial, common radio frequency excitation pulse 14, echoes 16 are generated after each refocusing pulse 15 and are read out; again, compare readout time windows 8.

[0056] It is evident that no phase encoding gradients are used in the phase correction scan, so the magnitude data provide comparable information about the signal waveform across the echo train. Thus, while the phase data from the phase correction scan are evaluated for eddy current correction as usual, the magnitude data can be used as measurement data in step S2 to determine the waveform dataset.

[0057] Fig. Figure 2 also shows the further course, for example, acquisition sections 9 and 10 of a first shot for the first and second layers, possibly continued for further layers, followed accordingly by acquisition sections 11 and 12 for the second shot. As can be seen from the sequence diagram 13 of the first shot for the first layer, phase encoding gradients are used there (see Graph 6) to sample different k-space lines.

[0058] Fig. Figure 3 shows, for clarification, a signal waveform 17 (intensity I) versus the echo number E. This is a signal waveform 17 averaged (with respect to the readout direction) in the k-space center. However, it is of course also possible to determine the signal with respect to the readout direction and / or even with resolution along the coil dimension.

[0059] If the sampling sequence for the subsequent recording sections 9, 10, 11, 12, ... is known, and thus the k-space trajectory, the signal profile 17 along the echo dimension can be transferred to k-space (in particular, the phase encoding direction(s)), since it is known which echo 16 of an echo train is read out along which k-space line. Therefore, in step S2, a decay matrix can be determined as the profile data set, in which each sampled k-space position can be assigned a measure of the signal intensity at the time of sampling, for example, a relative weighting.

[0060] If a magnetic resonance sequence is used that does not include a phase-corrected scan, for example a HASTE (half Fourier-acquisition single-shot turbo spin echo) sequence, non-phase-coded k-space lines can be interwoven into at least one echo train to obtain measurement data. This is possible, for example, in Fig. Figure 4 indicates a sequence diagram 18 of an echo train of a HASTE sequence. The k-space center, readout module 19, is recorded without phase-encoding gradients. Additionally, the first echo 16 (compare readout module 20) and the last echo 16 (compare readout module 21) are also recorded without phase-encoding gradients to obtain supplementary measurement data. From these three (and possibly more) data points, an exponential decay can be determined via the echo train by interpolation, for example, using a specified decay model, so that a signal waveform 17 can be obtained as in Fig. This results in 3. Alternatively or additionally, it is also conceivable to record a certain number of k-space lines twice and to determine the trend data set from the signal evolution.

[0061] It should be noted that it is also conceivable, in principle, to determine the progression dataset, at least partially, through simulation, calculation, and / or modeling, even without measurement data acquired with the magnetic resonance sequence. Instead or additionally, relaxation information can be used. For example, preliminary measurements of the examination procedure, particularly relaxometry or mapping techniques, can be utilized. The decay of the magnetic resonance signal across the entire k-space can also be calculated from, for example, a layer-by-layer T2 map. Furthermore, it is conceivable to determine a signal progression 17 from a lookup table of typical decay values ​​based on the body region and / or the set sequence parameters.

[0062] Preferably, the trend dataset is defined as a decay matrix in k-space, which assigns a decay value as a measure of the signal strength to each sampled k-space position. The dimensions of the trend dataset then correspond to those of the magnetic resonance dataset in k-space (where the corresponding k-space position is assigned the corresponding k-space value of the magnetic resonance data).

[0063] In step S3, a trained preparation function is applied as a processing function to input data comprising the original magnetic resonance (MRI) dataset and the history dataset, specifically data derived from these. As a result, the trained preparation function provides a corrected MRI dataset that corresponds to a reference history, particularly a single-echo measurement, where the differences in signal intensity from echo to echo are eliminated or at least reduced. In other words, the corrected MRI dataset is corrected for the decay of the MRI signal according to the signal history that deviates from the reference history. In a single-echo measurement, the same signal intensity is expected for each echo (and the same signal history along the readout direction each time).

[0064] Fig. Figure 5 shows a specific implementation in which the trained preparation function 22 is designed as a CNN, specifically as a U-Net. The input data consists of the original magnetic resonance data set 23 in k-space and the progress data set 24 in k-space. The output data is the corrected magnetic resonance data set 25 in k-space.

[0065] The U-Net, as schematically depicted and generally known, comprises an encoder arm 26 and a decoder arm 27. Skip connections 28 exist between the layers of arms 26 and 27.

[0066] Generally speaking, the U-Net comprises not only convolution layers but also pooling and upsampling layers. Along the encoder arm, the input data, in particular at least the original magnetic resonance data set 23, is first downsampled, after which it is upsampled again along the decoder arm 27 to obtain the corrected magnetic resonance data set 25. The resulting U-shaped architecture gives the U-Net its name.

