PET / MR attenuation correction method and system

Abstract

The invention relates to a PET / MR attenuation correction method and system, and the method comprises the steps: S1, obtaining a Dixon FLASH sequence, carrying out the noise reduction reconstruction of the Dixon FLASH sequence through a depth reconstruction technology, and obtaining an MRI image through the magnetic resonance reconstruction decoding; s2, based on the MRI image, lung and human body contour segmentation is carried out to obtain lung tissue and air tissue, sCT image synthesis is carried out by adopting a neural network to obtain bone tissue, and soft tissue and adipose tissue are obtained through a Dixon water-fat separation technology; s3, respectively endowing specific attenuation coefficients to the masks of the five tissues, and synthesizing into an attenuation coefficient graph [mu] MAP for attenuation correction; and S4, carrying out attenuation correction on the PET raw data by adopting the attenuation coefficient graph [mu] MAP to obtain a final PET / MR imaging result. Compared with the prior art, the method has the advantages of high attenuation correction precision and high image acquisition efficiency.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular, to a PET / MR attenuation correction method, system, device and medium. Background Art

[0002] PET / MR is a technology that combines positron emission tomography (PET) and magnetic resonance imaging (MRI). Among them, PET mainly undertakes the function of detecting biomolecules in the body, while MRI mainly plays the role of generating high-quality anatomical images, so as to better clarify the anatomical structure of abnormal uptake regions in PET images. In PET / MR, attenuation correction is a very crucial link, which can effectively improve the accuracy of PET quantification.

[0003] Currently, the tissue segmentation method based on Dixon Flash sequence is the mainstream technical route. However, this method faces many challenges, such as interference from respiratory motion artifacts, low imaging efficiency, inability to completely segment the whole body bones, and inability to distinguish bones with different densities. Summary of the Invention

[0004] The purpose of the present invention is to provide a PET / MR attenuation correction method and system with high attenuation correction accuracy and high image acquisition efficiency to overcome the defects of the above-mentioned existing technologies.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] According to the first aspect of the present invention, a PET / MR attenuation correction method is provided, and the method includes:

[0007] S1. Obtain the Dixon FLASH sequence, perform noise reduction and reconstruction on the Dixon FLASH sequence by using depth reconstruction technology, and obtain the MRI image through magnetic resonance reconstruction decoding;

[0008] S2. Based on the MRI image, perform lung and human body contour segmentation to obtain lung tissue and air tissue, use a neural network to synthesize the sCT image to obtain bone tissue, and obtain soft tissue and fat tissue through Dixon water-fat separation technology;

[0009] S3. Assign specific attenuation coefficients to the masks of the five obtained tissues respectively, and synthesize them into an attenuation coefficient map μMAP for attenuation correction;

[0010] S4. Use the attenuation coefficient map μMAP to perform attenuation correction on the PET raw data to obtain the final PET / MR imaging result.

[0011] Preferably, for obtaining the Dixon FLASH sequence, deep reconstruction technology is used to denoise and reconstruct the Dixon FLASH sequence. Specifically: The AI-assisted magnetic resonance compressed sensing reconstruction technology is used to obtain the original Dixon FLASH sequence, and deep reconstruction technology is used to denoise and reconstruct the Dixon FLASH sequence. Among them, during the process of denoising and reconstructing the Dixon FLASH sequence, a data fidelity module is embedded in the reconstruction chain. When performing network training, the loss function includes both image domain and k-space domain information, and the data fidelity module is constrained by the image domain and k-space domain information. At the same time, based on the k-space domain, intelligent verification and matrix fusion are performed on the output result and the original data, and iterative optimization is carried out through quadratic approximation.

[0012] Preferably, during the process of denoising and reconstructing the original Dixon FLASH sequence by the deep reconstruction technology, residual learning is performed using a CNN network.

