Method for enhancing nuclear medicine single photon emission computed tomography by using deep learning

Abstract

The invention relates to the technical field of medical images, in particular to a method for enhancing nuclear medicine single-photon emission computed tomography by using deep learning, which comprises the following steps of: acquiring projection data of nuclear medicine single-photon emission computed tomography; based on the projection data, constructing a deep learning enhanced counting rate sparse back projection algorithm neural network model; training the deep learning enhanced counting rate sparse back projection algorithm neural network model by using the projection data; reconstructing the projection data according to the trained deep learning enhanced counting rate sparse back projection algorithm neural network model; and the reconstructed three-dimensional nuclear medical image is output, the three-dimensional nuclear medical image comprises a forward projection three-dimensional nuclear medical image and a forward-back projection three-dimensional nuclear medical image, and when the low-dose and low-counting-rate nuclear medical image is processed, noise can be effectively suppressed, and the signal-to-noise ratio and the spatial resolution of the image can be improved.

Description

Technical Field

[0001] The present invention relates to the field of medical image technology, and more specifically, to a method for enhancing nuclear medicine single photon emission computed tomography imaging by using deep learning. Background Art

[0002] Nuclear medicine single photon emission computed tomography (SPECT) is an important medical imaging technology, which is widely used in the diagnosis of heart disease, tumors and nervous system diseases. However, traditional SPECT imaging technology faces many challenges, mainly including poor image quality, low reconstruction efficiency and high radiation dose.

[0003] In recent years, with the development of computer technology and image processing algorithms, the quality of SPECT imaging has been improved to a certain extent. Traditional reconstruction algorithms, such as filtered back projection (FBP) and maximum likelihood expectation maximization (MLEM), have improved image quality to a certain extent. However, these methods still have obvious shortcomings when processing low-dose and low-count rate nuclear medicine images, which often leads to large reconstructed image noise and low resolution, which is difficult to meet the needs of clinical diagnosis.

[0004] In addition, although the iterative reconstruction algorithm has improved image quality, it has high computational complexity and long reconstruction time, which makes it difficult to meet the needs of clinical real-time diagnosis. At the same time, in order to obtain high-quality images, traditional methods often require increasing the injection dose of radioactive tracers, which undoubtedly increases the risk of radiation exposure for patients.

[0005] Recently, deep learning technology has made significant progress in the field of medical image processing, bringing new opportunities for SPECT imaging. However, there are still many challenges in directly applying deep learning to SPECT image reconstruction. First, the special properties of nuclear medicine images, such as low signal-to-noise ratio and sparse sampling, make it difficult to directly apply traditional deep learning models. Second, how to effectively combine the advantages of deep learning and the physical model of traditional reconstruction algorithms to achieve higher quality and more efficient image reconstruction is still a problem to be solved.

[0006] In view of the above-mentioned shortcomings of the existing technologies, there is an urgent need for new methods that can comprehensively utilize the advantages of deep learning technology and traditional reconstruction algorithms to improve the quality and reconstruction efficiency of SPECT images while reducing radiation dose. Summary of the invention

[0007] The purpose of the present invention is to provide a method for enhancing nuclear medicine single photon emission computed tomography using deep learning, aiming to solve the above technical problems. Specifically, the present invention achieves a significant improvement in SPECT image quality, a substantial improvement in reconstruction efficiency, and an effective reduction in radiation dose through innovative deep learning model design and optimization strategies combined with the physical model of traditional reconstruction algorithms.

[0008] The present invention provides a method for enhancing nuclear medicine single photon emission computed tomography imaging using deep learning, comprising:

[0009] The acquisition steps include:

[0010] Acquire projection data for nuclear medicine single photon emission computed tomography;

[0011] Processing steps include:

[0012] Based on the projection data, a deep learning enhanced count rate sparse back-projection algorithm neural network model is constructed;

[0013] Using the projection data to train the deep learning enhanced count rate sparse back projection algorithm neural network model;

[0014] Reconstructing the projection data according to the trained deep learning enhanced count rate sparse back projection algorithm neural network model;

[0015] Output steps include:

[0016] The reconstructed three-dimensional nuclear medicine image is output, wherein the three-dimensional nuclear medicine image includes a forward projection three-dimensional nuclear medicine image and a forward-backward projection three-dimensional nuclear medicine image.

