CN111436936A - CT image reconstruction method based on MRI - Google Patents

Abstract

The invention discloses a CT image reconstruction method based on MRI, which comprises the following steps: 1) reconstructing MRI by using a deep learning network, training the deep learning network, acquiring undersampled k-space data of an object to be detected, and inputting the undersampled k-space data of the object to be detected into the trained deep learning network to acquire on-line MRI of the object to be detected; 2) CT images are reconstructed from MRI using a two-way generating countermeasure network. The invention has the following advantages: (1) MRI imaging speed is high; (2) the application range is wide, and the imaging device can be used for lung imaging and imaging of other parts of a human body; (3) CT images are obtained through MRI reconstruction, and ionizing radiation of CT examination is avoided; (4) the reconstructed CT images can also be used for radiation therapy planning, and PET attenuation correction.

Description

MRI-based CT image reconstruction method

technical field

The invention relates to the technical field of medical image processing, in particular to a CT image reconstruction method based on MRI.

Background technique

New coronary pneumonia is highly infectious and has a high fatality rate. Early detection, early diagnosis, early treatment, and early isolation are currently the most effective means of prevention and treatment. Compared with the limitations of nucleic acid examination, CT (computed tomography) examination is timely, accurate, fast, and has a high positive rate. The extent of lung lesions is closely related to clinical symptoms, so it has become the main reference for early screening and diagnosis of patients with new coronavirus pneumonia. in accordance with. According to the new coronavirus pneumonia diagnosis and treatment plan (trial version 6), multiple small patchy shadows and interstitial changes appeared in the early stage of new coronary pneumonia, especially in the extrapulmonary zone. It then develops into multiple ground-glass opacities and infiltration shadows in both lungs. In severe cases, pulmonary consolidation may occur, and pleural effusion is rare. From the initial evaluation of the CT scan at admission, to the understanding of the progression of the disease, to the cure and discharge, the patient had at least 2 CT examinations, and as many as 3 to 4 CT examinations. Due to the presence of ionizing radiation, people such as children and pregnant women are not suitable for CT examinations. Magnetic resonance imaging (MRI) has the advantages of high soft tissue contrast, no ionizing radiation, high resolution and arbitrary orientation tomography, and is an important technology in modern medical imaging. MRI is usually used as an important supplement to chest plain film and CT. It is of great help in identifying internal and external thoracic lesions, internal and external mediastinal lesions, and upper and lower diaphragmatic lesions, and to understand the origin of lesions. For the imaging examination of new coronary pneumonia, compared with CT, the main disadvantage of MRI lies in the slow imaging speed and poor display of the fine structures of the lungs.

SUMMARY OF THE INVENTION

In view of this, the present invention provides an MRI-based CT image reconstruction method with fast imaging speed and good display effect on lung fine structures, aiming at the above-mentioned problems of slow MRI imaging speed and poor display of lung fine structures.

The technical solution of the present invention is to provide an MRI-based CT image reconstruction method with the following steps, including the following steps:

1) Using deep learning network to reconstruct MRI, including the following steps:

Obtain the fully sampled offline k-space data of the sample object. The full sampling means that the k-space data collection satisfies the Nyquist sampling theorem, and the image of the sample object can be restored through the fully sampled k-space data. The offline k-space Data refers to k-space data obtained from magnetic resonance equipment;

Inverse Fourier transform is performed on the fully sampled offline k-space data to obtain a fully sampled offline multi-contrast MRI; the multi-contrast MRI refers to scanning with a variety of imaging sequences to obtain different contrasts, such as T1W, T2W, etc.;

Undersampling the fully sampled offline k-space data in k-space to obtain under-sampled offline k-space data. The undersampling means that the k-space data collection does not satisfy the Nyquist sampling theorem. Aliasing artifacts will be generated when image reconstruction is performed;

training a deep learning network according to the undersampled offline k-space data and the fully sampled offline multi-contrast MRI;

Obtain the undersampled k-space data of the object to be measured;

Input the under-sampled k-space data of the object to be measured into the trained deep learning network to obtain the online MRI of the object to be measured;

2) using a bidirectional generative adversarial network to reconstruct a CT image from an online MRI; the bidirectional generative adversarial network is composed of two generators and two discriminators, and the first generator G _A is mapped to the CT image by the online MRI, The second generator _GB maps the CT images to the online MRI, the discriminator includes a CT discriminator and an MRI discriminator, and the CT discriminator _{D CT} is used to distinguish the CT images generated by the first generator GA and real CT images, the MRI discriminator D _MRI is used to distinguish the MRI generated by the second generator G _B from the real MRI; the reconstruction of CT images from MRI using a bidirectional generative adversarial network includes the following steps:

Obtain unlabeled and unpaired MRI and CT images, respectively;

The real MRI is converted into the generated CT image GA _{( I MRI} ) by the generator GA;

Generate CT image GA ( _{I MRI} ) and then convert it into reconstructed MRI through generator _GB ;

The real CT image I _CT is converted into the generated MRI by the generator _GB ;

The generated MRI is then converted into _a reconstructed CT image GA ( _GB ( _{I CT} )) by the generator GA;

The generator network composed of the first generator G _A and the second generator G _B and the CT discriminator and the MRI discriminator form the discriminator network to confront each other, and constantly adjust the parameters. The final optimization makes the discriminant network unable to judge whether the output result of the generation network is not. True while minimizing the reconstruction loss ||G _B (G _A (I _MRI ))-I _MRI || and ||G _A (G _B (I _CT ))-I _CT ||.

By adopting the above method, the present invention has the following advantages compared with the prior art: (1) the imaging speed is fast; (2) the application range is wide, which can be used for lung imaging and imaging of other parts of the human body; (3) CT images obtained from MRI reconstruction avoid the ionizing radiation of CT examination; (4) CT images obtained by reconstruction can also be used for radiation therapy planning and PET (positron emission tomography) attenuation correction.

As an improvement, in step 2), the bidirectional generative adversarial network is a Wasserstein bidirectional generative adversarial network, and the Jensen–Shannon divergence in the bidirectional generative adversarial network is replaced by the Wasserstein distance, and the loss function is: λ ₁ ||G _B (G _A (I _MRI ))-I _MRI ||+λ ₂ ||G _A (G _B (I _CT ))-I _CT ||-D _MRI (G _B (I _CT ))-D _CT (G _A (I CT ) _MRI )), where λ ₁ and λ ₂ are regularization parameters that can be chosen empirically.

As an improvement, in step 2), the perceptual loss is added to the loss function, and the pre-trained VGG16 network is used as the feature extractor, and the loss function is:

其中λ₁、λ₂、λ₃和λ₄为正则化参数，可以根据经验选择。所述VGG16网络是图片分类任务中经典的深度学习模型，VGG是由Simonyan和Zisserman在文献《VeryDeepConvolutional Networks forLarge Scale Image Recognition》中提出的卷积神经网络模型，其名称来源于论文作者所在的牛津大学视觉几何组(Visual Geometry Group)的缩写，该模型参加2014年的ImageNet图像分类与定位挑战赛，取得了优异成绩：在分类任务上排名第二，在定位任务上排名第一。

Among them, λ ₁ , λ ₂ , λ ₃ and λ ₄ are regularization parameters, which can be selected according to experience. The VGG16 network is a classic deep learning model in image classification tasks. VGG is a convolutional neural network model proposed by Simonyan and Zisserman in the document "VeryDeepConvolutional Networks for Large Scale Image Recognition", whose name comes from the Oxford University where the author of the paper is located. Abbreviation for Visual Geometry Group, this model participated in the ImageNet Image Classification and Localization Challenge in 2014 and achieved excellent results: it ranked second in the classification task and first in the localization task.

As an improvement, in the step 1), the use of deep learning network to reconstruct MRI includes:

Obtain the fully sampled offline k-space data y ₀ of the sample object;

performing inverse Fourier transform on the fully sampled offline k-space data y ₀ to obtain a fully sampled offline multi-contrast magnetic resonance image x ₀ ;

Undersampling the fully sampled offline k-space data y ₀ in k-space to obtain under-sampled offline k-space data y ₁ ;

Perform high-pass filtering on the undersampled offline k-space data y ₁ to obtain y ₁ *h;

Train a deep learning network according to the high-pass filtered undersampled offline k-space data y ₁ *h and the fully sampled offline multi-contrast magnetic resonance image x ₀ ;

Obtain undersampled k-space data y ₂ of the object to be measured;

Perform high-pass filtering on the undersampled k-space data y ₂ to obtain y ₂ *h;

Input the high-pass filtered undersampled k-space data y ₂ *h of the object to be tested into the trained deep learning network, and obtain k-space filled k-space data y ₂ '*h;

Performing inverse high-pass filtering on the reconstructed k-space data y ₂ ′;

Inverse Fourier transform is performed on the reconstructed k-space data y ₂ ′ to obtain an online magnetic resonance image.

