PET image partial volume correction method based on conditional depth residual compensation

Abstract

The invention discloses a PET (positron emission tomography) image partial volume correction method based on conditional depth residual compensation, which is characterized in that fuzzy kernel priori is predicted through a conditional nerve blind deconvolution network, and partial volume correction (PVC) is carried out on a partial volume effect (PVE) of a PET image in combination with a conditional depth residual compensation network. According to the method, an anatomical MR prior image of a patient and a randomly generated Gaussian blurring kernel are used as input of a conditional nerve blind deconvolution network to perform blurring kernel estimation, and an error generated by inaccurate blurring kernel estimation is compensated through a conditional depth residual compensation network. According to the method, the potential complementary information of the multi-modal data is fully utilized, and the influence of inaccurate fuzzy kernel prediction on the PVC result is considered, so that the effectiveness and robustness of the neural network PVC are improved. According to the method, the accuracy of quantitative analysis of the PET image is remarkably improved, and particularly the accurate analysis of the brain PET image is improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning image processing, and in particular relates to a PET image partial volume correction method based on conditional deep residual compensation. Background Art

[0002] Positron emission tomography (PET) is an important molecular imaging technology, which is widely used in clinical diagnosis and scientific research. It plays an important role in medical diagnosis, especially in the early diagnosis of cancer.

[0003] However, due to the low spatial resolution of PET image scanners, partial volume effect (PVE) often occurs, which makes the radioactive uptake of specific voxels affected by surrounding tissues, thereby reducing the accuracy and resolution of PET images. This effect not only affects the early diagnosis of diseases, but also poses a severe challenge to the quantitative analysis of PET images. In recent years, although deep learning has made significant progress in partial volume correction (PVC) of PET images, most methods only rely on a large amount of PET image data for training, do not make full use of additional anatomical information, and do not consider the impact of inaccurate blur kernel estimation on PVC results. With the popularization of PET / MR scanners, it has become easier to simultaneously obtain high-precision MR images and PET images of patients, which provides valuable anatomical structure information for PVC of PET images. Therefore, it is an urgent problem in this field to effectively combine the complementary information between multimodal data to achieve partial volume correction of PET images, thereby improving the quality of PET images and the accuracy of quantitative analysis. Summary of the invention

[0004] The purpose of the present invention is to address the problems existing in the prior art and provide a PET image partial volume correction method based on conditional depth residual compensation, which makes full use of the potential complementary information of multimodal data and compensates for the error caused by inaccurate blur kernel estimation, thereby significantly improving the quality of PVC-PET images.

[0005] According to a first aspect of an embodiment of the present application, a method for partial volume correction of a PET image based on conditional depth residual compensation is provided, comprising:

[0006] (1) Obtain PVE-PET images and corresponding anatomical MR images of the patient;

[0007] (2) registering the PVE-PET image and the anatomical MR image, and preprocessing the registered anatomical MR image to obtain a registered PVE-PET image and an anatomical MR brain tissue image;

[0008] (3) using the registered PVE-PET images and the randomly generated Gaussian blur kernels, respectively pre-training the PVC-PET image generation network and the blur kernel generation network, thereby initializing the initial weight parameters of the PVC-PET image generation network and the blur kernel generation network;

[0009] (4) constructing a conditional neural blind deconvolution network based on the pre-trained PVC-PET image generation network and the blur kernel generation network, inputting the randomly generated Gaussian blur kernel and the anatomical MR brain tissue image into the conditional neural blind deconvolution network, and using the registered PVC-PET image as a training label to train the conditional neural blind deconvolution network;

[0010] (5) The anatomical MR brain tissue image is used as the input of the conditional deep residual compensation network, and combined with the blur kernel prior estimated by the conditional neural blind deconvolution network, the registered PVE-PET image is used as the training label to train the conditional deep residual compensation network, thereby obtaining the PVC-PET image and realizing partial volume correction.

[0011] Furthermore, in step (2), the registered anatomical MR image is preprocessed, specifically:

[0012] The skull part is removed from the registered anatomical MR image, and the brain tissue part is extracted to obtain the anatomical MR brain tissue image.

