Computer-implemented method and image processing device for determining magnetic resonance image dataset

Abstract

The invention relates to a computer-implemented method for determining a magnetic resonance image dataset from a measurement dataset of magnetic resonance imaging, in which the measurement dataset, in particular an undersampled measurement dataset, is acquired by parallel imaging using a receiving coil arrangement (11) having a plurality of receiving channels in an acquisition geometry, the determination of the magnetic resonance image dataset comprises a reconstruction process (1) using at least one trained image processing function, the reconstruction process comprising a data consistency step for increasing the consistency of the resulting image with the measurement dataset as a function of the regularization value, wherein the noise reduction intensity of the reconstruction process (1) is modified in a position-dependent manner in the data consistency step, the modification being carried out as a function of modification variables which locally describe the expected noise reduction intensity of the unmodified reconstruction process as a result of the acquisition geometry.

Description

Technical Field

[0001] This invention relates to a computer-implemented method for determining a magnetic resonance image dataset from a measurement dataset obtained from magnetic resonance imaging (MRI), wherein the measurement dataset, particularly an undersampled measurement dataset, is acquired by parallel imaging using a receiving coil arrangement with multiple receiving channels in an acquisition geometry, and the determination of the magnetic resonance image dataset includes a reconstruction process using at least one trained image processing function, wherein the reconstruction process includes a data consistency step for improving the consistency between the resulting image and the measurement dataset. Furthermore, this invention relates to an image processing apparatus, a computer program, and an electronically readable data carrier. Background Technology

[0002] In imaging, a receiving coil device with multiple coil elements, each assigned a receiving channel, is used to receive magnetic resonance signals in a time-parallel manner through different channels. The known spatial location and sensitivity of the coil elements are used to improve the spatial localization of the magnetic resonance signal. Parallel imaging is often combined with undersampling because this additional information allows for a reduction in the number of phase encoding steps during image acquisition. Therefore, acquisition time can be significantly reduced.

[0003] During reconstruction, the magnetic resonance image dataset of the acquisition area is reconstructed from the measurement data of each receiving channel. Traditional, classic reconstruction methods either use sensitivity information, particularly sensitivity maps, or use reconstruction weights to determine the magnetic resonance image dataset. See, purely exemplarily, Klaas P. Pruessmann et al., “SENSE: Sensitivity Encoding for Fast MRI”, Magnetic Resonance in Medicine 42 (1999), pp. 952–962, and MA Griswold et al., “Generalized autocalibrating partially parallel acquisitions (GRAPPA)”, Magnetic Resonance in Medicine 47 (2002), pp. 1202–1210.

[0004] In traditional reconstruction processes, such as GRAPPA, stringent parameters of the acquisition protocol, such as extremely high acceleration factors (i.e., severe undersampling), lead to a significant increase in noise in regions with high g-factors (geometric factors) due to the geometric arrangement of the receiving coils. However, iterative methods have recently been proposed that incorporate the results of a trained image processing function, the so-called prior image, into the reconstruction results in a regularized manner within an optimized framework. For example, the application of this approach is described in the paper "Enhancing gadoxetic acid-enhanced liver MRI: a synergistic approach with deep learning CAIPIRINHA-VIBE and optimized fat suppression techniques" by H. Wie et al., European Radiology 34 (2024), pp. 6712-6725. Typically, the optimized cost function includes, for example, a data consistency term (a mapping of the measurement data) and a second term, weighted according to the regularization value, that deviates as little as possible from the prior image. The trained image processing function, including neural networks, serves to reduce noise in this context. Therefore, in regions with high g-factors, stringent parameters of the acquisition protocol are more likely to lead to increased smoothing. This is counterintuitive to the observer, who perceives noise as a sign of low reliability of image regions. Furthermore, there is a risk that image features may be smoothed away, such as small lesions or similar structures that are no longer reliably visible.

