CN108896943B

CN108896943B - Magnetic resonance quantitative imaging method and device

Info

Publication number: CN108896943B
Application number: CN201810444262.5A
Authority: CN
Inventors: 黄峰
Original assignee: Neusoft Medical Systems Co Ltd
Current assignee: Neusoft Medical Systems Co Ltd
Priority date: 2018-05-10
Filing date: 2018-05-10
Publication date: 2020-06-12
Anticipated expiration: 2038-05-10
Also published as: CN108896943A

Abstract

The application discloses a magnetic resonance quantitative imaging method and a magnetic resonance quantitative imaging device. According to the method, the magnetic resonance quantitative value is obtained according to the second deep neural network, because the deep neural network is a data-driven model, accurate description of the real world can be achieved, and therefore compared with the existing imaging mathematical model for calculating the magnetic resonance quantitative value, the method can obtain the accurate magnetic resonance quantitative value according to the deep neural network model. In addition, the deep neural network model runs faster, so that the calculation time of magnetic resonance quantification can be reduced, and the acquisition efficiency is improved. In addition, the input data of the second deep neural network model for obtaining the magnetic resonance quantification is an image obtained by image reconstruction, and the image reconstruction speed is high by the method for reconstructing the image by using the deep neural network. Therefore, the image reconstruction method is also beneficial to improving the magnetic resonance quantitative imaging speed.

Description

Magnetic resonance quantitative imaging method and device

Technical Field

The application relates to the technical field of medical imaging, in particular to a magnetic resonance quantitative imaging method and device.

Background

Magnetic Resonance Imaging (MRI), which is a multi-parameter, multi-contrast Imaging technique, is one of the main Imaging modes in modern medical Imaging, can reflect various characteristics of tissues T1, T2, proton density and the like, and can provide information for detection and diagnosis of diseases.

Conventional magnetic resonance images mainly comprise qualitative images of different contrast properties, such as T₁Weighting, T₂Weighting, proton density weighting, diffusion weighting, susceptibility weighting, and the like. Magnetic resonance images can provide far more than this qualitative information, but they can also provide quantitative magnetic resonance information. The magnetic resonance quantitative information is more important for disease diagnosis, especially in the aspects of brain neuroscience research and clinical application. With the development of magnetic resonance technology, magnetic resonance-based quantitative imaging has been increasingly applied to clinical diagnosis and therapy guidance.

The existing magnetic resonance quantitative imaging is generally based on an imaging mathematical model of a magnetic resonance signal, a plurality of groups of magnetic resonance images are reconstructed, and then a data fitting method is utilized to obtain magnetic resonance quantitative values from the plurality of groups of magnetic resonance images. However, the existing magnetic resonance imaging has the defect of long scanning time, and an imaging mathematical model for calculating magnetic resonance quantification is often a simplified description of the real world, and the influence of many factors is ignored, so that the accuracy of a quantified value calculated based on the mathematical model is low, and in addition, the process of fitting calculation is often nonlinear, so that the calculation time is long.

Disclosure of Invention

In view of the above, the present disclosure provides a magnetic resonance quantitative imaging method to improve the magnetic resonance quantitative imaging speed and the accuracy of the quantitative value.

In order to solve the technical problem, the following technical scheme is adopted in the application:

a method of magnetic resonance quantitative imaging, the method comprising:

acquiring partial k-space data of a plurality of echoes according to a down-sampling mode to obtain k-space acquired data of the plurality of echoes;

respectively and successively adopting an image reconstruction method of a first deep neural network model and an explicit analytic solution imaging method or respectively and successively adopting the explicit analytic solution imaging method and the image reconstruction method of the first deep neural network model to carry out image reconstruction on the k-space acquisition data of each echo to obtain a magnetic resonance image of each echo;

inputting the magnetic resonance image of each echo into a second deep neural network model to obtain the magnetic resonance quantitative image of each echo;

the first deep neural network model is obtained by training a plurality of magnetic resonance images obtained by reconstructing full-acquisition or over-full-acquisition k-space data of a plurality of echoes as output training samples and specific part k-space data of the plurality of echoes or each magnetic resonance image obtained by partially reconstructing the specific part k-space data of the plurality of echoes as input training samples, wherein the specific part k-space data of the plurality of echoes are k-space data with specific proportion respectively selected from the full-acquisition or over-full-acquisition k-space data of each echo;

the second deep neural network model is obtained by training each subset of magnetic resonance quantitative values obtained by calculation according to a plurality of magnetic resonance images obtained by reconstruction of full-acquisition or over-full-acquisition k-space data of a plurality of echoes as an output training sample and each subset of each magnetic resonance image obtained by successive reconstruction of specific partial k-space data of the plurality of echoes by a partial reconstruction method and an image reconstruction method of the first deep neural network model as an input training sample.

An apparatus for magnetic resonance quantitative imaging, the apparatus comprising:

the acquisition unit is used for acquiring partial k-space data of a plurality of echoes according to a down-sampling mode to obtain k-space acquisition data of the plurality of echoes;

the image reconstruction unit is used for respectively and successively adopting an image reconstruction method of the first depth neural network model and an explicit analytic solution imaging method or respectively and successively adopting the explicit analytic solution imaging method and the image reconstruction method of the first depth neural network model to carry out image reconstruction on the k-space acquisition data of each echo so as to obtain a magnetic resonance image of each echo;

the magnetic resonance quantitative imaging unit is used for inputting the magnetic resonance image of each echo into the second deep neural network model and acquiring the magnetic resonance quantitative image of each echo;

the second deep neural network model is obtained by training, as input training samples, subsets of magnetic resonance quantitative values calculated according to a plurality of magnetic resonance images obtained by reconstructing full-acquisition or super-full-acquisition k-space data of a plurality of echoes, and the subsets of images obtained by successively reconstructing specific partial k-space data of the plurality of echoes by a partial reconstruction method and an image reconstruction method of the first deep neural network model.

Compared with the prior art, the method has the following beneficial effects:

based on the above technical solutions, the magnetic resonance quantitative imaging method provided in the embodiment of the present application obtains the magnetic resonance quantitative value according to the second deep neural network, because the deep neural network is a data-driven model, and accurate description of the real world can be achieved, compared with the existing imaging mathematical model for calculating the magnetic resonance quantitative value, the magnetic resonance quantitative value can be obtained according to the deep neural network model. In addition, the deep neural network model runs faster, so that the calculation time of magnetic resonance quantification can be reduced, and the acquisition efficiency is improved. In addition, the input data of the second deep neural network model for obtaining the magnetic resonance quantification is an image obtained by image reconstruction, and the image reconstruction speed is high by the method for reconstructing the image by using the deep neural network. Therefore, the image reconstruction method is also beneficial to improving the magnetic resonance quantitative imaging speed.

Drawings

FIG. 1 is a flowchart of a first deep neural network model training method provided by an embodiment of the present application;

FIG. 2 is a flowchart of a second deep neural network model training method provided by an embodiment of the present application;

FIG. 3 is a flowchart of a magnetic resonance quantitative imaging method according to an embodiment of the present application;

FIG. 4 is a flowchart of a magnetic resonance quantitative imaging method provided in the second embodiment of the present application;

FIG. 5 is a flowchart of a magnetic resonance quantitative imaging method provided in the third embodiment of the present application;

FIG. 6 is a flowchart of a magnetic resonance quantitative imaging method provided in the fourth embodiment of the present application;

fig. 7 is a flowchart of a magnetic resonance quantitative imaging method provided in the fifth embodiment of the present application;

FIG. 8 is a schematic structural diagram of a control apparatus for magnetic resonance quantitative imaging according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a magnetic resonance quantitative imaging apparatus provided in a sixth embodiment of the present application;

fig. 10 is a schematic structural diagram of a magnetic resonance quantitative imaging apparatus provided in a seventh embodiment of the present application;

fig. 11 is a schematic structural diagram of an image correction unit according to a seventh embodiment of the present application;

fig. 12 is a schematic structural diagram of another image correction unit provided in the seventh embodiment of the present application;

fig. 13 is a schematic structural diagram of a magnetic resonance quantitative imaging apparatus according to an eighth embodiment of the present application.

Detailed Description

Before describing the embodiments of the present application, technical terms used in describing the embodiments of the present application will be described.

The explicit analytical solution imaging method is a method of imaging using a display function having an analytical solution. The explicit analytical solution imaging method may include: a parallel imaging method, a k-t BLAST method, and a zero-padding method, and the like. And further, the Parallel imaging method includes two imaging methods, one is a self-calibrated Auto-calibrating Parallel reconstruction method (GRAPPA), and the other is a Sensitivity Encoding method (SENSE).

Deep Neural Networks (DNN) is a method for approximating a function that is difficult to express explicitly by using a multi-layer simple function, and parameters of the multi-layer simple function can be obtained by training a training data set.

The DNN model is a data-driven model, which is not a mathematical model fitted based on some existing data, but is obtained by training the relationship between a plurality of sets of input data and a desired prediction result (i.e., model training), so as to accurately describe the relationship between the input data and the desired prediction result. In the relation, any factor influencing the prediction result is not ignored, so that the DNN model can be regarded as a real description of the real world, and the real world can be more accurately reflected. In the magnetic resonance imaging process, k-space data, an ideal magnetic resonance image and an ideal magnetic resonance quantitative value can be acquired, so that the DNN model can be applied to the magnetic resonance imaging and magnetic resonance quantitative imaging processes.

Based on the background art, at present, magnetic resonance quantitative imaging is usually based on a magnetic resonance quantitative imaging mathematical model, and a data fitting method is used for obtaining magnetic resonance quantitative values from a plurality of groups of reconstructed magnetic resonance images. However, this method has the following drawbacks: firstly, the scan time of the magnetic resonance image is too long; secondly, a mathematical model for calculating the magnetic resonance quantification is usually a simplified description of the real world, and many factors influencing the quantitative value are ignored, so that the magnetic resonance quantification calculation precision is insufficient; third, the process of fitting the calculations tends to be non-linear, resulting in long calculation times. Due to the increasing demands of clinical medicine on the imaging speed and imaging quality of magnetic resonance quantitative imaging, the above defects can negatively affect the further development and application of magnetic resonance quantitative imaging in clinical medicine.