[0017] Preferably, the deep learning enhanced count rate sparse back projection algorithm neural network model comprises an image cascade layer and a back projection layer, wherein:

[0018] The image cascade layer includes a plurality of two-dimensional convolutional layers, a pooling layer and a two-dimensional deconvolutional layer;

[0019] The back-projection layer uses a count rate sparse back-projection algorithm to calculate and obtain reconstructed image data.

[0020] Preferably, the step of training the deep learning enhanced count rate sparse back projection algorithm neural network model specifically includes:

[0021] Constructing neural network structures for projection data of three-dimensional nuclear medicine images;

[0022] constructing an objective function, wherein the objective function considers both forward projection and back projection;

[0023] The parameters of the back-projection layer are fixed, and the parameters in the image cascade layer are iteratively updated.

[0024] Preferably, the iterative updating step comprises:

[0025] Perform data forward pass and calculate the loss of each layer;

[0026] Calculate the gradient of loss with respect to parameters;

[0027] Update the parameters using back-propagation.

[0028] Preferably, the step of acquiring projection data of nuclear medicine single photon emission computed tomography includes:

[0029] Initialize 3D nuclear medicine images;

[0030] Converting the initialized nuclear medicine image into a three-dimensional stereo model;

[0031] The photon detection probe is used to emit three-dimensional tomographic data to determine the detection event and the projection data.

[0032] Preferably, the step of initializing the three-dimensional nuclear medicine image uses a sinusoidal image for initialization, and the specific method is:

[0033] The sine function is used to generate initial values ​​of the forward projection image and the back projection image, wherein the initial value of the forward projection image is the amplitude of the sine image, and the initial value of the back projection image is the phase of the sine image.

[0034] Preferably, the step of converting the initialized nuclear medicine image into a three-dimensional stereoscopic model adopts a spherical coordinate back-projection method, specifically comprising:

[0035] Convert the pixel coordinates of the two-dimensional image into three-dimensional coordinates in the spherical coordinate system;

[0036] A three-dimensional model is constructed according to the three-dimensional coordinates in the spherical coordinate system.

[0037] As an advantage, the method further comprises the following steps:

[0038] Constructing a deep learning model based on a generative adversarial network (GAN), wherein the GAN includes a generator and a discriminator;

[0039] The GAN is used to enhance the quality of the reconstructed three-dimensional nuclear medicine images.

[0040] Preferably, the generator includes a deep generation subnet and a shallow generation subnet, wherein:

[0041] The input of the deep generation subnetwork is the preprocessed nuclear medicine scan image, and the output is the corresponding deep learning enhancement result;

[0042] The input of the shallow generation subnetwork is the preprocessed nuclear medicine scan image, and the output is the learning target of the corresponding deep learning method.

[0043] Preferably, the method further comprises the following steps:

[0044] An iterative algorithm is used to solve a count data optimization model, wherein the optimization model considers the back-projected three-dimensional image, the forward-projected three-dimensional image, and the count data obtained by the sampling probe;

[0045] Based on the solution results of the optimization model, the reconstructed three-dimensional nuclear medicine image is optimized.

[0046] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0047] The method of the present invention has many significant advantages and beneficial effects. First, through the deep learning enhanced count rate sparse back projection algorithm neural network model, the method effectively combines the feature extraction ability of deep learning and the physical model constraints of traditional algorithms, significantly improving the quality of image reconstruction. In particular, when processing low-dose, low-count rate nuclear medicine images, the method shows obvious advantages, can effectively suppress noise, and improve the signal-to-noise ratio and spatial resolution of the image.

