Deep Learning Based Noise Removal of MR Images

Abstract

The present invention relates to a method for MR imaging of an object placed in an examination volume of an MR system 1. The object of the present invention is to provide a deep learning based denoising approach that overcomes the Rice bias problem. As a solution, the present invention proposes a method comprising the steps of: a) subjecting the object to an imaging sequence with RF pulses and switched magnetic field gradients, whereby MR signals are generated; b) acquiring the MR signals; c) reconstructing a complex MR image from the acquired MR signals; d) denoising the MR images using a deep learning algorithm operating on real and imaginary parts of the MR images; and e) calculating an intensity MR image from the denoised complex MR images. According to an aspect of the present invention, the deep learning algorithm uses a set of denoising models that are trained with different loss functions. In this way, the trade-off between denoising and preserving small image details can be controlled. Furthermore, the present invention relates to an MR system 1 and a computer program for the MR system 1.

Description

[Technical field]

[0001] The present invention relates to the field of magnetic resonance (MR) imaging. The present invention relates to a method for MR imaging of an object placed in an examination volume of an MR system. The present invention also relates to an MR system and to a computer program executed on the MR system. [Background technology]

[0002] MR imaging methods, which utilize the interaction between a magnetic field and atomic nuclear spins to form two- or three-dimensional images, are widely used today, especially in the field of medical diagnostics, since they are superior in many ways to other imaging methods for imaging soft tissues, do not require ionizing radiation, and are usually non-invasive. Summary of the Invention [Problem to be solved by the invention]

[0003] Improving MR image quality is crucial for diagnostic reliability and quality. A low signal-to-noise ratio (SNR) can substantially reduce the quality of an MR image. One way to improve the SNR is to use multiple acquisitions at the same location in k-space and average the resulting signal data. Naturally, this process is time-consuming and expensive, creating a need for other solutions. Alternatively, denoising methods can be used to improve the SNR of acquired noisy MR images. In addition to traditional filtering techniques, artificial intelligence (AI) / deep learning-based methods for MR image denoising have recently gained popularity. (Jiang, D., Dou, W., Vosters, L., Xu, X., Sun, Y., Tan, T., 2018, "Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network", Japanese Journal of Radiology, 36, vol. 9, p. 566-574).

[0004] The noise of fully sampled MR imaging scans is usually modeled as Gaussian noise in the real and imaginary parts of complex-valued MR images. However, since MR images are usually displayed as magnitude images, the original Gaussian noise is transformed. The intensity distribution obtained in magnitude MR images follows an intensity-dependent Rician distribution. In areas of high signal intensity, this distribution resembles a Gaussian distribution. However, in image areas with low signal intensity, the distribution approaches a Rayleigh distribution with a nonzero mean. For low SNR images, this leads to the so-called Rician noise bias (see Cardenas-Blanco, A., Tejos, C., Irarrazaval, P., Cameron, I., 2008, "Noise in magnitude magnetic resonance images", Concepts in Magnetic Resonance Part A: An Educational Journal, 32, vol. 6, p. 409-416).

[0005] Rician noise bias is a significant issue for known deep learning based denoising methods because it cannot be completely removed, resulting in denoised but biased MR images.

[0006] US Patent Application Publication No. 2022 / 026516 relates to a method for correcting phase and reducing noise in magnetic resonance phase images. Phase information can be corrupted due to eddy current effects in spatial encoding, physiological noise, or gradient imperfections. A deep neural network is used to denoise the phase images. Complex images are denoised by multiplying the intensity images by phase-corrected pristine images.

[0007] From the above, it can be readily appreciated that there is a need for improved methods for denoising MR images. It is therefore an object of the present invention to provide a denoising approach that overcomes the Rice bias problem. [Means for solving the problem]

[0008] In this specification, the term noise removal is defined as noise reduction, which does not necessarily remove all noise, and therefore it is possible to discuss the degree of noise reduction achieved by noise removal.