The scan time of the existing magnetic resonance image is too long because the existing magnetic resonance imaging method adopts a single image reconstruction algorithm to realize the conversion from the acquisition reduction k-space data to the image. In this conversion process, there are many problems to be solved, such as: various artifacts, low image resolution, high noise, etc. If these problems are achieved by only one image reconstruction algorithm, it takes a long time. The inventor finds out through research that: the image quality problems resulting from the reconstruction of downsampled data can be divided into two main categories: one is that it can be expressed in display functions and has problems with display solutions, such as fold-over artifacts, image blurring due to half-fourier acquisitions, and gibbs artifacts; another class is the problem of difficult to express with display functions, or of not displaying solutions, such as high noise, low resolution and gibbs artifacts due to limited acquisition matrices, etc.

Furthermore, explicit solution imaging, such as parallel imaging, is based on redundant information provided by multi-channel coils, and DNN image reconstruction is based on the application of a priori knowledge. Thus, the information based on which the explicit resolution imaging and the DNN image reconstruction are different, therefore, the two imaging methods can be combined and used together to exert the respective advantages, improve the quality of the reconstructed image and the reconstruction speed,

in addition, the DNN model is a data-driven model, which is not a mathematical model fitted based on some existing data, but is obtained by training relationships between sets of input data and expected output results (i.e., model training), so as to accurately describe the relationships between the input data and the expected output results. Therefore, the DNN model can be regarded as a real description of the real world, and can reflect the real world more accurately. Therefore, the magnetic resonance quantitative value calculated by the DNN model is accurate.

In addition, when acquiring a magnetic resonance quantification using the DNN model, the DNN model for acquiring a magnetic resonance quantification may be trained in advance, and when applying the DNN model, the magnetic resonance quantification may be acquired directly using the trained DNN model, so that the calculation time for calculating a magnetic resonance quantification may be significantly reduced.

Based on this, this application embodiment provides a magnetic resonance quantitative imaging method. Firstly, acquiring partial k-space data of a plurality of echoes according to a down-sampling mode to obtain k-space acquired data of the plurality of echoes; then, respectively and successively adopting an image reconstruction method of a first deep neural network model and an explicit analytic solution imaging method or respectively and successively adopting an explicit analytic solution imaging method and an image reconstruction method of the first deep neural network model to carry out image reconstruction on the k-space acquisition data of each echo to obtain a magnetic resonance image of each echo; and finally, inputting the magnetic resonance image of each echo into a second deep neural network model to obtain the magnetic resonance quantitative image of each echo. According to the magnetic resonance quantitative imaging method, the magnetic resonance quantitative value is obtained according to the second deep neural network, because the deep neural network is a data-driven model which can accurately describe the real world, compared with the existing imaging mathematical model for calculating the magnetic resonance quantitative value, the magnetic resonance quantitative value can be obtained according to the deep neural network model. In addition, the deep neural network model runs faster, so that the calculation time of magnetic resonance quantification can be reduced, and the acquisition efficiency is improved. In addition, the input data of the second deep neural network model for obtaining the magnetic resonance quantification is the reconstructed image, and the image reconstruction speed is high by the method for reconstructing the image by using the deep neural network. Therefore, the image reconstruction method is also beneficial to improving the magnetic resonance quantitative imaging speed.

It should be noted that in the magnetic resonance quantitative imaging method provided in the embodiment of the present application, a first deep neural network Model (hereinafter, referred to as Model1) trained in advance is required to reconstruct an image, and a second deep neural network Model (hereinafter, referred to as Model2) is required to obtain magnetic resonance quantification. Therefore, before image reconstruction with the Model1, the Model1 needs to be trained; the Model2 needs to be trained before solving for magnetic resonance quantification using the Model 2. For ease of understanding, before describing the magnetic resonance quantitative method provided in the embodiment of the present application, a specific implementation of the training method of the Model1 and the Model2 provided in the embodiment of the present application is first described with reference to fig. 1 and fig. 2, respectively.

Fig. 1 is a flowchart of a Model1 training method according to an embodiment of the present disclosure. The training method of the Model1 comprises the following steps:

s101: the output training samples and the input training samples of the Model1 training set are obtained.

The output training samples and the input training samples are the data basis for the training set to train and estimate the Model1, and for this purpose, the output training samples and the input training samples need to be acquired first. The output training sample and the input training sample are required to be respectively related to output data and input data in the actual application scene after the Model1 is trained. In this way, the image quality of the magnetic resonance image obtained after DNN reconstruction using the Model1 can be ensured.

In the embodiment of the present application, the output data of the Model1 is required to be a magnetic resonance image with higher image quality, and a plurality of magnetic resonance images reconstructed from full-acquisition or super-full-acquisition k-space data of each echo have high image quality, and the embodiment of the present application takes the plurality of magnetic resonance images reconstructed from full-acquisition or super-full-acquisition k-space data of each echo as the output training samples of the Model 1.

To understand the concept of oversubscription k-space data, the concept of oversubscription k-space data is first introduced. Full acquisition k-space data is the data acquired on all rows of phase encoding lines in the practical application of magnetic resonance quantitative imaging. For example, in a practical application of magnetic resonance quantitative imaging, full acquisition k-space data includes data on 256 rows of phase encode lines, while over-full acquisition k-space data includes data on more than 256 rows of phase encode lines, such as 384 rows of phase encode lines. The scan time for the over-full acquisition k-space data is greater than the scan time for the full acquisition k-space data in terms of scan time consumed, e.g., the over-full acquisition k-space is 1.5 times, even 10 times, the full acquisition k-space scan time. As such, the signal-to-noise ratio and/or resolution of the magnetic resonance image reconstructed from the overfilled k-space data is higher than the signal-to-noise ratio and/or resolution of the magnetic resonance image reconstructed from the overfilled k-space data.

In the application scenario of the magnetic resonance quantitative imaging, the explicit resolution imaging and the DNN reconstruction are combined successively to perform image reconstruction on the k-space acquisition data of a plurality of echoes. In this embodiment of the present application, the order of explicit resolution imaging and DNN reconstruction is not limited, and thus, as an example, the explicit resolution imaging may be performed first, and then the DNN reconstruction may be performed by using the result of the explicit resolution imaging as an input of the DNN image reconstruction. As another example, DNN reconstruction may be performed first, followed by explicit resolution imaging reconstruction. Thus, for both examples, the input data to the Model1 may be the image obtained after reconstructing the down-sampled k-space acquired data portion of the multiple echoes using explicit resolution imaging, or the k-space acquired data of the multiple echoes.

Accordingly, the input training samples to the Model1 may be magnetic resonance images obtained by partially reconstructing the partial k-space data of the multiple echoes from explicit analytic imaging, or may be the partial k-space data of the multiple echoes. The partial k-space data of the multiple echoes can be partial k-space data selected from the over-full-acquisition k-space data of the multiple echoes corresponding to the output training sample. Specifically, the input training samples used by the present application to train the Model1 may be magnetic resonance images that are partially reconstructed or partially reconstructed from or particular portions of k-space data for multiple echoes, which is a proportion of k-space data selected from the oversubscription k-space data for the multiple echoes. The ratio may be 0 to 100% (both inclusive).

S102: the Model1 is run with the input training samples as input to the trained Model1 to obtain the prediction results of the input training samples.

In this step, the prediction result of the acquired input training sample is a predicted image corresponding to the input training sample.

S103: and judging whether the structural similarity of the prediction result and the output training sample meets a preset condition. If the preset condition is met, executing the step S104; if the preset condition is not satisfied, step S105 is performed.

The output training samples have the expected value meaning of the prediction result of the Model1 input training samples for the input training samples during Model1 training. The prediction result obtained by Model1 of the input training samples needs to be compared with the magnetic resonance image reconstructed by full-acquisition or over-full-acquisition k-space data, and is used as the basis for adjusting the parameters of the Model1 and how to adjust the parameters. The prediction result is also in the form of an image.

Structural Similarity Index (SSIM Index), originally proposed by image and video engineering laboratories of the austin division of texas university, is an Index that measures the Similarity of two images, often used to compare and measure an uncompressed undistorted image with a distorted image. In the process of training the Model1 in advance in this embodiment, the prediction result can be understood as a distorted image, and the output training samples can be understood as an undistorted image.

The preset condition in this step may be a preset condition for measuring whether the Model1 training is completed. For example, the preset condition may be one or multiple sub-preset conditions, and may be a prediction result of the input training sample and a lower threshold of the output training sample SSIM index. A more specific example is given below: the preset condition is that SSIM index in the rectangular regions (x1, y1) - (x2, y2) of the prediction result and the output training samples is not less than 87%. The predetermined conditions are critical conditions that limit the accuracy of the trained Model1, and are often set empirically.

S104: the iteration is stopped and the Model1 training is completed.

If the result of the determination in step S103 is yes, it indicates that the structural similarity between the prediction result of the input training sample and the output training sample has satisfied the preset condition, and meets the requirement of Model1 training. The Model1 is already available for magnetic resonance quantitative imaging.

S105: adjust the relevant parameters of the trained Model1 and return to execution S102.

If the result of the determination in step S103 is negative, it indicates that the structural similarity between the prediction result of the input training sample and the output training sample still fails to meet the training requirement of the Model 1. The output training samples of the Model1 are used as known variables for solving the parameters of the Model1 to adjust the relevant parameters of the current Model. After the adjustment, the Model1 needs to be trained further, so the iterative process is repeated until the Model1 training is completed finally, returning to the step S102.

To improve the accuracy of the Model1, the Model1 may be continuously trained with new training samples in practical applications to continuously update the Model1, improve the accuracy of the Model1, and further improve the image reconstruction quality.

The above is a specific implementation of the Model1 training method provided in the embodiments of the present application. The Model1 trained as described above can be applied to the reconstruction of magnetic resonance images during quantitative magnetic resonance imaging as provided in the following examples.