[0005] To address this issue, the current proposal is to teach the different behaviors of traditional reconstruction and those using trained image processing functions that employ smoothing or noise reduction, both in manuals and through training. Summary of the Invention

[0006] Therefore, this invention is based on the following objective: to propose a possibility for obtaining a magnetic resonance image dataset that is more intuitive and / or can reproduce smaller image features, even when using trained image processing functions during reconstruction in parallel imaging.

[0007] To address this task, according to the present invention, a computer-implemented method, an image processing apparatus, a computer program, and an electronically readable data carrier are proposed.

[0008] In one type of method mentioned at the beginning, according to the invention, the noise reduction intensity of the reconstruction process is modified in a location-related manner during the data consistency step, wherein the modification is based on a modification variable that locally describes the expected noise reduction intensity of the unmodified reconstruction process due to acquisition geometry.

[0009] Therefore, a method is proposed to reduce oversmoothing in high g-factor image regions, where the denoising intensity is position-dependently scaled across the entire reconstruction process. This spatially varied denoising intensity allows for significantly better identification of noisy regions in the resulting image, i.e., the magnetic resonance imaging dataset. Users obtain an image impression comparable to conventional methods, thus better judging whether the chosen protocol settings (e.g., acceleration factor, image resolution, etc.) are appropriate or excessively high. Furthermore, since denoising is less aggressive in regions exhibiting high noise amplification, smaller image features are smoothed less and are better captured in the final image. For example, experiments using a 3x3 accelerated 3D measurement dataset and reconstruction processes with and without position-dependent denoising modifications show that, for instance, small lesions remain clearly visible despite higher noise levels in the settings described in this paper, whereas they would be less easily identified without position-dependent denoising modifications.

[0010] The modifications rely on the acquisition geometry, particularly the arrangement of the coil elements in the receiving coil apparatus, which determines where particularly strong noise or low signal-to-noise ratios will occur, thus leading to noise amplification in many conventional methods (e.g., pure SENSE or GRAPPA). Therefore, the modifications are directly based on background knowledge about the causes of different noise behaviors in the measurement dataset. The modifications are performed outside of the trained image processing function, thus requiring no change or adjustment to that function. In the data consistency step, the consistency between the instantaneous image and the measurement data in the sense of the signal model is improved. This instantaneous image is the (unique) result image of the magnetic resonance dataset if the trained image processing function and the data consistency step are applied only once; otherwise, it is the result image of the corresponding iterative step. In the data consistency step, it can be said that the result of the trained image processing function is merged with the data consistency, where the regularization value describes the corresponding weighting. In other words, during reconstruction, at least one computed image is determined by at least one of at least one trained image processing function, where this computed image is incorporated into the reconstruction result in the data consistency step, taking into account the regularization value. Therefore, the regularization value can also be referred to as the weighting value or the step size of the data consistency step. By combining the acquisition geometry and signal model, the denoising process of the trained image processing function, and the data consistency step, denoising is achieved in the reconstruction process before the application of the proposed denoising intensity modification, whereas noise amplification occurs in traditional methods. This is due to the parallel imaging process, its specific implementation, and the signal model used; therefore, the modified variables precisely reflect these aspects.

[0011] In summary, by spatially varying the noise reduction intensity without altering the trained image processing function, the advantages of achieving a more intuitive image impression comparable to traditional reconstruction methods are realized. This increases the reliability for radiologists and other users, as well as improving the visibility of small image structures.

[0012] Typically, a trained function maps to cognitive functions that humans associate with other human brains. Through training based on training data (machine learning), a trained function can adapt to new situations and detect and infer patterns. Another way to express "trained function" is "trained machine learning model".

[0013] Generally, the parameters of a trained function can be tuned through training. Specifically, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be used. Furthermore, representation learning (also known as "feature learning") can also be used. The parameters of a trained function can be iteratively tuned, particularly through multiple training steps. Especially during training, a specific cost function can be minimized. For example, the backpropagation algorithm can be used when training a neural network.

