A method for obtaining attenuation correction coefficient images based on single-modality PET images

Abstract

The present invention relates to a method for obtaining an attenuation correction coefficient image based on a single-modality PET image. The method comprises: S10, obtaining a PET image of the detection data without attenuation correction for the detection data used for medical image reconstruction; S20, inputting a known attenuation coefficient image of a scanning bed and the PET image without attenuation correction into a pre-trained generator G, obtaining a pseudo attenuation coefficient image corresponding to the detection data output by G; the pseudo attenuation coefficient image is equivalent to a true attenuation coefficient image obtained by converting images from other modalities, and is used to reconstruct the PET image. The above method can achieve relatively accurate reconstruction based on single-modality PET images without the need for other modalities, reducing the number of scans required for the subject, and reducing attenuation artifacts in the reconstructed PET image, ensuring stability.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a method for obtaining an attenuation correction coefficient image based on a single-modality PET image, a PET image reconstruction method, and a PET system. Background Art

[0002] Currently, in the actual imaging process of multimodal (such as PET / CT) equipment, CT cannot accurately obtain attenuation information that matches the PET data. The reason is that CT scans are generally completed in a very short time, while PET scans of one body position usually take several minutes. Due to the patient's breathing movement and heartbeat, some organs in the body are not completely static, or the patient moves during the scan. There is a certain degree of mismatch between the PET reconstructed image and the CT reconstructed image. In this case, using CT data to perform attenuation correction for PET image reconstruction will produce attenuation artifacts in the PET image;

[0003] Secondly, PET's scanning range is generally larger than that of other modalities (such as CT or MRI). When scanning larger patients, other imaging modalities may not be able to provide a sufficiently large scanning range, which will result in truncation of the final attenuation coefficient image. Applying this image to PET image reconstruction will result in attenuation artifacts.

[0004] Although the radiation dose of PET / CT examinations is strictly controlled within an absolutely safe range, it is still necessary to minimize the radiation dose received by patients. In addition to undergoing internal injection of radioactive drugs, patients also need to undergo low-dose CT scans for anatomical positioning and attenuation correction. Under different scanning protocols, such as delayed scanning, multiple accompanying CT scans may be required, which further increases the radiation dose received by the subjects.

[0005] Finally, during the PET scan process, the scanning table cannot be detected by the PET scanner, but in fact it has a relatively large impact on the attenuation correction. In the absence of CT scanning, such as PET / MR equipment, additional processing methods are usually required to supplement the attenuation information of the scanning table into the PET attenuation correction to ensure the accuracy of the reconstructed tracer distribution image and avoid artifacts in the PET image due to ignoring the attenuation information of the scanning table.

[0006] In view of this, there is an urgent need for a method to perform attenuation correction on PET images without the need for other modalities to assist attenuation correction. Summary of the Invention

[0007] (1) Technical issues to be solved

[0008] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method for obtaining an attenuation correction coefficient image based on a single-modality PET image and a PET image reconstruction method.

[0009] (2) Technical solution

[0010] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0011] In a first aspect, an embodiment of the present invention provides a method for acquiring an attenuation correction coefficient image based on a single-modality PET image, comprising:

[0012] S10, acquiring a PET image of the detection data that is not attenuated corrected for the detection data used for medical image reconstruction;

[0013] S20, inputting a known attenuation coefficient image of the scanning bed and the unattenuated PET image into a pre-trained generator G to obtain a pseudo attenuation coefficient image corresponding to the detection data output by G; the pseudo attenuation coefficient image is used to reconstruct the PET image;

[0014] The generator G is generated by supervised learning training, and includes: a spatial transformation network and a synthesis network; the spatial transformation network is used to obtain scanning bed feature information based on a known attenuation coefficient image of the scanning bed and the PET image without attenuation correction, and input the information into the synthesis network; the synthesis network is used to extract first feature information of the PET image based on the PET image without attenuation correction, fuse the first feature information of the scanning bed feature information, and map and generate the pseudo attenuation coefficient image.

[0015] Optionally, before S10, the method further includes:

[0016] Based on the PET detection data and other modality images matched with the PET detection data, obtaining training samples for training the generator G;

[0017] Each training sample includes: a true linear attenuation correction coefficient image corresponding to PET detection data, a PET image without attenuation correction, and a known attenuation coefficient image of the scanning bed; the other modality images are used to obtain the true linear attenuation correction coefficient image;

[0018] During the training process, the unattenuated PET image and the attenuation coefficient image of the scanning bed in each training sample are input into G, and the output of G and the true linear attenuation correction coefficient image of the training sample are both input into the discriminator D to discriminate the output of G during training. The training of G and D is performed alternately so that the pseudo attenuation coefficient image generated by the trained G matches the true linear attenuation correction coefficient image.

