Denoising of MR images based on deep learning
By performing deep learning processing on the real and imaginary parts of the MR image, using the weighted training of complex value convolution neural network and different loss functions, the problems of Rice noise bias and oversmooth are solved, and efficient denoising and details are achieved Reserved, suitable for low-cost MR systems.
Patent Information
- Application Number
- CN202380032296.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-03-31
- Filing Date
- 2023-03-21
- Publication Date
- 2024-11-26
AI Technical Summary
The existing MR image denoising method based on deep learning cannot effectively overcome the Rice noise bias, resulting in denoising but biased image problems, and often lead to oversmoothing and damage diagnostic details.
By performing deep learning algorithm processing on the real and imaginary parts of the MR image, using the weighted training of complex value convolution neural network and different loss functions, the high-frequency and low-frequency parts are processed respectively to avoid Rice noise bias and accurately control the The balance between noise and retaining details.
Effectively remove noise in MR images, avoid Rice noise bias, preserve image details, and improve image quality. It is suitable for low-cost MR systems.
Smart Images

Figure CN119032288A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of magnetic resonance (MR) imaging. It relates to a method of MR imaging of an object placed in an examination volume of an MR system. The invention also relates to an MR system and a computer program to be run on the MR system. Background Art
[0002] Image formation MR methods which exploit the interaction between magnetic fields and nuclear spins in order to form two-dimensional or three-dimensional images are widely used today, in particular in the field of medical diagnostics, since they are superior in many respects to other imaging methods for imaging soft tissues, do not require ionizing radiation and are generally non-invasive.
[0003] MR image quality improvement is critical for diagnostic confidence and quality. A low signal-to-noise ratio (SNR) can significantly reduce the quality of MR images. One way to improve the SNR is to use multiple acquisitions at the same position in k-space and average the resulting signal data. Naturally, this process is time-consuming and expensive, which creates a need for other solutions. Alternatively, denoising methods can be used to improve the SNR of the acquired noisy MR images. In addition to conventional filtering techniques, artificial intelligence (AI) / deep learning-based methods have recently become increasingly popular in MR image denoising (see 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, pp. 566-574).
[0004] The noise of a fully sampled MR imaging scan is typically modeled as Gaussian noise in the real and imaginary parts of the complex-valued MR image. However, MR images are typically displayed as magnitude images, resulting in a transformation of the original Gaussian noise. The intensity distribution in the resulting magnitude MR image follows an intensity-dependent Rician distribution. For regions of high signal intensity, the distribution is similar to a Gaussian distribution. However, for regions of low signal intensity images, the distribution approaches a Rayleigh distribution with a non-zero mean. For low SNR images, this results in the so-called Rician noise bias (see Cárdenas-Blanco, A., Tjobs, C., Irarzavarval, P., Cameron, I., 2008, "Noise in magnitude magnetic resonance images", Concepts in Magnetic Resonance Part A: An Educational Journal, 32, Vol. 6, pp. 409-416).
[0005] For known deep learning-based denoising methods, the Rice noise bias is a critical issue because the Rice noise bias cannot be completely removed. As a result, denoised but biased MR images are produced.
[0006] US patent application US2022 / 026516 relates to a method for correcting phase and reducing noise in magnetic resonance phase images. Phase information may be corrupted due to eddy current effects, physiological noise or gradient defects in spatial encoding. A deep neural network is used to denoise the phase image. Denoising of complex images is accomplished by multiplying the amplitude image by the phase-corrected original image. Summary of the invention
[0007] From the foregoing it can be readily appreciated that there is a need for an improved method of denoising MR images. It is therefore an object of the present invention to provide a denoising method that overcomes the Rice bias problem. In this text, the expression denoising is defined as noise reduction, which does not necessarily result in the removal of all noise. Therefore, the degree of noise reduction achieved by denoising can be discussed.
[0008] According to the present invention, a method for MR imaging of an object positioned in an examination volume of an MR system is disclosed. The method comprises the following steps:
[0009] a) subjecting the object to an imaging sequence comprising RF pulses and switched magnetic field gradients, thereby generating MR signals,
[0010] b) acquiring the MR signal,
[0011] c) reconstructing a complex-valued MR image based on the acquired MR signals,
[0012] d) denoising the MR image using a deep learning algorithm operating on the real and imaginary parts of the MR image, and
[0013] e) Calculating a magnitude MR image from the de-noised complex-valued MR image.