[0014] The trained function can include, for example, neural networks, support vector machines (SVMs), decision trees, and / or Bayesian networks, and / or the trained function can be based on k-means clustering, Q-learning, genetic algorithms, and / or association rules. In particular, the 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).

[0015] A Convolutional Neural Network (CNN) is a type of neural network that uses convolution operations instead of general matrix multiplication in at least one of its layers, known as convolutional layers. Specifically, a convolutional layer can perform a dot product of one or more convolutional kernels with the input data / image, where the entries for these kernels are parameters or weights tuned through training. In particular, internal Frobenius products and ReLU activation functions can be used. CNNs can contain additional layers such as pooling layers, fully connected layers, and normalization layers.

[0016] CNNs can process input images with extremely high efficiency because convolutional operations based on different kernels can extract a wide variety of image features. By adjusting the weights of the convolutional kernels, relevant image features can be found during training. Furthermore, due to the shared weights of the kernels in the convolutional layers, fewer parameters need to be trained, thus avoiding overfitting during training and allowing for faster training or more layers in the CNN, thereby improving network performance.

[0017] Within the framework of this invention, the reconstruction process can, in particular, utilize sensitivity information (SENSE) describing the sensitivity of the receiving coil device, or use reconstruction weights, such as GRAPPA weights. The sensitivity information specifically includes a sensitivity map for each receiving channel or at least one coil element assigned to that receiving channel. In both methods, the reconstruction process corresponds to combining information from the individual receiving channels into a single overall image, with undersampling also compensated for, if present. The reconstruction result (and therefore the final determined image) is a magnetic resonance imaging dataset.

[0018] The method described herein is particularly suitable for iterative methods, where the trained image processing function is used particularly frequently. Therefore, suitable embodiments specify that, during the reconstruction process, in multiple iterative steps, the following is performed separately:

[0019] - For each iteration step after the first iteration step, a computational image is determined from the result image of the previous iteration step using a trained image processing function, specifically a prior image, and

[0020] - Taking into account the regularization values ​​assigned to the iterative step that describe the weighting, the resulting image of the iterative step is determined from the measurement dataset and the computational image of the iterative step.

[0021] The computational image can specifically correspond to the output data of a trained processing function, i.e., its result, particularly using the previous resulting image as input data. In particular, the computational image can be a so-called prior image (“imageprior”). The computational image for the first iteration step can, for example, be determined as a starting image containing zero and / or constant image values. Other ways of determining the computational image, particularly those dependent on the measurement dataset, are also conceivable.

[0022] The iteration is preferably constructed to minimize a cost function comprising: a data consistency term that describes the best consistency (or best correspondence) between the resulting image and the measurement dataset; and a second term weighted by a regularization value that describes the best consistency between the prior image and the measurement dataset.

[0023] A particular advantage is that the regularization values ​​and / or trained image processing functions can be at least partially different at different iteration steps. The regularization values ​​can be an integral part of the machine learning process throughout the reconstruction process. Therefore, for example, it can be specified that regularization values ​​are also learned while training at least one image processing function. This allows an optimal reconstruction rule to be determined holistically.

[0024] It should be noted that, in principle, it is also possible to consider using only a small number of iterative steps. For example, after using a defined starting image, only one additional iterative step using a trained image processing function could be performed. However, multiple iterative steps are typically used. The number can be preset, so for example, checking whether that preset number has been reached can be used as a condition for ending the iteration, where the last resulting image is used as a magnetic resonance imaging dataset.

[0025] As a specific example for explaining particular embodiments of the invention, a reconstruction process will be frequently used below, in which iterative regularized SENSE reconstruction is employed, during which a computed image is determined as a prior image for regularization by at least one trained image processing function with a U-Net architecture. In other words, alternatingly, a data consistency step is performed to correlate measurement data with the resulting image; and image enhancement measures are taken to determine a computed image, such as a prior image, for the next data consistency step using the trained image processing function. Of course, the invention can also be applied to other specific embodiments beyond this purely illustrative example.