[0019] Optionally, the loss function L during training is:

[0020]

[0021] L GAN (G,D) is the generation adversarial loss;

[0022] L GAN (G,D)=-E xy [D(x,z,y)]+E x [D(x,z,G(x,z))] (2)

[0023] L1 is the loss function used to ensure low-frequency information;

[0024]

[0025] For each training, x represents the PET image without attenuation correction, z represents the known attenuation coefficient image of the scanning bed, and y represents the true linear attenuation correction coefficient image; E xy [D(x,z,y)] represents the expectation under the input, D(x,z,y) represents the discrimination result of D on the true linear attenuation correction coefficient image, D(x,z,G(x,z)) represents the discrimination result of D on the image output by G; λ is a hyperparameter.

[0026] Optionally, the synthetic network includes: an encoder, a residual module, and a decoder; the three are connected in series in sequence; the decoder and the encoder have a symmetrical structure, and a jump connection is used between the decoder and the encoder to pass the convolutional features output by the encoder to the decoder;

[0027] The encoder includes: 1 convolution layer and 3 downsampling basic modules; the first convolution layer is used to receive the input PET image without attenuation correction and the scanning bed feature information input by the spatial transformation network, and perform convolution processing to obtain convolution features;

[0028] All downsampling basic modules include: a convolution layer with a stride of 2 and an activation function; three serially connected downsampling basic modules process the input convolution features in turn, and the output of the third downsampling basic module is fed into the residual module;

[0029] Each residual module includes two sets of convolution and activation operations connected in series. The output of each residual module is obtained by adding the input of the residual module and the output of the activation operation in the second set of convolution and activation operations at the pixel level.

[0030] Finally, the output of the residual module is sent to the decoder;

[0031] The decoder consists of three upsampling basic modules and one convolutional layer;

[0032] The upsampling basic module adopts deconvolution operation. The output of the upsampling basic module will be sent to the last convolution layer to perform convolution processing on the input convolution features and output a pseudo attenuation coefficient image.

[0033] Optionally, the spatial transformation network includes: a position positioning module and a feature resampling module;

[0034] The position positioning module is used to extract the spatial position relationship information between the input attenuation coefficient image of the scanning bed and the PET image without attenuation correction, and obtain affine transformation matrix parameters;

[0035] The position location module includes: 5 layers of downsampling modules, fully connected layers and activation functions; each downsampling module includes: convolution, pooling and activation operations, each downsampling module is used to halve the size of the extracted feature map; the feature map output by the 5 layers of downsampling modules is expanded into a one-dimensional array and passed through two sets of fully connected layers and activation functions to obtain the affine transformation matrix parameters;

[0036] The feature resampling module is used to resample the input attenuation coefficient image of the scanning bed plate according to the affine transformation matrix parameters to obtain a modified scanning bed plate feature image as the acquired scanning bed plate feature information.

[0037] In a second aspect, an embodiment of the present invention further provides a PET image reconstruction method, comprising:

[0038] P01. For the detection data to be reconstructed by PET, a pseudo attenuation coefficient image μ0 is obtained using the method described in the first aspect above;

[0039] P02. Reconstruct the PET radioactivity distribution x based on μ0 and the detection data.

[0040] Optionally, the P02 includes:

[0041] Based on μ0, an alternating iterative strategy is used to alternately iterate the attenuation correction coefficient μ and the PET radioactivity distribution x, to obtain an estimated value of x that satisfies the maximization objective function requirement as a reconstructed image of the detection data;

[0042] The alternating iterative strategy includes: first taking μ0 as the initial value, maximizing the objective function to solve x, taking the solved x as a constant, maximizing the objective function to solve μ, and alternating operations in sequence to maximize the objective function to obtain the final reconstructed PET radioactivity distribution x.

[0043] Optionally, the P02 includes:

[0044] Based on μ0, an alternating iterative strategy is used to alternately iterate the attenuation correction coefficient μ and the PET radioactivity distribution x, to obtain an estimated value of x that satisfies the maximization objective function requirement as a reconstructed image of the detection data;

[0045] Among them, the alternating iterative strategy includes: first taking μ0 as the initial value, maximizing the objective function to solve x, taking the solved x as a constant, maximizing the objective function to solve μ, using μ0 to adjust the solved μ, taking the adjusted μ as a constant, maximizing the objective function to solve the next x, and performing the operation alternately.