[0013] Further, step (3) comprises:

[0014] (3.1) Using the deep image prior network as the PVC-PET image generation network G x , using the fully connected layer network as the fuzzy kernel generation network G k ;

[0015] (3.2) Using the registered PVE-PET images and the randomly generated Gaussian blur kernels, the PVC-PET image generation network and the blur kernel generation network are pre-trained, respectively, wherein the loss functions of the PVC-PET image generation network and the blur kernel generation network are defined as:

[0016]

[0017] in and are the loss functions of the PVC-PET image generation network and the blur kernel generation network, respectively, petand z k are the registered PVE-PET images and randomly generated Gaussian blur kernels, respectively.

[0018] Furthermore, in step (4), the conditional neural blind deconvolution network includes a parallel PVC-PET image generation network and a blur kernel generation network, wherein the PVC-PET image generation network is used to preliminarily predict the PVC-PET image based on the anatomical MR image, and the blur kernel generation network is used to estimate the blur kernel prior based on a randomly generated Gaussian blur kernel.

[0019] Furthermore, in step (4), the optimization process during training of the conditional neural blind deconvolution network is:

[0020]

[0021] in is the loss function of the conditional neural blind deconvolution network, z a For anatomical MR brain tissue images, y pet and z k are the registered PVE-PET images and the randomly generated Gaussian blur kernel, G x ,G k They are PVC-PET image generation network and blur kernel generation network respectively.

[0022] Furthermore, in step (5), the conditional deep residual compensation network includes a parallel PVC-PET image generation network and a depth residual generation network, wherein the depth residual generation network is used to obtain a depth residual map based on the anatomical MR brain tissue image, and the PVC-PET image generation network is used to combine the depth residual map and the blur kernel prior to generate a PVC-PET image based on the anatomical MR brain tissue image.

[0023] Furthermore, in step (5), the loss function of the conditional deep residual compensation network during training is for:

[0024]

[0025] in and They are loss functions The fidelity term and regularization term in G x ,G r They are PVC-PET image generation network and deep residual generation network respectively. represents the blur kernel prior, z aFor anatomical MR brain tissue images, v and C represent the coefficient vector and DCT operator obtained after DCT transformation of the artifacts introduced by the blur kernel error, respectively, and λ1, λ2 and λ3 are hyperparameters that control the weights of each item.

[0026] According to a second aspect of an embodiment of the present application, a computer program product is provided. When the computer program / instructions are executed by a processor, the method described in the first aspect is implemented.

[0027] According to a third aspect of an embodiment of the present application, there is provided an electronic device, including:

[0028] one or more processors;

[0029] A memory for storing one or more programs;

[0030] When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in the first aspect.

[0031] According to a third aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which computer instructions are stored. When the instructions are executed by a processor, the steps of the method described in the first aspect are implemented.

[0032] The technical solution provided by the embodiments of the present application may have the following beneficial effects:

[0033] It can be seen from the above embodiments that the partial volume correction method of PET images based on conditional deep residual compensation in this application can effectively combine the potential complementary information in multimodal data and effectively compensate for the errors caused by inaccurate blur kernel prediction. This method realizes the prediction of blur kernel prior through the conditional neural blind deconvolution (CNBD) network, and compensates for the errors caused by inaccurate blur kernel prediction using the conditional deep residual compensation (CDRC) network, thereby overcoming the dependence of most previous PVC methods on single modality data and the influence of inaccurate blur kernel estimation on PET image PVC results. In addition, this method significantly improves the effectiveness and robustness of neural network PVC, especially when dealing with complex pathological changes and subtle structures, significantly improving the ability to restore PET image details and the accuracy of quantitative analysis.

[0034] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0036] Figure 1It is a flow chart of a method for partial volume correction of PET images based on conditional depth residual compensation of the present invention.

[0037] Figure 2 It is a schematic diagram of a model of the present invention.

[0038] Figure 3 PET images affected by PVE, where (a) is a simulated PET image affected by PVE, and (b) is a real clinical PET image affected by PVE.

[0039] Figure 4 The output result images of the present method and the existing method on the simulated data set are shown, where (a) is the preliminary PVC-PET image obtained using only the CNBD network, (b) is the final PVC-PET image obtained using the present method, (c) is the PVC-PET image obtained by processing using the existing method IY, and (d) is the PVC-PET image obtained by processing using the existing method RVC.

[0040] Figure 5 The output result images of the present method and the existing method on the real data set are shown, where (a) is the preliminary PVC-PET image obtained using only the CNBD network, (b) is the final PVC-PET image obtained using the present method, (c) is the PVC-PET image processed by the existing method IY, and (d) is the PVC-PET image processed by the existing method RVC.

