Cardiac magnetic resonance imaging artifact removal method and system based on phase circulation and deep learning

By using double phase cycle sequences and deep learning models in cardiac magnetic resonance imaging, combining reconstruction and perceived loss functions, the problems of low scanning efficiency and non-rosy artifact suppression in the prior art are solved, and efficient removal of cardiac magnetism is achieved Black belt and blood flow artifacts in resonant images.

CN119090991APending Publication Date: 2024-12-06SHANGHAI JIAOTONG UNIV

Patent Information

Application Number
CN202411256838.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In the prior art In cardiac magnetic resonance imaging, phase cycling methods lead to increased scanning time and low efficiency, and deep learning methods rely too much on training data, which is prone to illusions and is not robust enough to suppress artifacts.

Method used

Two images with different phase increments are collected in the k-space part by double phase cyclic sequence. Combined with the deep learning model, the model is trained by reconstructing the loss function of loss and perceived loss to achieve image recovery and artifact suppression.

Benefits of technology

The scanning time is shortened, the scanning efficiency is improved, the robustness of artifact suppression is enhanced, and the artifact removal effect of high-quality cardiac magnetic resonance images is achieved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090991A_ABST
    Figure CN119090991A_ABST
Patent Text Reader

Abstract

The invention provides a cardiac magnetic resonance imaging artifact removal method and system based on phase circulation and deep learning. Black band artifacts and blood flow artifacts in cardiac magnetic resonance film imaging can be inhibited at the same time. The phase cycle sequence collects a plurality of images with complementary artifact information, and scanning is accelerated by using an acceleration means, so that the plurality of images are completed in the same breath holding; the collected phase cycle image is recovered based on deep learning, on one hand, undersampling artifacts caused by an accelerated collection means are recovered, and on the other hand, artifacts of the image are restrained through complementary blood flow and black belt artifact information. Based on the sequence design and image reconstruction recovery, cardiac magnetic resonance imaging artifact removal combining phase circulation and deep learning is realized, and the imaging time is shortened.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning. Background Art

[0002] Cardiac magnetic resonance imaging is a powerful imaging modality in the diagnosis of cardiovascular diseases. In the past two decades, many important technologies have emerged in the research and clinical application of cardiac magnetic resonance. At present, in addition to being used for the diagnosis of cardiac function and structure, cardiac magnetic resonance imaging can also provide in-depth phenotypic analysis of myocardial tissue components for the evaluation of myocardial fibrosis, edema, perfusion and microstructure. Cardiac magnetic resonance imaging uses the principle of nuclear magnetic resonance to image the cardiac anatomical structure and cardiac function, and is a safe and radiation-free technology.

[0003] Cardiac cine imaging records the contraction and relaxation of the heart and performs dynamic imaging of the entire cardiac cycle. It has been widely used in the analysis of cardiac structure and function. At present, the balanced steady state free precession sequence (bSSFP) is one of the most commonly used sequences for clinical cardiac cine imaging. It has fast imaging speed, high signal-to-noise ratio, and good contrast, so it is widely used in cardiac cine imaging.

[0004] However, the bSSFP sequence is sensitive to the resonance frequency. When a specific location is affected by an uneven magnetic field, the resonance frequency shifts, presenting a low signal intensity, and further forming a black band artifact on the image. Since it is difficult to completely avoid magnetic field inhomogeneity, black band artifacts appear in most cardiac cine images, which brings certain difficulties to the analysis of cardiac structure and function.

[0005] At present, the traditional method for removing black band artifacts is achieved through phase cycle imaging. Phase cycling refers to cyclically changing the phase increment in the bSSFP sequence so that the position of the black band artifact on the image is shifted, thereby suppressing the black band artifact by combining multiple images. Although the phase cycle method can suppress black band artifacts, the scanning time is seriously increased and the scanning efficiency is reduced due to the need to collect multiple images at the same time; the phase cycle method requires changing the phase increment, and new blood flow artifacts are easily introduced into the collected images, resulting in serious blood flow artifacts in the combined image. Therefore, the effect of removing black band artifacts by the phase cycle method alone has application limitations.