[0046] In a third aspect, an embodiment of the present invention further provides a PET operating station, which includes: a memory and a processor; the memory stores computer program instructions, and the processor executes the computer program instructions stored in the memory, specifically performing any of the methods described in the first and second aspects above.

[0047] (3) Beneficial effects

[0048] The present invention obtains an attenuation correction coefficient image by means of a single-modality PET image, effectively solving the problem of inaccurate attenuation information in combined imaging of PET and other modalities.

[0049] In the embodiments of the present invention, the generator G is used to generate images that closely resemble true attenuation coefficient images. These images contain sufficient anatomical information. Furthermore, the generated attenuation coefficient images can be matched to PET data, enabling better PET attenuation correction and reconstructed PET images with reduced attenuation artifacts. In industrial applications, the methods of the embodiments of the present invention can replace CT in PET imaging, reducing the amount of radiation received by patients.

[0050] In the specific processing, the generator G is trained in a supervised manner. After a complex back-propagation and iterative training, the generator G can learn the mapping relationship between the input data features and the expected output. In this way, in the absence of an attenuation coefficient image, a pseudo attenuation coefficient image can be generated by the generator G for PET attenuation correction, ultimately achieving more quantitatively accurate PET imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic flow chart of a method for acquiring an attenuation correction coefficient image based on a single-modality PET image provided by one embodiment of the present invention;

[0052] Figure 2 A schematic diagram of a method for acquiring an attenuation correction coefficient image based on a single-modality PET image provided by another embodiment of the present invention;

[0053] Figure 3Schematic diagram of the synthesis network in the generator G;

[0054] Figure 4 Schematic diagram of the spatial transformation network in the generator G;

[0055] Figure 5 Schematic diagram comparing the attenuation coefficient images and corresponding reconstructed PET images of the method of the present invention and the existing method. DETAILED DESCRIPTION

[0056] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0057] Positron emission tomography (PET) is a noninvasive nuclear medicine imaging technique. It utilizes radioactive tracers to reveal metabolic activity within the human body and is used for tumor staging and metastasis detection in oncology, myocardial perfusion in cardiology, and neurological disease research. Unlike PET's functional metabolic imaging, computed tomography (CT) or magnetic resonance imaging (MRI) images the patient's anatomy, enabling precise localization of PET tracer distribution. The combination of these two techniques simultaneously captures both physiological changes and morphological information about lesions, improving the accuracy of disease diagnosis and treatment.

[0058] During the PET acquisition process, photons inevitably interact with various media before reaching the PET detector. There's a chance they'll collide with electrons, causing the photoelectric effect or Compton scattering, which can cause the gamma photon to disappear or its energy to decrease, rendering it undetectable by the PET detector. Photons can also change their direction, causing the PET detector to detect an anomalous coincidence event—scattering. Therefore, to obtain quantitative and accurate PET tracer distribution images, additional tissue linear attenuation coefficient images are required for accurate estimation of attenuation and scattering. Currently, anatomical imaging of the patient is typically performed using other modalities (such as CT and MRI). The tissue attenuation coefficient for gamma rays is then calculated from the anatomical images and applied to the reconstruction process to perform attenuation correction on the raw PET data, ultimately yielding an image of the actual radiotracer distribution within the patient.

[0059] However, during the actual imaging process of multimodal devices, differences in scanning time can lead to attenuation artifacts in PET images. Furthermore, there is also the issue of error information regarding the scanning bed. While the scanning bed cannot be detected during the PET scan, it can significantly impact attenuation correction. In non-CT scanning environments, such as PET / MR devices, additional processing is often required to incorporate the scanning bed's attenuation information into the PET attenuation correction process to ensure the accuracy of the reconstructed tracer distribution image and avoid artifacts in the PET image caused by ignoring the scanning bed's attenuation information.

[0060] The method of the embodiment of the present invention can achieve attenuation correction without the help of other modality images. The PET image is processed through the model, and the processing result is highly accurate without requiring multiple scans of the subject.

[0061] That is to say, in order to solve the problem of inaccurate attenuation information when PET is jointly imaged with other modalities, it can be processed with the help of a deep neural network trained in a supervised manner. Among them, the input images used for neural network training (PET images without attenuation correction and known attenuation coefficient images of the scanning bed) are paired with the network output images (real attenuation coefficient images). After a complex back propagation and iterative training, the deep neural network algorithm can learn the mapping relationship between the input data features and the expected output, so that in the absence of an attenuation coefficient image, a pseudo attenuation coefficient image can be generated through a deep neural network for PET attenuation correction, ultimately achieving more quantitatively accurate PET imaging. The deep neural network in this embodiment refers to the entire model framework, such as both the generator G and the discriminator D.