The following describes a specific implementation of the Model2 training method provided in the embodiments of the present application.

Fig. 2 is a flowchart of a Model2 training method according to an embodiment of the present invention. The training method of the Model2 comprises the following steps:

s201: the output training samples and the input training samples of the Model2 training set are obtained.

The output training samples and the input training samples are the data basis for the training set to train and estimate the Model2, and for this purpose, the output training samples and the input training samples need to be acquired first. The output training sample and the input training sample are required to be respectively related to output data and input data in the actual application scene after the Model2 is trained. Therefore, the accuracy of solving the magnetic resonance quantification by using the Model2 can be guaranteed.

In the embodiment of the present application, the output data of the Model2 is required to be the magnetic resonance quantitative values with higher accuracy, and the accuracy of the magnetic resonance quantitative values calculated from the plurality of magnetic resonance images obtained by reconstructing the full-acquisition or super-full-acquisition k-space data of each echo is higher, so that the embodiment of the present application takes each subset of the magnetic resonance quantitative values calculated from the plurality of magnetic resonance images obtained by reconstructing the full-acquisition or super-full-acquisition k-space data of each echo as the output training sample of the Model 2. In the application scenario of the magnetic resonance quantitative imaging, a magnetic resonance quantitative map is acquired according to the input image of the Model2 and the Model 2. In the embodiment of the present application, the input image of the Model2 is an image obtained by sequentially performing image reconstruction on the k-space acquired data of each echo by using a parallel imaging method and an image reconstruction method of the Model1, respectively. Accordingly, the input training samples to Model2 may be respective subsets of images that are successively reconstructed from the partial reconstruction method and the image reconstruction method of Model1 with the partial k-space data of the multiple echoes.

Optionally, to improve the accuracy of the trained Model2 in calculating the quantitative magnetic resonance value, the partial k-space data of the multiple echoes may be specific partial k-space data of the multiple echoes corresponding to the input training samples used by the training Model1, i.e., a proportion of k-space data selected from full-acquisition or over-full-acquisition k-space data of the multiple echoes.

It should be noted that, in this embodiment, a subset of a magnetic resonance image may be a part of the magnetic resonance image, and the subsets of a magnetic resonance image together constitute the complete magnetic resonance image.

In the magnetic resonance quantitative map, the calculation of the magnetic resonance quantitative value of one image point is independent of the magnetic resonance quantitative values of other peripheral points, so in order to increase the generalization ability of the Model2, the training set of the Model2 is expanded, in the embodiment, each subset of the magnetic resonance images can be used as the input training sample of the Model2, and the training set can be effectively expanded compared with the case that the whole image is used as the input training sample of the Model 2. The training effect of the Model2 can be improved by applying the expanded training set training samples.

The subset of the magnetic resonance image may be image points into which the image is divided point by point. For example, if the image is divided into subsets by each image point, there are 65536 subsets in the image, which is composed of 256 × 256 image points. In addition, the subset of the image may also be an image area obtained by dividing the image piece by piece, for example, an image area corresponding to a structure in the image whose quantitative value difference of tissues is smaller than a first threshold is taken as one piece, and each piece is referred to as a subset of the image. In addition, the subset of the image may also be an image region that is divided into classes for the image, for example, using a gray value as a reference for classifying the classes, and an image region in the image with a difference of gray values smaller than a second threshold value is used as a class, and each class is referred to as a subset of the image. The above description of the magnetic resonance image subset is only an exemplary way to describe the image subset, and the other subset forms of the image applied to the present application are not limited.

S202: and inputting the subsets of the magnetic resonance images of the echoes into the second deep neural network model one by one, and sequentially acquiring the predicted magnetic resonance quantification corresponding to the subsets of the magnetic resonance images of the echoes.

For example, the following steps are carried out: assuming that in the embodiment of the present application, 6 echoes are used, each echo includes 3 subsets, and thus, a total of 18 subsets of the magnetic resonance images are included, S202 specifically is: and inputting the 18 subsets into a second deep neural network model one by one, and sequentially acquiring the predicted magnetic resonance quantification corresponding to the 18 subsets respectively.

In this step, a magnetic resonance image may comprise a subset of the plurality of magnetic resonance images.

S203: respectively judging whether the average variances of the predicted magnetic resonance quantification corresponding to each subset of the magnetic resonance image of each echo and the magnetic resonance quantification of the corresponding subset in the output training sample are smaller than a preset threshold value, and if the average variances are smaller than the preset threshold value, executing a step S204; if there is at least one mean variance not less than the preset threshold, step S205 is performed.

The Model2 training method provided by the embodiment of the application adopts the Mean Square Error (MSE) as an index for measuring and predicting the volatility of the quantitative relative output training sample of the magnetic resonance. The output training samples have the expected value meaning of the Model2 input training samples' predicted magnetic resonance quantification for the input training samples during Model2 training. Thus, a smaller average variance of the predicted magnetic resonance quantification for a subset with the magnetic resonance quantification for a corresponding subset of the output training samples indicates a smaller volatility or degree of deviation of the predicted magnetic resonance quantification for the subset from the output training samples, and the more closely the predicted magnetic resonance quantification for the subset approaches the desired value. In this step, MSE between the predicted magnetic resonance quantification obtained by each subset of the magnetic resonance image of each echo through the Model2 and the magnetic resonance quantification of the corresponding subset in the output training sample is compared with a preset threshold, and the comparison result is used as a basis for whether the Model2 parameter is adjusted or not and how to adjust the parameter specifically.

The preset threshold in this step may be a preset condition for measuring whether the Model2 training is completed, and is also a key condition for restricting the accuracy of the trained Model2 in solving the magnetic resonance quantification, and is often set according to experience. For example, the preset threshold may be 0.06.

The description is given by way of example in S202. In the example 18 subsets, if the average variance between the preset mr quantification of the 18 subsets and the mr quantification of the corresponding subset in the output training sample is smaller than the preset threshold, step S204 is executed; if there is at least one mean variance not less than the preset threshold, step S205 is performed.

S204: the iteration is stopped and the Model2 training is completed.

If the MSE of the predicted magnetic resonance quantification obtained by Model2 of the input training samples corresponding to each subset and the MSE of the magnetic resonance quantification of the corresponding subset in the output training samples are both less than 0.06, it indicates that the Model2 has reached the desired target, and the Model2 is already available for magnetic resonance quantitative imaging.

S205: adjust the relevant parameters of the trained Model2 and return to execution S202.

If the MSE of the predicted magnetic resonance quantification obtained by the Model2 on the input training sample corresponding to each subset and the MSE of the magnetic resonance quantification of the corresponding subset in the output training samples are not less than 0.06, it indicates that the Model2 does not meet the expected target, and the output training sample of the Model2 is used as a known variable for Model2 parameter solution to adjust the relevant parameters of the current Model. After the adjustment, the Model2 needs to be trained further, so the iterative process is repeated until the Model2 training is completed finally, returning to the step S202.

To improve the accuracy of the Model2, the Model2 may be continuously trained with new training samples in practical applications to continuously update the Model2, improve the accuracy of the Model2, and further improve the image reconstruction quality.

The above is a specific implementation of the Model2 training method provided in the embodiments of the present application. The Model2 obtained by the above training can be applied to the magnetic resonance quantitative imaging method provided in the following embodiments.

It should be noted that, in the embodiment of the present application, the Model1 and the Model2 may use the same DNN Model, or may use different DNN models, for example, the Model1 may use the ResNET Model, and the Model2 may use the uet Model.

The following describes in detail a specific implementation of the magnetic resonance quantitative imaging method provided by the present application with reference to fig. 3.

The magnetic resonance quantitative imaging method provided by the embodiment of the application utilizes a mode of combining DNN and a display analysis solution imaging method to jointly complete the reconstruction of each echo magnetic resonance image. Also, to show the advantages of the analytical solution imaging method and the DNN reconstructed image, the analytical solution imaging method may complete a portion of the image reconstruction process, while the DNN reconstructed image method may complete the remainder of the image reconstruction process. For example, by setting the original k-space data acquired by down-sampling to 20% of the complete k-space data, 40% of the k-space data can be reconstructed by a display resolution imaging method using the 20% of the k-space data, and then 100% of the k-space data, i.e., the complete k-space data, can be reconstructed by DNN using the 40% of the k-space data.

More specifically, when the display resolution imaging method includes a parallel imaging method and a half fourier imaging method, the above example may be specifically: the method comprises the steps of reconstructing 40% of k-space data by using 20% of acquired original k-space data through a parallel imaging method, then obtaining 50% of k-space data by using the reconstructed 40% of k-space data through a half Fourier reconstruction method, and finally reconstructing 100% of k-space data and images thereof by using the 50% of k-space data through a DNN (digital noise network) model.

In the following specific implementation, an explicit analytic solution imaging method is taken as an example of a parallel imaging method, and an image obtained after input data of the Model1 is partially reconstructed from k-space acquired data of a plurality of echoes obtained by down-sampling by the parallel imaging method is taken as an example of a description. Thus, in the imaging methods provided in the following embodiments, parallel imaging is performed first, and then DNN reconstruction is performed.

Example one

It should be noted that in the embodiment of the present application, the reconstruction of each echo magnetic resonance image is jointly performed by using a parallel imaging method and by using the Model1 in a sequential combination manner. Moreover, in order to take advantage of the parallel imaging method and the DNN reconstructed image, the parallel imaging method may complete a part of the image reconstruction process, and the DNN reconstructed image method may complete the remaining part of the image reconstruction process.

Referring to fig. 3, a magnetic resonance quantitative imaging method provided in an embodiment of the present application includes the following steps:

s301: and acquiring partial k-space data of the multiple echoes according to a down-sampling mode to obtain k-space acquired data of the multiple echoes.

In order to realize the magnetic resonance quantitative imaging, the k-space data corresponding to a plurality of echoes needs to be acquired first. The down-sampling mode is a data sampling method for accelerating the magnetic resonance quantitative imaging speed, and the down-sampling mode is utilized to carry out partial acquisition on the k-space data of a plurality of echoes to obtain the k-space acquired data of the plurality of echoes for subsequent partial reconstruction of parallel imaging. The down-sampling mode can be various, such as random down-sampling, variable density down-sampling, equal density down-sampling, half-Fourier sampling, and the like.

