CORRECTION OF CHEMICAL SHIFT ARTIFACTS FROM BIPOLAR DIXON MR ACQUISITION DATA USING A CONVOLUTIONAL NEURAL NETWORK (CNN)

Description

[0001] The present invention relates to a method for correcting chemical shift artifacts (CSA), which occur in the magnetic resonance Dixon method when using bipolar readout gradients (bipolar Dixon-MR) to acquire the in-phase and anti-phase echoes.

[0002] The present invention relates to the field of magnetic resonance imaging (MRI), as used in medicine for patient examinations. Magnetic resonance imaging is based on the application of spin-echo and gradient-echo sequences, which enable excellent image quality within a given acquisition time.

[0003] The invention is particularly applicable to methods that use a binary readout gradient, such as magnetic resonance (MR) measurement according to the MR imaging method published by Dixon in 1984, which allows water and fat tissue to be displayed separately.

[0004] In his original implementation, Dixon acquired one image with water and fat signals in phase and another image with water and fat signals 180° out of phase. Dixon demonstrated that simple summation and subtraction of the two images could yield a pure water image or a pure fat image, respectively. The Dixon method or technique essentially encodes the chemical shift difference into the signal phase using modified data acquisition and then achieves water / fat separation through post-processing.

[0005] The Dixon magnetic resonance imaging method is widely known and is now used in routine clinical practice. Details on the Dixon method are known and are described, for example, in the 2008 review article by Jingfei Ma (Jingfei Ma, "Dixon Techniques for Water and Fat Imaging," Journal of Magnetic Resonance Imaging, 28, 543-558 (2008)).

[0006] It is known, and also described by Jingfei Ma (2008), that the Dixon MR technique leads to chemical shift artifacts caused by the influence of chemical shift at the interfaces between fat and water. Preventing or minimizing these artifacts is the subject of research, as described, for example, in DE 10 2008 057 294 A and EP 3 796 023 A1.

[0007] EP 3 796 023 A1, for example, describes a method that avoids a reduction in image quality and thus the occurrence of disturbing artifacts by switching the gradient polarity between individual repetitions during Dixon acquisition with bipolar readout gradients. However, this method is only applicable when multiple average values ​​are acquired.

[0008] An MRI scan can take up to 90 minutes, making it impractical for some applications or patients, as it requires lying still for such a long period. Typically, babies or people suffering from Parkinson's disease or claustrophobia cannot remain in the scanner for such a long time without undergoing general anesthesia. Therefore, the goal is to keep the scan time as short as possible while maintaining good quality.

[0009] The fast-DIXON method is a measurement based on the Dixon method and uses a bipolar readout gradient. It is a fast spin-echo (FSE) measurement.

[0010] The fast-DIXON method is well known, and an overview of its performance and advantages and disadvantages is described by Jingfei Ma et al. in "Fast Spin-Echo Triple-Echo Dixon (fTED) Technique for Efficient T2-Weighted Water and Fat Imaging" (Magnetic Resonance in Medicine 58; 103-109 (2007)) and the sources cited therein.

[0011] The present invention is based on the object of providing a method in which the information obtained after an MR measurement according to the Dixon method using a bipolar readout gradient (fast-Dixon) has the same or comparable quality as that obtained by an MR measurement according to the Dixon method with monopolar readout gradients. At the same time, the MR acquisition time should at least not be increased.

[0012] This object is achieved by an object according to independent claims 1, 11 and 13, respectively. Advantageous embodiments are the subject of the dependent claims, the description and the drawings.

[0013] In particular, the present invention relates to a deep learning-based correction method for removing chemical shift artifacts (CSA) from MR data. This method is defined in claim 1. Furthermore, the present invention relates to a system for implementing the method according to claim 11 and a computer program according to claim 13.

[0014] The technical advantage is the accelerated image data acquisition and the significantly improved image quality, which also has a beneficial effect on subsequent processes such as medical evaluations and diagnoses.