[0062] In this embodiment of the present invention, the known attenuation coefficient image of the scanning tabletop can be an attenuation coefficient image of an empty tabletop without a patient or user. During the training process of the generator G, general scanning tabletop dimensions, such as 2140 mm × 530 mm × 35 mm, can be used. This embodiment does not limit the scanning tabletop dimensions; these dimensions can be adjusted and selected based on actual needs.

[0063] Example 1

[0064] like Figure 1 As shown, an embodiment of the present invention provides a method for acquiring an attenuation correction coefficient image based on a single-modality PET image. The method of this embodiment may be performed by a control device / electronic device for a PET image reconstruction method. The control device may be integrated into an acquisition device of a PET system or a separate computer processing device. The method for acquiring an attenuation correction coefficient image based on a single-modality PET image includes the following steps:

[0065] S10, acquiring a PET image of the detection data that is not attenuated corrected for the detection data used for medical image reconstruction;

[0066] S20. A known attenuation coefficient image of the scanning bed and the uncorrected PET image are input into a pre-trained generator G to obtain a pseudo attenuation coefficient image corresponding to the detection data output by G. This pseudo attenuation coefficient image can be used as the attenuation coefficient image for PET image reconstruction, accelerating the iterative convergence path of PET image reconstruction and making the reconstructed PET image more accurate and artifact-free. The pseudo attenuation coefficient image generated in this embodiment is equivalent to an attenuation coefficient image converted from other modal images (e.g., an attenuation coefficient image converted from a CT image).

[0067] The generator G may be generated by supervised training. In specific use, the generator G may include a spatial transformation network and a synthesis network. The spatial transformation network is used to obtain scanning bed feature information based on a known attenuation coefficient image of the scanning bed and the PET image without attenuation correction, and input the information into the synthesis network. The synthesis network extracts the first feature information of the PET image based on the PET image without attenuation correction, fuses the scanning bed feature information, and maps to generate the pseudo attenuation coefficient image.

[0068] In practical applications, before the above step S10, Figure 1 The method shown further includes the following step S00 which is not shown in the figure:

[0069] S00, based on the PET detection data and other modality images matched with the PET detection data, obtaining training samples for training the generator G;

[0070] Each training sample includes: a true linear attenuation correction coefficient image corresponding to PET detection data, a PET image without attenuation correction, and a known attenuation coefficient image of the scanning bed; the other modality images are used to obtain the true linear attenuation correction coefficient image;

[0071] During the training process, the unattenuated PET image and the attenuation coefficient image of the scanning bed in each training sample are input into G, and the output of G and the true linear attenuation correction coefficient image of the training sample are both input into the discriminator D to discriminate the output of G during training. The training of G and D is performed alternately so that the pseudo attenuation coefficient image generated by the trained G matches the true linear attenuation correction coefficient image.

[0072] The method of this embodiment uses single-modality PET images to obtain attenuation correction coefficient images, effectively solving the problem of inaccurate attenuation information when PET is combined with other modalities for imaging. It can replace the role of CT in PET imaging, reduce the amount of radiation received by patients, and effectively reduce the generation of attenuation artifacts in PET reconstruction.

[0073] Example 2

[0074] To better understand the technical solution of the first embodiment, the following describes in detail a method for obtaining an attenuation correction coefficient image based on a single-modality PET image, combined with the training and use of the generator G. The method of this embodiment may include the following steps:

[0075] Training process:

[0076] Step 1: Obtain the real attenuation coefficient image of the PET scanning bed plate and use it as the scanning bed plate template; and obtain the real attenuation coefficient image of each training sample for training the generator G.

[0077] For example, PET scanning is always used in combination with other modality imaging. The scanning bed attenuation coefficient image of this embodiment and the actual attenuation coefficient image of each training sample can be obtained through other modality images.

[0078] Taking the PET / CT imaging system as an example, the high signal-to-noise ratio image obtained by the CT system can be used to convert the CT numerical value into a photon linear attenuation coefficient distribution image at 511KeV energy through the bilinear method to obtain the true attenuation coefficient image of the training sample; similarly, the scanning bed template image also comes from the CT image conversion of the bed.

