Magnetic resonance image processing method and device and computer readable storage medium

Abstract

The invention relates to the technical field of magnetic resonance image processing, and discloses a magnetic resonance image processing method and device and a computer readable storage medium. The processing method comprises the following steps: acquiring a to-be-processed magnetic resonance image; inputting the magnetic resonance image to be processed into a first artifact removal network to obtain a first artifact removal image; inputting the first artifact-removed image into a second artifact-removed network to obtain a second artifact-removed image, and taking the second artifact-removed image as a final magnetic resonance image; wherein the first artifact removal network and the second artifact removal network are respectively used for removing Gibbs artifacts generated by different factors. In this way, the quality of the magnetic resonance image is improved.

Description

Technical Field

[0001] This application relates to the field of magnetic resonance image processing technology, and in particular to methods, apparatus and computer-readable storage media for processing magnetic resonance images. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is increasingly used in medical diagnostics and research due to its superior imaging quality and safety, and is gradually becoming an indispensable tool. However, conventional MRI scans are too time-consuming, resulting in high costs. To shorten scan time, clinical applications often use methods such as reducing the acquisition spatial resolution and partially utilizing k-space to reduce the amount of scan data collected. However, this process can lead to Gibbs artifacts in the images.

[0003] Due to the limitations of sequence scanning time and K-space data acquisition, Gibbs artifacts are an unavoidable problem in magnetic resonance imaging. Summary of the Invention

[0004] The magnetic resonance image processing method, apparatus, and computer-readable storage medium provided in this application improve the quality of magnetic resonance images.

[0005] In a first aspect, this application provides a method for processing magnetic resonance images, the method comprising: acquiring a magnetic resonance image to be processed; inputting the magnetic resonance image to be processed into a first artifact removal network to obtain a first artifact removal image; inputting the first artifact removal image into a second artifact removal network to obtain a second artifact removal image, and using the second artifact removal image as the final magnetic resonance image; wherein the first artifact removal network and the second artifact removal network are respectively used to remove Gibbs artifacts caused by different factors.

[0006] The first artifact removal network is used to remove Gibbs artifacts caused by limited acquisition resolution; the second artifact removal network is used to remove Gibbs artifacts caused by some K-space acquisition; the magnetic resonance image to be processed is input into the first artifact removal network to obtain the first artifact-removed image, including: inputting the magnetic resonance image to be processed into the first artifact removal network to remove Gibbs artifacts caused by limited acquisition resolution to obtain the first artifact-removed image; inputting the first artifact-removed image into the second artifact removal network to remove Gibbs artifacts caused by some K-space acquisition to obtain the second artifact-removed image.

[0007] The first artifact removal network is used to remove some Gibbs artifacts generated by K-space acquisition; the second artifact removal network is used to remove Gibbs artifacts generated by limited acquisition resolution; the magnetic resonance image to be processed is input into the first artifact removal network to obtain a first artifact-removed image, including: inputting the magnetic resonance image to be processed into the first artifact removal network to remove some Gibbs artifacts generated by K-space acquisition to obtain a first artifact-removed image; inputting the first artifact-removed image into the second artifact removal network to remove Gibbs artifacts generated by limited acquisition resolution to obtain a second artifact-removed image.

[0008] The first and second artifact removal networks are trained in the following ways: acquiring a first training image set and acquiring a second training image set; wherein the artifact generation factors corresponding to the training images in the first training image set are different from those corresponding to the training images in the second training image set; the first artifact removal network is trained using the training images in the first training image set; and the second artifact removal network is trained using the training images in the second training image set.

[0009] The artifact generation factor corresponding to the training images in the first training image set is the finite acquisition resolution. Obtaining the first training image set includes: acquiring the original magnetic resonance image; wherein the original magnetic resonance image is free of Gibbs artifacts; converting the original magnetic resonance image to K space, and moving the low-frequency part of K space to the central region of the image to obtain the first K space image; cropping the first K space image in a centrally symmetrical manner to obtain the first cropped image; converting the first cropped image to the space of the original magnetic resonance image to obtain the first training image, thereby forming the first training image set.

[0010] The first cropped image is obtained by cropping the first K-space image in a centrally symmetric manner, which includes: cropping the first K-space image in a centrally symmetric manner along a first direction and / or a second direction of the first K-space image according to a cropping ratio, to obtain the first cropped image.

[0011] The artifacts in the training images of the second training image set are generated by partial K-space acquisition. The acquisition of the second training image set includes: acquiring the original magnetic resonance image; wherein the original magnetic resonance image has no Gibbs artifacts; converting the original magnetic resonance image to K-space and moving the low-frequency part of K-space to the central region of the image to obtain the second K-space image; cropping the second K-space image along the first direction and / or the second direction to obtain the second cropped image; converting the second cropped image to the space of the original magnetic resonance image to obtain the second training image, thereby forming the second training image set.

[0012] Both the first and second artifact removal networks are trained using the following loss function: Among them, O i GT represents the i-th pixel in the image after Gibbs artifact removal. i represents the i-th pixel in the original magnetic resonance image; N represents the total number of pixels in the image.

[0013] In a second aspect, this application provides a magnetic resonance image processing apparatus, which includes a processor and a memory coupled to the processor; wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the method provided in the first aspect.

[0014] Thirdly, this application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the method provided in the first aspect.