For ease of understanding, the down-sampling method will be described by taking an example in which the sampling target is a two-dimensional image. Taking the example of equal density sampling, if the acceleration multiple (also called acceleration factor) R of the sampling is 2, the data in the image can be acquired for every other row or column of a two-dimensional image. Due to the fact that the collected data volume is reduced, the collection speed is improved.

S302: and respectively and successively adopting a parallel imaging method and an image reconstruction method of the first depth neural network model to carry out image reconstruction on the k-space acquisition data of each echo so as to obtain a magnetic resonance image of each echo.

In the embodiment of the present application, in order to improve the accuracy of the finally obtained magnetic resonance quantitative value, the following two aspects are improved. In the first aspect, it is necessary to improve the quality of the magnetic resonance image of each echo. In the second aspect, the reliability of the pre-trained second deep neural network model needs to be ensured. The second aspect is ensured by the implementation manner of the Model2 training method provided in the foregoing embodiment of the present application.

The present step is mainly based on the first aspect, but it should be noted that the present step also aims to achieve a higher image reconstruction speed while improving the quality of the magnetic resonance image of each echo. Step S302 may specifically include:

a1: and performing partial reconstruction on the k-space acquired data of each echo by adopting a parallel imaging method to obtain a first image of each echo.

It should be noted that, because there are many choices for the down-sampling mode for acquiring the partial k-space data of the multiple echoes, and the k-space acquired data of the multiple echoes acquired by using different down-sampling modes have different data characteristics, different parallel imaging reconstruction methods can be adopted for different down-sampling modes in the step a1 in order to obtain a better image reconstruction effect.

As an example, when the random down-sampling manner or the variable density down-sampling manner is adopted to acquire the partial k-space data of the multiple echoes in step S301, GRAPPA may be used to perform partial reconstruction on the k-space acquired data of each echo to obtain the partial k-space data of each echo virtual channel; and then, carrying out image reconstruction on partial k-space data of each echo virtual channel to obtain a first image of each echo.

As another example, when the step S301 employs an iso-density down-sampling mode, such as equidistant sampling or crossed equidistant sampling to acquire partial k-space data, the sensitivity encoding method SENSE may be used to partially reconstruct the k-space acquired data of each echo, resulting in a first image of each echo. The SENSE algorithm is an image domain reconstruction algorithm, and transforms k-space acquired data into an image domain through Fourier transform, and then performs interpolation on the image domain to reconstruct a first image of each echo.

In step a1, the parallel imaging method is used to partially reconstruct the k-space acquired data of each echo, which improves the image reconstruction speed compared with the complete reconstruction. By partial reconstruction is meant that the reconstructed image is not an image in full k-space but in partial k-space. That is, the partial reconstruction fits only a portion of the non-acquired data in k-space, not all of the acquired data. For example, the acquired k-space is row 1, row 5, … …, and row 4n +1, and row 2, row 6, row … …, and row 4n +2 are fitted after partial reconstruction, but the data of row 4n +3 and row 4n +4 are not fitted.

Therefore, the time of magnetic resonance quantitative imaging can be effectively saved by partially reconstructing the k-space acquired data of each echo by adopting a parallel reconstruction method.

In order to take advantage of the parallel imaging method, the present step a1 performs partial reconstruction using a parallel imaging reconstruction method corresponding to a down-sampling method of acquiring partial k-space data of a plurality of echoes, and the main purpose is to improve the problem of aliasing artifacts in the magnetic resonance image.

A2: and according to the first image and the first deep neural network model of each echo, carrying out complete reconstruction through the first deep neural network model to obtain a second image of each echo, and taking the second image of each echo as a magnetic resonance image of each echo.

Although the first image of each echo is obtained after the k-space acquired data of each echo is partially reconstructed in the step a1, the acquisition time is effectively shortened, the magnetic resonance quantitative imaging speed is increased, and the problem of aliasing artifacts in the magnetic resonance quantitative imaging can be solved, the first image of each echo still has the problems of high noise and low resolution. In view of this, in combination with the advantages of the DNN reconstruction in terms of super resolution, denoising, and the like, in step a2, the first image of each echo is used as an input of the first deep neural network model, and complete reconstruction is performed through the first deep neural network model, so as to solve the problems of high noise and low resolution of the first image of each echo.

This step a2 uses a pre-trained Model1 to achieve a complete reconstruction of the first image of each echo.

Since the first images of the respective echoes obtained in steps S301 and a1 have extremely close correlation with the input training samples of the Model1, the output training samples are a plurality of magnetic resonance images reconstructed from full-acquisition or super-full-acquisition k-space data of the plurality of echoes, and the second image of the respective echoes output from the Model1 obtained from these training samples has higher reliability in terms of image quality, after the first image is used as the input of the Model1 for reconstructing the magnetic resonance image in step a2, the second image of the plurality of echoes output from the Model1 is finally compared with the first image of the corresponding echo, so that image noise is reduced and image resolution is improved. The second image of each echo is taken as the magnetic resonance image of each echo.

S303: and inputting the magnetic resonance image of each echo into the second deep neural network model to obtain the magnetic resonance quantitative image of each echo.

In this step, a magnetic resonance quantitative value is obtained from the magnetic resonance image of each echo obtained in step S302 and the Model2 trained in advance. As an example, the magnetic resonance quantification referred to herein may include a relaxation time T₁Relaxation time of

One or more of four groups, Proton Density (PD) and magnetic susceptibility mapping (QSM).

The foregoing is a magnetic resonance quantitative imaging method provided in an embodiment of the present application. According to the magnetic resonance quantitative imaging method, the magnetic resonance quantitative value is obtained according to the second deep neural network, because the deep neural network is a data-driven model, factors influencing a prediction result cannot be ignored by the data-driven model, and accurate description of the real world can be realized, so that compared with the existing imaging mathematical model for calculating the magnetic resonance quantitative value, the magnetic resonance quantitative value can be obtained more accurately according to the deep neural network model. In addition, the deep neural network model runs faster, so that the calculation time of magnetic resonance quantification can be reduced, and the acquisition efficiency is improved. In addition, the input data of the second deep neural network for obtaining the magnetic resonance quantification is the reconstructed image, and the image reconstruction speed is high by the method for reconstructing the image by using the deep neural network. Therefore, the image reconstruction method is also beneficial to improving the magnetic resonance quantitative imaging speed.

It should be noted that, in the magnetic resonance imaging method provided in the above embodiment, since the k-space acquired data obtained by the down-sampling mode is actually acquired, the authenticity of the data can be reflected, and the correction of the magnetic resonance image obtained in the above S302 is facilitated. Therefore, in order to further improve the imaging quality of magnetic resonance imaging, the embodiment of the present application further provides another implementation manner of magnetic resonance quantitative imaging, and specifically refer to embodiment two.

Example two

It should be noted that the second embodiment has many similarities with the first embodiment, and for the sake of brevity, only the differences are modified, and the similarities refer to the corresponding description of the first embodiment.

Please refer to fig. 4, which is a flowchart of a magnetic resonance quantitative imaging method according to the second embodiment of the present application. As shown in fig. 4, the magnetic resonance quantitative imaging method includes:

s401: and acquiring partial k-space data of the multiple echoes according to a down-sampling mode to obtain k-space acquired data of the multiple echoes.

S402: and performing partial reconstruction on the k-space acquired data of each echo by adopting a parallel imaging method to obtain a first image of each echo.

S403: and according to the first image and the first deep neural network model of each echo, carrying out complete reconstruction through the first deep neural network model to obtain a second image of each echo, and taking the second image of each echo as a magnetic resonance image of each echo.

In the present embodiment, the steps S401, S402, and S403 are respectively the same as the steps S301, a1, and a2 in the first embodiment, and for brevity, detailed description is omitted here, and please refer to the description in the first embodiment for detailed information.

S404: and respectively correcting the magnetic resonance image of the corresponding echo by using the k-space acquisition data of each echo, and taking each corrected echo image as a final magnetic resonance image of each echo.

S405: and inputting the final magnetic resonance image of each echo into a second deep neural network model to obtain the magnetic resonance quantitative image of each echo.

Step S405 in this embodiment is similar to step S303 in the first embodiment, but is different from this in that the input of the second deep neural network model is the final magnetic resonance image of each echo obtained after the correction in S404. Due to the correction, the input reliability of the second deep neural network model is enhanced, so that the accuracy of S405 magnetic resonance quantitative solution can be improved.

It should be noted that, in this embodiment, the input of the second deep neural network Model is changed from the input of the Model2 in the first embodiment, so the input training samples used in the pre-training process of the second deep neural network Model used in this embodiment are: and correcting the image obtained by partially reconstructing the specific partial k-space data of the multiple echoes and completely reconstructing the first deep neural network model in a corresponding correction mode S404 to obtain the final magnetic resonance image of each echo.

In the magnetic resonance quantitative imaging method provided in the second embodiment, there are various implementations of the correction of the magnetic resonance image of each echo in S404, and thus it is easily understood that various implementations of the magnetic resonance quantitative imaging method can be formed according to the implementation of S404. Several specific implementations of the magnetic resonance quantitative imaging method provided in the present application are described in detail below with reference to the embodiments and the accompanying drawings. It should be noted that the following embodiments are only exemplary illustrations and descriptions of the implementation of the provided magnetic resonance quantitative imaging method, and all other embodiments obtained by those skilled in the art without any creative effort based on the embodiments in the present application belong to the protection scope of the present application.

EXAMPLE III

It should be noted that, the third embodiment has many similarities with the second embodiment, and for the sake of brevity, only the differences are modified, and the similarities refer to the corresponding description of the second embodiment.

Please refer to fig. 5, which is a flowchart of a magnetic resonance quantitative imaging method provided in the third embodiment of the present application. As shown in fig. 5, the magnetic resonance quantitative imaging method includes:

s501: and acquiring partial k-space data of the multiple echoes according to a down-sampling mode to obtain k-space acquired data of the multiple echoes.