[0067] In other embodiments, other network architectures for the preparation function 22 are also conceivable.

[0068] Returning to Fig. 1. The corrected magnetic resonance dataset 25, optionally after a Fourier transform into image space, serves as input data for a trained application function, here exemplified by a resolution enhancement function. For example, this could be a super-resolution function. The resolution enhancement function was trained with training data exhibiting the reference curve, so that optimal processing results are obtained, which can then be made available.

[0069] Fig. Figure 6 schematically illustrates a possible procedure for training the preparatory function 22. In this procedure, 29 preparatory function training datasets are first compiled in a step group. Several methods exist for this, which can be used alternatively or in combination.

[0070] In a first approach, training input datasets are measured on the same object in a first training measurement (S11), and training output datasets are measured in a second training measurement. In this case, anatomical phantoms, fruits, and vegetables serve as the measurement objects because their decay properties and structure are very similar to recording areas of the human body, but they neither move nor change in any other way, for example, through physiological processes, between the first and second training measurements. In the example described above, a first training measurement can be performed using a TSE sequence with a turbo factor of 8 in order to derive a magnetic resonance dataset subjected to a signal curve (17) and a curve dataset as training input datasets.The second training measurement can then be performed with a turbo factor of 1, i.e., as a single-echo measurement, although the same echo time is used. In this case, the reference sequence is the single-echo measurement. However, it is also conceivable to target imperfect target k-spaces, for example, to perform the first training measurement with a turbo factor of 16 and the second training measurement with a turbo factor of 4, thus using a different reference. In the HASTE sequence, the echo train length can be shortened, for example, by parallel imaging.

[0071] In another approach to compiling preparatory function training datasets, a base dataset is provided in step S13. This base dataset contains no noise due to synthesis, or very little noise, for example, due to a prior denoising process. Because of the absence or extremely low noise, this base dataset can then be easily modified directly in step S14 with different historical datasets to obtain modification datasets. If the historical datasets are decay matrices with weights as decay values, a simple multiplication of the base dataset with the historical dataset can be performed.In step S15, each pair of basic data set and modification data set is then superimposed with Gaussian noise of the same distribution to obtain preparation function training data sets, each with a preparation function training output data set and a preparation function training input data set.

[0072] In step S16, the preparation function training datasets are used to train preparation function 22. The trained preparation function 22 can then be made available in step S17.

[0073] The schedule of the Fig. Figure 7 describes a second, modified embodiment of the processing method according to the invention. While steps S1 and S2 remain unchanged in their function of acquiring the magnetic resonance data set 23 / the measurement data and determining the trend data set 24, these are now directly provided to the correspondingly modified application function, in this case again a trained resolution enhancement function, in step S5. The resolution enhancement function thus uses the trend data set as additional input data and performs the correction for the trend internally. Suitable training data can also be obtained here, for example, by first and second training measurements.

[0074] Fig. Figure 8 finally shows an image processing device 30 according to the invention and the functional structure of its computing unit 31. The image processing device 31 can be integrated into a magnetic resonance device, in particular by integrating the computing unit 31 into its control unit.

[0075] The computing unit 31 comprises at least one processor and at least one storage medium 32. Magnetic resonance and measurement data can be obtained via an interface 33. In the case of a control unit for a magnetic resonance device, it can include a sequence unit that controls the recording operation and provides the original magnetic resonance data set 23 as well as the measurement data from which the trend data set 24 is to be determined via the interface 33. Further information can also be obtained via the interface 33, for example, sequence parameters, relaxation information, and the like.

[0076] In an investigation unit 34, the progress data record 24 is determined according to step S2. The present example illustrates the implementation of the procedure according to the first embodiment ( Fig. As shown in Figure 1, a processing unit 35 is used to apply the preparation function 22 to the original magnetic resonance data set 23 and the history data set 24 according to step S3. In an application unit 36, the application function can then be applied to the corrected magnetic resonance data set 25 according to step S4.

[0077] In alternative configurations that are suitable for carrying out the procedure in accordance with Fig. 7, the processing unit 35 is directly trained to apply the trained application function as a trained image processing function to the original magnetic resonance data set 23 and the trend data set 24 according to step S5.

[0078] Processing results, in particular the output data of the application function, can be provided via interface 33 or another interface.

[0079] The computing unit 31 can also include a training unit 37 as an additional functional unit, for example to prepare the preparatory function 22, as with regard to Fig. As described in section 6, training is possible. However, it is also possible to use a separate deployment system and training system.

[0080] It should be noted that the procedure described here can also be applied to other use cases besides multi-echo measurements, where a relevant signal waveform, particularly one that differs from other magnetic resonance sequences, is present over a recording period. One example is UTE sequences, where the waveform dataset can, for instance, refer to the k-space line of an echo.