[0009] According to the invention, a method for MR imaging of an object placed in an examination volume of an MR system is disclosed, comprising the steps of: a) subjecting the object to an imaging sequence including RF pulses and switched magnetic field gradients, whereby MR signals are generated, b) acquiring the MR signals, c) reconstructing a complex-valued MR image from the acquired MR signals, d) denoising the MR image using a deep learning algorithm operating on real and imaginary parts of the MR image, and e) calculating an intensity MR image from the denoised complex MR image.

[0010] According to the invention, an MR signal data set is acquired in a conventional manner. The MR signals are acquired in k-space using quadrature detection, which provides real and imaginary signal parts. Conventional image reconstruction techniques, such as Fourier transformation, generate complex MR images that indicate the phase and amplitude of the MR signal for each image location. The invention proposes to overcome the Rice bias problem by using a deep learning algorithm to perform complex noise removal. The deep learning algorithm does not operate on the intensity MR image (as in conventional approaches), but instead on the real and imaginary parts of the MR image. In this way, the invention exploits that the noise characteristics of the individual real and imaginary parts are Gaussian, which is significantly easier to remove. Potential noise is removed from the real and imaginary parts of the image, which is then finally converted into an intensity MR image for display to the user. As a result, the Rice noise bias is avoided before it exists.

[0011] The proposed complex denoising method processes the real and imaginary parts of the MR image. Therefore, the denoising procedure should be placed in the MR reconstruction pipeline (where the real and imaginary parts are available) before the final intensity MR image is generated. However, the technique proposed by the present invention can also be used independently, for example as a mere post-processing step unrelated to the actual MR signal acquisition, and for other types of imaging modalities other than MR imaging, where complex images are processed. Thus, the present disclosure encompasses a computer-implemented method for denoising an image, comprising the steps of providing an image with complex values ​​(complex image), denoising the complex image using a deep learning algorithm operating on the real and imaginary parts of the complex image, and calculating an intensity image from the denoised complex image.

[0012] In one embodiment, each of the real and imaginary parts is split into a high-frequency part and a low-frequency part before denoising the MR image. The terms "high-frequency" and "low-frequency" refer to image information related to spatial frequencies, i.e., the central and peripheral parts of k-space, respectively. To obtain better generalizability across different MR imaging contrasts, the complex denoising should operate only on the high-frequency parts of the real and imaginary parts of the MR image. In this way, low-frequency contrast-specific image information is preserved and the denoising deals only with high-frequency noise and structures in the MR image. The deep learning algorithm generates denoised versions of the high-frequency real and imaginary parts, respectively. After the denoising step, the denoised high-frequency real and low-frequency real parts are combined into the final real part, and the denoised high-frequency imaginary and low-frequency imaginary parts are combined into the final imaginary part. In other words, the low-frequency parts are added back to the denoised high-frequency real and imaginary parts, respectively. The intensity MR image is calculated from the final real and imaginary parts. The intensity MR image is then displayed to the user and / or stored in DICOM. Other image processing steps may be applied before the intensity MR image is shown to the user.

[0013] As mentioned above, the Rice noise bias is a significant challenge in known deep learning-based denoising methods. In particular, if the deep learning model operates only on the high frequency parts of the MR image, it cannot modify the low frequency parts of the MR image. For an intensity MR image with Rice noise bias, this means that the bias in the low frequency image parts cannot be removed. As a result, a denoised but biased image is generated. In contrast, the method of the present invention can apply noise reduction that only works on the high frequency parts without introducing bias into the final MR image. This is because the method of the present invention, in which the deep learning algorithm processes complex image data, is bias-free from the start.

[0014] In one embodiment of the present invention, the deep learning algorithm can use a complex-valued convolutional neural network. In this way, the neural network is able to operate on the real and imaginary parts of the MR images. This means that the neural network performs complex convolutions for processing the real and imaginary parts of the MR images.