S502: and performing partial reconstruction on the k-space acquired data of each echo by adopting a parallel imaging method to obtain a first image of each echo.

S503: and according to the first image and the first deep neural network model of each echo, carrying out complete reconstruction through the first deep neural network model to obtain a second image of each echo, and taking the second image of each echo as a magnetic resonance image of each echo.

Steps S501 to S503 in this embodiment are the same as steps S401 to S403 in the second embodiment, respectively, and for brevity, detailed description is omitted here, and please refer to the description in the first embodiment for detailed information. The following steps S504 to S506 will correspond to S404 in the second embodiment, and specifically describe how to modify the magnetic resonance image by using k-space acquired data, and use the modified magnetic resonance image as a final magnetic resonance image.

S504: and mapping the magnetic resonance image of each echo to k-space to obtain first complete k-space data of each echo.

The first complete k-space data of each echo obtained through mapping specifically includes: k-space data of the sampled region and k-space data of the non-sampled region. For the sampling region, compared with the k-space data of the sampling region in the first complete k-space data of each echo at present, the k-space acquisition data of a plurality of echoes obtained by actual sampling is obtained in the step S501 in the earlier stage; for the non-sampling region, the k-space data of the non-sampling region in the current first complete k-space data is the better data resource obtained after the image reconstruction of the region.

S505: and respectively replacing the k-space data of the sampling region in the first complete k-space data of the corresponding echo by using the k-space acquisition data of each echo to obtain second complete k-space data of each echo.

Compared with the k-space data of the corresponding sampling region after the second image of each echo is mapped to the k-space, the k-space acquisition data of the echoes is actually acquired instead of reconstructed mapping, so that the data can reflect higher authenticity and is beneficial to presenting a better magnetic resonance image. For this reason, in order to improve the imaging quality of the magnetic resonance quantitative imaging, in this step, the k-space data of the sampling region in the first complete k-space data of each echo is replaced with the k-space acquisition data of each echo obtained in step S501. The k-space data of the non-sampled region is not subject to any change. The k-space acquisition data displaced to the sampling region together with the k-space data of the non-sampling region constitute the second complete k-space data of the individual echoes.

S506: and carrying out image reconstruction according to the second complete k-space data of each echo to obtain a final magnetic resonance image of each echo.

The second complete k-space data of each echo obtained according to step S405 has more real data resources than the first complete k-space data of each echo. Therefore, the final magnetic resonance image obtained by reconstructing the image of the second complete k-space data of each echo has higher image quality and higher reliability than the magnetic resonance image of each echo obtained in step S503.

S507: and inputting the final magnetic resonance image of each echo into a second deep neural network model to obtain the magnetic resonance quantitative image of each echo.

The above is the magnetic resonance quantitative imaging method provided in the third embodiment of the present application. Compared with the k-space data of the sampling region after the magnetic resonance images of the multiple echoes are mapped to the k-space data, the k-space data of the sampling region is more fit to the actual situation. Therefore, the k-space data of the sampling region is replaced by the k-space acquisition data of a plurality of echoes, and the quality and the reliability of the final magnetic resonance image obtained after reconstruction are improved. Furthermore, the subset of the final magnetic resonance image is used as the final input image of the second deep neural network model, so that the quality and the resolution are higher, and the accuracy of the magnetic resonance quantitative solution in the step S507 can be correspondingly improved.

In addition, the present application provides another implementation manner of correcting a magnetic resonance image in the magnetic resonance quantitative imaging method on the basis of the magnetic resonance quantitative imaging method provided in the second embodiment. The problem that image quality is easily influenced in the magnetic resonance quantitative imaging process is respectively solved by adopting a parallel imaging method, a DNN reconstruction method and a half Fourier reconstruction method in a sequential combination mode. In order to take advantage of the parallel imaging method, the DNN reconstructed image, and the half-fourier reconstructed image, if the down-sampling multiple is set to 5, the original k-space data is 20% of the complete k-space data. 40% of k-space data are reconstructed by using 20% of the acquired original k-space data through a parallel imaging method, then 50% of k-space data are obtained by using the 40% of the reconstructed k-space data through a half Fourier reconstruction method, and finally 100% of k-space data and images thereof are reconstructed by using the 50% of the k-space data through a DNN model.

The following describes in detail a specific implementation of the magnetic resonance quantitative imaging method provided in this embodiment with reference to fig. 6.

Example four

It should be noted that, the fourth embodiment has many similarities with the second embodiment, and for the sake of brevity, only the differences are modified, and the similarities refer to the corresponding description of the second embodiment.

Please refer to fig. 6, which is a flowchart of a magnetic resonance quantitative imaging method according to the fourth embodiment of the present application. As shown in fig. 6, the magnetic resonance quantitative imaging method includes:

s601: and acquiring partial k-space data of the multiple echoes according to a half Fourier sampling mode to obtain k-space acquired data of the multiple echoes.

The half Fourier sampling mode only collects slightly more than half of data, so the scanning time is reduced by nearly half, and theoretically the acceleration multiple R is between 1 and 2. This sampling scheme takes the half-fourier interval as the sampling region of k-space, while k-space data outside the half-fourier interval is not acquired. For ease of understanding, the half-fourier sampling mode is exemplified below.

As an example of two-dimensional magnetic resonance imaging, if the half fourier coefficient is 0.8, an interval of 80% in which the phase encoding directions are continuous is taken as a sampling region. For two-dimensional magnetic resonance imaging, phase encoding only has one direction, so that partial sampling points are defined in one straight line direction of k-space, and the partial sampling points are acquired to obtain k-space acquisition data. In this example implementation scenario, some of the sample points may be equidistantly distributed, variably dense, or randomly distributed.

As an example of three-dimensional magnetic resonance imaging, half fourier sampling may be used in two phase encoding directions, for example, half fourier coefficients are 0.8 and 0.9, respectively, and then an interval of 72% in a continuous manner within an angle formed by the two phase encoding directions is used as a sampling region. Partial sampling points can be defined on a plane formed by two phase encoding directions in the k-space, and the partial sampling points are acquired to obtain k-space acquired data. In this example implementation scenario, some of the sampling points may be distributed equidistantly, relatively equidistantly, randomly, or in a variable density poisson distribution.

As an example of dynamic magnetic resonance imaging, partial sampling points may be defined in a two-dimensional or three-dimensional space formed by a phase encoding direction and a time dimension direction, and the partial sampling points are acquired to obtain k-space acquisition data.

S602: and performing partial reconstruction on the k-space acquired data of each echo by adopting a parallel imaging method to obtain a first image of each echo.

S603: and according to the first image and the first deep neural network model of each echo, carrying out complete reconstruction through the first deep neural network model to obtain a second image of each echo, and taking the second image of each echo as a magnetic resonance image of each echo.

In this embodiment, steps S602 and S603 are the same as steps S402 and S403 in the second embodiment, respectively, and for brevity, detailed description is omitted here, and please refer to the description in the second embodiment for detailed information. The following steps S604 to S609 will correspond to S404 in the second embodiment, and specifically describe how to modify the magnetic resonance image by using k-space acquired data, and use the modified magnetic resonance image as a final magnetic resonance image.

S604: and mapping the magnetic resonance image of each echo to k-space to obtain first complete k-space data of each echo.

The first complete k-space data of each echo obtained through mapping specifically includes: k-space data of the sampled region and first k-space data of the non-sampled region. For the sampling region, compared with the k-space data of the sampling region in the current first complete k-space data, the k-space acquisition data actually sampled according to the half-Fourier sampling mode is obtained in the step S601 in the earlier stage; for a non-sampling region, the first k-space data of the non-sampling region in the current first complete k-space data is the corresponding data resource of the region after image reconstruction.

However, in order to further optimize the k-space data for the final image reconstruction, the present embodiment also performs the following operations for the k-space data of the sampling region and the first k-space data of the non-sampling region in the first full k-space data, respectively:

s605: and setting the first k-space data of the non-sampling region in the first complete k-space data of each echo to be 0 to obtain third complete k-space data of each echo.

The first k-space data of the non-sampled region in the first complete k-space data of each echo is set to 0, mainly to facilitate reconstruction of the k-space data in a half fourier reconstruction method. The set k-space data of each echo is called third complete k-space data of each echo, which comprises k-space data of a sampling region and k-space data of a non-sampling region. The k-space data of the sampling region in the third complete k-space data of each echo is consistent with the k-space data of the sampling region in the first complete k-space data; and the k-space data of the non-sampled region in the third complete k-space data of each echo is 0.

S606: and respectively utilizing the third complete k-space data of each echo to carry out data reconstruction by a half Fourier reconstruction method so as to obtain the second k-space data of the non-sampling region of each echo.

The principle of reconstructing data by the half-Fourier reconstruction method is as follows: the data is replicated using mathematical symmetry of k-space. Therefore, in this step, after the half fourier reconstruction, the k-space data of each echo non-sampling region is reconstructed from 0 to data corresponding to the symmetric copy of the k-space data of the corresponding echo sampling region, which is referred to as the second k-space data of each echo non-sampling region.

Researches show that the data reconstruction is carried out by the half Fourier reconstruction method in the step, so that the problems of image blurring, Gibbs artifacts and the like caused by collecting partial k-space data in an S601 half Fourier sampling mode can be effectively solved.

S607: and carrying out weighted average on the first k-space data of the non-sampling region of each echo and the second k-space data of the non-sampling region to obtain third k-space data of the non-sampling region of each echo.

In this embodiment, in order to optimize the k-space data of each echo, which is finally used for image reconstruction, for the non-sampling region, the k-space data of the non-sampling region of the first complete k-space data and the half fourier reconstructed k-space data of the non-sampling region obtained after half fourier data reconstruction are calculated in a weighted average manner to obtain the k-space data of the non-sampling region of each echo as the final reconstructed k-space data of the non-sampling region.