[0015] The beneficial effects of the embodiments of this application are as follows: Unlike the prior art, the magnetic resonance image processing method, apparatus and computer-readable storage medium provided in this application use a first artifact removal network and a second artifact removal network to selectively remove Gibbs artifacts caused by different factors in the magnetic resonance image to be processed, thereby obtaining the final magnetic resonance image and improving the quality of the magnetic resonance image. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0017] Figure 1 This is a flowchart illustrating the first embodiment of the magnetic resonance image processing method provided in this application;

[0018] Figure 2 This is a flowchart illustrating the second embodiment of the magnetic resonance image processing method provided in this application;

[0019] Figure 3 This is a flowchart illustrating the third embodiment of the magnetic resonance image processing method provided in this application;

[0020] Figure 4 This is a flowchart illustrating the fourth embodiment of the magnetic resonance image processing method provided in this application;

[0021] Figure 5 This is a schematic flowchart of an embodiment of obtaining a first training image set provided in this application;

[0022] Figure 6This is a schematic flowchart illustrating another embodiment of obtaining the first training image set provided in this application;

[0023] Figure 7 This is a schematic diagram of an embodiment of obtaining a second training image set provided in this application;

[0024] Figure 8 This is a schematic diagram of another embodiment of obtaining a second training image set provided in this application;

[0025] Figure 9 Schematic diagram of Gibbs artifact removal results using different methods in artifact images with limited acquisition resolution;

[0026] Figure 10 Schematic diagram of Gibbs artifact removal results in partial K-space artifact images using different methods;

[0027] Figure 11 This is a schematic diagram of a magnetic resonance image after implementing the magnetic resonance image processing method provided in this application;

[0028] Figure 12 This is a schematic diagram of the structure of an embodiment of the magnetic resonance image processing apparatus provided in this application;

[0029] Figure 13 This is a schematic diagram of the structure of an embodiment of the computer-readable storage medium provided in this application;

[0030] Figure 14 This is a schematic diagram of a magnetic resonance image of Gibbs artifacts produced by the limited acquisition resolution provided in this application;

[0031] Figure 15 This is a schematic diagram of a magnetic resonance image with Gibbs artifacts generated by partial K-space acquisition, provided in this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only for explaining this application and not for limiting it. Furthermore, it should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all structures. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0033] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0034] Magnetic Resonance Imaging (MRI) is increasingly used in medical diagnostics and research due to its superior imaging quality and safety, and is gradually becoming an indispensable tool. However, conventional MRI scans are too time-consuming, resulting in high costs. To shorten scan time, clinical applications often use methods such as reducing the acquisition spatial resolution and partially utilizing k-space to reduce the amount of scan data collected. However, this process can lead to Gibbs artifacts in the images.

[0035] Due to the limitations of sequence scanning time and K-space data acquisition, Gibbs artifacts are an unavoidable problem in magnetic resonance imaging.

[0036] Traditional methods use filtering to suppress high-frequency information to remove Gibbs artifacts, but filtering-based methods can lead to overall image blurring and affect image detail. In addition, these methods are highly dependent on the personal experience of clinicians to adjust relevant parameters. Deep neural networks have shown great potential in removing artifacts while preserving effective image information. However, these methods [2] do not start from the mechanism of artifact generation and are difficult to deal with Gibbs artifacts caused by reducing the acquisition spatial resolution and part of the K space at the same time. Moreover, the methods for removing these two types of artifacts are the same, without distinguishing the specific cause. However, in reality, the causes of these two artifacts are different, and the image appearance is also different. It is difficult to obtain the best results if a unified method is used.

[0037] Based on this, this application proposes to use a first artifact removal network and a second artifact removal network to selectively remove Gibbs artifacts caused by different factors in the magnetic resonance image to be processed, thereby obtaining the final magnetic resonance image and improving the quality of the magnetic resonance image. See any of the following embodiments for specific technical solutions.

[0038] See Figure 1 , Figure 1 This is a flowchart illustrating a first embodiment of the magnetic resonance image processing method provided in this application. The processing method includes:

[0039] Step 11: Obtain the magnetic resonance image to be processed.

[0040] In some embodiments, the magnetic resonance image to be processed can be a magnetic resonance image acquired in real time.

[0041] In some embodiments, the magnetic resonance images to be processed may also be non-real-time acquired magnetic resonance images. For example, magnetic resonance images are acquired using a magnetic resonance imaging device during working hours, and then the acquired magnetic resonance images are processed uniformly according to the processing method of this application.

[0042] In some embodiments, the magnetic resonance image to be processed can be an image with Gibbs artifacts or an image without Gibbs artifacts.

[0043] In some embodiments, the magnetic resonance image to be processed may be an image with a factor that produces Gibbs artifacts.

[0044] In some embodiments, the magnetic resonance image to be processed can be an image with two factors that produce Gibbs artifacts.

[0045] Step 12: Input the magnetic resonance image to be processed into the first artifact removal network to obtain the first artifact removal image.

[0046] In some embodiments, the first artifact removal network can be trained in advance using training images.

[0047] In some embodiments, the magnetic resonance image to be processed is input into a first artifact removal network. The first artifact removal network removes Gibbs artifacts from the magnetic resonance image to be processed according to the factors that generate Gibbs artifacts that it has learned to remove, thereby obtaining a first artifact-removed image.

[0048] In some embodiments, if the magnetic resonance image to be processed has Gibbs artifacts that can be removed by the first artifact removal network, then inputting the magnetic resonance image to be processed into the first artifact removal network can remove the Gibbs artifacts and obtain the first artifact-removed image.

[0049] Step 13: Input the first artifact removal image into the second artifact removal network to obtain the second artifact removal image, and use the second artifact removal image as the final magnetic resonance image; wherein, the first artifact removal network and the second artifact removal network are used to remove Gibbs artifacts caused by different factors.

[0050] In some embodiments, the second artifact removal network can be trained in advance using training images.

[0051] In some embodiments, if the magnetic resonance image to be processed still has Gibbs artifacts that can be removed by the second artifact removal network, then after inputting the magnetic resonance image to be processed into the first artifact removal network, the first artifact removal image still contains Gibbs artifacts that can be removed by the second artifact removal network. Then, the first artifact removal image is input into the second artifact removal network to remove the Gibbs artifacts, resulting in a second artifact removal image, which is used as the final magnetic resonance image.

[0052] In some embodiments, the first artifact-removed image is input into the second artifact-removing network, which removes Gibbs artifacts from the first artifact-removed image according to the factors that generate Gibbs artifacts that it has learned to remove, thus obtaining the second artifact-removed image.