[0015] The term "direction" refers to the readout direction.

[0016] For example, in the TSE sequence, CSA artifacts only occur in the readout direction. The readout direction of the gradients can also be understood as the acquisition direction.

[0017] The terms "in-phase" and "opposed-phase" refer to the relative phase of the water and fat spins. The phase position (i.e., in-phase or opposed-phase) can be treated independently of polarity. For example, the phase position has nothing to do with "monopolar" (readout gradient for in-phase and anti-phase echoes in the same direction) and "bipolar" (readout gradients in different directions). The latter type is obtained with a monopolar image, the former with a bipolar image.

[0018] To train the network, one can run in- and opposed data from a monopolar image through a Dixon algorithm as described above, resulting in fat and water images. By shifting the images relative to each other in both directions, one can again simulate bipolar images and thus provide training data.

[0019] The input data of the network are images with CSA in different directions. The output data should be images with CSA in the same direction.

[0020] In a preferred embodiment of the invention, step (2) is performed after step (1). Alternatively, the sequence of steps can be varied. For example, it may be useful to perform step (2) before step (1), for example, by loading "old" data into the machine learning structure or CNN.

[0021] The images are those in which water and fat spins are in and out of phase. The CSAs occur at the interfaces between water and fat. However, fat and water are not already recorded separately in these images (although such methods exist; however, the suppression of the other species often does not always work with sufficient quality; this is why the Dixon method is necessary in the first place).

[0022] In a further preferred embodiment of the invention, the acquisition data in step (1) and / or in particular the training data and the input data of the CNN are raw image data in k-space. Alternatively or cumulatively, they can be reconstructed images in image space. Alternatively, the method can also be applied in a hybrid mode, so that the acquisition data (training data and CNN input data) are hybrid image data. "Hybrid image data" is data that is represented partly in k-space and partly in image space. Hybrid image data is, in particular, data for which a Fourier transformation from k-space to image space was performed only in one spatial direction, preferably the readout direction.

[0023] The output of the CNN is in-phase and anti-phase data, which is then fed into a Dixon algorithm. It is also possible, for example, in a method not falling within the scope of the invention, to integrate the Dixon algorithm directly into the CNN, so that the output of the CNN is the reconstructed image data (without CSA). Examples of such CNNs are described in "Magnitude and Complex Single- and Multi-echo Water Fat Separation via End-to-End Deep Learning," Goldfarb JW, Proc. Int. Soc. Magn. Res. in Med., ISMRM, Paris, France, June 16-21, 2018, No. 5608, June 1, 2018, and "Deep Neural Network for Single-Point Dixon Imaging with Flexible Echo Time," Son JB et al., Proc. Int. Soc. Magn. Res. in Med., ISMRM, May 11-16, 2019, Montreal, Canada, No. 4014, April 26, 2019.

[0024] In a further preferred embodiment of the invention, the CNN has a U-Net architecture or a GAN architecture.

[0025] The U-Net is a convolutional network architecture developed by researchers at the University of Freiburg for fast and precise image segmentation. Successful training of deep networks requires many thousands of annotated training samples. The U-Net and its associated training strategy rely heavily on data augmentation to more efficiently utilize the available annotated input data. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. It has been shown that such a network can be trained end-to-end from very few images, outperforming the current best method (a sliding-window convolutional network). Furthermore, the network is fast. Segmenting a 512x512 image takes less than a second on a current GPU.The complete implementation (based on Caffe) and the trained networks are available at http: / / lmb.informatik.uni-freiburg.de / people / ronneber / u-net. For technical details, please refer to the following publication: . http: / / lmb.informatik.uni-freiburg.de / Publications / 2015 / RFB15a. Meanwhile, the U-Net architecture has also been extended to image reconstruction and transformation (see, for example, https: / / www.ncbi.nlm.nih.gov / pmc / articles / PMC6785508 / pdf / qims -09-09-1516.pdf)