[0013] Preferably, obtaining the MRI image by magnetic resonance reconstruction decoding specifically includes:

[0014] Preprocessing the collected original Dixon FLASH sequence, including removing noise and correcting the phase;

[0015] Performing Fourier transform on the preprocessed signal to convert the signal from the time domain to the frequency domain;

[0016] Reconstructing the signal converted to the frequency domain into a first image for doctors to read using an image reconstruction algorithm;

[0017] Performing post-processing of enhancement and segmentation on the reconstructed first image to obtain the MRI image to be attenuation-corrected.

[0018] Preferably, the image reconstruction algorithm includes filtered back projection and iterative reconstruction algorithms.

[0019] Preferably, for performing lung and human body contour segmentation to obtain lung tissue and air tissue, specifically: Using the threshold segmentation method for lung and human body contour segmentation, and using a U-net network for lung and air segmentation to obtain lung tissue and air tissue.

[0020] Preferably, for synthesizing the sCT image using a neural network to obtain bone tissue, specifically:

[0021] Inputting the in-phase MRI image and the opposed-phase MRI image into a 3D U-net network for sCT image synthesis, outputting the sCT image to obtain bone tissue; among them, the loss function of the 3D U-net network is the residual between the sCT image and the CT image.

[0022] Preferably, the process of obtaining the training data of the 3D U-net network is specifically as follows:

[0023] Based on the AICS and DRecon technologies, MR images are acquired through the free-breathing method with the arms lowered.

[0024] CT images of the whole body are acquired through the free-breathing method with the arms lowered on a flat bed.

[0025] The mutual information algorithm is used to match the MR images and CT images for subsequent setting of the attenuation coefficient.

[0026] Preferably, the method of obtaining soft tissue and fat tissue through the Dixon water-fat separation technology is specifically as follows: Soft tissue and fat tissue are extracted from in-phase MRI images and opposed-phase MRI images through the Dixon water-fat separation technology.

[0027] According to the second aspect of the present invention, a system using the PET / MR attenuation correction method is provided, including:

[0028] An acquisition module, configured to obtain the Dixon FLASH sequence, perform noise reduction and reconstruction on the Dixon FLASH sequence using the depth reconstruction technology, and obtain MRI images through magnetic resonance reconstruction decoding.

[0029] A segmentation module, configured to segment the lungs and the human body contour based on the MRI images to obtain lung tissue and air tissue, synthesize sCT images using a neural network to obtain bone tissue, and obtain soft tissue and fat tissue through the Dixon water-fat separation technology.

[0030] An attenuation coefficient map synthesis module, configured to assign specific attenuation coefficients to the masks of the five obtained tissues respectively, and synthesize an attenuation coefficient map μMAP for attenuation correction.

[0031] An attenuation correction imaging module, configured to perform attenuation correction on the PET raw data using the attenuation coefficient map μMAP to obtain the final PET / MR imaging result.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) In the present invention, aiming at the different imaging characteristics of the five tissues, the lungs and the human body contour are segmented based on the threshold method to obtain lung tissue and air tissue, sCT images are synthesized using a neural network to obtain bone tissue, and soft tissue and fat tissue are obtained through the Dixon water-fat separation technology. The segmentation results of the five tissues are more accurate and reliable. Then, an attenuation coefficient map is synthesized by assigning specific attenuation coefficients to the masks of the five tissues respectively, and the PET raw data is corrected for attenuation using the attenuation coefficient map, which has the advantage of high attenuation correction accuracy.

[0034] (2) The present invention improves the AI-assisted magnetic resonance compressed sensing reconstruction technology and deep reconstruction technology, realizes the efficient acquisition of Dixon Flash sequence under free breathing, can complete the acquisition of Dixon FLASH sequence of the chest and abdomen within 14 s, and the quality of the generated μMAP image is equivalent to that of the μMAP obtained by traditional 300 s acquisition, meeting the requirements of high imaging rate and high imaging accuracy at the same time.

[0035] (3) The present invention uses the AI-assisted magnetic resonance compressed sensing reconstruction technology for image acquisition. On the basis of sparse image acquisition, it can further reduce the requirements for K-space data. Through the multi-channel sensitivity of parallel imaging technology and the phase constraint of the half-Fourier technique, iterative verification is carried out on image acquisition and reconstruction, so as to ensure the authenticity and accuracy of the image.