[0079] Taking the PET / MR imaging system as an example, the MR image is segmented into different regions (such as soft tissue, fat, lung, air, etc.), and then the corresponding theoretical linear attenuation coefficient values ​​are directly assigned (for example, the soft tissue region is assigned a value of 0.0975cm -1 , the fat area is assigned a value of 0.0864cm -1 , the lung area is assigned a value of 0.0224cm -1 , the air area is assigned to 0), and the real attenuation coefficient image of the training sample is obtained; the scanning bed plate template image can be derived from the mechanical drawing, and the theoretical scanning bed plate template image can be directly constructed according to the marked size and material information.

[0080] Step 2: Obtain a PET image that has not been attenuated and then reconstructed (hereinafter referred to as a PET NAC image); the PET NAC image comes from the reconstruction program of the PET system, and no attenuation correction is selected during reconstruction.

[0081] Step 3: Input both the scanning bed template image in step 1 and the PET NAC image in step 2 into the generator to be trained G at the same time, and output an accurate pseudo attenuation coefficient image.

[0082] The "pseudo" in the pseudo attenuation coefficient image of this embodiment is intended to express that the attenuation correction coefficient image generated by this neural network is different from that obtained by CT scan conversion (because it is actually obtained); the effect generated by the neural network is very realistic; for better distinction, the result output by the generator G in this embodiment is called a pseudo attenuation coefficient image.

[0083] The generator G of this embodiment includes: a synthesis network and a spatial transformation network; to better understand the above step 3, the following sub-steps 31 to 35 are used for explanation.

[0084] Step 31: The spatial transformation network extracts the spatial position relationship information between the input scanning bed template image and the PET NAC image to obtain a modified scanning bed feature image.

[0085] like Figure 4 As shown, the spatial transformation network includes: a position positioning module and a feature resampling module; the position positioning module is used to extract the spatial position relationship information between the input scanning bed template image and the PET NAC image to obtain the affine transformation matrix parameters;

[0086] The positioning module includes: a five-layer downsampling module, each of which contains convolution, pooling and activation operations to halve the size of the feature map; the feature map output by the five-layer downsampling module is expanded into a one-dimensional array, and passes through two sets of fully connected layers and activation functions to obtain the affine transformation matrix parameters; the feature resampling module resamples the input scanning bed feature image according to the obtained affine transformation matrix parameters to obtain a corrected scanning bed feature image.

[0087] Sub-step 32: The synthesis network extracts the PET NAC image feature information and fuses the corrected scanning bed feature image information, and generates a pseudo attenuation coefficient image through the network mapping.

[0088] Combine Figures 2 to 4 As shown, the above-mentioned generator G and the discriminator D involved in training the generator G are explained respectively.

[0089] Figure 2 shows the structure of the generator G, Figure 3 The structure of the synthesis network is shown. The synthesis network is a residual U-net network consisting of three parts: an encoder, a residual module, and a decoder, where a skip connection is used between the encoder and the decoder.

[0090] The encoder consists of multiple encoding units, which are: 1 convolutional layer, 3 downsampling basic modules;

[0091] The first convolutional layer is used to receive the input PET NAC image and the corrected scanning bed feature image, and perform convolution processing to obtain convolution features;

[0092] All downsampling basic modules include: a convolution layer with a stride of 2 and an activation function; three serially connected downsampling basic modules process the input convolution features in turn, and the output of the third downsampling basic module is fed into the residual module;

[0093] Each residual module consists of two sets of convolution and activation operations connected in series. The output of each residual module is the pixel-wise addition of the input of the residual module and the output of the activation operation in the second set of convolution and activation operations (i.e., the addition of pixels at corresponding positions).

[0094] The output of the residual module is then fed into the decoder, which has a symmetrical structure with the encoder.

[0095] The decoder consists of three upsampling basic modules and a final convolutional layer. The difference between the upsampling basic module and the downsampling basic module is that the convolution operation is replaced by a deconvolution operation. The output of the upsampling basic module is fed into the final convolutional layer, which convolves the input convolution features and outputs a pseudo-attenuation coefficient image.

[0096] Step 4: Input the true attenuation coefficient image and the pseudo attenuation coefficient image into the discriminator D, make judgments based on the loss function and adjust the training parameters of G. By alternately training G and D, the pseudo attenuation coefficient image finally output by G matches the true attenuation coefficient image, and the trained G is obtained.