[0048] Secondly, the present invention adopts an innovative training strategy, taking into account both the forward projection and back-projection processes, to achieve efficient optimization of model parameters. This strategy not only improves the convergence speed of the model, but also enhances the consistency and accuracy of the reconstructed image. By fixing the back-projection layer parameters and optimizing the image cascade layer parameters, the present invention fully exploits the advantages of deep learning while maintaining the accuracy of the physical model.

[0049] In addition, the image enhancement model based on the generative adversarial network (GAN) introduced in this invention further improves the visual quality of the reconstructed image. This method can effectively remove artifacts in the image, enhance detail features, make the reconstructed image closer to the real anatomical structure, and provide a more reliable imaging basis for clinical diagnosis.

[0050] It is worth noting that the method of the present invention significantly improves the reconstruction efficiency while improving image quality. Through the parallel computing capability of the deep learning model, the method greatly reduces the reconstruction time, making real-time or quasi-real-time SPECT image reconstruction possible, which is of great significance for clinical diagnosis and treatment decision-making.

[0051] Finally, the method of the present invention achieves the goal of maintaining high-quality image reconstruction while reducing radiation dose by optimizing data acquisition and processing strategies. This not only reduces the risk of radiation exposure to patients, but also expands the application scope of SPECT technology, especially for patients who need multiple follow-up examinations and special populations who are sensitive to radiation, which has important clinical value.

[0052] In summary, the method of enhancing nuclear medicine single photon emission computed tomography using deep learning provided by the present invention has made significant progress in many aspects such as image quality, reconstruction efficiency, and radiation dose control, opening up new avenues for the clinical application and further development of SPECT technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The figure is a flow chart of the method of the present invention.

[0054] Figure 2 A flow chart is constructed for the neural network model of the present invention.

[0055] Figure 3 This is a flow chart of model training of the present invention.

[0056] Figure 4 This is a flowchart of image reconstruction according to the present invention.

[0057] Figure 5 This is the GAN optimization flowchart of the present invention. DETAILED DESCRIPTION

[0058] Please refer to Figure 1-5 The present invention provides a method for enhancing nuclear medicine single photon emission computed tomography (SPECT) using deep learning. The method improves the image quality and reconstruction efficiency of nuclear medicine single photon emission computed tomography (SPECT) by deep learning technology. The specific embodiments of the present invention will be described in detail below.

[0059] First, the method of the present invention comprises an acquisition step, a processing step and an output step. In the acquisition step, projection data of nuclear medicine single photon emission computed tomography are acquired. These projection data are usually two-dimensional projection images collected at different angles by a SPECT scanner.

[0060] In the processing step, a deep learning enhanced count rate sparse back projection algorithm neural network model is constructed based on the acquired projection data. This model combines the feature extraction capabilities of deep learning with the advantages of the traditional count rate sparse back projection algorithm. Subsequently, the neural network model is trained using the acquired projection data. During the training process, the model learns how to extract useful features from the projection data and optimize the reconstruction process.

[0061] After training, the projection data is reconstructed using the trained deep learning enhanced count rate sparse back projection algorithm neural network model. This step converts the two-dimensional projection data into a three-dimensional nuclear medicine image.

[0062] In the output step, the reconstructed three-dimensional nuclear medicine image is output. It is worth noting that the method of the present invention simultaneously outputs a forward projection three-dimensional nuclear medicine image and a forward-backward projection three-dimensional nuclear medicine image. These two images provide complementary information, which helps doctors make more accurate diagnoses.

[0063] Next, we will introduce the structure of the neural network model of the deep learning enhanced count rate sparse back projection algorithm in detail. The model consists of two main parts: the image cascade layer and the back projection layer.