[0053] Since the first and second artifact removal networks are used to remove Gibbs artifacts caused by different factors, the final magnetic resonance image will have Gibbs artifacts caused by both factors removed.

[0054] In some embodiments, if the magnetic resonance image to be processed has a Gibbs artifact caused by a single factor, and the first artifact removal network can remove it, then the first artifact-removed image will be free of Gibbs artifacts. Even if the first artifact-removed image is input into a second artifact removal network to obtain a second artifact-removed image, the second artifact-removed image is not substantially different from the first artifact-removed image. Because the first artifact-removed image is free of Gibbs artifacts, even if the first artifact-removed image is input into the second artifact removal network, the second artifact removal network cannot remove artifacts from the first artifact-removed image.

[0055] If the second artifact removal network can remove it, the MRI image to be processed is input into the first artifact removal network, resulting in a first artifact removal image where the Gibbs artifact still exists. The first artifact removal image is then input into the second artifact removal network, which removes the Gibbs artifact from the first artifact removal image, resulting in a second artifact removal image, which is then used as the final MRI image.

[0056] In some embodiments, the magnetic resonance image to be processed can be an image with two factors that produce Gibbs artifacts. Then, one type of Gibbs artifact can be removed by a first artifact removal network and a second artifact removal network, respectively, to obtain the final magnetic resonance image.

[0057] In this embodiment, a first artifact removal network and a second artifact removal network are used to selectively remove Gibbs artifacts caused by different factors in the magnetic resonance image to be processed, thereby obtaining the final magnetic resonance image and improving the quality of the magnetic resonance image.

[0058] See Figure 2 , Figure 2This is a flowchart illustrating a second embodiment of the magnetic resonance image processing method provided in this application. The processing method includes:

[0059] Step 21: Obtain the magnetic resonance image to be processed.

[0060] In some embodiments, the magnetic resonance image to be processed typically has Gibbs artifacts resulting from limited acquisition resolution and Gibbs artifacts resulting from partial K-space acquisition.

[0061] Step 22: Input the magnetic resonance image to be processed into the first artifact removal network to remove the Gibbs artifacts caused by the limited acquisition resolution and obtain the first artifact-removed image.

[0062] In some embodiments, the first artifact removal network is used to remove Gibbs artifacts caused by limited acquisition resolution.

[0063] In some embodiments, the first artifact removal network can be trained using magnetic resonance images with Gibbs artifacts generated by limited acquisition resolution.

[0064] In some embodiments, since the first artifact removal network is trained only on magnetic resonance images with Gibbs artifacts generated by limited acquisition resolution, when the magnetic resonance image to be processed is input to the first artifact removal network, the first artifact removal network only removes the Gibbs artifacts generated by the limited acquisition resolution in the magnetic resonance image to be processed, thus obtaining the first artifact-removed image. That is, the first artifact-removed image still contains some Gibbs artifacts generated by K-space acquisition.

[0065] Step 23: Input the first artifact removal image into the second artifact removal network to remove some of the Gibbs artifacts generated by K-space acquisition, and obtain the second artifact removal image. Use the second artifact removal image as the final magnetic resonance image.

[0066] In some embodiments, the second artifact removal network is used to remove Gibbs artifacts generated by partial K-space acquisition.

[0067] Because the first artifact-removed image still contains some Gibbs artifacts generated by K-space acquisition, inputting the first artifact-removed image into the second artifact-removed network will remove some of the Gibbs artifacts generated by K-space acquisition from the first artifact-removed image, resulting in the second artifact-removed image. At this point, the second artifact-removed image no longer contains Gibbs artifacts generated by finite acquisition resolution or some Gibbs artifacts generated by K-space acquisition, and can therefore be used as the final magnetic resonance imaging (MRI) image.

[0068] In this embodiment, the first artifact removal network and the second artifact removal network are used to selectively remove Gibbs artifacts caused by the limited acquisition resolution and some Gibbs artifacts caused by K-space acquisition in the magnetic resonance image to be processed, so as to obtain the final magnetic resonance image and improve the quality of the magnetic resonance image.

[0069] See Figure 3 , Figure 3 This is a flowchart illustrating a third embodiment of the magnetic resonance image processing method provided in this application. The processing method includes:

[0070] Step 31: Obtain the magnetic resonance image to be processed.

[0071] In some embodiments, the magnetic resonance image to be processed typically has Gibbs artifacts resulting from limited acquisition resolution and Gibbs artifacts resulting from partial K-space acquisition.

[0072] Step 32: Input the magnetic resonance image to be processed into the first artifact removal network to remove some of the Gibbs artifacts generated by K-space acquisition, and obtain the first artifact-removed image.

[0073] In some embodiments, the first artifact removal network is used to remove Gibbs artifacts generated by partial K-space acquisition.

[0074] In some embodiments, the first artifact removal network can be trained using magnetic resonance images with Gibbs artifacts generated by partial K-space acquisition.

[0075] In some embodiments, since the first artifact removal network is trained only on magnetic resonance images with Gibbs artifacts generated by partial K-space acquisition, when the magnetic resonance image to be processed is input to the first artifact removal network, the first artifact removal network only removes the Gibbs artifacts generated by partial K-space acquisition in the magnetic resonance image to be processed, resulting in the first artifact-removed image. That is, the first artifact-removed image still contains Gibbs artifacts generated by the limited acquisition resolution.

[0076] Step 33: Input the first artifact removal image into the second artifact removal network to remove the Gibbs artifacts caused by the limited acquisition resolution, and obtain the second artifact removal image. Use the second artifact removal image as the final magnetic resonance image.

[0077] In some embodiments, the second artifact removal network is used to remove Gibbs artifacts caused by limited acquisition resolution.