[0036] (4) The present invention embeds a data fidelity module in the reconstruction chain through the deep reconstruction technology. The loss function includes both image domain and k-space domain information, and the two constrain the intelligent module to strictly ensure the authenticity of the network output result. In order to further ensure the accuracy of the result, based on the k-space domain, intelligent verification and matrix fusion are carried out on the output result and the original data, and iterative optimization is carried out through quadratic approximation, so as to achieve double optimization and fidelity.

[0037] (5) The present invention trains the sCT image synthesis neural network with the MR images collected by the free breathing method with the arms down and the whole-body CT images collected by the free breathing method with the arms down on a flat bed. The acquisition method with the arms down ensures the consistency of the positions of the MR images and the CT images, and only the mutual information algorithm can be used to achieve high-precision image registration subsequently.

[0038] (6) Since the gray values of the air in the lungs and the human periphery are much lower than those of human soft tissues, the threshold segmentation method is used to segment the lungs and the human body contour on the Dixon IP image. And the gas in the gastrointestinal tract and some calcified tissues also show low signals on magnetic resonance. Better lung segmentation and air segmentation effects can be obtained using the U-net network. Description of the Drawings

[0039] Figure 1 is the flowchart of the method of the present invention;

[0040] Figure 2 is the schematic diagram of the MRI image generation process for attenuation correction;

[0041] Figure 3 is the schematic diagram of the AICS neural network structure in the embodiment;

[0042] Figure 4 is the schematic diagram of the 3D U-net network structure in the embodiment;

[0043] Figure 5 Schematic diagram of the attenuation coefficient map generation process in the embodiment;

[0044] Figure 6 Schematic diagram of the acquisition process of the training data of the sCT image synthesis network in the embodiment;

[0045] Figure 7 Schematic diagrams of three MR Dixon FLASH sequence acquisition schemes for PET / MR attenuation correction in the embodiment;

[0046] Figure 8 An example of using the Dixon FLASH sequence in Scheme 2 for PET / MR attenuation correction in the embodiment. Specific implementation manners

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment

[0049] First, some English abbreviations used in this embodiment are explained.

[0050] PET: Positron Emission Tomography

[0051] MRI: Magnetic Resonance Imaging

[0052] PET / MR: Positron Emission Tomography / Magnetic Resonance

[0053] sCT: Synthetic Computed Tomography

[0054] Flash: Fast Low Angle Shot

[0055] Dixon: Dixon technique

[0056] DRecon: Deep Reconstruction

[0057] AICS: AI Artificial Intelligence Assisted Compressed Sensing

[0058] As Figure 1 and Figure 5 shown, this embodiment provides a PET / MR attenuation correction method, and the method includes:

[0059] S1. Obtain a Dixon FLASH sequence, perform noise reduction reconstruction on the Dixon FLASH sequence by using a deep reconstruction technique, and obtain an MRI image through magnetic resonance reconstruction decoding;

[0060] S2. Based on the MRI image, perform lung and human body contour segmentation to obtain lung tissue and air tissue, use a neural network to synthesize sCT images to obtain bone tissue, and obtain soft tissue and fat tissue through Dixon water-fat separation technology;

[0061] S3. Assign specific attenuation coefficients to the masks of the five obtained tissues respectively, and synthesize them into an attenuation coefficient map μMAP for attenuation correction;

[0062] S4. Use the attenuation coefficient map μMAP to perform attenuation correction on the PET raw data to obtain the final PET / MR imaging result.

[0063] Next, the method of this embodiment will be introduced in detail.

[0064] S1. Obtain the Dixon FLASH sequence, use the depth reconstruction technology to perform noise reduction reconstruction on the original Dixon FLASH sequence, and obtain the MRI image through magnetic resonance reconstruction decoding. The specific process is as Figure 2 shown.