As an alternative implementation manner of this step, coefficients may also be set for the first k-space data of the non-sampling region of each echo and the second k-space data of the non-sampling region of each echo, respectively, and then mathematical calculation is performed. For example, a coefficient q is set for the first k-space data of the non-sampling region, a coefficient p is set for the second k-space data of the non-sampling region, each data is multiplied by a corresponding coefficient, and then the k-space corresponding data multiplied by the coefficients are added to obtain the third k-space data of the non-sampling region of each echo.

S608: and adding the third k-space data of the non-sampling region of each echo and the k-space acquisition data of the corresponding echo to obtain fourth complete k-space data of each echo.

For the sampling region, similarly to the third embodiment of the present application, compared with the k-space data of the sampling region in the current first complete k-space data, the k-space acquisition data actually sampled according to the half fourier sampling manner has been obtained in step S601 earlier, so that the k-space acquisition data of each echo is used as the k-space data preferred by the sampling region for the final image reconstruction of each echo.

The third k-space data of the respective echo non-sampled region and the k-space acquisition data of the sampled region together thus constitute a fourth complete k-space data.

S609: and performing image reconstruction according to the fourth complete k-space data of each echo to respectively obtain a final magnetic resonance image of each echo.

The fourth complete k-space data of each echo includes k-space data for optimizing the final image reconstruction effect, and specifically includes: third k-space data optimized for non-sampled regions, and k-space acquisition data replaced for sampled regions. Therefore, after the final image reconstruction is performed on the fourth complete k-space data of each echo, a reconstructed image with higher quality than that of the magnetic resonance image can be obtained and used as the final magnetic resonance image of each echo.

S610: and inputting the final magnetic resonance image of each echo into a second deep neural network model to obtain the magnetic resonance quantitative image of each echo.

The magnetic resonance quantitative imaging method provided in the fourth embodiment is described above. According to the method, when partial k-space data of a plurality of echoes are collected, a half Fourier sampling mode is adopted, and based on the characteristics of the down-sampling mode, the first k-space data of a non-sampling area in the first complete k-space data of each echo is processed. Specifically, setting 0 to the first k-space data of a non-sampling region in the first complete k-space data, and then performing data reconstruction by using a half Fourier reconstruction method to obtain second k-space data of the non-sampling region; and then carrying out weighted average on the first k-space data of the non-sampling region and the second k-space data of the non-sampling region to obtain third k-space data of the non-sampling region, wherein the third k-space data is used as the k-space data of the non-sampling region in fourth complete k-space data for final image reconstruction.

In addition, for the sampling region, compared with the k-space data of the sampling region after the second image of each echo is mapped to the k-space data, the k-space acquisition data obtained by the half Fourier sampling mode of each echo is real, and the reliability of image reconstruction can be improved, so that the k-space acquisition data is used as the k-space data of the sampling region in the fourth complete k-space data for final image reconstruction. Therefore, the fourth complete k-space data formed by the third k-space data of the non-sampling region and the k-space acquisition data can be used for final image reconstruction, and the image obtained after the final reconstruction can remove various artifacts, has higher quality, resolution and definition, and can be used as a final magnetic resonance image. According to the final magnetic resonance image and the second deep neural network model, a magnetic resonance quantitative value with higher accuracy can be obtained.

In addition, when the down-sampling method is a half fourier sampling method, as an alternative to the fourth embodiment of the present application, half fourier reconstruction may be performed after parallel imaging and before DNN reconstruction. See example five for details.

EXAMPLE five

Referring to fig. 7, a flowchart of a magnetic resonance quantitative imaging method according to a fifth embodiment of the present application is provided. As shown in fig. 7, the magnetic resonance quantitative imaging method includes:

s701: and acquiring partial k-space data of the multiple echoes according to a half Fourier sampling mode to obtain k-space acquired data of the multiple echoes.

S702: and performing partial reconstruction on the k-space acquired data of each echo by adopting a parallel imaging method to obtain a first image of each echo.

In this embodiment, steps S701 and S702 are the same as steps S601 and S602 in the fourth embodiment, and for the sake of brevity, detailed description thereof is omitted.

S703: and performing half Fourier reconstruction on the first image of each echo to obtain a third image of each echo.

S704: and taking the third image of each echo as the input of the first deep neural network model, and completely reconstructing through the first deep neural network model to obtain the second image of each echo.

The second image of each echo is then taken as the magnetic resonance image.

S705: and inputting the magnetic resonance image of each echo into the second deep neural network model to obtain the magnetic resonance quantitative image of each echo.

The foregoing is a specific implementation manner of the magnetic resonance quantitative imaging method provided in the fifth embodiment of the present application.

According to the method, when partial k-space data of a plurality of echoes are collected, a half Fourier sampling mode is adopted, and based on the characteristics of the down-sampling mode, aiming at a first image of each echo obtained by partial reconstruction of a parallel imaging method, half Fourier reconstruction is firstly carried out to obtain a third image of each echo, and the third image is used as the input of a first deep neural network model; completely reconstructing according to the third image of each echo and the first deep neural network model to obtain a second image of each echo as the input of the second deep neural network model; and finally, inputting the final magnetic resonance image of each echo into a second deep neural network model to obtain the magnetic resonance quantitative image of each echo. In the embodiment, the parallel imaging method, the DNN reconstruction method and the half Fourier reconstruction method are combined successively to respectively solve the problem that the image quality is easily influenced in the magnetic resonance quantitative imaging process, so that the quality and the reliability of the final input image of the second deep neural network model are improved, and the accuracy of the magnetic resonance quantitative solution is improved.

The magnetic resonance quantitative imaging method provided by the above embodiments is exemplified by performing parallel imaging on k-space acquired data and then performing DNN reconstruction in this order, and performing image reconstruction successively by using the parallel imaging and the first deep neural network model. In fact, the magnetic resonance quantitative imaging method provided by the present application may also perform DNN reconstruction first, and then perform parallel imaging to realize reconstruction of the magnetic resonance image provided by the present application.

It should be noted that, when the reconstruction of the magnetic resonance image provided by the present application is implemented by using the sequence of first performing the reconstruction of the first deep neural network model and then performing the parallel imaging, the input data of the first deep neural network model is correspondingly partial k-space data, rather than the partially reconstructed image by the parallel imaging. Thus, the input training samples used to train the first deep neural network model are specific partial k-space data.

Based on the above-mentioned specific implementation of performing magnetic resonance quantitative imaging in the order of performing parallel imaging and then performing DNN reconstruction, those skilled in the art can easily think of the specific implementation of performing magnetic resonance quantitative imaging in the order of performing DNN reconstruction and then performing parallel imaging, and for the sake of brevity, the detailed description is omitted here.

In addition, the parallel imaging method in the above embodiment is an example of an explicit resolution imaging method, and should not be construed as a limitation to the embodiment of the present application. In fact, as an extension of the embodiments of the present application, the parallel imaging method can be replaced by other explicit analytic solution imaging methods such as k-t BLAST method or zero-padding method. In other words, the implementation manners of successively reconstructing the image of the k-space acquired data by using the explicit analytic solution imaging method and the image reconstruction method using the first deep neural network model to obtain the input image of the second deep neural network model and then obtaining the magnetic resonance quantification according to the input image and the second deep neural network model are all within the scope of the present application.

The magnetic resonance quantitative imaging method provided by the above embodiments can be executed by the control device shown in fig. 8. The control device shown in fig. 8 includes a processor (processor)810, a communication Interface (Communications Interface)820, a memory (memory)830, and a bus 840. Processor 810, communication interface 820, and memory 830 communicate with each other via bus 840.

The memory 830 may store logic instructions for magnetic resonance quantitative imaging, and the memory may be a non-volatile memory (non-volatile memory), for example. The processor 810 may invoke logic instructions to perform the magnetic resonance quantitative imaging in the memory 830 to perform the magnetic resonance quantitative imaging method described above. As an embodiment, the logic instruction of the magnetic resonance quantitative imaging may be a program corresponding to control software, and when the processor executes the instruction, the control device may correspondingly display a functional interface corresponding to the instruction on the display interface.

The functionality of the logic instructions of the magnetic resonance quantitative imaging, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present disclosure may be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the methods according to the embodiments of the present application. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.

The above-mentioned logic instructions for magnetic resonance quantitative imaging may be referred to as "magnetic resonance quantitative imaging apparatus", and the apparatus may be divided into various functional modules. See in particular the examples below.

The following describes a specific implementation of the magnetic resonance quantitative imaging apparatus provided in the embodiments of the present application with reference to the drawings.

EXAMPLE six

Referring to fig. 9, a schematic structural diagram of a magnetic resonance quantitative imaging apparatus of six is provided for an embodiment of the present application, the apparatus includes:

the acquisition unit 91 is configured to acquire partial k-space data of a plurality of echoes according to a down-sampling manner, so as to obtain k-space acquired data of the plurality of echoes;

the image reconstruction unit 92 is configured to perform image reconstruction on the k-space acquired data of each echo by respectively successively adopting an image reconstruction method and an explicit analytic solution imaging method of the first deep neural network model or respectively successively adopting an explicit analytic solution imaging method and an image reconstruction method of the first deep neural network model, so as to obtain a magnetic resonance image of each echo;

the magnetic resonance quantitative imaging unit 93 is configured to input each subset of the magnetic resonance image of each echo into the second deep neural network model one by one, sequentially obtain magnetic resonance quantification corresponding to each subset, and further obtain a magnetic resonance quantitative image; wherein the subset of the magnetic resonance image of each echo is a portion of the magnetic resonance image of the corresponding echo;

the first deep neural network model is obtained by training a plurality of magnetic resonance images obtained by reconstructing full-acquisition or super-full-acquisition k-space data of a plurality of echoes as output training samples and by using specific part k-space data of the plurality of echoes or each magnetic resonance image obtained by partially reconstructing the specific part k-space data of the plurality of echoes as input training samples, wherein the specific part k-space data of the plurality of echoes are k-space data with specific proportion respectively selected from the full-acquisition or super-full-acquisition k-space data of each echo;

When the image reconstruction unit 92 successively adopts the explicit analytic solution imaging method and the image reconstruction method of the deep neural network model to perform image reconstruction on the k-space acquisition data of each echo, the image reconstruction unit 92 may include:

a first image reconstruction subunit 921, configured to perform partial reconstruction on the k-space acquired data of each echo by using an explicit analytic solution imaging method, to obtain a first image of each echo;

and the second image reconstruction subunit 922 is configured to perform complete reconstruction through the first deep neural network model according to the first image and the first deep neural network model of each echo, and obtain a second image of each echo as a magnetic resonance image.