[0064] The image cascade layer consists of multiple 2D convolutional layers, pooling layers, and 2D deconvolutional layers. This structural design enables the model to effectively extract and process image features. Specifically, the 2D convolutional layer is used to extract local features, the pooling layer is used to reduce the spatial dimension of the feature map and improve the translation invariance of the model, and the 2D deconvolutional layer is used to restore the spatial resolution of the feature map.

[0065] The back-projection layer uses a count rate sparse back-projection algorithm to calculate the reconstructed image data. This layer takes advantage of the traditional algorithm and can effectively handle the low count rate problem in nuclear medicine imaging. By combining deep learning and traditional algorithms, the method of the present invention can improve the quality of the reconstructed image while maintaining computational efficiency.

[0066] During the training process, the present invention adopts an innovative training strategy. First, a neural network structure of projection data of three-dimensional nuclear medicine images is constructed. This structural design enables the model to learn the reconstruction process directly from the projection data without the need for pre-reconstructed images.

[0067] Secondly, construct the objective function, which takes both forward projection and back projection into account. The objective function can be expressed as:

[0068] L=λ s L s +λ i L i +λ t L t ,

[0069] Among them, L s , L i and L t They represent image similarity loss, projection data consistency loss and time consistency loss respectively. s , i and λ tis the corresponding weight coefficient. This objective function design can comprehensively consider various factors in the reconstruction process and help improve the quality of the reconstructed image.

[0070] During the training process, the parameters of the back-projection layer are fixed, and the parameters in the image cascade layer are iteratively updated. This strategy can effectively balance the advantages of deep learning and traditional algorithms. The parameter update adopts the back-propagation method, and the specific process includes forward data transmission, loss calculation, gradient calculation and parameter update.

[0071] Preferably, in one embodiment of the present invention, the process of forward data transmission can be expressed as:

[0072] h l =f l (W l h l-1 +b l ),

[0073] Among them, h l represents the output of the lth layer, f l is the activation function, W l and b l are weight and bias parameters respectively. The gradient calculation of loss to parameters can be expressed as:

[0074]

[0075] The parameter update adopts the stochastic gradient descent method, and the update formula is:

[0076]

[0077] Among them, η is the learning rate, which usually ranges from 0.001 to 0.1.

[0078] The method of the present invention adopts an innovative data acquisition method when acquiring projection data of nuclear medicine single photon emission computed tomography. First, the three-dimensional nuclear medicine image is initialized. The initialization method uses a sinusoidal graph, which can provide a good initial estimate and accelerate the convergence process.

[0079] Next, the initialized nuclear medicine image is converted into a three-dimensional model. This step uses the spherical coordinate back-projection method, which can effectively deal with the geometric distortion problem in SPECT imaging.

[0080] Finally, the photon detection probe is used to emit 3D tomographic data to determine the detection events and projection data. This data acquisition method based on photon counting can better handle low-dose imaging problems, improve image quality and reduce radiation dose.

[0081] The method of the present invention not only improves the image quality of nuclear medicine single photon emission computed tomography by combining deep learning with traditional algorithms, but also improves the reconstruction efficiency. The method has significant advantages in processing low-dose, low-count rate nuclear medicine images, providing a more reliable imaging basis for clinical diagnosis. In a preferred embodiment of the present invention, the iterative update process of the neural network model of the deep learning enhanced count rate sparse back projection algorithm is further refined. The process includes three main steps: data forward transmission, loss gradient calculation, and parameter update.

[0082] First, the data forward pass step calculates the loss of data at each layer of the neural network. The specific calculation formula is as follows:

[0083] L l =f(W l ·h l-1 +b l ),

[0084] Among them, L l represents the loss of the lth layer, f is the activation function, W l is the weight matrix, h l -1 is the output of the previous layer, b l Preferably, the present invention adopts ReLU (Rectified Linear Unit) as the activation function, and its expression is:

[0085] f(x)=max(0,x),

[0086] The choice of the RelU function is based on its wide application and excellent performance in deep learning, especially its advantages in solving the gradient vanishing problem.