[0065] Regarding the Dixon FLASH sequence:

[0066] The FLASH (Fast Low Angle Shot) sequence is a commonly used gradient echo sequence, which requires signal acquisition when the transverse magnetization vector reaches a stable equilibrium state. Dixon is a water-fat separation technology, and its mechanism is based on the chemical shift effect. There is a 3.5 ppm chemical shift between water and fat. This technology uses the images of hydrogen protons with different phases in water and fat, and then uses post-processing methods to obtain images that respectively reflect water signals and fat signals.

[0067] Image acquisition link: Use the AI-assisted magnetic resonance compressed sensing reconstruction technology AICS to obtain the Dixon FLASH sequence. The AI-assisted magnetic resonance compressed sensing reconstruction technology AICS combines compressed sensing technology, parallel acquisition technology, half-Fourier acquisition technology, and AI technology. On the basis of sparse image acquisition, it further reduces the requirements for K-space data, and can increase the acquisition acceleration multiple to 10 times. The deep learning module inside it pre-trains the undersampled K-space data to empower image reconstruction. Through the multi-channel sensitivity of the parallel imaging technology and the phase constraint of the half-Fourier technology, the acquisition and reconstruction of the image are iteratively verified to ensure the authenticity and accuracy of the image.

[0068] Specifically, the AI-assisted magnetic resonance compressed sensing reconstruction technology AICS integrates the deep learning network and the compressed sensing technology and is used for the reconstruction of magnetic resonance images, which can be understood as the following unconstrained optimization problem:

[0069]

[0070] Wherein: is a data fidelity term; F u x ∈ C M×N is an undersampled Fourier encoding matrix; R represents a regularization term on x; λ is a regularization parameter. The regularization term R usually involves an l q norm (0 ≤ q ≤ 1) in the sparsity domain of x. f cnn is the forward propagation of data through a convolutional neural network (CNN) parameterized by θ, and ζ is another regularization parameter. The image generated by the convolutional neural network is used as a reference image and also as an additional regularization term, where represents the optimized parameters of the trained convolutional neural network. In addition, is the reconstruction of the zero-padded undersampled k-space, where H represents the Hermitian transpose operation.

[0071] This embodiment gives a deep learning network structure of AICS, as Figure 3 shown: The overall structure is a conditional generative adversarial network, where the generator uses U-net, or an improved U-net can also be used. Four aspects of loss functions can impose strong constraints to control the quality of image generation: The Pixel-wise Loss ensures the visual effect of the human eye, the Frequency Domain Loss ensures the frequency domain consistency, the Perceptual Loss ensures the feature alignment of the generated image in the deep perception layer of the neural network, and the adversarial loss uses the way of network game to ensure the effect of the generated data.

[0072] Image reconstruction link: Adopt the deep reconstruction technology DRecon to suppress image artifacts and enhance image details, perform noise reduction reconstruction on the Dixon FLASH sequence, and decode the filtered raw data through magnetic resonance reconstruction to obtain the MRI image for attenuation correction.

[0073] To ensure the authenticity and reliability of the details in the images after deep learning, DRecon embeds a data fidelity module within the reconstruction chain. Compared with conventional deep learning algorithms, when training the network, the loss function of DRecon technology (i.e., the evaluation of the error rate for the training results) includes both image domain and k-space domain information, and these two constrain the intelligent module to strictly ensure the authenticity of the network output results. To further ensure the accuracy of the results, based on the k-space domain, DRecon uses the output result and the original data to conduct intelligent verification and matrix fusion, performing secondary approximation and iterative optimization, thus achieving double optimization and fidelity. Moreover, this technology conducts two-level fidelity control in terms of contrast and image structure to verify the reliability of the final output results again.

[0074] In this embodiment, DRecon selects a CNN network for residual learning, that is, the output of the network is not a noise-free image, but the noise of each pixel. Therefore, the setting of the network loss function is as follows:

[0075]

[0076] In the formula: represents N pairs of noise-free training images, and R(y) represents the image residual (noise).