[0041] Figure 6 It is a block diagram of a PET image partial volume correction device based on conditional depth residual compensation of the present invention.

[0042] Figure 7 It is a schematic diagram of an electronic device of the present invention. DETAILED DESCRIPTION

[0043] Here, exemplary embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0044] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0045] like Figure 1 As shown, the present invention is a PET image partial volume correction method based on conditional depth residual compensation, comprising the following steps:

[0046] (1) Obtain PVE-PET images and corresponding anatomical MR images of the patient;

[0047] Specifically, the PVE-PET image is a PET image affected by the partial volume effect.

[0048] (2) registering the PVE-PET image and the anatomical MR image, and preprocessing the registered anatomical MR image to obtain a registered PVE-PET image and an anatomical MR brain tissue image;

[0049] Specifically, the PVE-PET images were registered with the anatomical MR images using the ANTs registration tool, and the two were finally registered to the MNI152 standard space to ensure accurate spatial alignment of the multimodal images. At the same time, the skull was removed from the anatomical MR images using FSL's BET tool, and the brain tissue was extracted to obtain the anatomical MR brain tissue image.

[0050] (3) using the registered PVE-PET images and the randomly generated Gaussian blur kernels, respectively pre-training the PVC-PET image generation network and the blur kernel generation network, thereby initializing the initial weight parameters of the PVC-PET image generation network and the blur kernel generation network;

[0051] Specifically, this step may include the following sub-steps:

[0052] (3.1) Using the Deep Image Prior (DIP) network as the PET image generation network G x , using the fully connected network (FCN) as the fuzzy kernel generation network G k ;

[0053] Input the PVE-PET image into the network G x , we can get a PET image with potential internal information of the PVE-PET image. The randomly generated Gaussian blur kernel is input into the network G k , a roughly fitted Gaussian blur kernel can be obtained.

[0054] (3.2) Using the registered PVE-PET images and the randomly generated Gaussian blur kernel, we can x and G k Pre-training is performed, where G x and G kThe loss function of is set to mean square error (MSE) loss function, and each network is pre-trained separately by optimizing the loss function. x and G k The loss functions are defined as:

[0055]

[0056] in and G x and G k The loss function, y pet and z k are the registered PVE-PET images and randomly generated Gaussian blur kernels, respectively.

[0057] The pre-training method uses the registered PVE-PET images and randomly generated Gaussian blur kernels as the input of the DIP network and FCN, respectively, to autonomously optimize the network parameters. In an unsupervised environment, the network is self-trained by calculating the loss function between the input data and the network output data, without the need for additional labeled data. The pre-training method provides the network initialization weight parameters with information related to PVE-PET images and random Gaussian blur kernels for subsequent CNBD network training, significantly improving the accuracy of the CNBD network in blur kernel prediction and preliminary PVC-PET image estimation. At the same time, the generalization ability of the neural network is enhanced, and the unnecessary effects of random initialization parameters are reduced, so that it can effectively adapt to the individual differences of different patients, thereby improving the fitting effect of subsequent neural network training.

[0058] (4) constructing a conditional neural blind deconvolution (CNBD) network based on the pre-trained PVC-PET image generation network and blur kernel generation network, inputting the randomly generated Gaussian blur kernel and the anatomical MR brain tissue image into the conditional neural blind deconvolution network, and using the registered PVE-PET image as a training label to train the conditional neural blind deconvolution network;

[0059] Specifically, the conditional neural blind deconvolution (CNBD) network is used to perform blur kernel prior prediction and preliminary PVC-PET image estimation, and the pre-trained PVC-PET image generation network G x and the fuzzy kernel generation network G k The anatomical MR brain tissue image is input into the network G x Predict preliminary PVC-PET images and use randomly generated Gaussian blur kernels as input to the network G k, which is used to estimate the blur kernel prior. The two work together through a joint optimization framework to gradually refine the preliminary PVC-PET image and the blur kernel prior estimation results. The objective function is set to the mean square error (MSE) loss function, and the optimization process of the entire neural blind deconvolution network is:

[0060]

[0061] in is the loss function of the conditional neural blind deconvolution network, z a This is an anatomical MR image of brain tissue.