[0087] Next, we calculate the gradient of the loss with respect to the parameters θ and φ.

[0088]

[0089] Among them, G(I b ) and G(I f ) represent the back-projection image and the forward-projection image, respectively, and θ and φ are model parameters. This gradient calculation method can effectively capture the influence of model parameters on the final reconstruction result, thereby guiding the optimization direction of the parameters.

[0090] Finally, the back propagation method is used to update the values ​​of θ and φ. The update formula is as follows:

[0091] θ=θ-η·Grad(φ),

[0092] φ=φ-η·Grad(θ),

[0093] Wherein, η is the learning rate, which is an important hyperparameter. In the implementation of the present invention, the initial value of the learning rate is set to 0.001, and a learning rate decay strategy is adopted, where the learning rate is reduced to 90% of the original value after every 100 iterations. This strategy can quickly approach the optimal solution in the early stage of training, and carefully adjust the parameters in the later stage of training to improve the convergence performance of the model.

[0094] The method of the present invention adopts an innovative data acquisition and preprocessing method when acquiring projection data of nuclear medicine single photon emission computed tomography. First, the three-dimensional nuclear medicine image is initialized. The initialization adopts the sinogram method, which is specifically implemented as follows:

[0095] I init (x,y,z)=A·sin(ω x x+ω y y+ω z z+φ)

[0096] Among them, I init (x, y, z) is the initialized three-dimensional image, (x, y, z) is the spatial coordinate, A is the amplitude, ω x ,ω y ,ω z is the angular frequency in each direction, and φ is the phase. In the preferred embodiment of the present invention, A is set to 1, ω x =ω y =ω z = 2π / L (where L is the length of the image side), φ = 0. This initialization method can provide a good starting point and accelerate the convergence process of the model.

[0097] Next, the initialized nuclear medicine image is converted into a three-dimensional model. The present invention adopts a spherical coordinate back-projection method, which is specifically implemented as follows:

[0098]

[0099] Among them, (x, y, z) are the coordinates in the Cartesian coordinate system, and (r, θ, φ) are the corresponding spherical coordinates. This conversion method can effectively deal with the geometric distortion problem in SPECT imaging and improve the accuracy of reconstructed images.

[0100] Finally, the photon detection probe is used to emit three-dimensional tomographic data to determine the detection event and projection data. In this step, the present invention adopts a photon transmission simulation based on the Monte Carlo method. Specifically, for each emitted photon, its path is determined by the following probability distribution:

[0101] p(Δs)=μe -μΔs ,

[0102] Among them, p(Δs) is the probability of photons interacting after traveling a distance of Δs, and μ is the linear attenuation coefficient. Through this method, the transmission process of photons in human tissue can be more realistically simulated, thereby obtaining more accurate projection data.

[0103] The method of the present invention significantly improves the data quality of nuclear medicine single photon emission computed tomography by adopting these innovative data acquisition and preprocessing techniques. This not only provides more reliable input for subsequent image reconstruction, but also helps to reduce radiation dose and reduce the risk of radiation exposure to patients.

[0104] In general, the method of enhancing nuclear medicine single photon emission computed tomography using deep learning proposed in the present invention has made significant progress in improving image quality, reducing radiation dose and improving reconstruction efficiency through innovative model structure design, training strategy optimization, and data acquisition and preprocessing technology. This provides new possibilities for the application of nuclear medicine imaging in clinical diagnosis, and is expected to bring patients a more accurate and safer diagnostic experience. In another embodiment of the present invention, in order to further improve the quality of reconstructed images, a deep learning model based on a generative adversarial network (GAN) is introduced. The model includes two main components, a generator and a discriminator, which are used to enhance the quality of reconstructed three-dimensional nuclear medicine images.