[0026] CNN architectures are generally designed to work well with balanced datasets. A common problem with real-world datasets is that they are not balanced and suffer from class imbalance. For example, there may be only a small number of artifact images available for training the CNNs, while there are many good images (without artifacts), or vice versa. This leads to the problem that the networks tend to classify an image one-sidedly and according to the imbalance simply because of the imbalance in the data. This process usually led to correct classification during the training process and was thus rewarded with a small loss. Various methods are known to reduce or compensate for this imbalance. One frequently used method is data augmentation.This involves creating slightly modified copies of existing data and adding them to the corresponding class. Another option is adjusting the weighting. In this case, misclassifications of instances from classes with little data are given greater weight.

[0027] In contrast to the aforementioned approaches, Generative Adversarial Neural Networks (GANs) in this context aim to learn the underlying data distributions from the limited available images and then use the learned distributions to generate synthetic images. GANs can be used to generate synthetic images for sparse image classes from various imbalanced datasets. They could thus be used as an intelligent oversampling method. GANs are not only capable of generating a synthetic image but also offer a way to modify the original image.

[0028] GANs consist of two artificial neural networks. One network is the generator, which generates artificial data from a vector of latent variables. The second network is the discriminator, which evaluates the data and attempts to distinguish between artificial and real data. These two networks work "against each other," so to speak, and engage in a zero-sum game during the training process. The generator attempts to generate data so similar to the real data that the discriminator is unable to distinguish the artificially generated data from the real data. The first GAN architectures required a large amount of data for the training phase and could only generate images with limited resolution. However, newer architectures such as SinGAN7 or StyleGAN28, which work with Adaptive Discriminative Augmentation (ADA), allow high-quality images to be generated even with small datasets.

[0029] For the application of GANs and other architectures in medicine, please refer to the article: https: / / www.mdpi.com / 2076-3417 / 10 / 5 / 1816 / pdf.

[0030] In a further preferred embodiment of the invention, applying the trained CNN comprises providing nxm-channel image data to an input layer of the CNN, where n is the number of readout steps and m is the number of phase encoding steps. In particular, a 4-channel image or a 6-channel image with, for example, 256x256 pixels can be used as input.

[0031] In another preferred embodiment of the invention, a 3-point Dixon method can be used. In this case, the image data is then 6-channel, etc. This brings advantages for the network's performance.

[0032] In a further preferred embodiment of the invention, the input data comprises real and imaginary components of the acquisition data generated from the in-phase and out-of-phase echoes, which are concatenated with each other. The terms "real / imaginary component" refer to the fact that the MR signals are complex-valued. During processing in neural networks, these are generally divided into real and imaginary components (not magnitude and phase). "Concatenated" in this context means a concatenation of two tensors along one dimension.

[0033] The concatenation of input data serves to keep related data together or store it together. In real-world image recordings, this could be the RGB channels, for example.

[0034] The purpose of concatenating the extracting branch with data from the contracting branch is to restore the resolution. Interpolating from a very small to a very large matrix would otherwise be unstable. Therefore, the tensors with the same resolution from the contracting branch are included. Typically, concatenation occurs at every resolution level.

[0035] In a further preferred embodiment of the invention, training is performed with monopolar acquisition data and thus without assigning bipolar acquisition data. The learning method can be implemented as unsupervised learning. Bipolar in-phase and anti-phase acquisition data can be generated (estimated) from the monopolar acquisition data using an estimation algorithm. In a further embodiment of the invention, the estimation algorithm can add noise to simulate acquisition data with a wider bandwidth, whereby the bandwidth influences the position of the chemical shift.

[0036] In a further preferred embodiment of the invention, the acquisition data are generated from antiphase echoes shifted by a phase angle deviating from 180° by multiplying the fat proton data by e^(i*phase angle), where the phase angle is the actual dephasing in the range from 0 to 2pi.