[0062] (5) The anatomical MR brain tissue image is used as the input of the conditional deep residual compensation (CDRC) network, and the blur kernel prior estimated by the CNBD network is combined with the registered PVE-PET image as the training label to train the conditional deep residual compensation network, thereby obtaining the PVC-PET image and realizing partial volume correction;

[0063] Specifically, the PVC-PET image generation network G x and the deep residual generation network G r Combined with the conditional deep residual compensation (CDRC) network, the generated network G x and G r Both use anatomical MR brain tissue images as input, G r The output depth residual image effectively compensates for the PVC error caused by the inaccurate prior estimation of the blur kernel, significantly improving the fitting ability of the CDRC network to the PVE-PET image, thereby promoting the generation of the network G x Output the final high-quality PVC-PET image. Similarly, for the generated network G x and G r The joint optimization framework works together to gradually refine the depth residual compensation image and the high-quality PVC-PET image. The CDRC network can be constrained and optimized through the mean square error (MSE) loss function, total variation (TV) regularization, and sparsity regularization, which improves the ability to restore PVC-PET image details.

[0064] First, define the loss function and use the mean square error (MSE) loss function as the loss function of the fidelity term in the CDRC network loss function. Therefore, the fidelity term in the loss function is defined as:

[0065]

[0066] in represents the blur kernel prior; v and C represent the coefficient vector and DCT operator obtained by discrete cosine transform (DCT) of the artifact introduced by the blur kernel error Δk, respectively. The blur kernel error Δk is the error between the estimated blur kernel prior and the actual blur kernel of the PVE-PET image.

[0067] Total variation (TV) regularization is applied to the generative network G x , sparsity regularization is applied to the generative network G r And the coefficient vector v obtained by DCT transformation, the regularization term in the CDRC network loss function is defined as:

[0068]

[0069] where λ1, λ2, and λ3 are hyperparameters that control the weights of each regularization loss term, ||·|| TV and ||·||1 denote total variation regularization and sparsity regularization, respectively.

[0070] The total loss of the conditional deep residual compensation (CDRC) network is and The sum of the two:

[0071]

[0072] For the above CDRC network total loss function Optimization uses the alternating minimization optimization algorithm to decompose it into two sub-problems for iterative solution. The specific process is as follows:

[0073] 1. Fixed parameter v, generating network G x and G r The network parameters θ and ζ in are optimized and updated:

[0074]

[0075] The superscript i represents the number of iterations.

[0076] 2. Fix the parameters θ and ζ and optimize and update v:

[0077]

[0078] Use the approximate gradient descent method to optimize and update v in each iteration process. The specific form is:

[0079]

[0080] in, represents the soft threshold operator, and L is the proportional coefficient used to control the degree of sparsity.

[0081] Based on the total loss function of the above CDRC network The CDRC network is trained using an unsupervised learning method and an alternating minimization optimization algorithm. The training is stopped when the set number of training epochs is reached or when high-quality PVC-PET images can be output. Relevant indicators such as PSNR, SSIM, and RMSE can be used to determine whether high-quality PVC-PET images are obtained.

[0082] In this step, the conditional deep residual compensation (CDRC) network is introduced to compensate for the PVC error caused by the inaccurate blur kernel estimation of the CNBD network. The network combines the deep image prior (DIP) network and the deep residual prior (DRP) network, and uses total variation (TV) regularization and sparsity regularization to further constrain the optimization process of the neural network. By combining these prior information, the CDRC network can effectively improve the PVC performance of the neural network and obtain high-quality PVC-PET images.

[0083] With the continuous advancement of deep learning technology, PVC technology for PET images has gradually adopted unsupervised learning methods based on deep learning networks. Most PVC methods usually rely on single modality data and fail to fully tap the potential of complementary effects of multimodal data. In addition, the low resolution and noise problems of PET images pose challenges to clinical quantitative analysis and diagnosis. This method adopts an unsupervised learning framework to avoid dependence on a large amount of annotated data, and uses the patient's anatomical MR brain tissue images as prior information. Through the potential complementary information between multimodal data, the subtle structures in PET images can be more effectively restored.