[0097] Among them, the loss function L is:

[0098]

[0099] L GAN (G,D) is the generation adversarial loss;

[0100] L GAN (G,D)=-E xy [D(x,z,y)]+E x [D(x,z,G(x,z))] (2)

[0101] L1 is the loss function used to ensure low-frequency information;

[0102]

[0103] Where x represents the PET NAC image, z represents the scanning bed template image, and y represents the real attenuation coefficient image; E xy [D(x,z,y)] represents the expectation under the input, D(x,z,y) represents the discriminator's judgment result on the true attenuation coefficient image, and D(x,z,G(x,z)) represents the discriminator's judgment result on the image output by G. λ is a hyperparameter;

[0104] In this embodiment, the discriminator network structure of PatchGAN is used to distinguish the image output by the generator G from the real attenuation coefficient image, and a generative adversarial model is formed with the generator to encourage the generator to generate more realistic attenuation coefficient images.

[0105] The above scheme can be used for PET / CT equipment, and can also be extended to single PET or PET combined with other modalities, such as PET / MR, to train the generator G.

[0106] Usage process:

[0107] Step 5: For the detection data used for medical image reconstruction, obtain a PET image of the detection data that has not been attenuated corrected, that is, a PET NAC image;

[0108] Step 6: Based on the known attenuation coefficient image of the scanning bed and the uncorrected PET image, a pre-trained generator G is input to obtain a pseudo attenuation coefficient image corresponding to the detection data output by G. This pseudo attenuation coefficient image can be used as the attenuation coefficient image for PET image reconstruction, accelerating the iterative convergence path of PET image reconstruction and ensuring a more accurate and artifact-free reconstructed PET image. The pseudo attenuation coefficient image generated in this embodiment is equivalent to the attenuation coefficient image converted from other modal images (such as the attenuation coefficient image converted from CT images).

[0109] It is understandable that the pseudo attenuation coefficient image output by G is equivalent to the attenuation coefficient image converted from CT. In order to avoid inaccurate measurements of multiple modalities or patient radiation, G is used to obtain the attenuation coefficient image, replacing the dual-modality method of obtaining the attenuation coefficient image in the existing technology.

[0110] Reconstruction process:

[0111] Step 7: Reconstruct the PET radioactivity distribution x based on the pseudo attenuation coefficient image output from step 6 and the detection data from step 5.

[0112] It is understandable that various reconstruction methods can be used to reconstruct PET images in this embodiment, such as OSEM.

[0113] This embodiment provides experimental verification results, such as Figure 5 As shown, Figure 5 (a) is the pseudo attenuation coefficient image output by the neural network; Figure 5 (b) is the real attenuation coefficient image; Figure 5 (c) and Figure 5 (d) are all PET images reconstructed using the traditional OSEM algorithm, where Figure 5 The attenuation correction in (c) is based on the pseudo attenuation coefficient image, Figure 5 The attenuation correction in (d) is based on the true attenuation coefficient image.

[0114] from Figure 5 As can be seen, the present invention uses a deep neural network to generate images that closely resemble true attenuation coefficient images, including sufficient anatomical information. Furthermore, the generated attenuation coefficient images match the PET data, enabling better PET attenuation correction and reducing attenuation artifacts. This method holds great promise for replacing CT in PET imaging and reducing patient radiation exposure.

[0115] Example 3

[0116] An embodiment of the present invention provides a PET image reconstruction method, which includes:

[0117] P01. For the detection data to be reconstructed by PET, a pseudo attenuation coefficient image μ0 is obtained using the method described in the first or second embodiment above.

[0118] It is particularly noted that the pseudo attenuation coefficient image μ0 obtained in this embodiment is obtained by adding the attenuation information of the scanning bed, so there is no need to perform supplementary correction on the attenuation information of the scanning bed in the following reconstruction.

[0119] P02. Reconstruct the PET radioactivity distribution x based on μ0 and the detection data.

[0120] For example, in one implementation, based on μ0, an alternating iteration strategy may be used to alternately iterate the attenuation correction coefficient μ and the PET radioactivity distribution x to obtain an estimated value of x that satisfies the maximization objective function requirement as the reconstructed image of the detection data;

[0121] Among them, the alternating iteration strategy includes: first taking μ0 as the initial value, maximizing the objective function To solve x, take the solved x as a constant and maximize the objective function To solve μ, the operations are performed alternately in sequence to maximize the objective function and obtain the final reconstructed PET radioactivity distribution x.

[0122] For example, in another possible implementation, based on μ0, an alternating iteration strategy may be used to alternately iterate the attenuation correction coefficient μ and the PET radioactivity distribution x to obtain an estimated value of x that satisfies the maximization objective function requirement as the reconstructed image of the detection data.