[0006] Patent document CN116309524A discloses a cardiac magnetic resonance movie imaging suppression method and system based on deep learning, comprising the following steps: S1: collecting cardiac magnetic resonance movie images with black band artifacts or blood flow artifacts, preprocessing the images, and performing image joint calculation to obtain artifact-free images; S2: constructing a deep learning model, the input of the artifact suppression network in the model is the cardiac movie image with black band artifacts or blood flow artifacts, and the output is the cardiac movie image after the black band artifacts or blood flow artifacts are suppressed; S3: using the preprocessed artifact images and artifact-free images to train the deep learning model until the parameters of the artifact suppression network reach the target conditions, obtaining the trained deep learning model and saving it; S4: according to the trained deep learning model, performing artifact elimination on the cardiac movie image with black band artifacts or blood flow artifacts to be eliminated, and obtaining the cardiac movie image after the artifact elimination. The above technology only relies on deep learning post-processing technology, over-relies on training data, and is prone to produce illusions, resulting in artifact suppression that is not robust enough.

[0007] Although deep learning-based methods can suppress black band artifacts and blood flow artifacts, their shortcomings include over-reliance on training data, easy to produce hallucinations, and the suppression method is not robust enough. Therefore, we need to further explore and develop new methods to solve these problems. Summary of the invention

[0008] In view of the defects in the prior art, the object of the present invention is to provide a method and system for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning.

[0009] The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning provided by the present invention comprises:

[0010] Step S1: acquiring cardiac magnetic resonance movie images under multi-phase increments, uniformly preprocessing the images, obtaining training labels by using a phase cycle combination method, dividing the images into a training set and a test set according to a preset ratio, and performing k-space undersampling on the images in the training set and the test set according to a preset method;

[0011] Step S2: Build an image restoration deep learning model, construct a model training loss function, and define a loss function in the form of a combination of reconstruction loss and perception loss; select a model optimizer, train and save the image restoration deep learning model;

[0012] Step S3: using a double phase cycle sequence to acquire partial k-space data for two images with different phase increments in the same preset manner through multiple receiving coils;

[0013] Step S4: Based on the acquired double phase cycle partial k-space data, the image is restored using the trained image restoration deep learning model, including suppressing undersampling artifacts, black band artifacts and blood flow artifacts, and outputting the final artifact-free cardiac movie image.

[0014] Preferably, the following preprocessing operations are performed on the cardiac magnetic resonance movie images, including image cropping, image registration, and image normalization;

[0015] First, crop the image to a uniform size;

[0016] Secondly, for images with the same anatomical structure but different phase increments, a registration operation is performed;

[0017] Finally, the image intensity is normalized to a uniform range.

[0018] Preferably, the process of obtaining the training label is as follows: multiple groups of cardiac movie images with a cardiac movie phase increment of approximately 180° are selected, and a joint image is obtained by averaging to serve as an artifact-free training label.

[0019] Preferably, the process of building the image restoration deep learning model is as follows: the model input is 2 cardiac movie images with different phase increments, and the model output is an artifact-free cardiac movie image;

[0020] The deep learning model includes three stages: feature extraction, feature fusion and image restoration. The feature extraction stage is used to extract feature information of different scales in the input image. The feature fusion stage fuses feature information of different sizes of two images. The image restoration stage completes image restoration based on the output information of the first two stages to obtain an artifact-free cardiac movie image. The deep learning model is trained using the training labels obtained in step S1.

[0021] Preferably, the specific form of the model training loss function is as follows:

[0022] Input two undersampled three-dimensional double-phase cycle cardiac magnetic resonance movie images x1 and x2, the model output is F(x1,x2), the model training label is y, and the expression is:

[0023] L(x)=L recon (F(x1,x2),y)+λL perceptual (F(x1,x2),y)

[0024] Among them, L recon represents the image restoration loss, which is implemented by the two-norm difference between the model output and the corresponding label; L perceptualrepresents the perceptual loss, which is realized by the norm difference between the features output by the model and the corresponding labels output to each stage in the VGG19 network; λ is the weight for weighing the image restoration loss and the perceptual loss.