[0078] Because the first artifact-removed image still contains Gibbs artifacts caused by limited acquisition resolution, inputting the first artifact-removed image into the second artifact-removed network will remove the Gibbs artifacts caused by limited acquisition resolution from the first artifact-removed image, resulting in the second artifact-removed image. At this point, the second artifact-removed image no longer contains Gibbs artifacts caused by limited acquisition resolution or some Gibbs artifacts caused by K-space acquisition, and can therefore be used as the final magnetic resonance imaging (MRI) image.

[0079] In this embodiment, the first artifact removal network and the second artifact removal network are used to selectively remove Gibbs artifacts generated by some K-space acquisition and Gibbs artifacts generated by limited acquisition resolution in the magnetic resonance image to be processed, so as to obtain the final magnetic resonance image and improve the quality of the magnetic resonance image.

[0080] See Figure 4 , Figure 4 This is a flowchart illustrating the fourth embodiment of the magnetic resonance image processing method provided in this application. The first artifact removal network and the second artifact removal network are trained in the following ways:

[0081] Step 41: Obtain the first training image set and the second training image set.

[0082] In some embodiments, the artifact generation factors corresponding to the training images in the first training image set are different from those corresponding to the training images in the second training image set.

[0083] For example, the training images in the first training image set are magnetic resonance images with Gibbs artifacts caused by finite acquisition resolution. The training images in the second training image set are magnetic resonance images with Gibbs artifacts caused by partial K-space acquisition.

[0084] In some embodiments, the artifact generation factor corresponding to the training images in the first training image set is the finite acquisition resolution, see [reference]. Figure 5 Obtaining the first training image set can be done through the following process:

[0085] Step 51: Obtain the original magnetic resonance image; the original magnetic resonance image is free of Gibbs artifacts.

[0086] Since magnetic resonance images without Gibbs artifacts are difficult to obtain directly through magnetic resonance imaging equipment, the training images will not have corresponding real labels.

[0087] Based on this, this application utilizes raw magnetic resonance images without Gibbs artifacts for simulation, simulating magnetic resonance images with Gibbs artifacts at limited acquisition resolution and magnetic resonance images with partial K-space Gibbs artifacts to construct a first training image set and a second training image set. In this way, the raw magnetic resonance images can serve as the ground truth labels corresponding to the training images, facilitating the training of the first and second artifact removal networks.

[0088] In some embodiments, after acquiring the original magnetic resonance image, the image is preprocessed using traditional filters such as mean filtering, median filtering, or Gaussian filtering to effectively remove noise and artifacts. If the original magnetic resonance image has high quality and noise and artifacts are almost invisible, the preprocessing step can be omitted. The result of this step serves as the ground truth label (true value or true information) in the dataset.

[0089] Step 52: Convert the original magnetic resonance image to K space, and move the low-frequency part of K space to the center region of the image to obtain the first K space image.

[0090] In some embodiments, the original magnetic resonance image or the preprocessed original magnetic resonance image is converted to K space by fast Fourier transform, and the low-frequency part of K space is moved to the central region of the image by center transform to obtain a first K space image.

[0091] Step 53: Crop the first K-space image in a centrally symmetric manner to obtain the first cropped image.

[0092] In some embodiments, the first K-space image is cropped along a first direction and / or a second direction of the first K-space image in a centrally symmetrical manner according to a cropping ratio to obtain a first cropped image.

[0093] The first direction is the X-axis, such as the length direction of the image. The second direction is the Y-axis, such as the width direction of the image.

[0094] In some embodiments, the first K-space image can be cropped multiple times according to different cropping ratios to obtain multiple first cropped images.

[0095] Step 54: Convert the first cropped image to the space of the original magnetic resonance image to obtain the first training image, and then form the first training image set.

[0096] Furthermore, combined with Figure 6 Explanation:

[0097] like Figure 6As shown, the original image (raw magnetic resonance image) is first preprocessed using traditional filters such as mean filtering, median filtering, or Gaussian filtering to effectively remove noise and artifacts, resulting in a preprocessed image. If the original image quality is high and noise and artifacts are almost invisible, the preprocessing step can be omitted. The result of this step serves as the ground truth label (true value or true information) in the dataset.

[0098] The preprocessed image is first transformed to the K space by Fast Fourier Transform (FFT), and then the low-frequency part of the K space is moved to the center region of the image by center transformation.

[0099] Within a data range for the truncation ratio (which is determined empirically, such as within the interval [0.15, 0.3]), a truncation ratio is randomly selected, and a truncation direction (along the X-axis or along the Y-axis) is also randomly selected. The larger the truncation ratio, the heavier the Gibbs artifacts in the simulation. For example, the first K-space image is truncated in a centrally symmetric manner to obtain the first truncated image.

[0100] The K-space data is processed along the truncation direction. Based on the truncation ratio, the high-frequency information at both ends of the truncation direction is set to 0, retaining only the low-frequency information at the center of the image. Figure 6 As shown, the black areas of K-space after being truncated along the X-axis and K-space after being truncated along the Y-axis are both 0.

[0101] The truncated K-space is then subjected to inverse center transform and inverse Fast Fourier transform (IFFT) to obtain the simulated Gibbs artifact image (the first training image), as shown below. Figure 6 Each of the X and Y axes corresponds to a first training image. By performing the above processing on multiple original magnetic resonance images, or by changing the truncation ratio and performing the above processing on each original magnetic resonance image, a first training image set can be formed.

[0102] In some embodiments, the artifact generation factors corresponding to the training images in the second training image set are partial K-space acquisitions, see [reference]. Figure 7 The process of obtaining the second training image set can be as follows:

[0103] Step 71: Obtain the original magnetic resonance image; the original magnetic resonance image is free of Gibbs artifacts.

[0104] Since magnetic resonance images without Gibbs artifacts are difficult to obtain directly through magnetic resonance imaging equipment, the training images will not have corresponding real labels.