[0026] In schematic form, the optimization problem of this iterative regularized SENSE reconstruction can be written as:

[0027] (1)

[0028] Here, S represents the coil sensitivity, specifically the sensitivity information present in the form of a sensitivity matrix, which can be determined from the sensitivity map; m is the resulting image to be optimized; y is the measurement data from the acquired, undersampled measurement dataset; λ is the regularization value (regularization factor); and z is the computed image or prior image. Therefore, the first term is the data consistency term, the second term is the regularization term, and ultimately it also describes noise propagation in the sense of the signal model. This method can also be called Tikhonov regularization.

[0029] For regular undersampling with an acceleration factor of R, optimization can be performed in multiple blocks, each containing R aliased voxels. If the measurement data y are transformed to the image space, where they correspond to zero-padded, aliased coil images of the respective receive channels, each decoupling block can be decomposed by singular value decomposition into a matrix U ∑ V + To solve this, and for each iteration step n, obtain the data consistency step as a resolution step:

[0030] (2)

[0031] For the first iteration, the preferred choice is... For each further iteration step n > 0, we have ,in This indicates the application of at least one trained image processing function. Typically, a different trained image processing function is applied for each iteration. The parameters and regularization values ​​of the trained image processing function are also included. Then, through machine learning, they are jointly determined during a training process, thus forming a well-defined process.

[0032] The regularization value λ can be understood as the weighting value or step size of the data consistency steps. Specifically, when λ → 0, we have... When λ → ∞, the minimum value of unregularized SENSE is obtained.

[0033] Other specific formulaic expressions can also be considered, and the content described here is equally applicable. For example, gradient descent can be used, where updates are performed from the current computed image z of the corresponding trained image processing function along the negative direction of the gradient of the SENSE data term:

[0034] (2a)

[0035] Because of the calculation of image z and Since they are all located in the image space, position-dependent regularization or general noise reduction intensity modifications can be applied accordingly, as in the case of Equation (2).

[0036] In general, according to the present invention, the modifications are specifically and preferably implemented in the following ways:

[0037] - Scale the regularization value based on the modified variable according to its position, or

[0038] - Without scaling the regularization value, position-related scaling is applied to the sensitivity information describing the sensitivity of the receiving coil device used in the reconstruction process and the calculated image, based on the modified variables.

[0039] These two methods are particularly easy to implement when the variables are modified in a spatially resolved manner. This means that the aforementioned high correlation advantage can be obtained by making a small intervention in the algorithm outside the trained image processing function, i.e., by introducing a position-dependent (or spatially dependent) scaling factor.

[0040] Both methods ultimately achieve the same result, but the second method results in a scaling of the final image. However, this scaling can be easily eliminated. These two methods will be explained in more detail with the specific exemplary examples briefly described above.

[0041] For a given set of regularization values ​​λ nThe noise reduction intensity during the reconstruction process is modified by location-dependent scaling, for example:

[0042] (3)

[0043] Values ​​of 0 < α < 1 result in a greater emphasis on data consistency, thus reducing noise reduction; while α > 1 reduces the importance of data consistency, thus increasing noise reduction. This allows the noise reduction intensity to be adjusted to a locally acceptable level. This method is not limited to regular undersampling, but can also be applied, particularly beyond the specific exemplary examples, to incoherent or non-Cartesian sampling modes in k-spaces. In this case, the noise reduction intensity and scaling factor can, for example, depend on (or be determined according to) the bias field used for image normalization as a modification variable.

[0044] Regarding the second variant explained above, it is advantageous for specific examples to consider the optimization problem more explicitly as follows (1):

[0045] (4)

[0046] in These are the Fourier coefficients, and C describes the sensitivity information. Now, if we rescale:

[0047] (5)

[0048] This leads to the alternative optimization problem:

[0049] (6)

[0050] This means that the local weighted localization of the first variant can also be mapped to an optimization problem with constant regularization (for each iteration step), scaled coil sensitivity, and scaled prior image, where the optimization result is also rescaled.