[0025] Preferably, the specific form of the double phase cycle sequence is as follows: within one breath-hold, two movies are collected, wherein the two movies have different phase increments to provide different complementary structural information.

[0026] Preferably, the manner of acquiring k-space data of two images with different phase increments includes: partially acquiring k-space of one image and fully acquiring k-space of the other image; or partially acquiring k-space of both images.

[0027] Preferably, the method of acquiring partial data in k-space includes: partially acquiring k-space using the same trajectory for two images and reconstructing them into under-acquired images; or partially acquiring k-space using an interlaced k-space trajectory in two images and combining the two images to achieve accelerated scanning.

[0028] The cardiac magnetic resonance imaging artifact removal system based on phase cycle and deep learning provided by the present invention adopts the cardiac magnetic resonance imaging artifact removal method based on phase cycle and deep learning.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) The present invention collects partial k-space data of two phase cycle images in k-space when using a double phase cycle sequence for cardiac cine imaging, and uses fast imaging acceleration technology to accelerate scanning, thereby shortening the breath-holding time, improving scanning comfort, and reducing the risk of introducing motion artifacts;

[0031] (2) The present invention acquires two images with different phase increments in one scan, which can provide different complementary structures and artifact information for subsequent image restoration to achieve robust image artifact suppression;

[0032] (3) The deep learning model proposed in the present invention can realize joint image reconstruction and artifact suppression, which not only realizes the restoration of artifacts caused by partial acquisition of k-space, but also realizes the suppression of black band artifacts and blood flow artifacts of the image;

[0033] (4) The present invention can overcome the respective limitations of relying solely on the phase cycle method and the deep learning method, and combine the two methods together to realize a robust magnetic resonance cardiac movie imaging scheme to improve the quality of magnetic resonance cardiac movie images. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0035] Figure 1 is a flow chart of a method for jointly removing black band artifacts and blood flow artifacts in cardiac magnetic resonance movie imaging based on phase cycling and deep learning according to an embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of a double phase cycle sequence according to an embodiment of the present invention;

[0037] Figure 3 4 is a schematic diagram comparing reconstruction results of a prospective experiment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0039] Example

[0040] The present invention provides a method for removing artifacts from cardiac magnetic resonance imaging based on phase cycling and deep learning. Two images are collected in the k-space part using a double phase cycling sequence. The two images have different phase increments, so as to have complementary structures and artifact information. The collected images are subjected to artifact suppression based on the complementary structures and artifact information using a deep learning method. The present invention combines phase cycling and deep learning schemes, which can not only improve scanning efficiency, but also increase the robustness of deep learning artifact removal.

[0041] Specifically, Figure 1 , the method comprises the following steps:

[0042] Step S1, image acquisition: collect cardiac cine images of multiple healthy volunteers. For each volunteer, collect 12 sets of phase-cycled cardiac cine images by changing the phase increment of the bSSFP sequence. During scanning, the short-axis field of view of the heart is mainly used, and the long-axis field of view can be supplemented.

[0043] Step S2, image preprocessing: In order to make the input images have the same spatial size and intensity range distribution, the same preprocessing operation is performed on the images. First, the image size is cropped to 192×192; then, for images with different phase increments, the image with a phase increment of 180° is used as the reference image, and the remaining 11 groups of images are registered in a rigid registration manner to reduce image mismatches caused by respiratory motion, etc.

[0044] Step S3, retrospective undersampling to obtain low-resolution images: In order to obtain low-resolution images, the image is Fourier transformed to obtain k-space data. In the k-space, in the phase encoding direction, the central k-space data of one-fourth of the original size is retained, and the k-space corresponding to the periphery is filled with zeros. The undersampling k-space is inversely Fourier transformed to obtain a low-resolution phase cycle image.

[0045] Step S4, obtain training labels: In order to avoid cardiac blood flow artifacts caused by excessive bias frequency, select 5 groups of cardiac movie images with phase increments of [120, 150, 180, 210, 240]°, and obtain a joint image by averaging as an artifact-free training label. Here, the traditional phase cycle combination method is not used to reduce the risk of introducing cardiac blood flow artifacts. Since the artifact positions of each group of images are different, combining them together can suppress black band artifacts and does not introduce blood flow artifacts, which can be used as artifact removal labels.