[0014] According to the present invention, an MR signal data set is acquired in a conventional manner. MR signals are acquired in k-space using orthogonal detection that delivers real and imaginary signal portions. Conventional image reconstruction techniques (e.g., Fourier transforms) produce complex-valued MR images that indicate the phase and amplitude of the MR signal for each image position. The present invention proposes to overcome the Rice bias problem by performing complex denoising using a deep learning algorithm. The deep learning algorithm operates on the real and imaginary parts of the MR image, rather than on the amplitude MR image (as it is a conventional method). In this way, the present invention takes advantage of the fact that the noise characteristics of the individual real and imaginary parts are Gaussian, which is significantly easier to remove. Before the image is finally converted into an amplitude MR image for display to the user, potential noise is removed from the real and imaginary parts of the image. Therefore, Rice noise bias is avoided before Rice noise bias exists.
[0015] The proposed complex denoising method processes both the real and imaginary parts of the MR images. Therefore, the denoising process should be placed within the MR reconstruction pipeline (where both the real and imaginary parts are available) before generating the final amplitude MR image. However, the proposed technique can also be used independently, for example only as a post-processing step independent of the actual MR signal acquisition, and can also be used for other types of imaging modalities besides MR imaging, where complex images are processed. Therefore, the disclosure of the present invention also includes a computer-implemented method for denoising an image, comprising the following steps:
[0016] Provides a complex-valued image,
[0017] Denoising the image using a deep learning algorithm that operates on both the real and imaginary parts of the image, and
[0018] A magnitude image is computed from the denoised complex-valued image.
[0019] In one embodiment, each of the real and imaginary parts is separated into a high-frequency part and a low-frequency part before denoising of the MR image. The terms high frequency and low frequency refer to spatial frequencies, i.e., image information associated with the peripheral and central parts of k-space, respectively. In order to obtain better generalization capabilities across different MR imaging contrasts, complex denoising should operate only on the high-frequency part of the real part and the low-frequency part of the imaginary part of the MR image. In this way, low-frequency contrast-specific image information is preserved, and denoising only processes high-frequency noise and structure 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 each added back to the denoised high-frequency real and imaginary parts. The amplitude MR image is calculated based on the final real and final imaginary parts. The amplitude MR image is then displayed to the user and / or stored to Dicom. Additional image processing steps may be applied before the amplitude MR image is shown to the user.
[0020] As mentioned above, Rice noise bias is a key problem in known deep learning based denoising methods. In particular, if the deep learning model only operates on the high frequency part of the MR image, it cannot make changes to the low frequency part of the MR image. For amplitude MR images with Rice noise bias, this means that the bias in the low frequency image part cannot be removed. As a result, a denoised but biased image is produced. In contrast, the method of the present invention can apply denoising that only operates on the high frequency part without causing any bias in the final MR image. This is because the method of the present invention (in which the deep learning algorithm processes complex image data) is unbiased from the beginning.
[0021] In an embodiment of the present invention, the deep learning algorithm may use a complex-valued convolutional neural network. In this way, the neural network is enabled to operate on both the real and imaginary parts of the MR image. This means that the neural network performs a complex convolution for processing the real and imaginary parts of the MR image.
[0022] 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 often produce images with an unnatural appearance. In addition, strong smoothing can eliminate small anatomical features that may be important for diagnosis. Therefore, although noise removal helps to improve image quality, too much noise removal may also be harmful. For these reasons, the present invention proposes that a deep learning algorithm uses a collection of denoising models trained using different loss functions. Different loss functions within the meaning of the present invention can be different parameterized versions of the same dedicated loss function, where the denoising intensity can be set to a desired level. In a possible embodiment, this is achieved by a loss function that differs according to the corresponding weighting of their variance and bias, where the bias is a measure of the ability of the denoising model to retain image details, and the variance is a measure of the degree of denoising achieved by the denoising model. In this way, the trade-off between denoising and retaining small image details can be precisely controlled.