[0051] In the case of regular undersampling, it is particularly advantageous to specify that the g-factor of the g-factor plot assigned to the measurement dataset as a modification variable is used at least partially such that a lower g-factor leads to stronger noise reduction. What can be utilized here is that, for regular undersampling, the probability of determining the g-factor is known not only for the SENSE method (see the cited article by Pruessmann et al.) but also for techniques using reconstruction weights such as GRAPPA, for example, see Felix A. Breuer et al., “General Formulation for Quantitative g-factor Calculation in GRAPPA Reconstruction”, MagneticResonance in Medicine 62 (2009), pp. 739-746. Therefore, under regular sampling, the g-factor plot is known and / or determinable. Thus, in this implementation, these also constitute a voxel-by-voxel scaling plot or are used to derive the scaling plot.

[0052] Specifically, it can be specified that a position-dependent scaling factor is used during scaling, wherein this scaling factor is directly or inversely proportional to the power of the g-factor at the corresponding position, particularly at the corresponding image point, and / or, to determine the position-dependent scaling factor, the g-factor is constrained within a predetermined range or rescaled based on the g-factor map. For example, the scaling factor α can be chosen as... The power p can be chosen heuristically to balance the dependence on local conditions of the data consistency term. To avoid excessive influence, the g factor can also be restricted to a specific range of values, such as between 1 and 5.

[0053] If a g-factor map is unavailable in at least some regions, it can be specified that the sensitivity values ​​of the sensitivity map, particularly the sensitivity information, are used, at least partially, as a modification variable, such that lower sensitivity leads to lower noise reduction. Alternatively, a per-image-point scaled map can be derived from a sensitivity map predetermined for the receiving coil device, and if necessary, during the design phase. It must be noted here that sensitivity is an inverse measure of the g-factor; therefore, higher sensitivity is expected to result in a higher signal-to-noise ratio, thus lower noise, and weaker noise reduction for the trained image processing function. In a favorable improvement in this context, it can also be specified that at least one undersampling degree and / or sampling scheme be additionally considered during modification to incorporate the specific circumstances of the sampling process. Specifically, the undersampling degree and / or sampling scheme can be considered as follows: transforming the sensitivity map to k-space, applying the undersampling degree and / or sampling scheme, and then transforming it back to image space.

[0054] Finally, when neither a sensitivity plot nor a g-factor plot is available, or in cases where neither a sensitivity plot nor a g-factor plot is available, modification can be considered using at least one theoretically assumed sensitivity spatial distribution, particularly a sensitivity model. Specifically, a two-dimensional Gaussian function can be used as the sensitivity model. Since the coil elements of the receiving coil device are typically arranged around the body of the patient to be sampled, it can be assumed that the sensitivity outside the body is higher than that inside the body, away from the coil elements, which can be appropriately described by a Gaussian function. The scaling plot can again be appropriately derived from the sensitivity model.

[0055] For irregular undersampling, as explained above, the bias field used for image normalization can be used as a modification variable.

[0056] In addition to the method, the present invention also relates to an image processing apparatus for determining a magnetic resonance image dataset from a measurement dataset of magnetic resonance imaging, wherein the measurement dataset, particularly an undersampled measurement dataset, is acquired by parallel imaging using a receiving coil arrangement having multiple receiving channels in an acquisition geometry. The image processing apparatus includes: a reconstruction unit for determining the magnetic resonance image dataset during a reconstruction process using at least one trained image processing function, wherein the reconstruction process includes a data consistency step for improving the consistency between the resulting image and the measurement dataset based on a regularization value; and a modification unit for positionally modifying the noise reduction intensity of the reconstruction process in the data consistency step, wherein the modification depends on a modification variable that locally describes the expected noise reduction intensity of the unmodified reconstruction process due to the acquisition geometry. All descriptions of the method according to the invention can be similarly applied to the image processing apparatus according to the invention, thereby also obtaining the advantages already mentioned.