[0015] Another problem with deep learning based MR image denoising methods is that they often suffer from over-smoothing. For example, high-performance deep learning models that can remove almost all noise in MR images usually produce images with an unnatural appearance. Furthermore, strong smoothing can eliminate small anatomical features that may be important for diagnosis. Thus, while denoising helps to improve image quality, excessive denoising can also be harmful. For these reasons, the present invention proposes that the deep learning algorithm uses a set of denoising models that are trained using different loss functions. Different loss functions within the meaning of the present invention can be different parameterized versions of the same special loss function, where the denoising strength can be set to a desired level. In a possible embodiment, this is achieved by different loss functions with respective variance and bias weighting, where the bias is a measure of the ability of the denoising model to preserve image details and the variance is a measure of the degree of noise reduction achieved by the denoising model. In this way, the trade-off between noise reduction and preservation of small image details can be precisely controlled.

[0016] The proposed variable weighting of variance and bias can be implemented in combination with the above-mentioned denoising of real and imaginary parts of MR images, or independently of them, for example for conventional deep learning-based denoising operating directly on intensity images. To this end, the present disclosure encompasses a computer-implemented method for denoising an image, comprising the steps of providing an image, providing a set of deep learning-based denoising models trained using different loss functions, the loss functions differing in their individual variance and bias weighting, the bias being a measure of the ability of the denoising models to preserve image details, and the variance being a measure of the degree of noise reduction achieved by the denoising models, and denoising the image using a denoising model interactively selected by a user from the set. The user can try different denoising models and, after reviewing each denoised image, can select the model that he believes represents the best compromise between noise reduction and detail accuracy. The proposed denoising method can be applied to any type of imaging modality, in particular MR, CT, PET. In particular, low-cost MR systems, which are more severely affected by noise, benefit from the method of the present invention, since noise reduction and tuning of the noise reduction strength are of high practical relevance in these systems.

[0017] The inventive method described so far can be performed by an MR system having at least one main magnet coil for generating a uniform constant magnetic field B0 in an examination volume, a number of gradient coils for generating switched magnetic field gradients in different spatial directions in the examination volume, at least one body RF coil for generating RF pulses in the examination volume and / or receiving MR signals from a patient's body placed in the examination volume, a control unit for controlling the time sequence of the RF pulses and the switched magnetic field gradients, and a reconstruction unit for reconstructing MR images from the received MR signals. The inventive method can be realized by corresponding programs of the reconstruction unit and / or the control unit of the MR system.

[0018] The method of the invention can be advantageously implemented in most MR systems in clinical use. For this purpose, it is only necessary to utilize a computer program, which controls the MR system to execute the above-mentioned method steps of the invention. The computer program may be present on a data carrier or in a data network, so as to be downloaded for installation in the control unit of the MR system.

[0019] The accompanying drawings disclose preferred embodiments of the present invention. It is to be understood, however, that the drawings are designed for illustrative purposes only and are not designed as a definition of the limits of the invention. In the drawings: [Brief description of the drawings]

[0020] [Figure 1] FIG. 1 shows an MR system for implementing the method of the present invention. [Diagram 2] FIG. 2 illustrates the method of the present invention as a flow diagram. [Diagram 3] FIG. 1 illustrates our method compared to conventional deep learning based image denoising. [Figure 4] FIG. 4 illustrates, in a flow chart, another embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] Referring to Figure 1, an MR system 1 is shown as a block diagram. The apparatus comprises superconducting or resistive main magnet coils 2 such that a substantially uniform and time-constant main magnetic field B0 is generated along the z-axis through the examination volume. The apparatus further comprises a set of shimming coils 2' (primary, secondary and, if applicable, tertiary), the current through each individual shimming coil of the set 2' being controllable to minimize B0 deviations within the examination volume.

[0022] The magnetic resonance generation and manipulation system applies a series of RF pulses and switching magnetic field gradients to invert or excite nuclear magnetic spins, induce magnetic resonance, refocus magnetic resonance, manipulate magnetic resonance, spatially and otherwise encode magnetic resonance, saturate spins, and the like to perform MR imaging.