As an improvement, in the step 1), the use of deep learning network to reconstruct MRI includes:

Obtain the fully sampled multi-channel offline k-space data y ₀ of the sample object;

performing inverse Fourier transform on the fully sampled multi-channel offline k-space data y ₀ to obtain a fully sampled multi-channel offline multi-contrast magnetic resonance image x ₀ ;

Under-sampling the fully-sampled multi-channel offline k-space data y ₀ in k-space to obtain under-sampled multi-channel offline k-space data y ₁ ;

Perform high-pass filtering on the undersampled multi-channel offline k-space data y ₁ to obtain y ₁ *h;

According to the high-pass filtered undersampled multi-channel offline k-space data y ₁ *h and the fully sampled multi-channel offline multi-contrast magnetic resonance image x ₀ , a deep learning network is trained, as shown in FIG. 3 , Set up the deep learning network before and after parallel imaging respectively;

After two deep learning networks and parallel imaging processing, the k-space data is y ₁ '*h, inverse high-pass filtering and data consistency correction are performed on y ₁ '*h, and the k-space data obtained by sampling is replaced by the corresponding position. k-space data to ensure that the deep learning network only fills unsampled k-space data;

Perform inverse Fourier transform and root mean square operation on the k-space data after data consistency correction to obtain the final offline reconstructed magnetic resonance image;

Obtain multi-channel undersampled k-space data y ₂ of the object to be measured;

Perform high-pass filtering on the multi-channel undersampled k-space data y ₂ to obtain y ₂ *h;

Input the high-pass filtered multi-channel undersampled k-space data y ₂ *h of the object to be tested into the trained deep learning network to obtain k-space filled k-space data y ₂ '*h;

Performing inverse high-pass filtering on the reconstructed k-space data y ₂ ′;

Inverse Fourier transform and root mean square operation are performed on the reconstructed k-space data y ₂ ′ to obtain an online magnetic resonance image.

As an improvement, the parallel imaging method is one of GRAPPA or SPIRiT.

As an improvement, the deep learning network in the step 1) is composed of the k-space domain U-Net and the image domain U-Net, and the undersampled offline k-space data is first input into the k-space domain U-Net, and then the data consistency is carried out. Correction, obtain the magnetic resonance image through inverse Fourier transform, input the image domain U-Net, and then perform Fourier transform to obtain k-space data, and perform data consistency correction; the data consistency correction is obtained by sampling k The spatial data replaces the k-space data at the corresponding location, ensuring that the deep learning network is only populated with unsampled k-space data.

Description of drawings

Fig. 1 is the flow chart of the present invention utilizing deep learning network to reconstruct MRI;

2 is a schematic diagram of the present invention using a generative adversarial network to reconstruct CT images from MRI;

3 is a flowchart of reconstructing MRI using a deep learning network in Embodiment 1 of the present invention;

FIG. 4 is a structural diagram of a deep learning network in Embodiment 1 of the present invention.

1-MRI space, 2-CT space, 31-Generator G _A that generates CT from MRI, 32-Generator G _B that generates MRI from CT, 41- MRI discriminator, 42- CT discriminator, 11- Real MRI image, 12-generated CT image, 13-reconstructed MRI image; 21-real CT image, 22-generated MRI image, 23-reconstructed CT image.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited to these embodiments. The present invention covers any alternatives, modifications, equivalent methods and arrangements made within the spirit and scope of the present invention. In order to give the public a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details.

Compared with traditional MRI image reconstruction techniques, deep learning-based image reconstruction methods have great potential in shortening MRI scan time, accelerating imaging speed, and improving imaging quality. FIG. 1 is a flow chart of reconstructing MRI using a deep learning network according to the present invention. In the offline training process, the undersampled offline k-space data and fully sampled offline multi-contrast magnetic resonance images are used to train the deep learning network, and the reconstructed k-space data obtained by the deep learning network is subjected to inverse Fourier transform to obtain the reconstructed magnetic Resonance image. In the online testing process, the under-sampled k-space data of the object to be tested is input to the deep learning network, the reconstructed k-space data is output, and the reconstructed magnetic resonance image is obtained through inverse Fourier transform. The k-space is the Fourier-dual space of the rectangular coordinate space, that is, the frequency space of the Fourier transform, and is mainly used in the field of magnetic resonance imaging.