[0023] The proposed variable weighting of variance and bias can be implemented in conjunction with the above-described denoising of the real and imaginary parts of MR images, or independently thereof, such as for conventional deep learning-based denoising that operates directly on amplitude images. To this end, the disclosure of the present invention includes a computer-implemented method for denoising an image, comprising the following steps:
[0024] Provide images,
[0025] providing a collection of deep learning based denoising models trained using different loss functions, wherein the loss functions differ according to respective weightings of their variance and bias, wherein bias is a measure of the ability of the denoising model to preserve image details and variance is a measure of the degree of noise reduction achieved by the denoising model,
[0026] The image is denoised using a denoising model interactively selected by a user from the set. The user can try different denoising models and, after viewing the corresponding denoised images, select the model that he thinks 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 that are more severely affected by noise will benefit from the method of the invention, since denoising and tuning of the denoising strength are of high practical relevance in these systems.
[0027] The method of the present invention described so far can be performed by an MR system, the MR system comprising: at least one main magnet coil for generating a uniform and stable magnetic field B in an examination volume 0The method of the invention can be implemented by corresponding programming of the reconstruction unit and / or the control unit of the MR system.
[0028] The method of the present invention can be advantageously performed on most MR systems currently in clinical use. To this end, it is only necessary to use a computer program, through which the MR system is controlled so that it performs the above-mentioned method steps of the present invention. The computer program can exist on a data carrier or in a data network so as to be downloaded for installation in a control unit of the MR system. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The accompanying drawings disclose preferred embodiments of the present invention. However, it should be understood that the accompanying drawings are designed for illustrative purposes only and are not intended to limit the scope of the present invention. In the accompanying drawings:
[0030] Figure 1 An MR system for performing the method of the present invention is shown;
[0031] Figure 2 The method of the present invention is illustrated as a flow chart;
[0032] Figure 3 illustrates the method of the present invention compared with conventional deep learning based image denoising;
[0033] Figure 4 Further aspects of the invention are illustrated as flow charts. DETAILED DESCRIPTION
[0034] refer to Figure 1 , shows an MR system 1 as a block diagram. The device includes a superconducting or resistive main magnet coil 2 so that a substantially uniform, time-constant main magnetic field B is generated along the z-axis through the examination volume. 0 The apparatus further comprises a set 2' of (first, second and, where appropriate, third order) shim coils, wherein the currents through the individual shim coils of the set 2' are controllable in order to minimize B within the examination volume. 0 deviation.
[0035] Magnetic resonance generation and manipulation systems apply a series of RF pulses and switched 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, etc., to perform MR imaging.
[0036] More specifically, the gradient amplifier 3 applies current pulses or waveforms to selected ones of the whole body gradient coils 4, 5 and 6 along the x-axis, y-axis and z-axis of the examination volume. A digital RF frequency transmitter 7 transmits RF pulses or pulse packets to a body RF coil 9 via a transmit / receive switch 8 to transmit RF pulses into the examination volume. A typical MR imaging sequence consists of packets of short duration RF pulse segments that, together with any applied magnetic field gradients, achieve selected manipulation of the nuclear magnetic resonance signals. The RF pulses are used to saturate resonance, excite resonance, invert magnetization, refocus resonance or manipulate resonance and select a portion of a body 10 positioned in the examination volume. MR signals are also picked up by the body RF coil 9.
[0037] To generate MR images of limited regions of the body 10 by parallel imaging or for scan acceleration, a set of local array RF coils 11, 12, 13 are placed adjacent to the region selected for imaging. The array coils 11, 12, 13 may be used to receive MR signals induced by body coil RF transmissions.
[0038] The resulting MR signals are picked up by the body RF coil 9 and / or by the array RF coils 11,12,13 and demodulated by a receiver 14 which preferably comprises a preamplifier (not shown). The receiver 14 is connected to the RF coils 9,11,12 and 13 via a transmit / receive switch 8.
[0039] The main computer 15 controls the shim coils 2' and the gradient pulse amplifiers 3 and the transmitter 7 to generate any of a plurality 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 a single or multiple MR signals in rapid succession after each RF excitation pulse, wherein the MR signals are received orthogonally so that the MR signals are complex-valued and include real and imaginary parts. The 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 modern MR systems, the data acquisition system 16 is a separate computer dedicated to acquiring raw image data.
[0040] Ultimately, the digital raw image data is reconstructed into an image representation by a reconstruction processor 17, which applies a Fourier transform or other appropriate reconstruction algorithm. The MR image may represent a planar slice through the patient, an array of parallel planar slices, a three-dimensional volume, etc. The image is then stored in an image memory where it may be accessed to convert a slice, projection, or other portion of the image representation into an appropriate format for visualization, such as via a video monitor 18, which provides a human-readable display of the resulting MR image.