[0023] More specifically, gradient amplifiers 3 apply current pulses or waveforms to selected ones of whole-body gradient coils 4, 5, and 6 along the x-, y-, and z-axes of the examination volume. A digital RF frequency transmitter 7 transmits RF pulses or pulse packets to body RF coil 9 via transmit / receive switch 8, which transmits RF pulses into the examination volume. A typical MR imaging sequence consists of packets of short duration RF pulse segments which, together with any applied magnetic field gradients, achieve selected manipulation of the nuclear magnetic resonance signal. The RF pulses are used to saturate resonance, excite resonance, invert magnetization, refocus resonance, or manipulate resonance to select portions of a body 10 located within the examination volume. The MR signals are then picked up by the body RF coil 9.

[0024] To generate MR images of limited regions of the body 10, or for scan acceleration by parallel imaging, a set of local array RF coils 11, 12, 13 are positioned adjacent to the region selected for imaging. The array coils 11, 12, 13 can be used to receive MR signals induced by the body coil RF transmission.

[0025] The resulting MR signals are picked up by the body RF coil 9 and / or the array RF coils 11, 12, 13 and demodulated by a receiver 14, preferably having a preamplifier (not shown). The receiver 14 is connected to the RF coils 9, 11, 12, 13 via a transmit / receive switch 8.

[0026] A host computer 15 controls the shimming coil 2' as well as the gradient pulse amplifier 3 and the transmitter 7 to generate any of a number of MR imaging sequences, such as echo planar imaging (EPI), echo volume imaging, gradient and spin echo imaging, fast spin echo imaging, etc. For the selected sequence, the receiver 14 receives single or multiple MR signals in rapid succession following each RF excitation pulse, the MR signals being received in quadrature, the MR signals being complex-valued, having real and imaginary parts. A data acquisition system 16 performs analog-to-digital conversion of the received signals and converts the MR data samples into a digital format suitable for further processing. In today's MR systems, the data acquisition system 16 is a separate computer dedicated to collecting the raw image data. Finally, the digital raw image data is reconstructed into an image representation by a reconstruction processor 17, which applies a Fourier transform or other suitable reconstruction algorithm. The MR images can represent planar slices across the patient, an array of parallel planar slices, a three-dimensional volume, etc. The images are then stored in an image memory, which can be accessed to convert slices, projections, or other portions of the image representation into an appropriate format for visualization, for example, via a video monitor 18, which provides a human-readable display of the resulting MR image.

[0027] The reconstruction processor 17 is programmed to carry out the method of the invention as described herein above and below with further reference to Figures 2 to 4.

[0028] In one embodiment of the invention shown in Fig. 2, in step 21 an imaging sequence is applied, which comprises RF excitation pulses and switched magnetic field gradients in readout direction x and phase encoding direction y and slice selection direction z. MR signal data sets are acquired in multiple repetitions of the sequence using different gradient waveforms in x / y and / or z direction to completely cover the required region of k-space by an appropriate (e.g. radial or Cartesian) k-space sampling pattern. Step 21 also comprises reconstructing a complex MR image from the acquired MR signals, for example by means of the Fourier transformation described above.

[0029] In step 22, a frequency decomposition of the reconstructed MR image is performed. The real and imaginary parts of the MR image are respectively divided into low-frequency and high-frequency parts. This can be achieved, for example, by filtering (convolution) with a Gaussian kernel of appropriately selected width, which generates the low-frequency part. The high-frequency part is obtained by subtracting the respective low-frequency parts from the real and imaginary parts of the original MR image.

[0030] In step 23, denoising of the MR image is performed using a deep learning algorithm operating on the high frequency real and imaginary parts of the MR image.

[0031] The three different versions of the complex denoising step can be considered, all based on Convolutional Neural Networks (CNNs).

[0032] First, a basic channel stacking approach can be considered, where the high frequency real and imaginary parts are stacked along the channel dimension of the first CNN layer, similar to the color channel dimension in CNNs used to process natural images. Unlike traditional denoising CNNs, this model outputs two denoised images, each corresponding to the real and imaginary parts of the complex MR image at the input of the CNN.