[0123] At this time, the alternating iterative strategy includes: first taking μ0 as the initial value, maximizing the objective function to solve x, taking the solved x as a constant, maximizing the objective function to solve μ, using μ0 to adjust the solved μ, taking the adjusted μ as a constant, maximizing the objective function to solve the next x, and performing the operation alternately.

[0124] Specifically, taking μ0 as the initial value, the constructed log-likelihood function L(x,μ,y) is maximized to obtain the radioactivity distribution x; n represents the number of iterations, and the initial value of x is the set value;

[0125] Taking the iterative solution x as a constant, maximize the constructed log-likelihood function L(x,μ,y) to obtain the linear attenuation coefficient distribution

[0126] Use μ0 to adjust the solved μ and get the adjusted

[0127] Right now

[0128] d (n+0.5) is the displacement field change vector, n is a natural number greater than 1; Indicates that the linear attenuation coefficient prior result μ0 is transformed into the displacement field change vector d (n+0.5) Elastic transformation of scale.

[0129] The adjusted As a constant, maximize the constructed log-likelihood function L(x,μ,y) to obtain the radioactivity distribution x;

[0130] The operations are performed alternately to obtain an estimated value of x that satisfies the requirement of maximizing the constructed log-likelihood function. The above method can speed up the iterative path of the PET image reconstruction process and reduce artifacts in the reconstructed PET image.

[0131] The objective function in the above-mentioned PET reconstruction embodiment may be the known log-likelihood function L(x, μ, y).

[0132]

[0133] Where y=[y 1t ,y 2t,…,y it ,…,y NT ]' indicates the detected data, i.e., the detection data. represents the average value of the detection data, N represents the size of the detection data sinogram, T represents the size of the TOF discrete space, i represents the variable index of the detection data sinogram response line LOR (line of response), and t represents the variable index of the TOF discrete space. The single quote superscript represents the matrix transpose operation. x=[x1,x2,…,x j ,…,x M ]' represents the unknown radioactivity distribution image, M represents the size of the radioactivity distribution image space, j represents the variable index of the radioactivity distribution image space, and represents the point source at the corresponding spatial position. μ=[μ1,μ2,…,μ k ,…,μ K ]' represents the unknown linear attenuation coefficient image, K represents the size of the linear attenuation coefficient image space, and k represents the variable index of the linear attenuation coefficient image space, representing the point source at the corresponding spatial position. A=[A ijt ] is the system matrix, which expresses the probability that a point source j in the PET system is detected by a response line LORi with a time of flight TOF of t in mathematical form, reflecting the physical characteristics of the system, l = [l ik ] is the linear attenuation coefficient matrix, which represents the trajectory intersection length when LOR i passes through the spatial position point source k. r=[r 1t ,r 2t ,…,r it ,…,r NT ]' represents the average value of random noise and scattering noise. In practice, there is no attenuation effect and no scattering of gamma rays during the detection process.

[0134] In addition, an embodiment of the present invention also provides a PET operating station with a PET system, which includes: a memory and a processor; the memory stores computer program instructions, and the processor executes the computer program instructions stored in the memory, specifically executing the above-mentioned PET image reconstruction method or the method for obtaining an attenuation correction coefficient image based on a single-modality PET image, etc.

[0135] It should be noted that, in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims enumerating several means, several of these means may be embodied by one and the same hardware. The use of the words first, second, third etc. is for convenience only and does not indicate any order. These words may be understood as part of the component name.

[0136] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0137] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments after learning the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0138] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention shall also include such modifications and variations.

Claims

1. A method for obtaining an attenuation correction coefficient image based on a single-modality PET image, characterized in that: include: S10, acquiring a PET image of the detection data that is not attenuated corrected for the detection data used for medical image reconstruction; S20, inputting a known attenuation coefficient image of the scanning bed and the unattenuated PET image into a pre-trained generator G to obtain a pseudo attenuation coefficient image corresponding to the detection data output by G; the pseudo attenuation coefficient image is used to reconstruct the PET image; The generator G is generated by supervised learning training, and includes a spatial transformation network and a synthesis network. The spatial transformation network is used to obtain scanning bed feature information based on a known attenuation coefficient image of the scanning bed and the PET image without attenuation correction, and input the information into the synthesis network. The synthesis network extracts first feature information of the PET image based on the PET image without attenuation correction, fuses the first feature information of the scanning bed feature information, and generates the pseudo attenuation coefficient image through mapping. The spatial transformer network includes: Position positioning module and feature resampling module; The position positioning module is used to extract the spatial position relationship information between the input attenuation coefficient image of the scanning bed and the PET image without attenuation correction, and obtain affine transformation matrix parameters; The position location module includes: 5 layers of downsampling modules, fully connected layers and activation functions; each downsampling module includes: convolution, pooling and activation operations, each downsampling module is used to halve the size of the extracted feature map; the feature map output by the 5 layers of downsampling modules is expanded into a one-dimensional array and passed through two sets of fully connected layers and activation functions to obtain the affine transformation matrix parameters; The feature resampling module is used to resample the input attenuation coefficient image of the scanning bed plate according to the affine transformation matrix parameters to obtain a modified scanning bed plate feature image as the acquired scanning bed plate feature information.