[0084] In this application, the network weight parameters of the deep image prior (DIP) and the fully connected layer network (FCN) of the PVC-PET generation network are pre-trained using the registered PVE-PET image and the randomly generated Gaussian blur kernel. Subsequently, the network weight parameters with robustness and stability provided in the pre-training stage are used as the weight parameters of the conditional neural blind deconvolution (CNBD) network initialization, and the anatomical MR brain tissue image and the randomly generated Gaussian blur kernel are input into the DIP network and FCN in the CNBD network respectively to generate preliminary PVC-PET images and predict the blur kernel prior. Finally, the conditional deep residual compensation (CDRC) network is introduced to compensate for the PVC error generated by the inaccurate blur kernel prior generated by the CNBD network training, and the anatomical MR brain tissue image is used as the input of the deep residual prior (DRP) and DIP network in the CDRC network. The residual image generated by the DRP network compensates for the inaccurate blur kernel error, so that the DIP network generates the final high-quality PVC-PET image.

[0085] Example

[0086] The embodiments of the present invention were implemented using an NVIDIA 4090 graphics card on Python 3.8 and Pytorch 2.3.1 platforms. Simulation experiments were performed using a simulated dataset based on BrainWeb, and experiments were performed using an open access dataset provided by OpenNeuro for real clinical data. Both the simulated data and the real clinical data were finally registered to the standard MNI152 space, with an image size of 109×91×91;

[0087] like Figure 3 , Figure 4 As shown, compared with the CNBD method, the IY method (Kjell Erlandsson, Irene Buvat, PHendrik Pretorius, Benjamin A Thomas, and Brian F Hutton, "A review of partial volume correction techniques for emission tomography and their applications in neurology, cardiology and oncology," Physics in Medicine & Biology, vol. 57, no. 21, pp. R119, 2012.) and the RVC method (Jussi Tohka and Anthonin Reilhac, "Deconvolution-based partial volume correction in raclopride-pet and montecarlo comparison to mr-based method," Neuroimage, vol. 39, no. 4, pp. 1570–1584, 2008.), this method can obtain more reliable PVC-PET images on simulated data, while more effectively retaining the local details of the PET images and more clearly displaying the brain structure area. Similarly, if Figure 5 As shown in the figure, on the real clinical dataset, this method also obtains more accurate and reliable PVC-PET images.

[0088] Corresponding to the aforementioned embodiment of the method for partial volume correction of PET images based on conditional depth residual compensation, the present application also provides an embodiment of a network framework model device for partial volume correction of PET images based on conditional depth residual compensation.

[0089] Figure 2The present invention is a block diagram of a network framework model device for partial volume correction of PET images based on conditional depth residual compensation according to an exemplary embodiment. Figure 2 , the device may include: FCN and DIP network pre-training models, conditional neural blind deconvolution model, conditional deep residual compensation model, and finally generate high-quality PVC-PET images.

[0090] An image acquisition module 21 is used to acquire PVE-PET images and corresponding anatomical MR images of the patient;

[0091] A registration and preprocessing module 22, used for registering the PVE-PET image and the anatomical MR image, and preprocessing the registered anatomical MR image to obtain a registered PVE-PET image and an anatomical MR brain tissue image;

[0092] A pre-training module 23 is used to pre-train the PVC-PET image generation network and the blur kernel generation network respectively using the PVE-PET image processed by the registration and pre-processing module and the randomly generated Gaussian blur kernel, so as to initialize the initial weight parameters of the PVC-PET image generation network and the blur kernel generation network;

[0093] A CNBD network training module 24 is used to construct a conditional neural blind deconvolution network based on the pre-trained PVC-PET image generation network and the blur kernel generation network, input the randomly generated Gaussian blur kernel and the anatomical MR image processed by the registration and pre-processing module into the conditional neural blind deconvolution network, and use the registered PVE-PET image as a training label to train the conditional neural blind deconvolution network;

[0094] The CDRC network training module 25 is used to use the anatomical MR image processed by the registration and preprocessing module as the input of the conditional deep residual compensation network, and combine the blur kernel prior estimated by the conditional neural blind deconvolution network, use the registered PVE-PET image as a training label, and train the conditional deep residual compensation network to obtain the PVC-PET image and realize partial volume correction.

[0095] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0096] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application scheme. A person of ordinary skill in the art can understand and implement it without paying any creative work.

[0097] Accordingly, the present application also provides a computer program product, including a computer program / instruction, which implements the method described in the first aspect when the computer program / instruction is executed by a processor.