[0033] As a second option, a complex-valued CNN can be considered to process the high-frequency real and imaginary parts of the complex MR images. In this case, the convolutions can be performed similarly to complex multiplications. At each convolutional layer, the feature maps are TIFF2025510836000002.tif16140In the above formula, x is a complex-valued feature map and w is a complex-valued convolution kernel.

[0034] As another alternative, an independent processing scheme can be considered, in which the high frequency real part and the high frequency imaginary part are processed by the same CNN (e.g., in sequence).

[0035] All CNN models can be trained in residual mode. This means that the CNN receives a noisy training image as input and predicts the noise present in the image. The noise is then subtracted from the input image and compared to the noise-free ground truth for the calculation of the training loss function used. The CNN is then trained using a gradient descent optimization algorithm, and the weights in the CNN are updated using an error backpropagation technique. Gradient descent refers to the gradient of the training loss function that the training algorithm tries to minimize by modifying the weights of the CNN, which means that the optimization algorithm is navigating downwards (descending) the gradient of the loss function error. Noisy training images for this procedure can be generated by adding Gaussian noise to the real and imaginary parts of existing MR images. In the last variant mentioned above (independent processing of high frequency real and imaginary parts by the same CNN), the CNN can also be trained with natural (i.e., non-MR) images, so that access to high quality MR images is not required. This is possible because the noise in the real and imaginary parts of the MR images can be considered uncorrelated.

[0036] Processing step 24 is the recombination to produce a final denoised MR image: the denoised high frequency real and imaginary parts are added to their respective low frequency counterparts.

[0037] The final MR image is then obtained in step 25 by taking the intensity of the denoised complex MR image.

[0038] In one embodiment of the present invention, the described denoising method can be advantageously implemented in the reconstruction processor 17 of the MR system 1, since the real and imaginary parts of the MR image need to be available. First, the frequency division described in step 22 is applied to obtain low and high frequency parts for each of the real and imaginary parts of the MR image. Then, the high frequency part of the image is processed by the complex deep learning model described in step 23. The output of the model is the noise present in the high frequency parts of the real and imaginary parts of the MR image. The predicted noise is then subtracted from the original high frequency parts. If necessary, the user can interactively control the amount of noise to be removed. Finally, the synthesis described in step 24 is applied to calculate the final denoised MR intensity image in step 25. The intensity MR image can be displayed to the user and stored in DICOM.

[0039] Figure 3 shows the results of an exemplary implementation of the method of the present invention compared to conventional deep learning-based intensity denoising. A noisy intensity MR image is shown on the left. In the top row, the denoising results of the present invention (complex CNN) are shown. In the bottom row, conventional intensity CNN denoising is shown. Intensity CNN denoising results in visible biases inside and outside the anatomical structures, while complex CNN does not show any visible bias. The residual bias map shown on the right refers to the absolute difference between the denoised image and the corresponding noiseless ground truth image.

[0040] 4 shows another aspect of the present invention that provides for controlling the denoising strength of the deep learning based denoising method. Noise simulation is performed in step 41 to generate realistic noise scenarios from noise-free ground truth images. For Rician noise scenarios (e.g., SENSE scans), Gaussian noise can be added to the real and imaginary parts of the MR images. For colour noise scenarios (e.g., C-SENSE), the noise-free ground truth images can be undersampled in k-space to generate noise realisations. In this way, a set of training MR images is generated.