[0077] In this embodiment, a possible implementation of a CNN network with residual layers is given. The network structure includes a convolutional input layer with a rectified linear unit (ReLU), eight convolutional layers with batch normalization and ReLU, and a convolutional output layer. The image generated by the output layer has the same size as the input image. This network does not contain any max pooling layers because the output should have the same size as the input.

[0078] The MRI image is obtained through magnetic resonance reconstruction decoding, specifically including:

[0079] 1) Preprocess the collected Dixon FLASH sequence, including removing noise and correcting the phase;

[0080] 2) Perform Fourier transform on the preprocessed signal to convert the signal from the time domain to the frequency domain;

[0081] 3) Reconstruct the signal in the frequency domain into the first image for doctors to read using an image reconstruction algorithm; the image reconstruction algorithm includes filtered back projection and iterative reconstruction algorithms.

[0082] 4) Perform post-processing of enhancement and segmentation on the reconstructed first image to obtain the MRI image to be attenuation-corrected.

[0083] S2. Based on the MRI images, perform lung and human body contour segmentation to obtain lung tissue and air tissue, use a neural network to synthesize sCT images to obtain bone tissue, and obtain soft tissue and fat tissue through Dixon water-fat separation technology, as Figure 5 shown, specifically including:

[0084] S21. Perform lung and human body contour segmentation through Network 1 to obtain lung tissue and air tissue, specifically including:

[0085] Since the gray values of the lungs and the air outside the human body are much lower than those of human soft tissues, in this embodiment, the threshold segmentation method is used to segment the lungs and the human body contour on the in-phase MRI image (Dixon IP). However, the gas in the gastrointestinal tract and some calcified tissues also show low signals on magnetic resonance. Using a U-net or improved U-net network can achieve better lung segmentation and air segmentation effects.

[0086] S22. Input the in-phase MRI image and the opposed-phase MRI image into a 3D U-net network (Network 2) to perform sCT image synthesis and output the sCT image to obtain bone tissue (including cortical bone and cancellous bone); among them, the loss function of the 3D U-net network is the residual between the sCT image and the CT image.

[0087] In this embodiment, the sCT image synthesis network selects a 3D U-net network, whose input is the in-phase IP image and the opposed-phase OP image of the Dixon sequence, and the output is the sCT image.

[0088] Define the loss function of the sCT image synthesis network as the residual between the sCT image and the CT image:

[0089]

[0090] In the formula: represents N pairs of noise-free training images, represents the generated sCT.

[0091] Figure 4 shows the structural schematic of the 3D U-net network. Input the in-phase MRI image and the opposed-phase MRI image, which are integrated by the first convolutional layer with 32 different kernels. Each of the encoder and decoder paths contains 3×3×3 kernels, followed by batch normalization (BN) and rectified linear unit (ReLU) activation functions for faster convergence. At the end of each convolutional operation, a similar convolutional layer with a stride of 2 is connected for downsampling.

[0092] The above process of obtaining the training data of the 3D U-net network, as Figure 6As shown in the figure, specifically: based on AICS and DRecon technologies, MR images are acquired through the free breathing method with the arms lowered; CT images of the whole body are acquired through the free breathing method with the arms lowered on a flat bed; the mutual information algorithm is used to match the MR images and CT images for subsequent setting of the attenuation coefficient. Since the free breathing method with the arms lowered is used for image acquisition in both cases, and the bed boards are flat bed boards, the image registration link in network training is greatly simplified, ensuring the image quality of sCT in the ribs and scapulae.

[0093] Specifically, since the bed surface of the MR examination bed is flat while the bed surface of the CT examination bed is arc-shaped, large-area non-rigid registration of the chest and abdomen is required to achieve registration of the two modalities, and it is difficult to ensure the accuracy of image registration. In this embodiment, the radiotherapy simulation positioning CT images with the arms lowered are used. Since a flat radiotherapy bed board is used, the back of this CT image is flat and has the same body position as the MR image. On this basis, the classic mutual information algorithm can be used to achieve accurate matching of MR and CT. By continuously adjusting the transformation parameters, the mutual information of the corresponding anatomical structures in the two images is maximized, thereby achieving accurate registration.