[0057] In particular, the image processing device includes at least one processor and at least one storage device. Functional units are formed by hardware and / or software to perform the steps of the method according to the invention. In addition to the functional units already mentioned—the modification unit and the reconstruction unit—other functional units may also be considered to implement the method according to the invention and / or provide other functions. The image processing device may have a first interface for receiving measurement datasets and possibly other information, such as sensitivity information and g-factor plots. The reconstruction results, i.e., the magnetic resonance image dataset, can be output through a second interface.

[0058] The image processing device can be integrated into the control device of the magnetic resonance imaging (MRI) device. This means that the MRI device can include a control device configured, in particular, to perform the method according to the invention by integrating the image processing device according to the invention. Thus, the function of reconstructing and determining the MRI image dataset is directly integrated into and available there, for example, to enable a preliminary review of the MRI image dataset.

[0059] The computer program according to the invention can be directly loaded into the storage device of an image processing apparatus and has program means such that when the computer program is executed on the image processing apparatus, it causes the image processing apparatus to perform the steps of the method according to the invention. The computer program can be stored on an electronically readable data carrier according to the invention, which therefore contains control information stored thereon, including at least one computer program according to the invention, and is configured to, when the data carrier is used in the image processing apparatus, cause the image processing apparatus to be configured to perform the method according to the invention.

[0060] It should also be noted that, for the sake of purely optimizing the observer's image impression, post-hoc modifications to the magnetic resonance image dataset can also be considered, which can be performed as an alternative to or supplement to the method according to the invention. In particular, this method can be employed when the reconstruction process does not allow or directly permits spatial variation in the noise reduction intensity. Specifically, a method for determining the output display image from a measurement dataset of magnetic resonance imaging, wherein the measurement dataset, particularly an undersampled measurement dataset, is acquired through parallel imaging using a receiving coil device with multiple receiving channels in an acquisition geometry, and the magnetic resonance image dataset is determined during reconstruction using a trained image processing function with noise reduction capabilities, wherein the display image is determined from the magnetic resonance image dataset obtained as a reconstruction result:

[0061] - Generate a randomized or pseudo-randomized noisy image with the resolution of a magnetic resonance imaging dataset.

[0062] - Based on at least one modified variable, specifically using a g-factor plot, perform position-dependent scaling on the noisy image, and

[0063] - Add the noisy image to the magnetic resonance image dataset to obtain the displayed image.

[0064] Therefore, one could consider first determining a "noise map," or noisy image, randomly or pseudo-randomly. Then, each point of this noisy image could be scaled according to modified variables, particularly the local g-factor, and added to the reconstruction result in the final step. Alternatively, the scaled noisy image could also be added to the intermediate result during reconstruction to directly obtain the display image as the reconstruction result.

[0065] Other advantages and details of the invention will become apparent from the embodiments described below and from the accompanying drawings. Attached Figure Description

[0066] Figure 1 A flowchart of one embodiment of the method according to the present invention is shown.

[0067] Figure 2 The functional structure of the image processing apparatus according to the present invention is shown, and

[0068] Figure 3 A magnetic resonance device is shown. Detailed Implementation

[0069] Figure 1 A flowchart of an embodiment of the method according to the present invention is shown. Part of it refers to the specific example of iterative regularized SENSE reconstruction with Tikhonov regularization mentioned above to illustrate a setting, but does not preclude the applicability of this setting to other specific settings known in the prior art, particularly iterative ones.

[0070] In step S1, a measurement dataset is acquired, or, if acquisition has been completed, a measurement dataset is provided. This measurement dataset is a measurement dataset from parallel imaging, for which a receiving coil device with multiple receiving channels arranged in the acquisition geometry is used. This measurement dataset is typically undersampled. Purely by example, a SPACE sequence can be used to acquire the measurement dataset. For a three-dimensional measurement dataset with three-dimensional acquisition, for example, there might be 3 x 3 undersampling, hence the speedup factor R = 9. The measurement data (y in the formula shown above and equally valid here) here still exists separately by receiving channel. In a reconstruction process 1 that now begins, a magnetic resonance image dataset summarizing the measurement results from the different receiving channels should be reconstructed from the measurement data.