[0037] The use of the Dixon magnetic resonance technique with bipolar readout gradients, referred to below as "fast Dixon MR" or "bipolar Dixon MR," has the advantage over the Dixon magnetic resonance technique with monopolar readout gradients, referred to below as "Dixon MR" or "monopolar Dixon MR," that the acquisition time is at least halved and that disturbing motion effects during the acquisition are minimal. MRI examinations are frequently performed on patients, and the shortened acquisition time contributes significantly to the acceptance of this technique.

[0038] When acquiring bipolar DIXON MR data, a plurality of k-space echoes are acquired, preferably bidirectionally or in different spatial directions. At least one echo is an in-phase (even) echo acquired in a first direction, and at least one echo is an anti-phase (odd) echo acquired in a second direction, e.g., opposite to the first direction. The k-space echo alignment is corrected between the even and odd echoes.

[0039] The fast-DIXON MR method is a two- or multi-point method, where two-point means that both the in-phase and the opposite-phase measurements are taken. This measurement can be extended to multiple phases, e.g., to three phases in the sequence of opposite-phase, in-phase, and opposite-phase measurements.

[0040] The term "artifact" refers here only to artifacts that are attributable to the MRI procedure, but not to those that occur due to incorrect selection of various parameters. Artifacts that are attributable to the MRI procedure include, for example, motion artifacts and chemical shift artifacts (also called "chemical shift artifacts" or "CSA").

[0041] The occurrence of such chemical shift artifacts is pronounced in an MR measurement using the fast-DIXON method, since the echoes are recorded with different polarities of the readout gradient.

[0042] "Chemical shift" refers to the property that the resonance frequency shifts slightly proportional to the field strength, depending on the type of chemical bond in which a signaling nucleus is located. Due to their concentration in the human body, hydrogen nuclei in free water and fat primarily contribute to the image. Their relative resonance frequency difference is approximately 3 ppm (parts per million).

[0043] This results in a modulation of the signal intensities depending on the echo time when using spin echo and gradient echo sequences.

[0044] CSA occurs at the interfaces between water and fat and is caused by the different resonance frequencies. Due to the slightly different spin frequencies of water and fat protons, the positions of the differently bound protons can be shifted relative to each other in the direction of the readout gradient in the MR image. After Dixon reconstruction, the interfaces are sharply defined in the water image, while they are blurred in the fat image. In the in- and opposed-phase images, the artifacts, as described above, are only found on the opposite side of the interface.

[0045] In this case, quality is understood to mean that the information in the corresponding images allows clear statements about the recorded structures.

[0046] In-phase means that the magnetization vectors of water (W) and fat (F) are parallel and point in the same direction (W+F) and anti-phase means that these magnetization vectors are parallel but point in the opposite direction (WF).

[0047] The invention relates to the use of one or more neural networks, in particular a CNN (convolutional neural network), to modify bipolar DIXON-MR acquisition data in such a way that they correspond to monopolar DIXON-MR acquisition data and thus have minimal, preferably no, CSA in different spatial directions.

[0048] The advantage of the faster acquisition time of the fast-DIXON method can thus be used advantageously without reducing the quality of the acquired data.

[0049] In a preferred embodiment of the invention, a U-Net network architecture is used. After the convolutional layers, a connection is made between a contraction path (down-sampling path) and an expansion path (up-sampling path) of the network, which allows the propagation of the original data into the up-sampling paths.

[0050] In an alternative advantageous embodiment of the invention, a different network architecture, such as a GAN (generative adversarial network) network, is used instead of the U-Net.

[0051] The input layer consists of the real and imaginary parts of the in-phase and anti-phase data of the bipolar Dixon MR image in a four-channel display. Concatenation is necessary to obtain an estimate of the non-shifted images from the two data sets (in-phase and anti-phase), with shifts in the opposite direction.

[0052] The output layer has the same dimensionality and contains in-phase and anti-phase data with reduced or eliminated chemical shifts.