Optionally, the multi-contrast image includes a T1-weighted image, a T2-weighted image, and a proton density image, and the multi-contrast images have the same field of view and matrix size. Wherein, the T1-weighted image mainly highlights the longitudinal relaxation difference of the tissue in the sample object, and minimizes the influence of other characteristics of the tissue, such as lateral relaxation, on the image. The T2-weighted image primarily highlights lateral relaxation differences of tissue in the sample object. The proton density image mainly reflects the proton content difference of the tissue in the sample object.

FIG. 2 is a schematic diagram of reconstructing CT images from MRI using a generative adversarial network according to the present invention. 1-MRI space, 2-CT space, 31-Generator G _A that generates CT from MRI, 32-Generator G _B that generates MRI from CT, 41- MRI discriminator, 42- CT discriminator, 11- Real MRI image, 12-generated CT image, 13-reconstructed MRI image; 21-real CT image, 22-generated MRI image, 23-reconstructed CT image.

According to one embodiment, unlabeled unpaired MRI and CT images are acquired separately;

The real MRI image I _MRI is converted into _a generated CT image GA ( _{I MRI} ) by the generator GA;

Generate a CT image GA (I MRI ) and then convert it into _a reconstructed _{MRI image GB (GA (I MRI ) ) through} the generator _GB ;

Similarly, the real CT image I _CT is converted into a generated MRI image _GB (I _CT ) by the generator _GB ;

_The MRI image _GB ( _ICT ) is generated and then converted into _a reconstructed _CT image GA ( _GB (ICT)) by the generator GA;

The generator network and the discriminator network confront each other and continuously adjust the parameters, and the final optimization makes the discriminant network unable to judge whether the output result of the generation network is real.

During the online test, the MRI image is input, and the corresponding CT image is obtained through the generator GA _.

According to one embodiment, during the online test, the under-sampled k-space data of the object to be tested is input, the reconstructed magnetic resonance image is obtained through a deep learning network for reconstructing the magnetic resonance image, and the reconstructed magnetic resonance image is input into the generating adversarial network to obtain the final CT images.

According to one embodiment, the CT image reconstruction method includes the following steps:

Obtain the fully sampled offline k-space data y ₀ of the sample object;

performing inverse Fourier transform on the fully sampled offline k-space data y ₀ to obtain a fully sampled offline multi-contrast magnetic resonance image x ₀ ;

Undersampling the fully sampled offline k-space data y ₀ in k-space to obtain under-sampled offline k-space data y ₁ ;

Perform high-pass filtering on the undersampled offline k-space data y ₁ to obtain y ₁ *h;

Train a deep learning network according to the high-pass filtered undersampled offline k-space data y ₁ *h and the fully sampled offline multi-contrast magnetic resonance image x ₀ ;

Obtain undersampled k-space data y ₂ of the object to be measured;

Perform high-pass filtering on the undersampled k-space data y ₂ to obtain y ₂ *h;

Input the high-pass filtered undersampled k-space data y ₂ *h of the object to be tested into the trained deep learning network, and obtain k-space filled k-space data y ₂ '*h;

Performing inverse high-pass filtering on the reconstructed k-space data y ₂ ′;

Inverse Fourier transform is performed on the reconstructed k-space data y ₂ ′ to obtain an online magnetic resonance image.

The online magnetic resonance image is input into a generative adversarial network to obtain a final CT image.

According to one embodiment, for multi-channel magnetic resonance imaging, the MRI imaging step combining parallel imaging and deep learning includes:

Obtain the fully sampled multi-channel offline k-space data y ₀ of the sample object;

performing inverse Fourier transform on the fully sampled multi-channel offline k-space data y ₀ to obtain a fully sampled multi-channel offline multi-contrast magnetic resonance image x ₀ ;

Under-sampling the fully-sampled multi-channel offline k-space data y ₀ in k-space to obtain under-sampled multi-channel offline k-space data y ₁ ;

Perform high-pass filtering on the undersampled multi-channel offline k-space data y ₁ to obtain y ₁ *h;

According to the high-pass filtered undersampled multi-channel offline k-space data y ₁ *h and the fully sampled multi-channel offline multi-contrast magnetic resonance image x ₀ , a deep learning network is trained, as shown in FIG. 3 , Set up the deep learning network before and after parallel imaging respectively;

After two deep learning networks and parallel imaging processing, the k-space data is y ₁ '*h, inverse high-pass filtering and data consistency correction are performed on y ₁ '*h, and the k-space data obtained by sampling is replaced by the corresponding position. k-space data to ensure that the deep learning network only fills unsampled k-space data;