[0105] Based on this, this application utilizes raw magnetic resonance images without Gibbs artifacts for simulation, simulating magnetic resonance images with Gibbs artifacts at limited acquisition resolution and magnetic resonance images with partial K-space Gibbs artifacts to construct a first training image set and a second training image set. In this way, the raw magnetic resonance images can serve as the ground truth labels corresponding to the training images, facilitating the training of the first and second artifact removal networks.

[0106] In some embodiments, after acquiring the original magnetic resonance image, the image is preprocessed using traditional filters such as mean filtering, median filtering, or Gaussian filtering to effectively remove noise and artifacts. If the original magnetic resonance image has high quality and noise and artifacts are almost invisible, the preprocessing step can be omitted. The result of this step serves as the ground truth label (true value or true information) in the dataset.

[0107] Step 72: Convert the original magnetic resonance image to K space, and move the low-frequency part of K space to the center region of the image to obtain the second K space image.

[0108] In some embodiments, the original magnetic resonance image or the preprocessed original magnetic resonance image is converted to K space by fast Fourier transform, and the low-frequency part of K space is moved to the central region of the image by center transform to obtain a second K space image.

[0109] Step 73: Crop the second K-space image along the first direction and / or the second direction to obtain the second cropped image.

[0110] In some embodiments, the second K-space image can be cropped along the first direction and / or the second direction of the second K-space image according to different cropping ratios to obtain a plurality of second cropped images.

[0111] Step 74: Convert the second cropped image to the space of the original magnetic resonance image to obtain the second training image, thereby forming the second training image set.

[0112] Furthermore, combined with Figure 8 Explanation:

[0113] The simulation of some K-space Gibbs artifact images is as follows:

[0114] like Figure 8 As shown, the original image (raw magnetic resonance image) is first preprocessed using traditional filters such as mean filtering, median filtering, or Gaussian filtering to effectively remove noise and artifacts, resulting in a preprocessed image. If the original image quality is high and noise and artifacts are almost invisible, the preprocessing step can be omitted. The result of this step serves as the ground truth label (true value or true information) in the dataset.

[0115] The preprocessed image is first transformed to the K space by Fast Fourier Transform (FFT), and then the low-frequency part of the K space is moved to the center region of the image by center transformation.

[0116] Within a data range for the truncation ratio (which is determined empirically, such as within the interval [0.15, 0.3]), a truncation ratio is randomly selected, and a truncation direction (along the X-axis or along the Y-axis) is also randomly selected. The larger the truncation ratio, the more pronounced the Gibbs artifacts in the simulation.

[0117] The K-space data is processed along the truncation direction, and the high-frequency information corresponding to the end of the truncation direction is set to 0 according to the truncation ratio, such as... Figure 8 As shown, the black areas of K-space after being truncated along the X-axis and K-space after being truncated along the Y-axis are both 0.

[0118] The truncated K-space is then subjected to inverse center transform and inverse Fast Fourier transform (IFFT) to obtain the simulated Gibbs artifact image (the second training image), as shown below. Figure 8 Each of the X and Y axes corresponds to a second training image. By performing the above processing on multiple original magnetic resonance images, or by changing the truncation ratio and performing the above processing on each original magnetic resonance image, a second training image set can be formed.

[0119] Step 42: Train the first artifact removal network using training images from the first training image set.

[0120] In some embodiments, the training images in the first training image set correspond to real labels (real information) formed based on the original magnetic resonance images.

[0121] The first artifact removal network can be trained using training images from the first training image set to remove Gibbs artifacts caused by the limited acquisition resolution of the training images, thus obtaining the first target artifact-removed image. Then, loss calculation is performed using the information in the first target artifact-removed image and the real label. The calculated loss value is used to determine whether to adjust the network parameters of the first artifact removal network, and training ends when the loss value meets the requirements.

[0122] Step 43: Train the second artifact removal network using training images from the second training image set.

[0123] In some embodiments, the training images in the second training image set correspond to real labels (real information) formed based on the original magnetic resonance images.

[0124] The training images in the second training image set can be used to train the second artifact removal network, so that the second artifact removal network removes some of the Gibbs artifacts generated by K-space acquisition in the training images, and obtains the second target artifact removal image. Then, the information in the second target artifact removal image and the real label is used to calculate the loss. The calculated loss value is used to determine whether to adjust the network parameters of the second artifact removal network, and the training ends when the loss value meets the requirements.

[0125] In some embodiments, both the first artifact removal network and the second artifact removal network are trained using the following loss function: Among them, O i GT represents the i-th pixel in the image after Gibbs artifact removal. i represents the i-th pixel in the reference image (original magnetic resonance image); N represents the total number of pixels in the image.

[0126] In some embodiments, the first and second artifact removal networks can employ any deep network architecture, such as UNet, DenseNet, ResNet, etc. Furthermore, the first and second artifact removal networks can both employ the same network architecture, or they can each employ different network architectures.

[0127] In some embodiments, the first artifact removal network can be trained using training images from the second training image set, and the second artifact removal network can be trained using training images from the first training image set. In this application, it is only necessary to ensure that the trained first and second artifact removal networks can remove Gibbs artifacts from different factors.

[0128] In one application scenario, after obtaining the first training image set and the second training image set, the first training image set and the second training image set are divided into training data and test data respectively according to a 9 / 1 ratio for training.

[0129] The structures of the first artifact removal network and the second artifact removal network can be the same or different. For example, both the first artifact removal network and the second artifact removal network adopt the DnCNN network structure. The DnCNN network consists of a series of convolutional layers connected in series, totaling 20 convolutional layers. Layers 2-19 consist of convolution, activation, and normalization operations; the first convolutional layer consists of convolution and activation operations; and the last convolutional layer only has convolution operations. The convolutional layers use 3x3 kernels with zero-padding boundaries of size 1; the normalization layer uses ReLU operations; and the activation layer uses batch normalization (BatchNorm). The first and second artifact removal networks are trained using the PyTorch deep learning framework. During training, the image is randomly cropped into patches with a height and width of 128, and data augmentation methods are used, including random rotation (0, 90, 180, 270 degrees) and flipping. The training batch size was set to 32, the Adam optimizer was used for optimization, the epoch was set to 200, the initial learning rate was set to 1e-3, and the learning rate was gradually reduced using cosine annealing so that the learning rate change curve resembled a cosine function.