[0041] The reconstruction processor 17 is programmed to perform the above and further referenced below Figure 2-Figure 4 The method of the present invention is described.
[0042] exist Figure 2 In the embodiment of the invention shown, an imaging sequence is applied in step 21, which includes RF excitation pulses and switched magnetic field gradients in the readout and phase encoding directions x and y and the slice selection direction z. MR signal data sets are acquired in multiple repetitions of the sequence using different gradient waveforms in the x / y direction and / or the z direction so as to completely cover the desired region of k-space by a suitable (e.g., radial or Cartesian) k-space sampling pattern. Step 21 also includes reconstructing a complex-valued MR image from the acquired MR signals, for example by Fourier transformation as described above.
[0043] In step 22, frequency segmentation of the reconstructed MR image is performed. The real and imaginary parts of the MR image are each divided into a low-frequency part and a high-frequency part. This can be implemented, for example, by filtering (convolution) with a Gaussian kernel of an appropriately selected width that produces the low-frequency part. The high-frequency part is obtained by subtracting the corresponding low-frequency part from the real and imaginary parts of the original MR image.
[0044] In step 23, denoising of the MR image is performed using a deep learning algorithm that operates on the high frequency real and imaginary parts of the MR image.
[0045] Three different versions of the complex denoising step can be considered, all based on convolutional neural networks (CNNs).
[0046] First, a basic channel stacking approach can be envisioned, 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 conventional denoising CNNs, the model outputs two denoised images, each corresponding to the real and imaginary parts of the complex MR image at the input of the CNN.
[0047] As a second option, one can envision a complex-valued CNN for processing the high-frequency real and imaginary parts of the complex MR images. In this case, convolution can be performed similarly to multiplication of complex numbers. In each convolutional layer, the feature map can be calculated as follows:
[0048]
[0049] Among them, x is the complex-valued feature map, and w is the complex-valued convolution kernel.
[0050] As a further alternative, independent processing schemes can be envisaged, where the high frequency real and imaginary parts are processed by the same CNN (e.g., sequentially).
[0051] 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 calculating the training loss function used. The CNN is then trained using a gradient descent optimization algorithm, in which 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 seeks to minimize by changing the weights of the CNN, which means that the optimization algorithm is navigating downward (lowering) the gradient of the loss function error. Noisy training images for this process can be generated by adding Gaussian noise to the real and imaginary parts of existing MR images. In the last variant described 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.
[0052] The processing step 24 is the reassembly into a denoised final MR image.The denoised high frequency real and imaginary parts are added to their corresponding low frequency counterparts.
[0053] Then, in step 25, a final MR image is obtained by taking the amplitude of the denoised complex-valued MR image.
[0054] In an 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 segmentation described in step 22 is applied, thereby producing a low-frequency part and a high-frequency part for the real and imaginary parts of the MR image, respectively. The high-frequency part of the image is then processed by the complex deep learning model described in step 23. The output of the model is the noise present in the high-frequency part of the real and imaginary parts of the MR image. The predicted noise is then subtracted from the original high-frequency part. If necessary, the user can interactively control the amount of noise removed. Finally, the combination described in step 24 is applied, and the final denoised MR amplitude image is calculated in step 25. The amplitude MR image is displayed to the user and can be stored to Dicom.
[0055] Figure 3 Results of an exemplary embodiment of the method of the present invention compared to conventional deep learning-based amplitude denoising are shown. A noisy amplitude MR image is shown on the left. In the upper row, the denoising results of the present invention (complex CNN) are shown. In the lower row, conventional amplitude CNN denoising is shown. Although amplitude CNN denoising results in visible biases inside and outside the anatomical structure, the 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 noise-free ground truth image.
[0056] Figure 4 Another aspect of the invention is illustrated, which provides a control of the denoising strength of a deep learning based denoising method. In step 41, a noise simulation is performed, which can generate realistic noise scenes from noise-free ground truth images. For Rice noise scenes (e.g., in SENSE scans), Gaussian noise can be added to the real and imaginary parts of the MR images. For colored noise scenes (e.g., C SENSE), the noise-free ground truth images can be undersampled in k-space to generate noise realizations. In this way, a training MR image set is generated.