[0094] S23. Soft tissue and adipose tissue are obtained through the Dixon water-fat separation technique.

[0095] S3. Specific attenuation coefficients are assigned to the masks of the five obtained tissues respectively, and they are synthesized into an attenuation coefficient map μMAP for attenuation correction.

[0096] In this embodiment, the Dixon FLASH sequence obtained by scanning is not used for direct diagnosis, but for generating the attenuation coefficient map μMAP.

[0097] The process of obtaining the attenuation coefficient μ value is specifically as follows: The MR image is registered to the CT image for acquisition. An example of the μ value is given in this embodiment: air 0 cm -1 , lungs 0.024 cm -1 , fat 0.086 cm -1 , soft tissue 0.101 cm -1 , bone 0.146 cm -1 .

[0098] S4. The attenuation coefficient map μMAP is used to perform attenuation correction on the PET raw data to obtain the final PET / MR imaging result.

[0099] In this embodiment, the device model is uPMR890 integrated TOF PET / MR. Among them, the phase angle between the two echoes differs by 150°, the phase encoding direction is the anteroposterior direction AP, the in-plane resolution is 1.5 mm * 1.5 mm, and the slice thickness THK is 2 mm. For specific parameter settings, please refer to Table 1 below.

[0100] Table 1

[0101] Repetition time TR 4.88 First echo time TE1 2.24 Second echo time TE2 3.19 Bandwidth BW 1080Hz Flip angle FA 10 ACS 10 Matrix MATRIX 329*256 Field of view FOV 500*350 Slice thickness THK 2mm

[0102] Figure 7 Schematic diagrams of three Dixon FLASH sequence acquisition schemes for PET / MR attenuation correction, where:

[0103] Scheme 1 uses a breath-holding method for acquisition, with good image quality, but is not suitable for patients who cannot cooperate with breath-holding; Scheme 2 uses the AI fast acquisition method of the present invention, without requiring the patient to hold their breath, but the contour of the liver top is clear; Scheme 3 uses a respiratory gating method for acquisition, with the best image quality, but takes too long. Figure 8 It is an example of using the Dixon FLASH sequence in Scheme 2 for PET / MR attenuation correction.

[0104] This embodiment also provides a PET / MR attenuation correction system, including:

[0105] An acquisition module for obtaining the Dixon FLASH sequence, denoising and reconstructing the Dixon FLASH sequence using depth reconstruction technology, and obtaining an MRI image through magnetic resonance reconstruction decoding;

[0106] A segmentation module for segmenting the lungs and human body contours based on the MRI image to obtain lung tissue and air tissue, synthesizing sCT images using a neural network to obtain bone tissue, and obtaining soft tissue and fat tissue through Dixon water-fat separation technology;

[0107] An attenuation coefficient map synthesis module for respectively assigning specific attenuation coefficients to the masks of the five obtained tissues and synthesizing an attenuation coefficient map μMAP for attenuation correction;

[0108] An attenuation correction imaging module for performing attenuation correction on the PET raw data using the attenuation coefficient map μMAP to obtain the final PET / MR imaging result.

[0109] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A PET / MR attenuation correction method, characterized in that: The method includes: S1. Obtain a Dixon FLASH sequence, use deep reconstruction technology to reduce noise and reconstruct the Dixon FLASH sequence, and obtain an MRI image through magnetic resonance reconstruction and decoding; S2, based on MRI images, lung and human body contour segmentation is performed to obtain lung tissue and air tissue, sCT image synthesis is performed using a neural network to obtain bone tissue, and soft tissue and fat tissue are obtained through Dixon water-fat separation technology; S3, assigning specific attenuation coefficients to the masks of the five tissues respectively, and synthesizing them into an attenuation coefficient map μMAP for attenuation correction; S4. Use the attenuation coefficient map μMAP to perform attenuation correction on the PET raw data to obtain the final PET / MR imaging results.