[0130] Furthermore, both the first and second artifact removal networks are trained using the following loss function:

[0131]

[0132] Where O is the image after removing Gibbs artifacts, GT is the reference image, N represents the total number of pixels in the image, and i represents the pixel at the current position.

[0133] The trained first and second artifact removal networks can effectively learn features related to Gibbs artifacts, thereby enabling the first and second artifact removal networks to accurately remove artifacts with limited acquisition resolution and some K-space acquisition artifacts.

[0134] Furthermore, the sufficiency of the sample size is one of the key factors affecting the performance of the network model when constructing the first and second artifact removal networks. The training data for the first and second artifact removal networks in this application contains 9829 images, of which 8953 are the training set and 876 are the test set. This dataset covers the head, abdomen, cervical spine, and lumbar spine. All data were acquired by scanning on corresponding magnetic resonance imaging (MRI) devices. After preprocessing, the acquired images were used for training, testing, and performance analysis of the first and second artifact removal networks.

[0135] like Figure 9 The results of different methods for removing Gibbs artifacts in images with limited acquisition resolution are shown in the figure. Figure 9 The first column is the original image. Figure 9The second column shows the results of the traditional Fermi filtering algorithm. Figure 9 The third column shows the artifact removal results (removal of artifacts at limited acquisition resolution) of the proposed method. The images clearly demonstrate that the proposed method efficiently removes Gibbs artifacts while ensuring that image details are not significantly blurred or lost. Figure 9 The image below each column is a magnified image of the corresponding boxed area above it.

[0136] like Figure 10 This demonstrates the results of different methods for removing Gibbs artifacts in partially K-space artifact images. Figure 10 The first column is the original image. Figure 10 The second column shows the results of the traditional Fermi filtering algorithm. Figure 10 The third column shows the artifact removal (partial K-space artifact removal) results of the scheme in this application. From Figure 10 As can be seen from the images in the third column, the method proposed in this application can efficiently remove some Gibbs artifacts caused by K-space while preserving more image details without significant blurring or loss. Figure 10 The image below each column is a magnified image of the corresponding boxed area above it.

[0137] like Figure 11 The application also demonstrates the combined processing of the first and second artifact removal networks to handle partial K-space artifacts and finite acquisition resolution artifact images, such as... Figure 11 The first column of the image contains some K-space Gibbs artifacts and finite acquisition resolution Gibbs artifacts, such as... Figure 11 The second column of images shows the first artifact-removed image after artifact removal using the first artifact removal network, such as... Figure 11 The image in the third column is the second artifact-removed image after artifact removal using the second artifact removal network. For example... Figure 11 As shown, the combination of the first and second artifact removal networks in this application can effectively remove both types of Gibbs artifacts. Among them, Figure 11 The image below each column is a magnified image of the corresponding boxed area above it.

[0138] See Figure 12 , Figure 12 This is a schematic diagram of an embodiment of the magnetic resonance image processing apparatus provided in this application. The processing apparatus 120 includes a processor 121 and a memory 122 coupled to the processor 121; wherein the memory 122 is used to store a computer program, and the processor 121 is used to execute the computer program to implement the following method:

[0139] A magnetic resonance image to be processed is acquired; the magnetic resonance image to be processed is input into a first artifact removal network to obtain a first artifact removal image; the first artifact removal image is input into a second artifact removal network to obtain a second artifact removal image, and the second artifact removal image is used as the final magnetic resonance image; wherein, the first artifact removal network and the second artifact removal network are used to remove Gibbs artifacts caused by different factors.

[0140] In some embodiments, a first artifact removal network is used to remove Gibbs artifacts generated by limited acquisition resolution; a second artifact removal network is used to remove Gibbs artifacts generated by partial K-space acquisition; the processor 121 is further used to execute a computer program to implement the following method: inputting the magnetic resonance image to be processed into the first artifact removal network to remove Gibbs artifacts generated by limited acquisition resolution, obtaining a first artifact-removed image; inputting the first artifact-removed image into the second artifact removal network to remove Gibbs artifacts generated by partial K-space acquisition, obtaining a second artifact-removed image.

[0141] In some embodiments, a first artifact removal network is used to remove Gibbs artifacts generated by partial K-space acquisition; a second artifact removal network is used to remove Gibbs artifacts generated by limited acquisition resolution; the processor 121 is further used to execute a computer program to implement the following method: inputting the magnetic resonance image to be processed into the first artifact removal network to remove Gibbs artifacts generated by partial K-space acquisition, obtaining a first artifact-removed image; inputting the first artifact-removed image into the second artifact removal network to remove Gibbs artifacts generated by limited acquisition resolution, obtaining a second artifact-removed image.

[0142] In some embodiments, the processor 121 is further configured to execute a computer program to implement the following methods: acquiring a first training image set and acquiring a second training image set; wherein the artifact generation factors corresponding to the training images in the first training image set are different from those corresponding to the training images in the second training image set; training a first artifact removal network using the training images in the first training image set; and training a second artifact removal network using the training images in the second training image set.

[0143] In some embodiments, the artifact generation factor corresponding to the training images in the first training image set is a finite acquisition resolution. The processor 121 is further configured to execute a computer program to implement the following method: acquiring an original magnetic resonance image; wherein the original magnetic resonance image is free of Gibbs artifacts; converting the original magnetic resonance image to K-space and moving the low-frequency part of K-space to the central region of the image to obtain a first K-space image; cropping the first K-space image in a centrally symmetric manner to obtain a first cropped image; converting the first cropped image to the space of the original magnetic resonance image to obtain a first training image, thereby forming a first training image set.