Perform inverse Fourier transform and root mean square operation on the k-space data after data consistency correction to obtain the final offline reconstructed magnetic resonance image;

Obtain multi-channel undersampled k-space data y ₂ of the object to be measured;

Perform high-pass filtering on the multi-channel undersampled k-space data y ₂ to obtain y ₂ *h;

Input the high-pass filtered multi-channel undersampled k-space data y ₂ *h of the object to be tested into the trained deep learning network to obtain k-space filled k-space data y ₂ '*h;

Performing inverse high-pass filtering on the reconstructed k-space data y ₂ ′;

Inverse Fourier transform and root mean square operation are performed on the reconstructed k-space data y ₂ ′ to obtain an online magnetic resonance image.

The online magnetic resonance image is input into a generative adversarial network to obtain a final CT image.

According to one embodiment, as shown in Figure 4, the deep learning network includes a k-space domain U-Net and an image domain U-Net, and the undersampled offline k-space data is first input into the k-space domain U-Net, and then data consistency is performed Correction, obtain the magnetic resonance image through inverse Fourier transform, input the image domain U-Net, and then perform Fourier transform to obtain k-space data, and perform data consistency correction; the data consistency correction is obtained by sampling k The spatial data replaces the k-space data at the corresponding location, ensuring that the deep learning network is only populated with unsampled k-space data.

Based on the above reconstruction method, an MRI-based CT image reconstruction system can be formed.

The above only describes the preferred embodiments of the present invention, but should not be construed as limiting the claims. The present invention is not limited to the above embodiments, and the specific structure thereof can be changed. In a word, all changes made within the protection scope of the independent claims of the present invention are all within the protection scope of the present invention.

Claims (7)

1. An MRI-based CT image reconstruction method, comprising the steps of:

1) reconstructing an MRI using a deep learning network, comprising the steps of:

acquiring fully sampled offline k-space data of a sample object, wherein the fully sampling means that the k-space data acquisition satisfies the Nyquist sampling theorem, and an image of the sample object can be restored by the fully sampled k-space data, and the offline k-space data means k-space data acquired from a magnetic resonance device;

performing inverse Fourier transform on the fully sampled offline k-space data to obtain fully sampled offline multi-contrast MRI, wherein the multi-contrast MRI refers to scanning by using multiple imaging sequences to obtain different contrasts;

undersampling the fully sampled offline k-space data in k-space to obtain undersampled offline k-space data, wherein undersampled means that the k-space data acquisition does not meet the Nyquist sampling theorem and aliasing artifacts are generated when the undersampled k-space data are directly used for image reconstruction;

training a deep learning network according to the undersampled offline k-space data and the fully sampled offline multi-contrast MRI;

acquiring undersampled k-space data of an object to be detected;

inputting the undersampled k-space data of the object to be detected into a trained deep learning network to obtain an on-line MRI of the object to be detected;

2) reconstructing a CT image from an on-line MRI using a two-way generative countermeasure network; the bidirectional generation countermeasure network is composed of two generators and two discriminators, the first generator G_AFor mapping from an on-line MRI to a CT image, a second generator G_BFor mapping from a CT image to an on-line MRI, the discriminators include a CT discriminator and an MRI discriminator_CTFor distinguishing by the first generator G_AGenerated CT image and real CT image, MRI discriminator D_MRIFor distinguishing by the second generator G_BGenerated MRI and real MRI; reconstructing a CT image from MRI using a bi-directional generation countermeasure network includes the steps of:

respectively acquiring unmarked and unpaired MRI and CT images;

true MRI pass generator G_AConversion to generate CT image G_A(I_MRI)；

Generating a CT image G_A(I_MRI) Re-pass generator G_BConversion to reconstructed MRI;

true CT image I_CTThrough generator G_BConversion to generating an MRI;

generating MRI and then passing through generator G_AConversion into reconstructed CT image G_A(G_B(I_CT))；

First generator G_AAnd a second generator G_BThe formed generator network, the CT discriminator and the MRI discriminator form the discriminator network to resist each other and continuously adjust parameters, finally, the discrimination network can not judge whether the output result of the generator network is real or not through optimization, and meanwhile, the reconstruction loss G is minimized_B(G_A(I_MRI))-I_MRII and G_A(G_B(I_CT))-I_CT||。

2. The MRI-based CT image reconstruction method of claim 1, wherein: in step 2), the bidirectional generation countermeasure network is a Wasserstein bidirectional generation countermeasure network, Wasserstein distance is used for replacing Jensen-Shannon divergence in the bidirectional generation countermeasure network, and the loss function is as follows:

λ₁||G_B(G_A(I_MRI))-I_MRI||+λ₂||G_A(G_B(I_CT))-I_CT||-D_MRI(G_B(I_CT))-D_CT(G_A(I_MRI) Where λ) is₁And λ₂Is a regularization parameter.