Because in the magnetic resonance quantitative imaging apparatus provided in the embodiment of the present application, the image needs to be reconstructed by using the pre-trained first deep neural network model, optionally, the apparatus may further include: the first model training unit 94 is configured to train the first deep neural network model in advance, and the first model training unit 94 specifically includes:

a first obtaining subunit 941, configured to obtain an output training sample and an input training sample of a training set, where the output training sample is a plurality of magnetic resonance images reconstructed from full acquisition or super full acquisition k-space data of each echo; the input training sample is a magnetic resonance image obtained by partially reconstructing k-space data of a specific part or k-space data of the specific part; the specific partial k-space data is a specific proportion of k-space data selected from the full-acquisition or over-full-acquisition k-space data;

a first iteration subunit 942, configured to iterate parameters in the first deep neural network model using the input training sample and the output training sample, and in each iteration process, use the input training sample as an input of the first deep neural network model, and obtain a prediction result of the input training sample after passing through the first deep neural network model; and judging whether the structural similarity between the prediction result and the output training sample meets a preset condition, if so, stopping iteration, finishing the training of the first deep neural network model, and if not, adjusting the parameters of the first deep neural network model and continuing the next iteration process.

In addition, because in the magnetic resonance quantitative imaging apparatus provided in the embodiment of the present application, it is necessary to image the magnetic resonance quantitative using the second deep neural network model trained in advance, optionally, the apparatus may further include: a second model training unit 95, configured to train the second deep neural network model in advance, where the second model training unit 95 specifically includes:

a second obtaining subunit 951, configured to obtain an output training sample and an input training sample of a training set, where the output training sample is each subset of a magnetic resonance quantitative value calculated from multiple magnetic resonance images obtained by reconstructing full-acquisition or super-full-acquisition k-space data of multiple echoes, and the input training sample is each subset of an image obtained by partially reconstructing specific partial k-space data of the multiple echoes and completely reconstructing the first deep neural network model, and is used as an input training sample;

a second iteration subunit 952, configured to iterate parameters in a second deep neural network model using the input training sample and the output training sample, in each iteration process, use each subset of each magnetic resonance image as an input of the second deep neural network model, obtain a prediction result of the subset through the second deep neural network model, respectively determine whether an average variance between the prediction result of each subset of each magnetic resonance image and a magnetic resonance quantification of a corresponding subset in the output training sample is smaller than a preset threshold, stop the iteration if each average variance is smaller than the preset threshold, complete training of the second deep neural network model, and if at least one average variance is not smaller than the preset threshold, adjust parameters of the second deep neural network model, and continue to perform the next iteration process.

It should be noted that, in order to improve the accuracy of the first model training unit 94 and the second model training unit 95 in training the first deep neural network model and the second deep neural network model respectively, new training samples may be continuously used to train the first model training unit 94 and the second model training unit 95 in practical applications, so as to continuously update the first model training unit 94 and the second model training unit 95, improve the accuracy of the first model training unit 94 and the second model training unit 95, and apply the training samples to image reconstruction and magnetic resonance quantitative imaging respectively, thereby improving the accuracy of magnetic resonance quantitative solution.

The magnetic resonance quantitative imaging device provided by the fifth embodiment of the present application is described above. This magnetic resonance quantitative imaging device obtains magnetic resonance quantitative value according to second degree of depth neural network, because degree of depth neural network is the data drive type model, can realize the accurate description to the real world, consequently, compare in the current imaging mathematical model that is used for calculating the magnetic resonance ration, the magnetic resonance quantitative imaging device that this application embodiment five provided can obtain comparatively accurate magnetic resonance quantitative value according to degree of depth neural network model. In addition, the deep neural network model runs faster, so that the calculation time of magnetic resonance quantification can be reduced, and the acquisition efficiency is improved. In addition, the input data of the second deep neural network for acquiring the magnetic resonance quantification is an image obtained by image reconstruction, and the image reconstruction speed of the device for reconstructing the image by using the deep neural network is high. Therefore, the magnetic resonance quantitative imaging speed is also improved.

In the magnetic resonance quantitative imaging apparatus provided by the sixth embodiment, the k-space acquisition data obtained by the down-sampling unit is actually acquired, so that higher reality can be embodied, and the correction of the magnetic resonance image obtained by the six-image reconstruction unit in the sixth embodiment is facilitated. Therefore, in order to improve the imaging quality of the magnetic resonance quantitative imaging, the present application provides another implementation manner of the magnetic resonance quantitative imaging apparatus on the basis of the sixth embodiment, and specifically, the magnetic resonance image is corrected by using the k-space acquired data, and the corrected magnetic resonance image is taken as the final magnetic resonance image. The following describes in detail a specific implementation of the embodiments of the magnetic resonance quantitative imaging apparatus provided in the present application with reference to fig. 10 to 12.

EXAMPLE seven

Fig. 10 is a schematic structural diagram of the magnetic resonance quantitative imaging apparatus provided in this embodiment. As shown in fig. 10, the magnetic resonance quantitative imaging apparatus provided in this embodiment may further include, on the basis of the sixth embodiment:

and an image correction unit 96, configured to, after obtaining the magnetic resonance image of each echo and before acquiring the magnetic resonance quantification, respectively correct the magnetic resonance image of the corresponding echo by using the k-space acquisition data of each echo, and use each corrected echo image as a final magnetic resonance image of each echo.

Fig. 11 is a schematic diagram of a configuration of the image correction unit 96. As an optional implementation, the image correction unit 96 may specifically include:

a first mapping subunit 9601, configured to map the magnetic resonance image of each echo to k-space, so as to obtain first complete k-space data of each echo; the first full k-space data comprises k-space data of a sampled region and k-space data of a non-sampled region;

a replacement subunit 9602, configured to replace, by the acquired k-space data of each echo, k-space data of a sampling region in first complete k-space data of a corresponding echo, to obtain second complete k-space data of each echo;

and the first reconstruction subunit 9603 is configured to perform image reconstruction according to the second complete k-space data of each echo, so as to obtain a final magnetic resonance image of each echo.

The above is an optional implementation of the image correction unit of the magnetic resonance quantitative imaging apparatus provided in this embodiment. For the sampling region, the k-space acquisition data of a plurality of echoes obtained by down-sampling is more suitable for the actual situation compared with the k-space data of the sampling region after the magnetic resonance images of the plurality of echoes are mapped to the k-space data. Therefore, the k-space data of the sampling region is replaced by the k-space acquisition data of a plurality of echoes, and the quality and the reliability of the final magnetic resonance image obtained after reconstruction are improved. Furthermore, the subset of the final magnetic resonance image is used as the final input image of the second deep neural network model, so that the quality and the resolution are higher, and the accuracy of the magnetic resonance quantitative solution can be correspondingly improved.

Fig. 12 is a schematic diagram of another configuration of the image correction unit 96. When the down-sampling mode adopted by the acquisition unit 91 is a half fourier sampling mode, as another optional implementation, the image modification unit 96 may specifically include:

a second mapping subunit 9611, configured to map the magnetic resonance image of each echo to k-space, so as to obtain first complete k-space data of each echo; the first full k-space data comprises k-space data of a sampled region and first k-space data of a non-sampled region;

a setting subunit 9612, configured to set the first k-space data of the non-sampling region in the first complete k-space data of each echo to 0, to obtain third complete k-space data of each echo;

a second reconstruction subunit 9613, configured to perform data reconstruction by using the third complete k-space data of each echo through a half fourier reconstruction method, respectively, to obtain second k-space data of a non-sampling region of each echo;

a first calculating subunit 9614, configured to perform weighted average on the first k-space data of the non-sampling region of each echo and the second k-space data of the non-sampling region to obtain third k-space data of the non-sampling region of each echo;

a second calculating subunit 9615, configured to add the third k-space data of the non-sampling region of each echo and the k-space acquisition data of the corresponding echo to obtain fourth complete k-space data of each echo;

and a third reconstruction subunit 9616, configured to perform image reconstruction according to the fourth complete k-space data of each echo, to obtain a final magnetic resonance image of each echo respectively.

The image correction unit of the magnetic resonance quantitative imaging apparatus provided in the above embodiment is another optional implementation manner. According to the embodiment, when the partial k-space data of a plurality of echoes are acquired, a half Fourier sampling mode is adopted, and based on the characteristic of the down-sampling mode, the first k-space data of a non-sampling area in the first complete k-space data of each echo is processed. Specifically, setting 0 to the first k-space data of a non-sampling region in the first complete k-space data, and then performing data reconstruction by using a half Fourier reconstruction method to obtain second k-space data of the non-sampling region; and then carrying out weighted average on the first k-space data of the non-sampling region and the second k-space data of the non-sampling region to obtain third k-space data of the non-sampling region, wherein the third k-space data is used as the k-space data of the non-sampling region in fourth complete k-space data for final image reconstruction.

In addition, as an alternative embodiment to the above embodiment, the present application further provides an eighth embodiment of a magnetic resonance quantitative imaging apparatus, specifically referring to fig. 13. Fig. 13 is a schematic structural diagram of a magnetic resonance quantitative imaging apparatus according to the eighth embodiment.

Example eight

The magnetic resonance quantitative imaging apparatus provided in this embodiment includes an acquisition unit 91 and an image reconstruction unit 92, wherein the image reconstruction unit 92 may further include, in addition to the first image reconstruction subunit 921 and the second image reconstruction subunit 922:

the third image reconstruction subunit 1301 is configured to perform half fourier reconstruction on the first image of each echo after the first image reconstruction subunit 921 obtains the first image of each echo and before the second image reconstruction subunit 922 performs complete reconstruction through the first deep neural network model, so as to obtain a third image of each echo;

in this embodiment, the second image reconstruction subunit 922 is specifically configured to use the third image of each echo obtained by the third image reconstruction subunit 1301 as an input of the first deep neural network model, and perform complete reconstruction through the first deep neural network model to obtain the second image of each echo.