[0071] The reconstruction process 1 includes iterative steps, each comprising steps S2 to S4. First, in step S2, a prior image (z) is determined as the computation image for the corresponding iterative step (n). In the first iterative step, the prior image is, for example, determined as a zero image with all image values ​​equal to zero, or a trained image processing function is applied to such a zero image. In further iterative steps, the prior image is determined by applying a trained image processing function, here specific to the iterative step and with noise reduction capabilities, to the result image of the previous iterative step. This means that the result image of the previous iterative step constitutes the input data for the trained image processing function, while the prior image of the corresponding iterative step constitutes the output data.

[0072] In step S3, the data consistency step, the resulting image for the current iteration step is determined from the prior image and measurement data, for example, according to formula (2) listed above. Here, a constant and specific regularization value λ is used for each iteration step. n Regularization is applied, which results in a weighted inclusion of the prior image. The regularization value is λ. n It is learned along with the parameters of the trained image processing function. Then, in step S4, it is checked whether the iteration should end, i.e., whether the predetermined number of iteration steps has been reached.

[0073] Here, the noise reduction intensity of reconstruction process 1 is modified location-dependently, as illustrated in step S5. This is done outside the trained image processing function, here by setting a scaling dependent on the modified variable in the data consistency step. In the specific embodiment discussed here, the regularization value λ n It is scaled position-dependently, referring to formula (3) above. However, it is also possible to equivalently scale the prior image and sensitivity, referring to formula (5) above.

[0074] The scaling here is based on a g-factor plot provided along with the measurement dataset, which thus serves as the basis for the scaling plot used for location-related scaling. For example, the selection can be specifically specified. The power p is heuristically determined, for example, in the range of 0.5 to 4. The value of the factor g can be restricted to a range, for example, 1 to 5.

[0075] As an alternative to using the g-factor as a modification variable, the scaling map can also be determined from sensitivity values, specifically, for example, sensitivity maps and / or sensitivity models. In general, the scaling map describes where, due to the acquisition geometry of the receiving coil device, over-denoising or smoothing is expected during reconstruction process 1 due to the action of the trained image processing function. Location-dependent modifications to the denoising intensity at least partially compensate for this.

[0076] After reconstruction process 1 is completed, the reconstruction result is the final image, which is provided as a magnetic resonance imaging dataset in step S6. For example, it can then be displayed to a user, such as a radiologist or other evaluator, who can better judge the corresponding displayed image based on intuitive expectations of noise behavior and can also better identify smaller image features in areas with poor signal-to-noise ratio.

[0077] Figure 2 A schematic diagram illustrating the functional principle of the image processing device 2 according to the present invention is shown. This image processing device firstly has a storage device 3, capable of storing various information, particularly measurement datasets, result images, prior images, and g-factor plots. Through a first interface 4, the measurement dataset, particularly along with other information such as g-factor plots, can be received according to step S1. Figure 1 Start receiving.

[0078] The measurement dataset is then forwarded to reconstruction unit 5 for reconstruction, which is configured to perform reconstruction process 1 according to steps S2, S3, and S4. By modifying unit 6, the location-dependent modification of the noise reduction intensity can be implemented according to step S5, for example, by providing a suitable scaling image. The reconstruction result, i.e., the magnetic resonance image dataset, is output through the second interface 7 according to step S6.

[0079] Figure 3 A schematic diagram of the magnetic resonance imaging (MRI) device 8 is shown. This MRI device, as is generally known, includes a main magnet unit 9 with a patient receiving cavity 10 into which the patient can be moved via a patient bed (not shown in detail here) to acquire measurement data. In addition to the superconducting main magnet, an array of gradient coils and a high-frequency coil array may also be arranged around the patient receiving cavity 10 within the main magnet unit 9. A receiving coil device 11, which can be used for parallel imaging, is schematically shown here.