[0053] Since chemical shifts occur in the read-encode direction, it is preferable to propagate this dimension in image space, i.e., to use either the full image-space representation (x, y) or a hybrid space representation (x, ky). In an alternative embodiment of the invention, a pure k-space representation of the data (kx, ky) can also be used as input data (both for training and in the inference phase).

[0054] In another advantageous embodiment of the invention, it is also possible to use water and fat images after Dixon reconstruction instead of in-phase and anti-phase images. However, since these images already contain a nonlinear mixture of in-phase and anti-phase images, the results with this approach may be inferior.

[0055] In another embodiment not belonging to the invention, the input layer contains the in-phase and anti-phase images and the output layer contains fat and water images and optionally a phase map (field map) that has been removed from the input data, ie an inclusion of the DIXON reconstruction in the neural network.

[0056] To train the network, bipolar in-phase and anti-phase data (e.g., TSE-DIXON with the fast-DIXON-MR option) can be used as input data, and monopolar in-phase and anti-phase data (e.g., TSE-DIXON without fast-DIXON-MR) can be used as target data. A supervised learning method can be applied. In an implementation based on the PyTorch platform, a U-Net can preferably be specified as the network architecture. This is then used to create the tensors with all input and output data, split the data into training and validation data, and then train the network. The network can then be trained, for example, using backpropagation based on a metric (e.g., L2 norm between the desired target image and the image generated by the network).

[0057] An alternative preferred embodiment of the invention does not require separately labeled data (with associated target data) as training data, but can generate the training data itself. For example, a large number of bipolar Dixon MR acquisition data sets can be obtained from each monopolar Dixon MR acquisition data set as training data sets. The algorithm can provide an assignment of which of the images reconstructed from the Dixon algorithm contains fat or water.

[0058] First, multiple chemical shifts are simulated by applying different (sub-)pixel shifts, each with a different sign, to fat images in the readout direction. This is technically necessary to simulate the bipolar readout of one image at a time.

[0059] Second, a phase map is extracted either from the conventional DIXON reconstruction or a corresponding pre-scan or synthesized from the images.

[0060] From this phase map, a multitude of phase maps are generated, e.g., by adding offsets, multiplying with scalars, or superimposing an eigenfunction, in particular a spherical harmonic.

[0061] The fat images are then added to and subtracted from the water image, followed by multiplication of the complex exponential of the synthesized or artificial field maps to generate synthesized or artificial bipolar acquisition data.

[0062] In addition, all resulting images can be further multiplied with additional phase maps. Additionally, artificial noise can be injected to simulate higher-bandwidth acquisition data.

[0063] In addition, phase-shifted data with a phase angle other than 180° for the antiphase case can be generated by multiplying the fat data by e^(i*phase angle), where the angle is the actual dephasing in the range 0 to 2pi.

[0064] Target images for training are obtained by proceeding as described, but 1. omitting the pixel shifting step, 2. introducing less noise, and 3. eliminating eddy current-induced effects on the phase map. The latter could be achieved by subtracting a field map obtained with conventional approaches (i.e., a 2- or 3-echo gradient-echo acquisition) from the field map obtained with the Dixon algorithm on a bipolar dataset.

[0065] Preferably, the phase shift (dephasing) is 180°. In this case, the vectors act in exactly opposite directions, and the signals cancel each other out. Alternative embodiments provide a phase shift other than 180°. The network must then have been trained accordingly with data from this phase shift.

[0066] The solution to the problem was described above using the method. Features, advantages, or alternative embodiments mentioned therein are also to be transferred to the other claimed subject matter, and vice versa. In other words, the independent claims directed to a system or a computer program product can also be developed with the features described or claimed in connection with the method. The corresponding functional features of the method are implemented by corresponding material modules, in particular by hardware modules or microprocessor modules, of the system or product, and vice versa.

[0067] According to a further aspect, the invention relates to a system according to claim 11 and a computer program according to claim 13.