The foregoing is a specific implementation manner of the magnetic resonance quantitative imaging apparatus provided in the embodiments of the present application.

For the device embodiments, since they substantially correspond to the method embodiments, reference may be made to the partial description of the method embodiments for relevant points. The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

The foregoing detailed description is directed to a magnetic resonance quantitative imaging method and apparatus provided in the embodiments of the present application, and specific examples are used herein to explain the principles and implementations of the present application, and the description of the foregoing embodiments is only used to help understand the method and its core ideas of the present application; meanwhile, for a person skilled in the art, according to the idea of the present application, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present application.

Claims

1. A method of magnetic resonance quantitative imaging, the method comprising:

2. The method of claim 1, wherein after obtaining the magnetic resonance image of each echo and before acquiring the magnetic resonance quantitative map of each echo, further comprising:

and respectively correcting the magnetic resonance image of the corresponding echo by using the k-space acquisition data of each echo, and taking each corrected echo image as a final magnetic resonance image of each echo.

3. The method according to claim 2, wherein the step of correcting the magnetic resonance image of the corresponding echo by using the k-space acquisition data of each echo, and using each corrected echo image as a final magnetic resonance image of each echo comprises:

mapping the magnetic resonance image of each echo to k-space to obtain first complete k-space data of each echo; the first full k-space data comprises k-space data of a sampled region and k-space data of a non-sampled region;

replacing k-space data of a sampling region in first complete k-space data of a corresponding echo by using the k-space acquisition data of each echo to obtain second complete k-space data of each echo;

and carrying out image reconstruction according to the second complete k-space data of each echo to obtain a final magnetic resonance image of each echo.

4. The method according to claim 2, wherein the down-sampling mode is a half-fourier sampling mode, and the modifying the magnetic resonance image of the corresponding echo by using the k-space acquisition data of each echo respectively and using each modified echo image as the final magnetic resonance image of each echo specifically comprises:

mapping the magnetic resonance image of each echo to k-space to obtain first complete k-space data of each echo; the first full k-space data comprises k-space data of a sampled region and first k-space data of a non-sampled region;

setting the first k-space data of the non-sampling region in the first complete k-space data of each echo to be 0 to obtain third complete k-space data of each echo;

respectively utilizing the third complete k-space data of each echo to carry out data reconstruction by a half Fourier reconstruction method so as to obtain second k-space data of a non-sampling region of each echo;

carrying out weighted average on the first k-space data of the non-sampling region of each echo and the second k-space data of the non-sampling region to obtain third k-space data of the non-sampling region of each echo;

adding the third k-space data of the non-sampling region of each echo and the k-space acquisition data of the corresponding echo to obtain fourth complete k-space data of each echo;

and performing image reconstruction according to the fourth complete k-space data of each echo to respectively obtain a final magnetic resonance image of each echo.

5. The method of claim 1, wherein the input training samples of the first deep neural network model are magnetic resonance images partially reconstructed from specific partial k-space data of a plurality of echoes,

the method for respectively and successively reconstructing the images of the k-space acquisition data of each echo by adopting an explicit analytic solution imaging method and an image reconstruction method of a first deep neural network model to obtain the magnetic resonance image specifically comprises the following steps:

partially reconstructing k-space acquired data of each echo by adopting an explicit analytic solution imaging method to obtain a first image of each echo;

and according to the first image and the first deep neural network model of each echo, carrying out complete reconstruction through the first deep neural network model to obtain a second image of each echo, wherein the second image of each echo is used as a magnetic resonance image of each echo.

6. The method of claim 5, wherein the down-sampling mode is a half-Fourier sampling mode,

after the obtaining of the first image of each echo and before the complete reconstruction by the first deep neural network model, the method further includes:

performing half Fourier reconstruction on the first image of each echo to obtain a third image of each echo;

the completely reconstructing the first deep neural network model according to the first image and the first deep neural network model of each echo to obtain the second image of each echo specifically includes:

and taking the third image of each echo as the input of a first deep neural network model, and completely reconstructing through the first deep neural network model to obtain a second image of each echo.

7. The method of any one of claims 1-6, wherein the subset of the magnetic resonance image is an image point in the magnetic resonance image, an image region corresponding to a structure with a tissue quantitative value difference smaller than a first threshold value, or an image region with a gray value difference smaller than a second threshold value.

8. The method according to any one of claims 1-6, further comprising:

pre-training the first deep neural network model,

the pre-training of the first deep neural network model specifically includes:

acquiring an output training sample and an input training sample of a training set;

iterating parameters in a first deep neural network model by using the input training sample and the output training sample, taking the input training sample as the input of the first deep neural network model in each iteration process, and obtaining a prediction result after passing through the first deep neural network model; and judging whether the structural similarity between the prediction result and the output training sample meets a preset condition, if so, stopping iteration, finishing the training of the first deep neural network model, and if not, adjusting the parameters of the first deep neural network model and continuing the next iteration process.

9. The method according to any one of claims 1-6, further comprising:

pre-training the second deep neural network model,

the pre-training of the second deep neural network model specifically includes:

and iterating parameters in the second deep neural network model by using the input training sample and the output training sample, taking each subset of each magnetic resonance image as the input of the second deep neural network model in each iteration process, obtaining a prediction result of the subset after passing through the second deep neural network model, respectively judging whether the average variance between the prediction result of each subset of each magnetic resonance image and the magnetic resonance quantification of the corresponding subset in the output training sample is smaller than a preset threshold value, stopping iteration if each average variance is smaller than the preset threshold value, finishing the training of the second deep neural network model, if at least one average variance is not smaller than the preset threshold value, adjusting the parameters of the second deep neural network model, and continuing the next iteration process.

10. An apparatus for quantitative magnetic resonance imaging, the apparatus comprising:

11. The apparatus of claim 10, further comprising:

and the image correction unit is used for correcting the magnetic resonance image of the corresponding echo by respectively utilizing the k-space acquisition data of each echo after the magnetic resonance image of each echo is obtained and before the magnetic resonance quantitative map of each echo is acquired, and taking each corrected echo image as the final magnetic resonance image of each echo.

12. The apparatus according to claim 11, wherein the image modification unit specifically comprises:

the first mapping subunit is used for mapping the magnetic resonance image of each echo to k-space to obtain first complete k-space data of each echo; the first full k-space data comprises k-space data of a sampled region and k-space data of a non-sampled region;

the replacing subunit is used for replacing the k-space data of the sampling region in the first complete k-space data corresponding to the echo by using the k-space acquisition data of each echo to obtain second complete k-space data of each echo;

and the first reconstruction subunit is used for reconstructing an image according to the second complete k-space data of each echo to obtain a final magnetic resonance image of each echo.

13. The apparatus of claim 11, wherein the down-sampling mode is a half-Fourier sampling mode,

the image correction unit specifically includes:

the second mapping subunit is used for mapping the magnetic resonance image of each echo to k-space to obtain first complete k-space data of each echo; the first full k-space data comprises k-space data of a sampled region and first k-space data of a non-sampled region;

the setting subunit is used for setting the first k-space data of the non-sampling region in the first complete k-space data of each echo to 0 to obtain third complete k-space data of each echo;

the second reconstruction subunit is used for respectively utilizing the third complete k-space data of each echo to carry out data reconstruction through a half Fourier reconstruction method so as to obtain second k-space data of a non-sampling area of each echo;

the first calculating subunit is used for performing weighted average on the first k-space data of the non-sampling region of each echo and the second k-space data of the non-sampling region to obtain third k-space data of the non-sampling region of each echo;

the second calculating subunit is used for adding the third k-space data of the non-sampling region of each echo and the k-space acquisition data of the corresponding echo to obtain fourth complete k-space data of each echo;

and the third reconstruction subunit is used for reconstructing an image according to the fourth complete k-space data of each echo to respectively obtain a final magnetic resonance image of each echo.

14. The apparatus according to any one of claims 10-13, further comprising:

a first model training unit for training the first deep neural network model in advance,

the first model training unit specifically includes:

the device comprises a first acquisition subunit, a second acquisition subunit and a third acquisition subunit, wherein the first acquisition subunit is used for acquiring an output training sample and an input training sample of a training set, and the output training sample is a plurality of magnetic resonance images obtained by reconstructing full-acquisition or super-full-acquisition k-space data of each echo; the input training sample is a magnetic resonance image obtained by partially reconstructing k-space data of a specific part or k-space data of the specific part; the specific partial k-space data is a specific proportion of k-space data selected from the full-acquisition or over-full-acquisition k-space data;

the first iteration subunit is used for iterating parameters in the first deep neural network model by using the input training sample and the output training sample, taking the input training sample as the input of the first deep neural network model in each iteration process, and obtaining a prediction result after passing through the first deep neural network model; and judging whether the structural similarity between the prediction result and the output training sample meets a preset condition, if so, stopping iteration, finishing the training of the first deep neural network model, and if not, adjusting the parameters of the first deep neural network model and continuing the next iteration process.

15. The apparatus according to any one of claims 10-13, further comprising:

a second model training unit for training the second deep neural network model in advance,

the second model training unit specifically includes:

the second acquisition subunit is used for acquiring an output training sample and an input training sample of the training set;

and the second iteration subunit is used for iterating parameters in the second deep neural network model by using the input training sample and the output training sample, taking a subset of a magnetic resonance image as the input of the second deep neural network model in each iteration process, obtaining a prediction result of the subset through the second deep neural network model, judging whether the average variance between the prediction result of the subset and the magnetic resonance quantification of the corresponding subset in the output training sample is smaller than a preset threshold value, if so, stopping iteration, finishing the training of the second deep neural network model, if not, adjusting the parameters of the second deep neural network model, and continuing the next iteration process.