[0057] In step 42, the generated training MR images are used to train a collection of deep learning denoising models. Each model receives a noisy training image as its input and predicts a corresponding noise-free ground-truth image at its output. Alternatively, the model can be trained to predict the noise present in the image. This predicted noise is then subtracted from the input, also resulting in a denoised image in each case. The model can have any architecture, a typical choice would again be a CNN. Each denoising model of the collection is associated with a different loss function used to train the model. The loss functions differ according to their respective weightings of variance and bias, where the bias is a measure of the ability of the denoising model to preserve image detail 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 across multiple processed noisy training images and the corresponding noise-free ground-truth image. For additive Gaussian noise, the noisy training image can be defined as Where x is the noise-free image and n is the added noise, where n ≈ N(0,σ). Accordingly, the variance can be expressed as the pixel-by-pixel standard deviation across multiple processed noisy training images. Therefore, the loss function to be optimized (minimized) in the training step 42 can be expressed as:
[0058]
[0059] in
[0060]
[0061] and
[0062]
[0063] in, is the L1 norm (optionally, the L2 norm or any other distance metric can also be used), is a single denoised training image, and is a tensor containing all M denoised training images. D is the total number of pixels or voxels in the image. σ is a function that computes the pixel-wise standard deviation across different training images. As an alternative, the function can also compute the variance. The parameter α is adjustable and allows precise control over the trade-off between preserving details and denoising strength. Therefore, a collection of denoising models is trained for different values of α. A different value of α is associated with each denoising model in the collection.
[0064] In step 43, an MR image to be denoised is provided. The MR image may be a complex MR image or a magnitude MR image acquired as described above.
[0065] In step 44, the MR image is denoised using one of the previously trained denoising models for the selected alpha value, and the denoised image is displayed to the user.
[0066] In step 45, the user evaluates the displayed image and decides whether the results of the denoising process are satisfactory. If the results are not satisfactory, step 44 is repeated for different α values.
[0067] If the result is considered satisfactory, the process ends in step 46. The final de-noised MR image can be stored in Dicom for further diagnosis.
[0068] Various aspects and embodiments of the present invention are summarized in the following clauses:
[0069] Clause 1. A method of MR imaging of an object (10) positioned in an examination volume of an MR system (1), the method comprising the following steps:
[0070] a) subjecting the object (10) to an imaging sequence comprising RF pulses and switched magnetic field gradients, thereby generating MR signals,
[0071] b) acquiring the MR signal,
[0072] c) reconstructing a complex-valued MR image based on the acquired MR signals,
[0073] d) denoising the MR image using a deep learning algorithm operating on the real and imaginary parts of the MR image, and
[0074] e) Calculating a magnitude MR image from the de-noised complex-valued MR image.
[0075] Clause 2. The method of clause 1, wherein prior to the denoising of the MR image, each of the real part and the imaginary part is separated into a high frequency part and a low frequency part.
[0076] Clause 3. The method according to clause 2, wherein the denoising operates only on the high frequency portion of the real part and the high frequency portion of the imaginary part of the MR image.
[0077] Item 4. A method according to Item 3, wherein the denoised high-frequency real part and the low-frequency real part are combined into a final real part, and the denoised high-frequency imaginary part and the low-frequency imaginary part are also combined into a final imaginary part, wherein the amplitude MR image is calculated based on the final real part and the final imaginary part.
[0078] Clause 5. A method according to any one of clauses 1-4, wherein the deep learning algorithm uses a convolutional neural network.
[0079] Clause 6. A method according to clause 5, wherein the convolutional neural network is complex-valued.
[0080] Clause 7. A method according to any one of clauses 1-6, wherein the deep learning algorithm uses a collection of denoising models trained using different loss functions.
[0081] Clause 8. A method according to clause 7, wherein the loss functions differ according to the respective weightings of their variance and bias, wherein 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.
[0082] Clause 9. The method according to clause 7 or 8, wherein one of the denoising models from the set is interactively selected by a user for denoising the MR image.
[0083] Clause 10. A computer-implemented method for denoising an image, the method comprising the steps of:
[0084] Provides a complex-valued image,
[0085] denoising the image using a deep learning algorithm that operates on the real and imaginary parts of the image, and
[0086] Compute a magnitude image from the denoised complex-valued image.