[0041] In step 42, a set of deep learning denoising models is trained using the generated training MR images. Each model receives as its input a noisy training image and predicts at its output the corresponding noise-free ground truth image. Alternatively, the models can be trained to predict the noise present in the image. This predicted noise is then subtracted from the input, resulting in a denoised image in each case. The models can have any architecture, a typical choice would again be a CNN. Each denoising model of the set is associated with a different loss function used to train the model. The loss functions differ by their respective variance and bias weightings, where the bias is a measure of the ability of the denoising model to preserve image details and the variance is a measure of the degree of noise reduction achieved by the denoising model. The bias can be expressed as the error between the average model output over multiple processed, noisy training images and the respective noise-free ground truth image. In the case of additive Gaussian noise, the noisy training images are TIFF2025510836000003.tif931 where x is the noise-free image and n is the additive noise with n ∼ N(0,σ). Correspondingly, the variance can be expressed as the pixel-wise standard deviation across multiple processed noisy training images. Thus, the loss function that is optimized (minimized) in the training step 42 can be expressed as TIFF2025510836000004.tif978 Where: TIFF2025510836000005.tif2365 and TIFF2025510836000006.tif2453 where L1 is the L1 norm (optionally the L2 norm or other distance measures can be used) and x i ^ is a single denoised training image, x^ is a tensor containing all M denoised training images, D is the total number of pixels or voxels in the image, and σ is a function that calculates the pixel-wise standard deviation across the various training images. Alternatively, the function can also calculate the variance. The parameters are adjustable, allowing precise control of the trade-off between detail retention and denoising strength. Thus, a set of denoising models is trained for different values ​​of α. A different value of α is associated with each denoising model in the set.

[0042] In step 43, a denoised MR image is provided. The MR image may be a complex or intensity MR image acquired as described above.

[0043] In step 44, the MR image is denoised using one of the pre-trained denoising models for the selected value of α, and the denoised image is displayed to the user.

[0044] In step 45, the user evaluates the displayed image and determines whether the results of the denoising procedure are satisfactory. If not, step 44 is repeated for a different value of α.

[0045] If the results are deemed satisfactory, the procedure ends at step 46. The final denoised MR image can be stored in DICOM for further use in diagnosis.

[0046] Various aspects and embodiments of the present invention are summarized under the following headings:

[0047] Item 1. A method for MR imaging of a subject (10) placed in an examination volume of an MR system (1), comprising the steps of: a) subjecting the subject (10) to an imaging sequence including RF pulses and switched magnetic field gradients, whereby MR signals are generated; b) acquiring the MR signals; c) reconstructing a complex MR image from the acquired MR signals; d) denoising the MR image using a deep learning algorithm operating on real and imaginary parts of the MR image; and e) calculating an intensity MR image from the denoised complex MR image.

[0048] Item 2. 2. The method of claim 1, wherein each of the real and imaginary parts is split into a high frequency part and a low frequency part prior to denoising of the MR image.

[0049] Item 3. 3. The method according to claim 2, wherein the denoising operates only on the high frequency parts of the real part and the imaginary part of the MR image.

[0050] Item 4. 4. The method of claim 3, wherein a final real part is obtained by combining the denoised high frequency part of the real part and the low frequency part of the real part, and a final imaginary part is obtained by combining the denoised high frequency part of the imaginary part and the low frequency part of the imaginary part, and the intensity MR image is calculated from the final real part and the final imaginary part.

[0051] Item 5. 5. The method according to any one of items 1 to 4, wherein the deep learning algorithm uses a convolutional neural network.

[0052] Item 6. 6. The method of claim 5, wherein the convolutional neural network processes complex values.

[0053] Item 7. 7. The method according to any one of items 1 to 6, wherein the deep learning algorithm uses a set of denoising models trained using different loss functions.

[0054] Item 8. 8. The method of claim 7, wherein the loss functions differ by their respective variance and bias weightings, the bias being a measure of the ability of the denoising model to preserve image detail and the variance being a measure of the degree of noise reduction achieved by the denoising model.

[0055] Item 9. 9. The method of claim 7 or 8, wherein one of the denoising models in the set is interactively selected by a user to denoise the MR image.

[0056] Item 10. 1. A computer-implemented method for denoising an image, comprising the steps of: providing a complex image; denoising the complex image using a deep learning algorithm that operates on real and imaginary parts of the complex image; and calculating an intensity image from the denoised complex image.

[0057] Item 11. 1. A computer-implemented method for denoising an image, comprising: providing an image; providing a set of deep learning based denoising models trained using different loss functions, the loss functions differing by their respective variance and bias weightings, the bias being a measure of the ability of the denoising model to preserve image detail and the variance being a measure of the degree of noise reduction achieved by the denoising model; and denoising the image using a denoising model selected from the set interactively by a user.