[0144] In some embodiments, the processor 121 is further configured to execute a computer program to implement the following method: cropping the first K-space image in a centrally symmetric manner along a first direction and / or a second direction of the first K-space image according to a cropping ratio to obtain a first cropped image.

[0145] In some embodiments, the artifact generation factors corresponding to the training images in the second training image set are partially acquired in K-space. In some embodiments, the processor 121 is further configured to execute a computer program to implement the following method: acquiring an original magnetic resonance image; wherein the original magnetic resonance image is free of Gibbs artifacts; converting the original magnetic resonance image to K-space and moving the low-frequency part of K-space to the central region of the image to obtain a second K-space image; cropping the second K-space image along a first direction and / or a second direction to obtain a second cropped image; converting the second cropped image to the space of the original magnetic resonance image to obtain a second training image, thereby forming a second training image set.

[0146] In some embodiments, both the first artifact removal network and the second artifact removal network are trained using the following loss function: Among them, O i GT represents the i-th pixel in the image after Gibbs artifact removal. i represents the i-th pixel in the original magnetic resonance image; N represents the total number of pixels in the image.

[0147] In some embodiments, the processor 121 is also configured to execute a computer program to implement the method of any embodiment of this application.

[0148] See Figure 13 , Figure 13 This is a schematic diagram of an embodiment of the computer-readable storage medium provided in this application. The computer-readable storage medium 130 is used to store a computer program 131, which, when executed by a processor, implements the following method:

[0149] A magnetic resonance image to be processed is acquired; the magnetic resonance image to be processed is input into a first artifact removal network to obtain a first artifact removal image; the first artifact removal image is input into a second artifact removal network to obtain a second artifact removal image, and the second artifact removal image is used as the final magnetic resonance image; wherein, the first artifact removal network and the second artifact removal network are used to remove Gibbs artifacts caused by different factors.

[0150] In some embodiments, the first artifact removal network is used to remove Gibbs artifacts generated by limited acquisition resolution; the second artifact removal network is used to remove Gibbs artifacts generated by partial K-space acquisition; when executed by the processor, the computer program 131 is also used to implement the following method: inputting the magnetic resonance image to be processed into the first artifact removal network to remove Gibbs artifacts generated by limited acquisition resolution, obtaining a first artifact-removed image; inputting the first artifact-removed image into the second artifact removal network to remove Gibbs artifacts generated by partial K-space acquisition, obtaining a second artifact-removed image.

[0151] In some embodiments, a first artifact removal network is used to remove Gibbs artifacts generated by partial K-space acquisition; a second artifact removal network is used to remove Gibbs artifacts generated by limited acquisition resolution; when executed by a processor, computer program 131 is further used to implement the following method: inputting the magnetic resonance image to be processed into the first artifact removal network to remove Gibbs artifacts generated by partial K-space acquisition, obtaining a first artifact-removed image; inputting the first artifact-removed image into the second artifact removal network to remove Gibbs artifacts generated by limited acquisition resolution, obtaining a second artifact-removed image.

[0152] In some embodiments, when the computer program 131 is executed by a processor, it is further configured to implement the following methods: acquiring a first training image set and acquiring a second training image set; wherein the artifact generation factors corresponding to the training images in the first training image set are different from those corresponding to the training images in the second training image set; training a first artifact removal network using the training images in the first training image set; and training a second artifact removal network using the training images in the second training image set.

[0153] In some embodiments, the artifact generation factor corresponding to the training images in the first training image set is finite acquisition resolution. When the computer program 131 is executed by the processor, it is further configured to implement the following method: acquiring the original magnetic resonance image; wherein the original magnetic resonance image is free of Gibbs artifacts; converting the original magnetic resonance image to K space and moving the low-frequency part of K space to the central region of the image to obtain a first K space image; cropping the first K space image in a centrally symmetric manner to obtain a first cropped image; converting the first cropped image to the space of the original magnetic resonance image to obtain a first training image, thereby forming a first training image set.

[0154] In some embodiments, when the computer program 131 is executed by a processor, it is also used to implement the following method: to crop the first K-space image in a centrally symmetric manner along a first direction and / or a second direction of the first K-space image according to a cropping ratio, thereby obtaining a first cropped image.

[0155] In some embodiments, the artifact generation factors corresponding to the training images in the second training image set are partially acquired in K-space. In some embodiments, when the computer program 131 is executed by the processor, it is also used to implement the following method: acquiring the original magnetic resonance image; wherein the original magnetic resonance image has no Gibbs artifacts; converting the original magnetic resonance image to K-space and moving the low-frequency part of K-space to the central region of the image to obtain a second K-space image; cropping the second K-space image along a first direction and / or a second direction to obtain a second cropped image; converting the second cropped image to the space of the original magnetic resonance image to obtain a second training image, thereby forming a second training image set.

[0156] In some embodiments, both the first artifact removal network and the second artifact removal network are trained using the following loss function: Among them, O i GT represents the i-th pixel in the image after Gibbs artifact removal. i represents the i-th pixel in the original magnetic resonance image; N represents the total number of pixels in the image.

[0157] In some embodiments, when executed by a processor, computer program 131 is also used to implement the method of any embodiment of this application.

[0158] In some embodiments, the generation of Gibbs artifacts in images is mainly due to the following two reasons: (1) Limited acquisition of K-space data. MRI imaging fits signals using limited K-space data, resulting in missing high-frequency data (filled with zeros). After inverse Fourier transform, Gibbs artifacts appear at signal jumps in the image due to the lack of high-frequency components. This type of limited acquisition resolution artifact is as follows: Figure 14 As shown. (2) K-space asymmetric partial sampling: When undersampling imaging is used in certain modes, Gibbs artifacts will be generated during image reconstruction due to the lack of signals of certain frequencies. These partial K-space artifacts are as follows. Figure 15 As shown. The technical solution of this application fully considers the generation mechanism of Gibbs artifacts, specifically introduces data features into model training, and designs end-to-end deep convolutional networks. The models can accurately capture features related to Gibbs artifacts. Concatenating these two models can effectively remove artifacts caused by partial K-space acquisition and limited acquisition resolution. That is, this application can effectively remove Gibbs artifacts in MRI images while preserving image details and avoiding image blurring.