[0068] The following detailed description of the figures discusses non-limiting embodiments, their features, and other advantages based on the drawings. These show: Fig. 1 shows a sequence diagram of a conventional bipolar TSE Dixon (Fast DIXON) acquisition; Fig. 2 shows example images of a monopolar DIXON-MR acquisition and a bipolar fast-DIXON MR acquisition; Fig. 3 shows a block diagram according to a preferred embodiment of the invention of a U-Net architecture for CSA correction, which is prominently present in fast-DIXON methods that use bipolar readout gradients to acquire the in-phase and anti-phase echoes; Fig. 4 shows an overview of a training of the network according to a preferred embodiment of the invention; Fig. 5 shows a flowchart of a method according to a preferred embodiment of the invention; and Fig. 6 shows a block diagram of modules of a magnetic resonance imaging device. Detailed description of the characters

[0069] Fig. 1 shows a typical sequence diagram of a conventional fast Dixon acquisition. In Dixon imaging, data acquisition occurs at at least two different time points, so that the signal from water and fat spins is in phase and out of phase (antiphase). A post-processing algorithm can then separate the fat and water signals and generate two images representing only the fat and water portions, respectively. To speed up acquisition, in-phase and anti-phase echoes are often acquired bipolarly. This is done, for example, in VIBE imaging and with the Fast Dixon function for TSE, where the acquisition of an in-phase and an anti-phase echo occurs after a refocusing pulse (see Figure 1denoted by "Refoc"). The interval between an in-phase and an out-of-phase state decreases linearly with increasing field strength. At 3 T, this interval is only about 1.2 ms. In-phase and anti-phase echoes are read out using a bipolar gradient. Since the position of the chemical shift artifacts depends on the polarity of the gradient (cf. Figure 2 ), the artifacts in the antiphase and single-phase echoes occur at different spatial positions. (In Fig. 1 only the first part of the echo chain is shown).

[0070] For conventional monopolar DIXON: IP = W(δ) + F(δ+chs), where IP denotes in-phase, W stands for water, and F for fat; OP = W(δ) - F(δ+chs), where OP denotes opposite phase.

[0071] For bipolar DIXON: IP = W δ + F δ + chs OP = W δ − F δ − chs

[0072] Here, IP and OP are the in-phase and anti-phase data sets, respectively, W is the water signal, F is the fat signal, δ is the voxel position, and chs is the chemical shift introduced by the difference between the Larmor frequencies of water and fat protons.

[0073] By adding and subtracting single-phase and counter-phase images, the fat and water images for conventional DIXON imaging can be obtained: 0 , 5 ∗ IP + OP = W δ 0 , 5 ∗ IP − OP = F δ + chs

[0074] Here, the chemical shift appears as a global shift for all voxels only in the fat image. Fat and water images can then be aligned, if desired, by simple spatial registration.

[0075] Fig. 2shows an example of the differences in quality between a monopolar and bipolar Dixon image. The chemical shift artifact (CSA) is caused by different Larmor frequencies for fat and water protons and can be seen in the images as a line artifact. As the readout gradient increases from left to right, fat is shifted to the left; as the readout gradient decreases from left to right, fat is shifted to the right. The effect of the different chemical shifts on the final fat and water images is shown on the right. Ringing artifacts are particularly evident at the fat-water boundaries in the water images. (Source: Leinhard et al, ISMRM 2008).

[0076] A possible network architecture is shown in Fig. 3Here, a U-net architecture with skipped connections is chosen, which allows the propagation of the original data into the upsampling path of the network.

[0077] The input and output data consist of real and imaginary parts of single- and anti-phase data, which are concatenated in the channel dimension to produce a 4-channel image. A state-of-the-art CNN architecture can be used, as developed by Ronneberger et al. (see Ronneberger, Olaf; Fischer, Philipp; Brox, Thomas (2015). "U-Net: Convolutional Networks for Biomedical Image Segmentation". arXiv:1505.04597). After upsampling, the 4-channel data is split again into single- and anti-phase images and fed into a conventional Dixon algorithm to obtain fat and water images in a final step.