[0058] Item 12. 12. An MR system comprising at least one main magnet coil (2) for generating a uniform constant magnetic field B0 in an examination volume, a plurality of gradient coils (4, 5, 6) for generating switched magnetic field gradients in various spatial directions in the examination volume, at least one RF coil (9) for generating RF pulses in the examination volume and / or receiving MR signals from a subject (10) placed in the examination volume, a control unit (15) for controlling the time sequence of the RF pulses and the switched magnetic field gradients, and a reconstruction unit (17) for reconstructing MR images from the received MR signals, wherein the MR system (1) is configured to perform the method according to any one of items 1 to 11.

[0059] Item 13. A computer program comprising instructions that, when the computer program is executed by a computer, preferably by a reconstruction unit (17) of an MR system (1), cause the computer to carry out the method according to any one of items 1 to 11.

Claims

1. 1. A method for MR imaging of an object disposed within an examination volume of an MR system, comprising: a) subjecting the subject to an imaging sequence including RF pulses and switched magnetic field gradients, whereby MR signals are generated; b) acquiring said MR signals; c) reconstructing a complex MR image from the acquired MR signals; d) denoising the complex MR image using a deep learning algorithm operating on real and imaginary parts of the complex MR image; e) calculating an intensity MR image from the denoised complex MR image; The method according to claim 1,

2. The method of claim 1 , wherein each of the real and imaginary parts is split into a high frequency part and a low frequency part prior to denoising of the complex MR image.

3. The method of claim 2 , wherein the denoising operates only on the high frequency parts of the real part and the high frequency parts of the imaginary part of the complex MR image.

4. 4. The method of claim 3, wherein a final real part is obtained by combining the high frequency part of the denoised real part and the low frequency part of the real part, a final imaginary part is obtained by combining the high frequency part of the denoised imaginary part and the low frequency part of the imaginary part, and the intensity MR image is calculated from the final real part and the final imaginary part.

5. The method of claim 1 , wherein the deep learning algorithm uses a convolutional neural network.

6. The method of claim 5 , wherein the convolutional neural network processes complex values.

7. The method of claim 1 , wherein the deep learning algorithm uses a set of denoising models that are trained using different loss functions.

8. 8. The method of claim 7, wherein the loss function varies with respective variance and bias weightings, the bias being a measure of the ability of the denoising model to preserve image detail, and the variance being a measure of the degree of noise reduction achieved by the denoising model.

9. The method of claim 7 or 8, wherein one of the denoising models in the set is interactively selected by a user to denoise the complex MR image.

10. 1. A computer-implemented method for denoising an image, comprising: Providing a complex image; denoising the complex image using a deep learning algorithm operating on real and imaginary parts of the complex image; calculating an intensity image from the denoised complex image; The method according to claim 1,

11. 1. A computer-implemented method for denoising an image, comprising: Providing an image; providing a set of deep learning based denoising models that are trained using different loss functions, the loss functions differing by their respective variance and bias weightings, the bias being a measure of the ability of the denoising models to preserve image detail and the variance being a measure of the degree of noise reduction achieved by the denoising models; denoising the image using a denoising model interactively selected from the set by a user; The method according to claim 1,

12. at least one main magnet coil for generating a uniform steady magnetic field within the examination volume; a plurality of gradient coils which are switched to generate magnetic field gradients in different spatial directions within the examination volume; at least one RF coil for generating RF pulses in the examination volume and / or receiving MR signals from a subject disposed in the examination volume; a control unit for controlling the time sequence of the RF pulses and the switched magnetic field gradients; a reconstruction unit for reconstructing an MR image from the received MR signals; 12. An MR system comprising: a first input / output tuner (120), the first input / output tuner (120) being configured to perform the method according to any one of claims 1 to 11.

13. Computer program comprising instructions which, when said computer program is executed by a computer or by a reconstruction unit of an MR system, cause the computer or the reconstruction unit to carry out the method according to any one of claims 1 to 11.

Images