[0105] The generator consists of a deep generative subnet and a shallow generative subnet. The input of the deep generative subnet is the preprocessed nuclear medicine scan image, and the output is the corresponding deep learning enhancement result. The input of the shallow generative subnet is also the preprocessed nuclear medicine scan image, but its output is the learning target of the corresponding deep learning method. This dual subnet structure design can capture both the deep and shallow features of the image, thereby achieving a more comprehensive image enhancement effect.

[0106] Preferably, the deep generation subnet adopts a U-Net structure, wherein the encoder part is composed of multiple convolutional layers and downsampling layers, and the decoder part is composed of corresponding deconvolutional layers and upsampling layers. Specifically, the encoder structure can be expressed as:

[0107] E(x)=Conv n (Conv n-1 (...Conv1(x))),

[0108] Among them, Conv i represents the i-th convolutional layer, and x is the input image. The decoder structure can be expressed as:

[0109] D(E(x))=DeConv1(DeConv2(...DeConv n (E(x)))),

[0110] Among them, DeConv i represents the i-th deconvolution layer. This structural design can effectively retain the detailed information of the image while achieving high-quality image reconstruction. The shallow generation subnet adopts a residual network structure, which contains multiple residual blocks. The structure of each residual block can be expressed as:

[0111]

[0112] Where x is the input, is the residual mapping. This structural design can effectively alleviate the gradient vanishing problem in deep network training and improve the learning ability of the model. The discriminator adopts a design based on PatchGAN, which can distinguish the authenticity of local areas of the image. The output of the discriminator can be expressed as:

[0113] D(x)=σ(Conv n (Conv n-1 (...Conv1(x)))),

[0114] Among them, σ is the sigmoid activation function, which is used to map the output to the [0,1] interval, indicating the probability that the discriminator considers the input image to be a real image.

[0115]

[0116] in, To combat losses, is the reconstruction loss, λ adv and λ rec is the corresponding weight coefficient. In a preferred embodiment of the present invention, λ adv Set to 0.001,λ rec Set to 1.0. The discriminator loss function uses the standard binary cross entropy loss:

[0117]

[0118] Among them, p data represents the real data distribution, p z represents the noise distribution, G and D represent the generator and discriminator respectively.

[0119] The method of the present invention also introduces an iterative algorithm to solve the counting data optimization model to further improve the quality of the reconstructed image. The optimization model takes into account three aspects: the back-projected three-dimensional image, the forward-projected three-dimensional image, and the counting data obtained by the sampling probe. Its objective function can be expressed as:

[0120]

[0121] Wherein, f is the 3D image to be reconstructed, A is the system matrix, g is the observed data, R(f) is the regularization term, and λ is the regularization parameter. In a preferred embodiment of the present invention, total variation regularization is adopted, that is:

[0122]

[0123] in, and Represent the gradient operators in the x, y and z directions respectively.

[0124] To solve this optimization problem, the present invention adopts the alternating direction multiplier method (ADMM). Specifically, an auxiliary variable v is introduced to transform the original problem into:

[0125]

[0126] in, Represents the gradient operator. The iterative steps of ADMM are as follows:

[0127] 1. Update f:

[0128]

[0129] 2. Update v:

[0130]

[0131] 3. Update the dual variable u:

[0132]

[0133] Wherein, ρ is a penalty parameter, and k represents the number of iterations. In the embodiment of the present invention, the initial value of ρ is set to 0.1, and an adaptive strategy is used for adjustment.

[0134] Through this iterative optimization method, the present invention can effectively deal with the low signal-to-noise ratio and sparse sampling problems in nuclear medicine imaging, and significantly improve the quality of reconstructed images. Combined with the deep learning enhanced count rate sparse back projection algorithm neural network model and the GAN-based image enhancement model introduced above, the method proposed in the present invention has achieved many innovations and breakthroughs in the field of nuclear medicine single photon emission computed tomography.

[0135] This comprehensive optimization method not only improves the quality and efficiency of image reconstruction, but also effectively reduces radiation dose, providing patients with a safer and more accurate diagnostic experience. At the same time, the method of the present invention has good adaptability and scalability, and can be flexibly adjusted and optimized according to specific application scenarios and needs, providing new ideas and possibilities for the further development of nuclear medicine imaging technology.