[0078] The input layer consists of the real and imaginary parts of the in-phase and anti-phase data of the bipolar Dixon images in a 4-channel representation. Concatenation is necessary to enable the estimation of the unshifted images from the two datasets with shifts in opposite directions.

[0079] Instead of a U-network, other network architectures such as a Generative Adversarial Network (GAN) could be used.

[0080] Fig. 4 Using example images, shows a training approach for training the network that allows the creation of a variety of training datasets from a single monopolar Dixon reconstruction.

[0081] Reference numeral 4.1 denotes the monopolar fat uptake (fat image data) and 4.2 denotes the monopolar water uptake (water image data). Reference numeral 4.3 represents the application of a variety of pixel shifts along the readout direction to simulate the CSA with different bandwidths. Reference numeral 4.4 denotes the use of the phase map (field map) Φ B from the Dixon reconstruction. Optionally, various other modified phase maps can be generated. In 4.5, a variety of artificial in-phase and anti-phase images are generated according to the equation as in Fig. 4 From this, the simulated bipolar in-phase images shown in 4.6 and the simulated bipolar anti-phase images shown in 4.7 can then be generated.

[0082] The proposed method enables the rapid acquisition of bipolar Dixon data with improved image quality, as the chemical shift in various directions is removed prior to Dixon reconstruction, resulting in less blur. Furthermore, ringing artifacts (caused by the Fourier transform-based image processing calculations of a sampled signal line) can be reduced.

[0083] In addition to Dixon imaging, the method can also be used for other bipolar acquisition schemes such as mapping sequences or TGSE sequences.

[0084] Fig. 5is a flowchart of the method according to a preferred embodiment of the invention. After starting, the trained CNN is provided in step S1. The network may have been trained in preprocessing. For this purpose, training data may have been generated synthetically, for example by generating a portion of the images, e.g. the images containing CSA with CSA in different directions, from existing image data using image processing procedures. In step S2, at least two bipolar Dixon images are created by acquiring the in-phase and anti-phase echoes of the water and fat protons. In step S3, the trained CNN is applied to the image data created in step S2 to minimize or completely remove the CSA and to calculate corrected image data.

[0085] Fig. 6is a block diagram of a system for reconstructing MR images with CSA correction. The MR scanner or magnetic resonance imaging device (MR) is used to acquire bipolar Dixon MR images and is connected to a processing unit R, which can be configured as a reconstruction computer, via a data interface. The reconstruction computer is further connected via a network connection to a memory in which the trained CNN is stored, in particular a U-network, as described in more detail above.

[0086] Finally, it should be noted that the description of the invention and the exemplary embodiments are not intended to be restrictive with regard to a specific physical implementation of the invention. All features explained and shown in connection with individual embodiments of the invention can be provided in various combinations in the subject matter of the invention in order to simultaneously realize their advantageous effects.

[0087] The scope of the present invention is given by the following claims and is not limited by the features explained in the description or shown in the figures.

[0088] It is particularly obvious to a person skilled in the art that the invention can be applied not only to specific MR devices from one manufacturer, but also to other MR devices from other manufacturers. Furthermore, the components of the magnetic resonance imaging device can be distributed across multiple physical products. In particular, the memory and / or the processing unit can be outsourced to a separate or additional electronic module.