[0098] Accordingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned PET image partial volume correction method based on conditional depth residual compensation. Figure 7 As shown, it is a hardware structure diagram of any device with data processing capability in which a deep learning data set access system provided by an embodiment of the present invention is located, except Figure 7 In addition to the processor, memory and network interface shown, any device with data processing capability in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capability, which will not be described in detail.

[0099] Accordingly, the present application also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed by a processor, the PET image partial volume correction method based on conditional depth residual compensation as described above is implemented. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Further, the computer-readable storage medium can also include both an internal storage unit of any device with data processing capabilities and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

[0100] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the contents disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not disclosed in the present application.

Claims

1. A method for partial volume correction of PET images based on conditional depth residual compensation, characterized in that: include: (1) Obtain PVE-PET images and corresponding anatomical MR images of the patient; (2) registering the PVE-PET image and the anatomical MR image, and preprocessing the registered anatomical MR image to obtain a registered PVE-PET image and an anatomical MR brain tissue image; (3) using the registered PVE-PET images and the randomly generated Gaussian blur kernels, respectively pre-training the PVC-PET image generation network and the blur kernel generation network, thereby initializing the initial weight parameters of the PVC-PET image generation network and the blur kernel generation network; (4) constructing a conditional neural blind deconvolution network based on the pre-trained PVC-PET image generation network and the blur kernel generation network, inputting the randomly generated Gaussian blur kernel and the anatomical MR brain tissue image into the conditional neural blind deconvolution network, and using the registered PVC-PET image as a training label to train the conditional neural blind deconvolution network; (5) The anatomical MR brain tissue image is used as the input of the conditional deep residual compensation network, and combined with the blur kernel prior estimated by the conditional neural blind deconvolution network, the registered PVE-PET image is used as the training label to train the conditional deep residual compensation network, thereby obtaining the PVC-PET image and realizing partial volume correction.

2. The method according to claim 1, characterized in that In step (2), the registered anatomical MR image is preprocessed, specifically: The skull part is removed from the registered anatomical MR image, and the brain tissue part is extracted to obtain the anatomical MR brain tissue image.

3. The method according to claim 1, characterized in that Step (3) includes: (3.1) Using the deep image prior network as the PVC-PET image generation network G x , using the fully connected layer network as the fuzzy kernel generation network G k ; (3.2) Using the registered PVE-PET images and the randomly generated Gaussian blur kernels, the PVC-PET image generation network and the blur kernel generation network are pre-trained, respectively, wherein the loss functions of the PVC-PET image generation network and the blur kernel generation network are defined as: in and are the loss functions of the PVC-PET image generation network and the blur kernel generation network, respectively, pet and z k are the registered PVE-PET images and randomly generated Gaussian blur kernels, respectively.

4. The method according to claim 1, characterized in that: In step (4), the conditional neural blind deconvolution network includes a parallel PVC-PET image generation network and a blur kernel generation network, wherein the PVC-PET image generation network is used to preliminarily predict the PVC-PET image based on the anatomical MR image, and the blur kernel generation network is used to estimate the blur kernel prior based on a randomly generated Gaussian blur kernel.

5. The method according to claim 4, characterized in that In step (4), the optimization process during training of the conditional neural blind deconvolution network is: in is the loss function of the conditional neural blind deconvolution network, z a For anatomical MR brain tissue images, y pet and z k are the registered PVE-PET images and the randomly generated Gaussian blur kernel, G x ,G k They are PVC-PET image generation network and blur kernel generation network respectively.

6. The method according to claim 1, characterized in that In step (5), the conditional deep residual compensation network includes a parallel PVC-PET image generation network and a depth residual generation network, wherein the depth residual generation network is used to obtain a depth residual map based on the anatomical MR brain tissue image, and the PVC-PET image generation network is used to combine the depth residual map and the blur kernel prior to generate a PVC-PET image based on the anatomical MR brain tissue image.

7. The method according to claim 6, characterized in that In step (5), the loss function of the conditional deep residual compensation network training is for: in and They are loss functions The fidelity term and regularization term in G x ,G r They are PVC-PET image generation network and deep residual generation network respectively. represents the blur kernel prior, z a For anatomical MR brain tissue images, v and C represent the coefficient vector and DCT operator obtained after DCT transformation of the artifacts introduced by the blur kernel error, respectively, and λ1, λ2 and λ3 are hyperparameters that control the weights of each item.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images