[0136] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for enhancing nuclear medicine single photon emission computed tomography using deep learning, characterized in that: include: The acquisition steps include: Acquire projection data for nuclear medicine single photon emission computed tomography; Processing steps include: Based on the projection data, a deep learning enhanced count rate sparse back-projection algorithm neural network model is constructed; Using the projection data to train the deep learning enhanced count rate sparse back projection algorithm neural network model; Reconstructing the projection data according to the trained deep learning enhanced count rate sparse back projection algorithm neural network model; Output steps include: The reconstructed three-dimensional nuclear medicine image is output, wherein the three-dimensional nuclear medicine image includes a forward projection three-dimensional nuclear medicine image and a forward-backward projection three-dimensional nuclear medicine image.

2. The method according to claim 1, characterized in that The deep learning enhanced count rate sparse back-projection algorithm neural network model includes an image cascade layer and a back-projection layer, wherein: The image cascade layer includes a plurality of two-dimensional convolutional layers, a pooling layer and a two-dimensional deconvolutional layer; The back-projection layer uses a count rate sparse back-projection algorithm to calculate and obtain reconstructed image data.

3. The method according to claim 2, characterized in that The step of training the deep learning enhanced count rate sparse back projection algorithm neural network model specifically includes: Constructing neural network structures for projection data of three-dimensional nuclear medicine images; constructing an objective function, wherein the objective function considers both forward projection and back projection; The parameters of the back-projection layer are fixed, and the parameters in the image cascade layer are iteratively updated.

4. The method according to claim 3, characterized in that The iterative updating step comprises: Perform data forward pass and calculate the loss of each layer; Calculate the gradient of loss with respect to parameters; Update the parameters using back-propagation.

5. The method according to claim 1, characterized in that The step of acquiring projection data of nuclear medicine single photon emission computed tomography includes: Initialize 3D nuclear medicine images; Converting the initialized nuclear medicine image into a three-dimensional stereo model; The photon detection probe is used to emit three-dimensional tomographic data to determine the detection event and the projection data.

6. The method according to claim 5, characterized in that The step of initializing the three-dimensional nuclear medicine image is performed by using a sinusoidal graph, and the specific method is as follows: The sine function is used to generate initial values ​​of the forward projection image and the back projection image, wherein the initial value of the forward projection image is the amplitude of the sine image, and the initial value of the back projection image is the phase of the sine image.

7. The method according to claim 5, characterized in that The step of converting the initialized nuclear medicine image into a three-dimensional stereoscopic model adopts a spherical coordinate back-projection method, specifically comprising: Convert the pixel coordinates of the two-dimensional image into three-dimensional coordinates in the spherical coordinate system; A three-dimensional model is constructed according to the three-dimensional coordinates in the spherical coordinate system.

8. The method according to claim 1, characterized in that The following steps are also included: Constructing a deep learning model based on a generative adversarial network (GAN), wherein the GAN includes a generator and a discriminator; The GAN is used to enhance the quality of the reconstructed three-dimensional nuclear medicine images.

9. The method according to claim 8, characterized in that The generator includes a deep generation subnet and a shallow generation subnet, wherein: The input of the deep generation subnetwork is the preprocessed nuclear medicine scan image, and the output is the corresponding deep learning enhancement result; The input of the shallow generation subnetwork is the preprocessed nuclear medicine scan image, and the output is the learning target of the corresponding deep learning method.

10. The method according to claim 1, characterized in that The method further comprises the following steps: An iterative algorithm is used to solve a count data optimization model, wherein the optimization model considers the back-projected three-dimensional image, the forward-projected three-dimensional image, and the count data obtained by the sampling probe; Based on the solution results of the optimization model, the reconstructed three-dimensional nuclear medicine image is optimized.

Images