2. The method according to claim 1, characterized in that Before S10, the following is also included: Based on the PET detection data and other modality images matched with the PET detection data, obtaining training samples for training the generator G; Each training sample includes: a true linear attenuation correction coefficient image corresponding to PET detection data, a PET image without attenuation correction, and a known attenuation coefficient image of the scanning bed; the other modality images are used to obtain the true linear attenuation correction coefficient image; During the training process, the unattenuated PET image and the attenuation coefficient image of the scanning bed in each training sample are input into G, and the output of G and the true linear attenuation correction coefficient image of the training sample are both input into the discriminator D to discriminate the output of G during training. The training of G and D is performed alternately so that the pseudo attenuation coefficient image generated by the trained G matches the true linear attenuation correction coefficient image.

3. The method according to claim 2, characterized in that The loss function L during training is: (1) To generate adversarial loss; (2) is the loss function used to ensure low-frequency information; (3) For each training session, represents the PET image without attenuation correction, represents the known attenuation coefficient image of the scanning bed, represents the true linear attenuation correction coefficient image; represents the expectation under the input, It represents the discrimination result of D on the true linear attenuation correction coefficient image, Denotes the discrimination result of D on the image output by G; is a hyperparameter.

4. The method according to claim 2, characterized in that The synthetic network includes: an encoder, a residual module and a decoder; the three are connected in series in sequence; the decoder and the encoder have a symmetrical structure, and a jump connection is used between the decoder and the encoder to pass the convolutional features output by the encoder to the decoder; The encoder includes: 1 convolution layer and 3 downsampling basic modules; the first convolution layer is used to receive the input PET image without attenuation correction and the scanning bed feature information input by the spatial transformation network, and perform convolution processing to obtain convolution features; All downsampling basic modules include: a convolution layer with a stride of 2 and an activation function; three serially connected downsampling basic modules process the input convolution features in turn, and the output of the third downsampling basic module is fed into the residual module; Each residual module includes two sets of convolution and activation operations connected in series. The output of each residual module is obtained by adding the input of the residual module and the output of the activation operation in the second set of convolution and activation operations at the pixel level. Finally, the output of the residual module is sent to the decoder; The decoder consists of three upsampling basic modules and one convolutional layer; The upsampling basic module adopts deconvolution operation. The output of the upsampling basic module will be sent to the last convolution layer to perform convolution processing on the input convolution features and output a pseudo attenuation coefficient image.

5. A PET image reconstruction method, characterized in that: include: P01. For the detection data to be reconstructed by PET, a pseudo attenuation coefficient image is obtained using the method described in claim 1. ; P02, based on Using the detection data, reconstructing the PET radioactivity distribution x specifically includes: Based on this , alternately iterating the attenuation correction coefficient μ and the PET radioactivity distribution x using an alternating iterative strategy to obtain an estimated value of x that satisfies the maximization objective function requirement as a reconstructed image of the detection data; Among them, the alternating iteration strategy includes: first As the initial value, maximize the objective function to solve x, take the solved x as a constant, maximize the objective function to solve μ, and perform alternating operations in sequence to maximize the objective function to obtain the final reconstructed PET radioactivity distribution x.

6. The method according to claim 5, characterized in that The P02 includes: Based on this , alternately iterating the attenuation correction coefficient μ and the PET radioactivity distribution x using an alternating iterative strategy to obtain an estimated value of x that satisfies the maximization objective function requirement as a reconstructed image of the detection data; Among them, the alternating iteration strategy includes: first As the initial value, maximize the objective function to solve x, take the solved x as a constant, maximize the objective function to solve μ, and use The solved μ is adjusted, and the adjusted μ is used as a constant to maximize the objective function to solve the next x, and the operation is performed alternately.

7. A PET operating station, characterized in that: include: A memory and a processor; the memory stores computer program instructions, and the processor executes the computer program instructions stored in the memory, specifically performing the method described in any one of claims 1 to 6 above.

Images