[0159] In summary, the magnetic resonance image processing method, apparatus, and computer-readable storage medium provided in this application utilize a first artifact removal network and a second artifact removal network to selectively remove Gibbs artifacts caused by different factors in the magnetic resonance image to be processed, thereby obtaining the final magnetic resonance image and improving the quality of the magnetic resonance image.

[0160] Furthermore, the magnetic resonance image processing method, apparatus, and computer-readable storage medium provided in this application can solve Gibbs artifacts caused by the loss of high-frequency components in MRI. Traditional filtering algorithms reduce ringing near image edges by eliminating high-frequency components, but such methods blur image details and rely on the personal experience of clinicians to adjust relevant parameters. Deep learning-based algorithms can learn the mapping relationship between Gibbs artifact images and clear images, but the performance of such supervised end-to-end algorithms is highly dependent on the dataset used for training, and the training data lacks an understanding of the artifact generation mechanism, making it difficult to handle artifacts arising from different mechanisms. To address the shortcomings of related technologies, this application starts from the causes of Gibbs artifact generation and, based on the two artifact generation principles of limited acquisition resolution and partial K-space, specifically trains end-to-end deep networks (a first artifact removal network and a second artifact removal network), enabling the deep model (the first artifact removal network and the second artifact removal network) to learn more Gibbs artifact-related features and remove artifacts in a targeted manner.

[0161] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0162] If the integrated units in the other embodiments described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processing circuit component (processor) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0163] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for processing magnetic resonance images, characterized in that, The processing method includes: Acquire the magnetic resonance image to be processed; The magnetic resonance image to be processed is input into the first artifact removal network to obtain the first artifact-removed image; The first artifact-removed image is input into the second artifact-removed network to obtain the second artifact-removed image, and the second artifact-removed image is used as the final magnetic resonance image; wherein, the first artifact-removed network and the second artifact-removed network are respectively used to remove Gibbs artifacts caused by different factors.

2. The method according to claim 1, characterized in that, The first artifact removal network is used to remove Gibbs artifacts caused by limited acquisition resolution; the second artifact removal network is used to remove Gibbs artifacts caused by partial K-space acquisition. The step of inputting the magnetic resonance image to be processed into the first artifact removal network to obtain the first artifact-removed image includes: The magnetic resonance image to be processed is input into the first artifact removal network to remove the Gibbs artifacts caused by the limited acquisition resolution, and the first artifact-removed image is obtained. The first artifact-removed image is input into the second artifact-removing network to remove some of the Gibbs artifacts generated by K-space acquisition, thus obtaining the second artifact-removed image.

3. The method according to claim 1, characterized in that, The first artifact removal network is used to remove Gibbs artifacts generated by some K-space acquisition; the second artifact removal network is used to remove Gibbs artifacts generated by limited acquisition resolution. The step of inputting the magnetic resonance image to be processed into the first artifact removal network to obtain the first artifact-removed image includes: The magnetic resonance image to be processed is input into the first artifact removal network to remove some of the Gibbs artifacts generated by K-space acquisition, thus obtaining the first artifact-removed image; The first artifact-removed image is input into the second artifact-removing network to remove Gibbs artifacts caused by the limited acquisition resolution, thus obtaining the second artifact-removed image.

4. The method according to claim 1, characterized in that, The first artifact removal network and the second artifact removal network are trained in the following ways: Obtain a first training image set and obtain a second training image set; wherein the artifact generation factors corresponding to the training images in the first training image set are different from those corresponding to the training images in the second training image set. The first artifact removal network is trained using training images from the first training image set. The second artifact removal network is trained using training images from the second training image set.

5. The method according to claim 4, characterized in that, The artifact generation factors corresponding to the training images in the first training image set are limited acquisition resolution. The acquisition of the first training image set includes: Obtain the original magnetic resonance image; wherein the original magnetic resonance image is free of Gibbs artifacts; The original magnetic resonance image is converted to K-space, and the low-frequency part of the K-space is moved to the central region of the image to obtain a first K-space image; The first K-space image is cropped in a centrally symmetric manner to obtain the first cropped image; The first cropped image is converted to the space of the original magnetic resonance image to obtain the first training image, thereby forming the first training image set.

6. The method according to claim 5, characterized in that, The step of cropping the first K-space image in a centrally symmetric manner to obtain the first cropped image includes: The first K-space image is cropped along the first direction and / or the second direction of the first K-space image in a centrally symmetrical manner according to the cropping ratio to obtain the first cropped image.

7. The method according to claim 4, characterized in that, The artifacts in the training images of the second training image set are generated by partial K-space acquisition. The acquisition of the second training image set includes: Obtain the original magnetic resonance image; wherein the original magnetic resonance image is free of Gibbs artifacts; The original magnetic resonance image is converted to K-space, and the low-frequency part of the K-space is moved to the central region of the image to obtain a second K-space image; The second K-space image is cropped along a first direction and / or a second direction to obtain a second cropped image; The second cropped image is converted to the space of the original magnetic resonance image to obtain the second training image, thereby forming the second training image set.

8. The method according to claim 4, characterized in that, Both the first and second artifact removal networks are trained using the following loss function: Among them, O i GT represents the i-th pixel in the image after Gibbs artifact removal. i represents the i-th pixel in the original magnetic resonance image; N represents the total number of pixels in the image.

9. A magnetic resonance image processing apparatus, characterized in that, The processing device includes a processor and a memory coupled to the processor; The memory is used to store a computer program, and the processor is used to execute the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program, which, when executed by a processor, is used to implement the method as described in any one of claims 1-8.

Images