[0080] The operation of the magnetic resonance imaging (MRI) device 8 is controlled by a control device 12, which also includes an image processing device 2. In this way, a suitable MRI image dataset can be determined directly on-site for the acquired measurement dataset, and it can also be displayed, particularly on the display device of the MRI device 8.

Claims

1. A computer-implemented method for determining a magnetic resonance image dataset from a measurement dataset of magnetic resonance imaging, wherein, The measurement dataset, particularly an undersampled measurement dataset, is acquired by parallel imaging using a receiving coil device (11) with multiple receiving channels arranged in a shooting geometry. The determination of the magnetic resonance image dataset includes a reconstruction process (1) using at least one trained image processing function, wherein the reconstruction process includes a data consistency step for improving the consistency between the resulting image and the measurement dataset based on a regularization value. The feature is that the noise reduction intensity of the reconstruction process (1) is modified in a position-related manner during the data consistency step, wherein the modification is based on a modification variable that locally describes the expected noise reduction intensity of the unmodified reconstruction process due to the acquisition geometry.

2. The method according to claim 1, characterized in that, During the reconstruction process, at least one computational image is determined by at least one of the at least one trained image processing functions, wherein the computational image is incorporated into the reconstruction result in the data consistency step, taking into account the regularization value.

3. The method according to claim 2, characterized in that, In the reconstruction process (1), in multiple iterative steps respectively - For each iteration step following the first iteration step, the computed image is determined from the result image of the previous iteration step using at least one of the at least one trained image processing function, and - The resulting image of the iterative step is determined from the measurement dataset and the computational image of the iterative step, taking into account the regularization values ​​assigned to the iterative step that describe the weighting.

4. The method according to any one of the preceding claims, characterized in that, The modifications are implemented in the following ways: - Scale the regularization value position-dependently based on the modified variable, or - Without scaling the regularization value, position-related scaling is performed on the sensitivity information describing the sensitivity of the receiving coil device (11) used in the reconstruction process (1) and the computed image, in particular the prior image, determined by at least one of the at least one trained image processing functions, based on the modified variable.

5. The method according to any one of the preceding claims, characterized in that, As a modification variable, the g-factor of the g-factor graph assigned to the measurement dataset is used at least partially, such that a lower g-factor results in stronger noise reduction.

6. The method according to claim 5, characterized in that, When scaling, a position-dependent scaling factor is used, wherein the scaling factor is directly or inversely proportional to the power of the g factor at the corresponding position, particularly at the corresponding image point, and / or, in order to determine the position-dependent scaling factor, the g factor is constrained or rescaled to a predetermined interval according to the g factor graph.

7. The method according to any one of the preceding claims, characterized in that, As a modification variable, the sensitivity values ​​of the sensitivity map, particularly the sensitivity values ​​of the sensitivity information, are used at least partially such that lower sensitivity results in lower noise reduction.

8. The method according to claim 7, characterized in that, When making the modifications, at least the degree of undersampling and / or the sampling scheme should be taken into account.

9. An image processing apparatus (2) for determining a magnetic resonance image dataset from a measurement dataset of magnetic resonance imaging, wherein, The measurement dataset, particularly an undersampled measurement dataset, is acquired by parallel imaging using a receiving coil device (11) with multiple receiving channels arranged in an acquisition geometry. This includes a reconstruction unit (5) for determining the magnetic resonance image dataset during a reconstruction process (1) using at least one trained image processing function. The reconstruction process includes a data consistency step to improve the consistency between the resulting image and the measurement dataset based on a regularization value. The image processing device (2) is characterized in that it further includes a modification unit (6) for modifying the noise reduction intensity of the reconstruction process (1) in a position-related manner during the data consistency step, wherein the modification is performed based on a modification variable that locally describes the expected noise reduction intensity of the unmodified reconstruction process due to the acquisition geometry.

10. A computer program product, when executed on an image processing device (2), causes the device to perform the steps of the method according to any one of claims 1 to 8.

11. An electronically readable data carrier having a computer program product according to claim 10 stored thereon.

Images