2. A PET / MR attenuation correction method according to claim 1, characterized in that: The method of obtaining a Dixon FLASH sequence and using a deep reconstruction technique to reduce the noise and reconstruct the Dixon FLASH sequence is specifically as follows: using an AI-assisted magnetic resonance compressed sensing reconstruction technique to obtain a Dixon FLASH sequence, and using a deep reconstruction technique to reduce the noise and reconstruct the Dixon FLASH sequence; wherein, in the process of reducing the noise and reconstructing the Dixon FLASH sequence, a data fidelity module is embedded in the reconstruction chain, and when performing network training, a loss function includes both image domain and k-space domain information, and the data fidelity module is constrained by the image domain and k-space domain information, and based on the k-space domain, the output result and the original data are intelligently checked and matrix fused, and iterative optimization is performed through quadratic approximation.

3. A PET / MR attenuation correction method according to claim 2, characterized in that: In the process of denoising and reconstructing the original Dixon FLASH sequence using the deep reconstruction technology, a CNN network is used for residual learning.

4. A PET / MR attenuation correction method according to claim 1, characterized in that: The MRI image is obtained by magnetic resonance reconstruction and decoding, specifically comprising: Preprocess the acquired original Dixon FLASH sequence, including removing noise and correcting phase; Perform Fourier transform on the preprocessed signal to convert the signal from time domain to frequency domain; Reconstructing the signal converted into the frequency domain into a first image for a doctor to read by using an image reconstruction algorithm; The reconstructed first image is enhanced and segmented to obtain an MRI image to be attenuated corrected.

5. A PET / MR attenuation correction method according to claim 4, characterized in that: The image reconstruction algorithm includes filtered back projection and iterative reconstruction algorithm.

6. A PET / MR attenuation correction method according to claim 1, characterized in that: The lung and human body contour segmentation is performed to obtain lung tissue and air tissue, specifically: a threshold segmentation method is used to segment the lung and human body contour, and a U-net network is used to segment the lung and air to obtain lung tissue and air tissue.

7. A PET / MR attenuation correction method according to claim 1, characterized in that: The use of a neural network to synthesize sCT images to obtain bone tissue is specifically as follows: The in-phase MRI image and the anti-phase MRI image are input into the 3D U-net network, sCT image synthesis is performed, and the sCT image is output to obtain bone tissue; wherein the loss function of the 3D U-net network is the residual of the sCT image and the CT image.

8. A PET / MR attenuation correction method according to claim 7, characterized in that: The 3DU-net network training data acquisition process is specifically as follows: Based on AICS and DRecon technology, MR images were acquired by free breathing with the arms lowered; Whole-body CT images were collected on a flat bed by free breathing with the arms down; The mutual information algorithm was used to match the MR image and the CT image for the subsequent setting of the attenuation coefficient.

9. A PET / MR attenuation correction method according to claim 1, characterized in that: The soft tissue and fat tissue are obtained by using the Dixon water-fat separation technique, specifically: the soft tissue and fat tissue are extracted from the in-phase MRI image and the anti-phase MRI image by using the Dixon water-fat separation technique.

10. A system using the PET / MR attenuation correction method according to claim 1, characterized in that: include: The acquisition module is used to obtain the Dixon FLASH sequence, use the deep reconstruction technology to reduce the noise and reconstruct the Dixon FLASH sequence, and obtain the MRI image through magnetic resonance reconstruction and decoding; The segmentation module is used to segment the lung and human body contours based on MRI images to obtain lung tissue and air tissue, synthesize sCT images using neural networks to obtain bone tissue, and obtain soft tissue and fat tissue through Dixon water-fat separation technology; An attenuation coefficient map synthesis module is used to assign specific attenuation coefficients to the masks of the five tissues obtained, respectively, and synthesize them into an attenuation coefficient map μMAP for attenuation correction; The attenuation correction imaging module is used to perform attenuation correction on the PET raw data using the attenuation coefficient map μMAP to obtain the final PET / MR imaging results.

Images