[0046] Step S5, image division: the pre-processed cardiac movie image is divided according to a ratio to obtain training data and verification data; specifically, among the pre-processed 28-bit magnetic resonance data of the subjects, the scan data of the first 18 volunteers are selected as the training set, and the scan data of the last 10 volunteers are selected as the test set. Of course, in other embodiments of the present invention, the division can also be performed according to other ratios, which does not affect the purpose of the present invention.

[0047] Step S6, construct a deep learning model, and the model parameters are randomly initialized.

[0048] The deep learning model is based on the visual attention network, and constructs the feature extraction stage, feature fusion processing stage, and image restoration stage. The feature extraction stage is mainly based on the convolution module, which is used to extract shallow features and process each image independently; the feature fusion processing stage is mainly based on the self-attention module and the mutual attention module. The self-attention module is used for feature extraction and learning of a single image, and the mutual attention module is used for mutual feature extraction and learning of two phase-cycled images; the image restoration stage is mainly based on the convolution module, which is used to further restore the features learned in the first two stages to obtain artifact-suppressed and under-sampled artifact-suppressed images.

[0049] The specific formula of the feature extraction stage is as follows, where Stage1 represents the feature extraction network of the first stage, x1 and x2 represent the two undersampled double-phase cycle cardiac movie images input, and F1 and F2 represent the features extracted from the two images;

[0050] F1=Stage1(x1)

[0051] F2=Stage1(x2)

[0052] The specific formula of the feature fusion processing stage is as follows, where self-attention represents the self-attention module, which is used for feature extraction and learning of a single image, and cross-attention represents the mutual attention mechanism, which further extracts and learns features from two images;

[0053] SE1=Self-attention(x1)

[0054] SW2=Self-attentiom(x2)

[0055] CW=Cross-attention(SW 1, SW2)

[0056] The specific formula of the image restoration stage is as follows, where Stage3 represents the image restoration network of the image restoration stage, Concat represents merging the outputs of the first two stages along the channel dimension as the input of the image restoration stage, and output represents the final output;

[0057] Output=Stage3(Concat(F1,F2,CW))

[0058] Step S7, construct the model loss function:

[0059] Input two undersampled three-dimensional double-phase cycle cardiac magnetic resonance movie images x1 and x2, the model output result is F(x1,x2), and the model training label is y;

[0060] L(x)=L recon (F(x1,x2),y)+λL perceptual (F(x1,x2),y)

[0061] Among them, L recon represents the image restoration loss, which is implemented by the two-norm difference between the model output and the corresponding label; L perceptual represents the perceptual loss, which is realized by the difference between the model output and the corresponding label output to the features output at each stage in the VGG19 network; λ is the weight for weighing the image restoration loss and the perceptual loss. In this embodiment, λ=0.01.

[0062] Step S8, training model: input the under-sampled double-phase cycle cardiac movie image into the deep learning model, use the training labels obtained above, and rely on the loss function defined above for training. When the number of training iterations reaches a certain number, the validation set loss no longer decreases, then the training is stopped and the model parameters are saved. In this embodiment, the Adam optimizer is used, the learning rate is maintained at 0.0001, and the model is iterated 400 times.

[0063] Step S9, using a double phase cycle sequence to collect prospective data: using a double phase cycle sequence, in one breath hold, collect two movies with different phase cycles, such as Figure 2 In this embodiment, the acquisition reduces the resolution in the phase encoding direction by 4 times to achieve accelerated scanning; and two different phase cycle increments are set to 90° and 270°.

[0064] Step S10, obtaining an artifact-removed image: removing black band artifacts and blood flow artifacts in cardiac movie imaging according to the trained joint super-resolution and artifact suppression network, and improving the resolution in the phase encoding direction, such as Figure 3 .