[0081] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

Claims

[1] Computer-implemented method for processing a magnetic resonance data set (23) of an acquisition area based on an acquisition in an examination procedure in a magnetic resonance device with a magnetic resonance sequence in which, in particular, after a common radio frequency excitation pulse (14), several echoes (16) are acquired in an echo train, characterized by , that - a signal progression data set (24) describing the signal progression (17) of the measured magnetic resonance signal over a recording time period, in particular the echo train, in the magnetic resonance sequence is determined in k-space for the magnetic resonance data set (23), and - is passed to a trained image processing function to determine a result data set together with the magnetic resonance data set (23) as input data, characterized by, that the progress data set (24) is determined at least partially by a calculation and / or modeling and / or simulation, using at least one relaxation information about the relaxation in the recording area. [2] Method according to claim 1, characterized by , that the progress data set (24) is determined at least partially from measurement data of a measurement of the examination process. [3] Method according to claim 2, characterized by that the measurement data - Magnitude data from a phase-corrected scan using the magnetic resonance sequence acquired without phase-encoding gradients, and / or - Additional measurement data of certain echoes (16), in particular comprising the first and last echo (16) of an echo train, recorded without phase-encoding gradients, and / or - Repeat measurement data for at least one k-space section that was measured several times during the echo train, be used. [4] Method according to any one of the preceding claims, characterized by , that the relaxation information includes at least one T2 and / or T2* relaxation time and / or is determined at least partially from a map and / or relaxometry measurement of the recording area previously measured, in particular during the examination procedure, and / or from an assignment rule that assigns at least part of the relaxation information to information about the anatomical location of the recording area and / or the material composition of the recording area. [5] Method according to any of the preceding claims, characterized by , that the trend data set (24), in particular from the measurement data, is determined along a readout direction of the magnetic resonance sequence and / or for different local coil elements of a local coil arrangement used. [6] Method according to any of the preceding claims, characterized by, that the trained image processing function is a preparation function (22) which outputs as output data a corrected magnetic resonance data set (25) whose signal profile corresponds to a reference profile, in particular the signal profile of a single echo recording. [7] Method according to claim 6, characterized by , that the corrected magnetic resonance data set (25) is used as input data for a trained application function, in particular a resolution enhancement function, wherein the application function is trained with application function training data sets whose application function training input sets have the reference profile. [8] Method according to claim 6 or 7, characterized by , that the preparation function (22) includes a Convolutional Neural Network, in particular a U-Net. [9] Method according to any one of claims 6 to 8, characterized by , that for training the preparatory function (22) - to determine preparatory function training datasets * a synthesized and / or at least essentially noise-free basic dataset of the reference curve in k-space is provided, * at least one training history dataset is applied to the base dataset to determine a modification dataset, and * the modification data set and the basic data set are provided with noise, in particular Gaussian noise, to obtain a preparation function training input data set and a preparation function training output data set respectively, and / or - Preparation function training input datasets are recorded in a first training measurement and preparation function training output datasets are recorded with a second training measurement using the reference curve, in particular as a single echo measurement, and / or - at least some of the preparatory function training data should be determined by simulation, in particular Bloch simulation using the reference curve. [10] Method according to any one of claims 1 to 5, characterized by that the trained image processing function is an application function, in particular a resolution enhancement function, for determining the result data set as input data. [11] Method according to claim 9 or 10, characterized by , that for carrying out training measurements to obtain training input data and training output data in successive measurement processes, stationary measurement objects, in particular phantoms and / or fruits and / or vegetables and / or dead tissue, are used. [12] Image processing device (30) comprising a computing device (31) with at least one processor and at least one storage means (32), wherein the computing device (31) is designed to process a magnetic resonance data set (23) of an imaging area based on an image acquired during an examination in a magnetic resonance device with a magnetic resonance sequence in which, in particular, several echoes (16) are acquired in an echo train following a common radio frequency excitation pulse (14): - an investigation unit (34) for determining a trend data set (24) in k-space describing the signal progression (17) of the measured magnetic resonance signal over a recording time period, in particular the echo train, in the magnetic resonance sequence to the magnetic resonance data set (23), and - a processing unit (35) for determining a result data set using a trained image processing function that uses the trend data set (24) together with the magnetic resonance data set (23) as input data, wherein the trend data set (24) is determined at least partially by a calculation and / or modeling and / or simulation, using at least one relaxation information about the relaxation in the recording area. [13] Computer program which, when executed on a computing device (31) of an image processing device (30), causes the latter to perform the steps of a method according to any one of claims 1 to 11. [14] Electronically readable data carrier on which a computer program according to claim 13 is stored.

Images