Claims

1. Method for correcting chemical shift artifacts, CSA, which arise in the magnetic resonance DIXON method when using bipolar readout gradients, fast DIXON MR, to capture the in-phase and opposed-phase echoes, wherein in-phase and opposed-phase relate to a phase relationship between water spins and fat spins, comprising the following method steps: (1) providing (S1) a trained convolutional neural network, CNN, which has been trained using acquisition data acquired in phase and in phase opposition by the DIXON MR method, comprising acquisition data that contains CSA in mutually opposite directions of the readout gradient, and acquisition data that contains CSA only in one direction of the readout gradient, so that the CNN is trained to transform the acquisition data obtained by the fast DIXON MR such that it exhibits CSA that now arise only in the same direction of the readout gradient; (2) producing control instructions for acquiring (S2) at least two instances of fast DIXON MR acquisition data, by acquiring the in-phase and opposed-phase echoes of the water protons and fat protons; (3) applying (S3) the trained CNN to the acquisition data produced in the preceding step (2) in order to minimise the CSA or remove it entirely from the acquisition data in mutually opposite directions of the readout gradient and calculate corrected acquisition data, feed the corrected acquisition data to a conventional Dixon algorithm to generate a fat image and a water image, wherein fat and water images are aligned by spatial registration and the CSA only arising in the same direction of the readout gradient only arises as a global shift for all voxels in the fat image.

2. Method according to claim 1, in which step (2) is performed before step (1).

3. Method according to one of the preceding claims, wherein the acquisition data is the acquisition data in which water spins and fat spins are in phase and out of phase, and the CSA are present at the boundary surfaces between water and fat.

4. Method according to one of the preceding claims, wherein the acquisition data in step (1) is raw image data in k-space, reconstructed images in image space, or hybrid image data.

5. Method according to one of the preceding claims, wherein the CNN has a U-net architecture or a GAN architecture.

6. Method according to one of the preceding claims, in which applying the trained CNN comprises providing 4-channel or 6-channel image data into an input layer of the CNN.

7. Method according to one of the preceding claims, in which the input data of the CNN comprises real and imaginary components of the acquisition data generated from the in-phase and opposed-phase echoes, which are concatenated with each other.

8. Method according to one of the preceding claims, in which the training is performed with monopolar acquisition data without assigning bipolar acquisition data, wherein an estimation algorithm is used to generate bipolar in-phase and opposed-phase acquisition data from the monopolar acquisition data, wherein monopolar relates to one readout gradient for in-phase and opposed-phase echo in the same direction, and wherein bipolar relates to readout gradients with opposite directions.

9. Method according to claim 8, in which the estimation algorithm adds noise in order to simulate acquisition data having a lower signal-to-noise ratio.

10. Method according to one of claims 8 or 9, in which acquisition data is generated from opposed-phase echoes offset by a phase angle other than 180° by multiplying the data for the fat protons by e^(i*phase-angle), where the phase angle is the actual dephasing in the range 0 to 2pi.

11. System with a processing unit (R) and a magnetic resonance imaging device (MR) for performing the method according to one of the preceding claims, wherein the processing unit is connected to the magnetic resonance imaging device via a data interface, having: - a memory (MEM), in which is stored a trained convolutional neural network, CNN that has been trained with acquisition data acquired in phase and in phase opposition by the DIXON method, comprising acquisition data that contains CSA in mutually opposite directions of the readout gradient, and acquisition data that contains CSA only in one direction of the readout gradient, so that the CNN is trained to transform the acquisition data obtained by the fast DIXON MR such that it exhibits CSA that now arise only in the same direction of the readout gradient; - wherein the magnetic resonance imaging device (MR) is configured to produce bipolar DIXON MR acquisition data by acquiring the in-phase and opposed-phase echoes of the water protons and fat protons; - wherein the processing unit (R) is intended to apply the trained CNN to the acquisition data obtained in step (2) in order to minimise the CSA or remove it entirely from the acquisition data in mutually opposite directions of the readout gradient and to calculate the corrected acquisition data, as well as to feed the corrected acquisition data to a conventional Dixon algorithm to generate a fat image and a water image, wherein fat and water images are aligned by spatial registration and the CSA only arising in the same direction of the readout gradient only arises as a global shift for all voxels in the fat image.

12. System according to claim 11, wherein the processing unit (R) is embodied as a reconstruction computer, with a network connection; and - wherein the memory (MEM), in which the trained CNN is stored, has a network connection for the connection to the reconstruction computer (R).

13. Computer program comprising program code means, which cause the system according to claim 11 or 12 to carry out the method stops according to one of claims 1-10.