[0065] The present invention also provides a cardiac magnetic resonance imaging artifact removal system based on phase cycle and deep learning, which adopts the cardiac magnetic resonance imaging artifact removal method based on phase cycle and deep learning.

[0066] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0067] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning, characterized in that: include: Step S1: acquiring cardiac magnetic resonance movie images under multi-phase increments, uniformly preprocessing the images, obtaining training labels by using a phase cycle combination method, dividing the images into a training set and a test set according to a preset ratio, and performing k-space undersampling on the images in the training set and the test set according to a preset method; Step S2: Build an image restoration deep learning model, construct a model training loss function, and define a loss function in the form of a combination of reconstruction loss and perception loss; select a model optimizer, train and save the image restoration deep learning model; Step S3: using a double phase cycle sequence to acquire partial k-space data for two images with different phase increments in the same preset manner through multiple receiving coils; Step S4: Based on the acquired double phase cycle partial k-space data, the image is restored using the trained image restoration deep learning model, including suppressing undersampling artifacts, black band artifacts and blood flow artifacts, and outputting the final artifact-free cardiac movie image.

2. The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning according to claim 1, characterized in that: The following preprocessing operations were performed on cardiac magnetic resonance cine images, including image cropping, image registration, and image normalization; First, crop the image to a uniform size; Secondly, for images with the same anatomical structure but different phase increments, a registration operation is performed; Finally, the image intensity is normalized to a uniform range.

3. The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning according to claim 1, characterized in that: The process of obtaining training labels is as follows: multiple groups of cardiac movie images with cardiac movie phase increments near 180° are selected, and a joint image is obtained by averaging as an artifact-free training label.

4. The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning according to claim 1, characterized in that: The process of building the image restoration deep learning model is as follows: the model input is 2 cardiac movie images with different phase increments, and the model output is an artifact-free cardiac movie image; The deep learning model includes three stages: feature extraction, feature fusion and image restoration. The feature extraction stage is used to extract feature information of different scales in the input image. The feature fusion stage fuses feature information of different sizes of two images. The image restoration stage completes image restoration based on the output information of the first two stages to obtain an artifact-free cardiac movie image. The deep learning model is trained using the training labels obtained in step S1.

5. The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning according to claim 1, characterized in that: The specific form of the model training loss function is as follows: Input two undersampled three-dimensional double-phase cycle cardiac magnetic resonance movie images x1 and x2, the model output is F(x1,x2), the model training label is y, and the expression is: L(x)=L recon (F(x1,x2),y)+λL perceptual (F(x1,x2),y) Among them, L recon represents the image restoration loss, which is implemented by the two-norm difference between the model output and the corresponding label; L perceptual represents the perceptual loss, which is realized by the norm difference between the features output by the model and the corresponding labels output to each stage in the VGG19 network; λ is the weight for weighing the image restoration loss and the perceptual loss.

6. The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning according to claim 1, characterized in that: The specific form of the double phase cycle sequence is as follows: within one breath hold, two movies are collected, wherein the two movies have different phase increments to provide different complementary structural information.

7. The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning according to claim 1, characterized in that: Methods of acquiring k-space data of two images with different phase increments include: partially acquiring k-space of one image and fully acquiring k-space of the other image; or partially acquiring k-space of both images.

8. The method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning according to claim 1, characterized in that: Methods of acquiring partial data in k-space include: partially acquiring k-space using the same trajectory for two images and reconstructing them into under-acquired images; or partially acquiring k-space using an interlaced k-space trajectory in two images and combining the two images to achieve accelerated scanning.

9. A cardiac magnetic resonance imaging artifact removal system based on phase cycling and deep learning, characterized in that: A method for removing cardiac magnetic resonance imaging artifacts based on phase cycling and deep learning as described in any one of claims 1 to 8 is adopted.

Citation Information

Patent Citations

  • Cardiac magnetic resonance film imaging artifact suppression method and system based on deep learning

    CN116309524A

Cited By

  • B0 robust heart T1 quantitative imaging method

    CN120938399A

  • Dual phase-cycled cardiac magnetic resonance imaging artifact removal method guided by magnetic susceptibility field map

    CN122929030A