[0087] Clause 11. A computer-implemented method for denoising an image, the method comprising the steps of:
[0088] Provide images,
[0089] providing a collection of deep learning based denoising models trained using different loss functions, wherein the loss functions differ according to respective weightings of their variance and bias, wherein bias is a measure of the ability of the denoising model to preserve image details and variance is a measure of the degree of noise reduction achieved by the denoising model,
[0090] The image is denoised using a denoising model interactively selected by a user from the set.
[0091] Clause 12. A MR system comprising:
[0092] At least one main magnet coil (2) for generating a uniform and stable magnetic field B in the examination volume 0; a plurality of gradient coils (4, 5, 6) for generating switching magnetic field gradients in different spatial directions within the examination volume; at least one RF coil (9) for generating RF pulses within the examination volume and / or for receiving MR signals from an object (10) positioned in the examination volume; a control unit (15) for controlling the temporal succession of RF pulses and switching magnetic field gradients; and a reconstruction unit (17) for reconstructing an MR image from the received MR signals, wherein the MR system (1) is arranged to perform a method according to any of clauses 1-11.
[0093] Clause 13. A computer program comprising instructions which, when executed by a computer, preferably by a reconstruction unit (17) of an MR system (1), causes the computer to perform the method according to any of clauses 1-11.
Claims
1. A method for MR imaging of an object (10) positioned in an examination volume of an MR system (1), the method comprising the following steps: a) subjecting the object (10) to an imaging sequence comprising RF pulses and switched magnetic field gradients, thereby generating MR signals, b) acquiring the MR signal, c) reconstructing a complex-valued MR image based on the acquired MR signals, d) denoising the MR image using a deep learning algorithm operating on the real and imaginary parts of the MR image, and e) Calculating a magnitude MR image from the de-noised complex-valued MR image.
2. The method according to claim 1, wherein: Prior to the denoising of the MR image, each of the real part and the imaginary part is separated into a high frequency part and a low frequency part.
3. The method according to claim 2, wherein: The denoising operates only on the high frequency portion of the real part and the high frequency portion of the imaginary part of the MR image.
4. The method according to claim 3, wherein: The denoised high-frequency real part and the low-frequency real part are combined into a final real part, and the denoised high-frequency imaginary part and the low-frequency imaginary part are also combined into a final imaginary part, wherein the magnitude MR image is calculated based on the final real part and the final imaginary part.
5. The method according to any one of claims 1 to 4, wherein: The deep learning algorithm uses a convolutional neural network.
6. The method according to claim 5, wherein: The convolutional neural network is complex-valued.
7. The method according to any one of claims 1 to 6, wherein: The deep learning algorithm uses an ensemble of denoising models trained using different loss functions.
8. The method according to claim 7, wherein: The loss functions differ according to their respective weightings of variance and bias, where bias is a measure of the ability of the denoising model to preserve image details and variance is a measure of the degree of noise reduction achieved by the denoising model.
9. The method according to claim 7 or 8, wherein: One of the denoising models from the set is interactively selected by a user for denoising the MR image.
10. A computer-implemented method for denoising an image, the method comprising the steps of: Provides a complex-valued image, denoising the image using a deep learning algorithm that operates on the real and imaginary parts of the image, and Compute a magnitude image from the denoised complex-valued image.
11. A computer-implemented method for denoising an image, the method comprising the steps of: Provide images, providing a collection of deep learning based denoising models trained using different loss functions, wherein the loss functions differ according to respective weightings of their variance and bias, wherein bias is a measure of the ability of the denoising model to preserve image details and variance is a measure of the degree of noise reduction achieved by the denoising model, The image is denoised using a denoising model interactively selected by a user from the set.
12. An MR system, comprising: At least one main magnet coil (2) for generating a uniform, stable magnetic field (B0) within an examination volume; a plurality of gradient coils (4, 5, 6) for generating switching magnetic field gradients in different spatial directions within the examination volume; at least one RF coil (9) for generating RF pulses within the examination volume and / or for receiving MR signals from an object (10) positioned in the examination volume; a control unit (15) for controlling the temporal succession of RF pulses and switching magnetic field gradients; and a reconstruction unit (17) for reconstructing an MR image from the received MR signals, wherein the MR system (1) is arranged to perform the method according to any one of claims 1 to 11.
13. A computer program comprising instructions which, when executed by a computer, preferably by a reconstruction unit (17) of an MR system (1), causes the computer to perform the method according to any one of claims 1-11.
Citation Information
Patent Citations
Systems and methods of generating robust phase images in magnetic resonance images
US20220026516A1