3. The MRI-based CT image reconstruction method according to claim 2, wherein: in step 2), adding the perceptual loss in a loss function, and using a pre-trained VGG16 network as a feature extractor, wherein the loss function is as follows:

wherein λ₁、λ₂、λ₃And λ₄Is a regularization parameter.

4. The MRI-based CT image reconstruction method of claim 1, wherein: the step 1) of reconstructing MRI by using the deep learning network comprises the following steps:

acquiring fully sampled off-line k-space data y of a sample object₀；

For the fully sampled offline k-space data y₀Performing inverse Fourier transform to obtain fully sampled offline multi-contrast magnetic resonance image x₀；

K-space data y under a line of k-space for said full sampling₀Undersampling to obtain undersampled offline k-space data y₁；

For the undersampled offline k-space data y₁High-pass filtering to obtain y₁*h；

From the high-pass filtered undersampled offline k-space data y₁H and the fully sampled under-line multi-contrast magnetic resonance image x₀Training a deep learning network;

obtaining undersampled k-space data y of an object to be measured₂；

For the undersampled k-space data y₂High-pass filtering to obtain y₂*h；

The undersampled k-space data y after the high-pass filtering of the object to be detected₂Inputting h into the trained deep learning network to obtain k space data y after k space filling₂’*h；

Carrying out inverse high-pass filtering on the k space data to obtain reconstructed k space data y₂’；

For the reconstructed k-space data y₂' inverse Fourier transform to obtain an in-line magnetic resonance image.

5. The MRI-based CT image reconstruction method according to claim 1 or 4, wherein the reconstructing MRI using the deep learning network in the step 1) comprises:

acquiring fully sampled multi-channel offline k-space data y of a sample object₀；

For the fully sampled multi-channel offline k-space data y₀Performing inverse Fourier transform to obtain fully sampled multi-channel offline multi-contrast magnetic resonance image x₀；

K-space data y under k-space versus the fully sampled multi-channel line₀Undersampling to obtain undersampled multi-channel offline k-space data y₁；

For the undersampled multi-channel offline k-space data y₁High-pass filtering to obtain y₁*h；

From the high-pass filtered undersampled multi-channel offline k-space data y₁H and the fully sampled multi-channel offline multi-contrast magnetic resonance image x₀Training a deep learning network, and respectively arranging the deep learning networks before and after parallel imaging as shown in FIG. 3;

the k space data after two deep learning networks and parallel imaging processing is y₁' h, pair y₁' inverse high-pass filtering and data consistency correction are carried out, and k space data at corresponding positions are replaced by k space data obtained by sampling, so that the deep learning network is ensured to be only filled with the k space data which are not sampled;

performing inverse Fourier transform and root mean square operation on the k space data after data consistency correction to obtain a final offline reconstructed magnetic resonance image;

acquiring multi-channel undersampled k-space data y of object to be detected₂；

For the multi-channel undersampled k-space data y₂High-pass filtering to obtain y₂*h；

High-pass filtering the multi-channel undersampled k-space data y of the object to be detected₂Inputting h into the trained deep learning network to obtain k space data y after k space filling₂’*h；

Carrying out inverse high-pass filtering on the k space data to obtain reconstructed k space data y₂’；

For the reconstructed k-space data y₂' inverse Fourier transform and root mean square operations are performed to obtain an in-line magnetic resonance image.

6. The MRI-based CT image reconstruction method of claim 5, wherein: the parallel imaging method is one of GRAPPA or SPIRiT.

7. The MRI-based CT image reconstruction method according to any one of claims 1 to 3, wherein: the deep learning network in the step 1) is composed of a k-space domain U-Net and an image domain U-Net, undersampled off-line k-space data is firstly input into the k-space domain U-Net, then data consistency correction is carried out, a magnetic resonance image is obtained through inverse Fourier transformation, the image domain U-Net is input, then Fourier transformation is carried out to obtain k-space data, and data consistency correction is carried out; the data consistency correction ensures that the deep learning network only fills the non-sampled k-space data by replacing the k-space data at the corresponding position with the k-space data obtained by sampling.

Images